**In-Situ Yeast Fermentation Medium in Fortifying Protein and Lipid Accumulations in the Harvested Larval Biomass of Black Soldier Fly**

**Chung Yiin Wong 1, Yeek Chia Ho 2, Jun Wei Lim 1,\*, Pau Loke Show 3, Siewhui Chong 3, Yi Jing Chan 3, Chii Dong Ho 4,\*, Mardawani Mohamad 5, Ta Yeong Wu 6,7, Man Kee Lam <sup>8</sup> and Guan Ting Pan <sup>9</sup>**


Received: 2 January 2020; Accepted: 7 February 2020; Published: 14 March 2020

**Abstract:** Recently, worldwide researchers have been focusing on exploiting of black soldier fly larval (BSFL) biomass to serve as the feed mediums for farmed animals, including aquaculture farming, in order to assuage the rising demands for protein sources. In this study, yeast was introduced into coconut endosperm waste (CEW) whilst serving as the feeding medium to rear BSFL in simultaneously performed in situ fermentation. It was found that at a 2.5 wt% yeast concentration, the total biomass gained, growth rate and rearing time were improved to 1.145 g, 0.085 g/day and 13.5 days, respectively. In terms of solid waste reduction, the inoculation of yeast over 0.5 wt% in CEW was able to achieve more than 50% overall degradation, with the waste reduction indexes (WRIs) ranging from 0.038 to 0.040 g/day. Disregarding the concentration of yeast introduced, the protein productivity from 20 BSFL was enhanced from only 0.018 g/day (the control) to 0.025 g/day with the presence of yeast at arbitrary concentrations. On the other hand, the larval protein yield was fortified from the control (28%) to a highest value of 35% with the presence of a mere 0.02 wt% yeast concentration. To summarize, the inclusion of a minimal amount of yeast into CEW for in situ fermentation ultimately enhanced the growth of BSFL, as well as its protein yield and productivity.

**Keywords:** black soldier fly; yeast; fermentation; protein; larvae; organic waste; coconut endosperm waste

#### **1. Introduction**

The black soldier fly (BSF) thrived in North America before it migrated to tropical other countries during WWII. It mimics the appearance of a wasp, confusing the public with its appearance. BSF larvae (BSFL) are intrinsically polyphagous as well as saprophagous, since the larvae only consume organic matter during this stage and can ingest different kinds of decaying organic matters such as animal manure, animal carcasses or sometimes even decaying wood matters. Unlike houseflies, the BSF does not carry any transmitted diseases, as the adult fly does not feed and only relies on body fat or the energy accumulated during the larval stage for metabolism. Upon maturing sexually, the female BSF will oviposit eggs at the cracks near to food sources to ensure the newly ecloded BSF larvae (neonates) have enough food to complete their life cycle [1]. Generally, after the copulation process, the female black soldier fly will oviposit the eggs after two to three days. The whole life cycle of a black soldier fly from egg to adult will take up to around 40 to 44 days [2].

Owing to its high protein content, the direct introduction of BSFL biomass into animal feed has been explored as an alternative fishmeal, which is growing in cost. From previous research studies, the inclusion of BSFL biomass at 17%, 33%, 49%, 64% and 75% into aquaculture feed was found to decrease feed consumption due to its low digestibility. In this case, the highest protein retention in fed fish was obtained when 33% of BSFL biomass was used, thereafter decreasing as BSFL biomass was incorporated. From the study, the inclusion of BSFL biomass into aquaculture feed was feasible at low percentages, and it has been suggested that the presence of chitin in BSFL biomass contributes certain benefits to the growth performance of the turbot from the feed intake, including the availability and digestibility of nutrients [3]. The BSFL protein was also introduced to rainbow trout as a replacement meal with the partial inclusion at 25% and 50%, and the outcome showed that the BSFL biomass degraded the lipid health indexes of the rainbow trout while negatively impacting the contents of polyunsaturated fatty acids with increases of BSFL biomass. In order to prevent the negative impacts of BSFL inclusion on trout, it was suggested that a 40% inclusion level of BSFL biomass could be used without impacting the survival, growth performance, condition factor and so on [4]. Apart from the aquaculture field, BSFL biomass can also be introduced as animal feed for broilers in either a partial or highly defatted form. From the past study, an inclusion of partially defatted BSFL biomass into broilers' feed showed higher digestibility by the chicken. [5]. According to Schiavone et al. [6], an inclusion of defatted BSFL in broiler chicken diets at 10% showed improvements in carcass and meat quality parameters as well as the heavy metal contents, and there were no negative consequences. Moreover, when the BSFL biomass was incorporated into quail feed to replace fishmeal, the outcome showed a similar result as with the fishmeal. When 25% to 50% BSFL biomass at 25% and 50% was included, no impact on the palatability of ration or quail appetites was detected. In short, the 50% replacement of fishmeal with BSFL biomass was generally recommended, as no negative impact was demonstrated on the growth performance of most of the farmed animals [7].

The study by Loponte et al. [8] showed that the corn-soybean meal diet used for *Barbary partridge* rearing could be replaced with *Tenebrio molitor* and *Hermetia illucens* biomass at 25% and 50%. Even though the control group had heavier weight of partridges fed and longer intestinal and caecal lengths, the live weights of the birds that were fed *T. molitor* and *H. illucens* meals were significantly higher than the control due to improved nutrient digestibility. Apart from these, several studies were carried out to determine the impacts of insect meal on the egg characteristics of laying hens. With the inclusion of *H. illucens* into laying hens' diets, lay percentage and egg mass were found to be affected only at 25% replacement, owning to higher methionine and lysine. A replacement by insect meal more than 50% negatively impacted dry matter, organic matter and crude protein digestibility due to the presence of chitin; hence, a 25% insect meal replacement was recommended for the diets of laying hens [9]. A 100% soybean meal replacement by *H. illucens* was found feasible in Lohmann Brown Classic laying hens during 21 weeks of rearing. Eggs laid

by the hens fed with the insect diet were found to possess higher quality of yolks than the control group, which was fed soybean meal. Also, the red index of the eggs laid was found to be higher in the insect treatment group (5.63) compared with the control (1.36). Moreover, the insect treatment group laid eggs with higher γ-tocopherol (4.0 against 2.4 mg/kg), lutein (8.6 against 4.9 mg/kg), β-carotene (0.33 against 0.19 mg/kg) and total carotenoids (15 against 10.5 mg/kg) than the control. Nonetheless, the insect treatment group eggs contained 11% less cholesterol than the control group, and no differences were found in fatty acid composition [10].

Recently, worldwide researchers have focused on exploiting BSFL biomass to serve as a feed medium for farmed animals, including aquaculture farming, in order to sustain the rising demands for a protein source. In this regard, various low-cost organic wastes had been employed to farm BSFL without truly optimizing its larval protein content. It has been hypothesized that increasing the protein content of BSFL would directly permit a higher inclusion of larval biomass in animal feeds whilst reducing the costs attributed mostly as a result of the unsustainable use of fishmeals. BSFL is currently proposed as the best protein source for animal farming and aquafarming, since the cost of animal feed and fishmeal continue increasing year after year due to marine overexploitation and a limited availability of lands. Animal feeds consist mainly of fishmeal and soybean, which serve as the protein alimentation, in addition to fish oils, seed cakes and other grains [11]. Thus, the main objective of this study was to enhance the protein content of BSFL by introducing yeast to execute fermentation on low-cost organic waste for larval feeding (i.e., coconut endosperm waste). The presence of yeast to ferment coconut endosperm waste would improve the nutritional content of larval feeding medium and eventually the larval protein content upon feeding. The degree of fermented coconut endosperm waste valorization by BSFL has also been reported to unveil organic waste treatment potentiality.

#### **2. Materials and Methods**

#### *2.1. Acquisition of Coconut Endosperm Waste*

The grated fresh coconut endosperm waste (CEW) was initially acquired from a local stall selling coconut milk and kept within 2 to 4 ◦C in a refrigerator. The moisture content of the CEW was determined through a gravimetric method and adjusted to 70% by homogenizing with sterile distilled water as calculated using Equation (1) prior to being used in the experiment.

$$\mathbf{V\_{H\_2O}} = \frac{(\%\_{\rm H\_2O})(\mathbf{M\_S})}{1 - (\%\_{\rm H\_2O})} - \mathbf{M\_{H\_2O}} \tag{1}$$

where VH2O represents the total volume of sterile distilled water to be added (in g considering the density of water 1 g/mL), %H2O represents the percentage of desired moisture (which was 70% (0.7 was inserted into the equation) in this study), MS represents the total dry weight of the CEW (in g) and MH2O represents the initial moisture content of the CEW (in g).

#### *2.2. Attainment of Black Soldier Fly Larvae (BSFL)*

We weighed 200 g of fresh CEW and transferred it into a plastic container with a size of 35 × 25 cm (height × diameter). We left the ventilated container in a sun-shaded area, serving as a bait to lure female BSFs. Several pieces of paper box cardboard with a size of 8 cm × 3 cm (length × width) were attached to the inner wall of the plastic container about 3 to 5 cm above the CEW medium, acting as a platform for the female BSF to oviposit her eggs. This cardboard was checked daily for BSF eggs. The attained eggs were then transferred into sterile Petri dishes and incubated until the larvae emerged. The new BSFL (neonate) were reared on CEW until 6 days old prior to being used in the experiments [12].

#### *2.3. Rearing of BSFL Using CEW Inoculated with Yeast*

Figure 1 presents the schematic flow of the reported works. Different quantities of dry yeast powder (commercial brand: Bunga Raya) with 0.02, 0.1, 0.5, 1.0 and 2.5 wt% were separately homogenized with CEW to serve as an initial inoculum for fermentation to take place. A 10 g, dry weight basis of each CEW that had been inoculated with yeast medium was then immediately administered to 20 six-day-old BSFLs. The larval rearing using each CEW medium inoculated with different percentages of yeast was stopped once the BSFL reached its fifth instar, as determined by head size and body color [1,13]. Each batch of harvested BSFL was deactivated at 105 ◦C for 5 min then dried at 60 ◦C until reaching a constant weight. This was followed by grinding the BSFL into powder and storing it at −20 ◦C prior to the chemical analyses [14]. All CEW residues were also separately collected and dried at 105 ◦C until reaching a constant weight. All setups were (at least) duplicated to verify the statistical reproducibility.

**Figure 1.** Schematic flow of the experimental procedures.

#### *2.4. Growth Performance of the BSFL*

Upon the completion of experiments, growth of the BSFL was evaluated using Equation (2) for the total biomass gained and Equation (3) for the BSFL growth rate [15], as shown below:

$$\text{Total biomass gained (g)} = \text{Final BSFL dried mass (g)} - \text{Initial BSFL dried mass (g)} \tag{2}$$

$$\text{BSFL growth rate (g/day)} = \text{Total biomass gained (g)/Rearing time (day)} \tag{3}$$

#### *2.5. Treatment of CEW Via Valorization by BSFL*

In order to determine the degree of CEW reduction, two parameters were measured including Equation (4) for overall degradation (OD) and Equation (5) for the waste reduction index (WRI) [16], as shown below:

$$\text{Overall degradation} = \text{Total feed consumed (g)} / \text{Total feed off/g} \tag{4}$$

$$\text{WRI (g/day)} = \text{Total feed consumed (g)/Rearing time (day)}\tag{5}$$

#### *2.6. Nitrogen, Chitin and Protein Analyses*

Nitrogen contents of dried BSFL biomass were determined through the Dumas combustion method (Perkin Elmer, CHNS/O 2400). The sample was weighed in the range of 1 to 1.5 mg then transferred into a tin capsule, wrapped and combusted at 925 ◦C. The nitrogen compounds were then converted into NOx, further reduced to nitrogen gas at 640 ◦C and detected by a thermal conductivity detector (TCD) [17]. In this study, the larval protein contents were estimated with a multiplication factor of 6.25 [18]. However, the presence of chitin in BSFL biomass will influence the larval protein content and, hence, nitrogen from chitin has to be deducted from the total larval nitrogen content prior

to protein conversion in order to avoid over-estimation [19]. Chitin is a polysaccharide that can be found in yeast, fungi, crustaceans and insects [20], as well as being present in the exoskeleton of BSFL, where it accounts for 6.89% of the nitrogen content [16]. The formic acid method was applied for chitin determination in this study [19,21], with modification to suit a small sample size. We mixed 10 mL of 90% formic acid with 1 g of BSFL dried fat-free biomass (the initial mass prior to being defatted had been recorded) at room temperature for 24 h. Then, the mixture was centrifuged, and the supernatant was decanted. The residue was washed with 10 mL of 100% acetone, followed by 10 mL of 70% acetone before being recentrifuged to separate the acetone. The residue was refluxed with 5% of 10 mL sodium hydroxide for 90 min before being filtered and washed with distilled water on ashless filter paper (Whatman No. 1 with a 55 mm diameter). Next, the residue was dried in the oven to a constant weight at 105 ◦C, then later it was ashed at 550 ◦C for 24 h. The final weight of the sample was recorded and assumed to be intact chitin.

Chitin content (%) = Mass of residues after ashing (g)/Initial mass of BSFL (g) × 100% (6)

TNChitin (%) = [Chitin content (%) × Nitrogen content in chitin (%) (which is 6.89%)]/100% (7)

Corrected protein yield for BSFL (%) = [TNBSFL (%) − TNChitin (%)] × 6.25 (8)

Protein productivity (g/day) = Protein content (g)/Rearing time (day) (9)

where TNBSFL is the total nitrogen from the BSFL biomass and TNChitin is the total nitrogen from the chitin.

#### **3. Results and Discussion**

#### *3.1. Growth Performances of BSFL*

Initially, 10 g of yeast-inoculated feed was introduced to 20 BSFL at different concentrations. The total biomasses gained for the BSFL were recorded once every setup had reached the fifth instar, as shown in Table 1. Under the control condition, the total biomass gained by the BSFL was attained at only 0.998 g from a total of 20 BSFL. This value increased with the increment of yeast concentrations rising from 0.02 to 2.5 wt%, and it attained its highest point at 1.145 g. As compared with a previous study by Zheng et al. [22], the performance of in situ yeast fermentation at the highest concentration in this study was comparable to the best RID-X dosage (w/w), which was equivalent to 1.228 g per 20 BSFL with a difference of merely 0.08 g per 20 BSFL. RID-X was the active bacterial product introduced into the larval feeding medium in the study by Zheng et al. [22]. On the other hand, besides changing the nutritional properties of larval feed by introducing microorganisms, the growth of the BSFL could also be altered by feeding with a protein-rich medium, as suggested by Rehman et al. [23]. At a 1:4 ratio of dairy manure to protein-rich soybean curd residue, the total dry larval mass that could be attained was 28.1 g, which is equivalent to 0.56 g from 20 BSFL. This showed that the performance of BSFL growth through the co-digestion treatment was still lower compared to the microorganism inoculation treatment (i.e., yeast in this study). Thus, the inoculation of microorganisms into larval feed is strongly recommended for better BSFL growth.

**Table 1.** Growth performances of BSFL fed with CEW having been inoculated with different yeast concentrations.


Moreover, the growth rate of the BSFL also increased in parallel to the increasing concentrations of yeast from an initial 0.065 g/day to a maximum of 0.085 g/day. This phenomenon can be explained by the shortening of the rearing time of the BSFL. The in situ yeast fermentation of feeding medium had a reduced rearing time from 15.5 days to 13.5 days. This occurrence could have been due to the introduction of yeast that favored the digestibility of carbohydrate compounds in CEW [24] and thus improved the assimilation of nutrients into the BSFL body mass in the form of lipids. Also, Yoon et al. [25] reported that the yeast was capable of breaking down carbohydrates through fermentation, especially common monosaccharides such as D-glucose, D-fructose, D-mannose and D-galactose. On the other hand, it has been proven that the BSFL was also able to convert additional glucose into lipids upon excess availability [26]. Indeed, the measured lipid content increased from about 40% for the control to 50% for a 1.0 wt% yeast concentration. The lipids could later serve as a potential source for biodiesel production, which is something that could be explored further.

#### *3.2. CEW Valorization by BSFL*

Due to its polyphagous nature, BSFL is able to reduce solid organic wastes during the rearing process. In this study, the overall degradation of CEW was 0.48 under the control, and this value was maintained for low yeast concentrations of 0.02 and 0.1 wt%. With the addition of yeast at more than 0.5 wt%, the overall degradation of CEW increased to a range of 0.51 to 0.53. Thus, it could be concluded that the 20 BSFL were able to degrade about half of the CEW upon completion of the rearing process, disregarding the concentrations of yeast inoculated. With the introduction of yeast at different concentrations in the feeding medium, it was shown that the WRI increased from 0.31 g/day under the control, to 0.33 g/day with a 0.02 wt% of yeast and 0.38 g/day with a 0.5 wt% yeast concentration. At last, the WRI reached its highest point of 0.40 g/day with a 2.5 wt% yeast concentration. The WRI increment was about 15% faster in 0.5 wt% compared to the 0.02 wt%. This could plausibly be because the addition of 0.5 wt% yeast reached the concentration threshold for maximizing the in situ fermentation to spur the ingestion of CEW by BSFL [27]. Also, it can be observed from Table 1 that the rearing duration for BSFL decreased from 15.5 days and reached a plateau at 13.5 days when the 0.5 wt% yeast concentration (and beyond) were employed for in situ fermentation. Above the 0.5 wt% yeast concentration, the effect on WRI was not significant, if not deteriorating, as reported by Palma et al. [28]. In their study of managing high fiber food waste using BSFL, incremental larval growth led to a decrease in almond hull consumption and vice-versa. The authors presumed that the occurrence was the result of a competition for resources between the BSFL and microbial communities, or because of enhanced synergy between the larvae and their associated microbiota.

#### *3.3. Protein Contents in BSFL*

The chitin content from the BSFL was determined to be around 8%, and the nitrogen from the chitin was deducted from the total nitrogen of the BSFL to prevent the over-estimation of BSFL protein content. Figure 2 shows that the corrected protein of the BSFL was only attained around 28% under the control system, and that this value increased to its peak at about 35% when the lowest yeast concentration was used for fermentation. The corrected protein value dropped to around 30% and remained at that level with yeast concentrations from 0.5 to 2.5 wt%. Looking into the protein productivity from 20 BSFL, the value was attained at around 0.02 g/day under the control system and increased to around 0.025 g/day with the introduction of yeast at 0.02 wt%. The value fluctuated within the range of 0.023 to 0.025 g/day with higher yeast concentrations from 0.5 to 2.5 wt%.

As reported by Diener et al. [19], a daily feeding rate of 100 mg of chicken feed per larva was proposed to produce better larval quality and higher waste reduction in the shortest period of time. At this rate, the corrected protein content of BSFL was 34.4%, which is comparable with the current study in which an average of 34.0 ± 3.4% was attained. This result shows that it is possible to attain an output with a similar larval protein content through the initial "one-off feeding method" by using microorganisms to execute fermentation. The introduction of microorganisms into larval feeding media

has been widely practiced as a means to improve the growth of BSFL. According to Gao et al. [29], the addition of *Aspergillus oryzae* into maize straw for fermentation ultimately improved the growth of BSFL and was able to obtain approximately 42% of larval crude protein. At the same time, the BSFL reared on fermented maize straw were found to contain higher amounts of monounsaturated fatty acids and polyunsaturated fatty acids, and were lower in saturated fatty acids as opposed to the control medium without exo-microorganisms. Concisely, it could be confidently deduced that the introduction of microorganisms into BSFL media through larval farming systems could promisingly enhance larval growth and, eventually, achieve more harvested larval biomasses.

**Figure 2.** Impact of different yeast concentrations inoculating CEW on corrected protein yields and protein productivities from BSFL.

The introduction of BSFL biomass into animal feed could plausibly replace the exploitation of unsustainable soybean and fishmeal. Indeed, BSFL could serve as the sole protein source, since larval biomass is generally fortified with high protein as well as fat [30]. The inclusion of BSFL into animal feed for laying hens had been found to significantly increase the production of both day and house eggs. At the same time, it has also positively impacted the characteristics of eggs and the growth of laying hens [31]. In the case of aquaculture cultivation, a partial inclusion of BSFL into feed at 25%, serving as fishmeal protein has been shown to increase the growth performances of yellow catfish by 21.7%, while also improving their immune indexes [32]. Moreover, it was reported that the replacement of fishmeal by BSFL between 28.4% and 50% into the diets of juvenile barramundi could promote fish growth, fish whole body proximate and amino acid composition [33]. A 100% replacement of fishmeal by BSFL was also possible in Jian carp cultivation, as it had been reported that there was no unfavorable impact on the growth of Jian carp. BSFL meal could be an economic and sustainable replacement for current fish diets that could circumvent both feed shortages and the increasing price of fishmeal [34]. Thus, it is recommended that BSFL biomass meal be utilized as a substitution for protein alimentation in animal feed and fishmeal in the long-term, whilst also advocating for the green and sustainable farming of land and aquatic animals, respectively.

#### **4. Conclusions**

The inoculation of yeast at different concentrations into CEW to serve as the feeding medium for BSFL rearing enhanced larval growth. For a setup initially containing 20 neonates of BSFL, a final weight of 1.145 g, a growth rate at 0.085 g/day and a rearing period of 13.5 days were achieved when BSFL were fed with fermented CEW inoculated with 2.5 wt% yeast. With an increase in yeast concentrations, the overall degradation of CEW was found to improve from 0.48 to 0.53, with the waste reduction indexes fluctuating between 0.38 and 0.40 g/day. Likewise, the protein yield from BSFL was boosted from the control (28%) to its highest value of 35% in the presence of merely 0.02 wt% yeast concentration. On the other hand, protein productivity was increased from 0.018 g/day for the control to around 0.025 g/day across 0.02 to 2.5 wt% yeast concentrations. To conclude, the growth of BSFL was promoted with the inclusion of yeast as the fermentation precursor, and the harvested larval biomass can potentially be used as a replacement of protein sources in animal feeds and fishmeals.

**Author Contributions:** Conceptualization, C.Y.W. and J.W.L.; methodology, C.Y.W. and Y.C.H.; formal analysis, P.L.S. and S.C.; investigation, Y.J.C. and M.M.; resources, C.D.H.; writing—original draft preparation, C.Y.W.; writing—review and editing, T.Y.W. and G.T.P.; supervision, J.W.L. and M.K.L.; project administration, C.Y.W.; funding acquisition, J.W.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by The Murata Science Foundation with the cost center of 015ME0-104 and the Ministry of Education Malaysia under HICoE with the cost center of 015MA0-052.

**Acknowledgments:** The administrative and technical supports provided by the members from the HICoE-Centre for Biofuel and Biochemical Research, Universiti Teknologi PETRONAS are greatly acknowledged.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Hygro-Thermo-Mechanical Responses of Balsa Wood Core Sandwich Composite Beam Exposed to Fire**

#### **Luan TranVan 1, Vincent Legrand 2, Pascal Casari 2, Revathy Sankaran 3, Pau Loke Show 4,\*, Aydin Berenjian 5,\* and Chyi-How Lay <sup>6</sup>**


Received: 15 December 2019; Accepted: 9 January 2020; Published: 13 January 2020

**Abstract:** In this study, the hygro–thermo–mechanical responses of balsa core sandwich structured composite was investigated by using experimental, analytical and numerical results. These investigations were performed on two types of specimen conditions: dry and moisture saturation sandwich composite specimens that are composed of E-glass/polyester skins bonded to a balsa core. The wet specimens were immersed in distilled water at 40 ◦C until saturated with water. The both dry and wet sandwich composite specimens were heated by fire. The mass loss kinetic and the mechanical properties were investigated by using a cone calorimeter following the ISO 5660 standard and three-point bending mechanical test device. Experimental data show that the permeability and fire resistance of the sandwich structure are controlled by two composite skins. Obtained results allow us to understand the Hygro–Thermo–Mechanical Responses of the sandwich structured composite under application conditions.

**Keywords:** sandwich composite fire; mechanical responses; moisture content; balsa core; mass loss kinetic; buckling failure

#### **1. Introduction**

The use of organic matrix composite materials has been continuously growing since the 1960s. As known to all, the material undergoes important physical and/or chemical modifications under extreme conditions, such as an appearance of metastable states or phase transitions [1,2]. Measurements in extreme conditions are facing scientific challenges to spot the properties of materials, and a technical challenge to apply new materials. Obviously, the increasing use of composites has reached a level that these materials compete with conventional materials such as steel and aluminium alloys in diverse areas, particularly aeronautics, aerospace and the shipbuilding industry due to their advantages in physical, chemical and mechanical properties [3–6].

Compared to other materials, organic matrix composites have low density, high specific stiffness and strength, good fatigue endurance and outstanding resistance to corrosion. However, there are several disadvantages compared to metals that include low impact tolerance, low fire performance and anisotropic properties [5–8]. Hence, the study on the effect of composite material properties under extreme conditions is needed to understand the behaviour and to optimize their properties. In the marine industry, the use of a sandwich structure consisting of a lightweight core made of polymer foam or balsa wood surrounded by thin stiff composite skins made of fiberglass and a major polymer as vinyl ester, epoxy or polyester, is common [5–8]. These combinations allowed a construction of an extraordinarily lightweight, durable and rigid structure. However, this type of material structure in the naval industry requires the special precaution of fire resistance. Composite sandwich materials are subject to strict regulation, and it is important to predict their thermomechanical properties as handling any applications [9–15]. The thermal degradation of materials as a composite sandwich have been widely described in detail [1,2,8,16–21]. There is limited study on the evaluation of the losses of the mechanical properties under the coupling effect of heat flux and moisture absorption. It is important to know the residual mechanical properties at room temperature of a burnt sandwich composite material in order to estimate the fire resistance of this structure after a fire exposure [9–13,15,21–25]. In this context, we were analysing the h hygro–thermo–mechanical responses of a sandwich structure composed of fiber-E-glass embedded in a polyester matrix, for the composition of the skins bonded to a balsa core. Previous studies focused on the thermal degradation of sandwich composite materials [15,23–25], and there are limited studies performed on the hygroscopically aged materials exploring the coupling hygro–thermo properties [20,22,24]. Officially, the water and temperature simultaneously could cause extreme degradation on the skins of this sandwich structure, thus the weak core would be exposed to application conditions. The resulting mechanical states of the core material can eventually induce the geometrical stability damage of such sandwich structure [13–15,23–25]. Balsa wood is widely used for cores of sandwich structures, especially in the shipbuilding industry due to its microstructure composing of long cells aligned in the axial direction which could provide the required axial strength and stiffness. However, only a few detailed studies were conducted on this subject [13–15,23–25]. Some research works focused specifically on the mechanical properties of balsa at high temperature and axial response failure under compression [1,7,8,16].

In the present work, we focus on determining the mass loss kinetic and flexural behaviour under fire of the two types of dry and wet composite sandwich samples by using a cone calorimeter and a Zwick universal testing machine. The dry sandwich samples that were obtained from the shipbuilding industry were immersed in distilled water at 40 ◦C until water saturation. Fire tests were processed with a heat flux of 50 kW/m<sup>2</sup> at different pyrolysis times. Additionally, a multi-layer analysis (skins and core) was conducted based on experimental results of the composite sandwich structure to estimate the hygro–thermo–mechanical properties of the global sandwich structure. This study enables the evaluation of the elastic modulus E and flexural load of the remaining sandwich structure material after enduring harsh working conditions such as exposure in water-fire.

#### **2. Experiment Set-Up**

#### *2.1. Sandwich Composite Materials*

The sandwich composite samples were made up of E-glass/polyester skins bonded to a balsa core by a direct infused process. These samples were cut from commercial plates that are used in the naval structure. Figure 1 indicated a studied E-glass/polyester/balsa sandwich specimen. The skins consist of E-glass fabric M450/QX868 made into a 2-plies layer surrounding the balsa core. The core was ordered to balsa wood pieces concocted in the form of about 50.0 <sup>×</sup> 30.0 <sup>×</sup> 16.0 mm<sup>3</sup> blocks. The wood fiber direction (D3) of the core is perpendicular to the composite skins. The average value of the balsa wood's density was 126 <sup>±</sup> 30 kg m−3. The coefficients es = 1.2 (mm), l = 40.1 (mm), l\* = 111.0 (mm) and e = 18.5 (mm), respectively, stand for the average skin thickness, average width, length and thickness of the sandwich specimens.

**Figure 1.** The studied E-glass/polyester/balsa sandwich composite beam specimen (**a**). Dimensions were illustrated the chosen directions of the moisture and heat flux. D1, D2 and D3 are, respectively, the transverse directions and the thickness direction (**b**).

#### *2.2. Experimental Measurements*

#### 2.2.1. Water Absorption Measurements

The sandwich specimens for water uptake were dehydrated in an oven at 50 ◦C until the weight loss was stabilized, and they were then placed into a container at room temperature for 24 h. The second step involves the complete immersion of specimens in distilled water at 40 ◦C. In order to measure the quantity of the moisture absorption, the specimens were periodically removed out of the water bath one at a time, wiped off with an absorbent cloth, and immediately their mass was weighed. Sartorius, an MCBA 100 balance with precision ranging of (0–60) g ± 0.1 mg, (61–110) g ± 0.2 mg and (111–210) g ± 0.5 mg, was used. The moisture absorbed by the material at a specific time M (t) was experimentally determined by using the following equation:

$$\mathbf{M}\_{\mathbf{l}} = \frac{\mathbf{W}\_{\mathbf{l}} - \mathbf{W}\_{0}}{\mathbf{W}\_{0}} \times 100\tag{1}$$

where, Wt is the weight of the specimen at the immersion time t and W0 is the initial weight of the specimen. The weight measurements were taken initially at an interval of one quantification per day during the first 30 days, and later with a longer periodicity, since the mass fluctuations were not as large as during these days.

#### 2.2.2. Mass Loss Kinetic Measurements

Fire-induced mass loss of the sandwich structured composite specimens were carried out using cone calorimetry (ISO 5660 standard). A radiative heat source was emitted from the cone constructed by winding an electrical resistance. The radiative source was kept at a uniform heat flux of 50 kW m−2. Gas flux was diluted with fresh air and drawn into a chimney. The ignition of the sandwich composite material was caused by a pilot spark. The temperature of the flame during a fire exposure of the material was 750 ◦C. A surface of the test sample was positioned at a distance of 26 mm from the radiative source. Figure 2 shows the diagram of the position of the cone calorimetry heating setup. The direction of the heat flux was perpendicular to a surface of the sandwich sample test. During a fire resistance test time, the mass of the test sample was recorded as a function of the combustion time. For the purpose of quickly stopping the degradation of the combustion sample, we quickly removed a holder of the sample from the calorimeter cone and placed it into a chamber under nitrogen atmosphere. A residual mechanical property measurement was performed by using a Zwick universal testing machine. (Zwick Roell Group, Ulm, Germany). The mass loss kinetic measurement was conducted for the skin alone and the sandwich structured composite samples.

**Figure 2.** Diagram of the ATLAS cone calorimetry: holder of the test specimen (**a**), conical heater (**b**).

#### 2.2.3. Mechanical Property Measurement

In order to measure a remaining mechanical property of the material after an exposure time to fire, a three-point bending flexural test was performed on the samples that underwent the thermal degradation processes. Both unaged and thermally-aged sample types were measured for the maximum flexural force reached before failure as a function of degradation time. The Zwick universal testing machine with 15 mm radius supports was used with a displacement speed of 10 mm min<sup>−</sup>1. During the test, both the flexural force and displacement of the test specimen were recorded. Figure 3 shows the supporting and the device for the three-point bending tests.

**Figure 3.** Sketch of the supporting (**a**) and the device (**b**) for the three-point bending tests of sandwich beams.

#### **3. Experimental Results**

#### *3.1. Moisture Di*ff*usion*

If balsa wood immersed in water, the moisture will absorb in two ways: the first water floods in free volumes (cracks, hollow fibers) and the second moisture diffuses the dense material. Thus, in the single balsa wood, the moisture is quickly absorbed, and the moisture content obtained is high [6,7]. In the case of balsa as the core materialin sandwich composite structures, the skins acted as waterproof barriers, and infusion resin made full an important amount of hollow fibers for the moisture's viable passage. Figure 4 showed the moisture absorption characterization in both the composite sandwich structure and the bi-blade (balsa core + skin) specimens. The polyester/E-glass fibers skins played as a boundary which limits the water diffusion in the longitudinal direction of the wood fibers. Thus, the sandwich structure absorbed the relative moisture slowly.

**Figure 4.** Diffusion kinetics of moisture at 40 ◦C in a sandwich beam (**a**) and a bi-blade (one skin and a core) (**b**).

The moisture content obtained in the sandwich material was lower than the bi-blade (balsa core + skin). This is due to the presence of one skin which limits moisture diffusion in one side of the sandwich structure. Figure 4a shows that the moisture diffusion was saturated in the sandwich structure after 400 days of immersion. For the bi-blade, the moisture saturation obtained after 75 days. This difference in the saturation time is due to the skins conducting a high restriction to the moisture penetration inside the material. Note that there is a large difference in the moisture content between both of the specimen types. The water absorption limitation corresponds to the value of the moisture content gain (Figure 4), which is approximately 180% for the bi-blade structure, and only 160% for the sandwich structure, even though the water diffusion takes place in three sides of the specimen. This measured value is consistent with values in the written works of the composite sandwich structure and single balsa [6,8,19,26].

#### *3.2. Mass Loss Kinetic*

In order to understand the hygro–thermo–mechanical responses of the sandwich structured composite, it is necessary to know the mass loss kinetic property of this structure. Thus, the two specimen types were measured: single skin and the entire sandwich structured composite. The skin consists of polyester resin and E-glass fibers, and the core is made of balsa wood impregnated with resin. The skin is practically insensitive to water uptake, whereas the water content in the core is very high, reaching more than 400% [6,7]. So, it is expected to know a fire response between the dry and the moisture saturation sandwich structured composite specimens.

In the first approximation for this paper, we consider that the fire response of the dry composite skin and the skin hygroscopically aged are similar because the skin absorbed a negligible amount of water. We therefore separated the skin from a dry sandwich composite sample (dimensions are indicated in Section 2.1), and measured its fire resistance property under 50 KW m−<sup>2</sup> heat flux and 750 ◦C. The corresponding mass loss kinetic curve is shown in Figure 5. The entire combustion time was about 200 s, and it induced a mass loss of 60% (initial mass = 10.99 g). In the first 100 s, the single skin lost about 50% of its weight. This result provides an estimation of the average mass loss rate of the composite skin of about vs = 0.05 g s<sup>−</sup>1. At the start of the combustion, the skin burned and emitted a white smoke, and subsequently released a thick, black smoke. At the end of the combustion, only the fiberglass fabrics skin remained; the resin completely disappeared.

**Figure 5.** Mass loss curves as a function of fire exposure time for dry E-glass polyester single skin (**a**) and sandwich composite beam (**b**) from cone calorimetry measurements at 750 ◦C.

For the dry sandwich specimen as shown on the Figure 5b, (initial mass = 38.06 g), the material combustion curve was investigated in two successive portions: Portion I presents the pyrolyzed top skin, the exposure time to fire was 100 s, induced loss mass 20% and the mass loss rate was equal to vts = 0.08 g s−1. Portion II respectively presents the char of the balsa core and the bottom skin, induced the loss mass about 45% during 775 s and the mass loss rate was about 0.02 g s<sup>−</sup>1. The incombustible mass was about 35% of the initial mass. Combustion products of the sandwich structured composite at the end of the degradation were left with glass fiber fabric and wood charcoal. The glass fiber fabric and wood charcoal did not degrade due to the pyrolysis temperature that was conducted at 750 ◦C, which is lower than the melting temperature of the glass fibers and the wood charcoal [2,8,11]. The obtained results of the sandwich structured composite was pyrolyzed by the cone calorimetry, and it was found that the kinetic mass loss characteristic of three elements (top skin–core–bottom skin) was discontinuous. There was a superposition of the combustion at interface between top skin–core and core–bottom skin. During the first 100 s, the top skin was degraded and retracted by the heat flux, followed by the balsa core degradation. The top skin–core progressively induced delamination, an inflammation of the balsa core on the edges of the test specimen, and on resin between the balsa wood blocks. It made the balsa core lose its mass while the degradation in the top skin continues. The equivalent mass loss curve of the sandwich structure was obtained, which is the continuous curve in Figure 5b. The overview of the degradation processes on the cross section at different time intervals of the sandwich specimens during fire exposure is illustrated in Figure 6.

Finally, with the same measurements with above dry sandwich specimens, the determination of moisture saturation specimens (immersed in water at 40 ◦C) were performed, (initial mass = 78.93 g). Based on the observation, the fire behavior of the wet sandwich sample was similar to the dry sandwich sample. During the first 100 s, the combustion of the top skin induced a mass loss about 5% and the mass loss rate was equal to vts = 0.04 g s<sup>−</sup>1. Then, the combustion of the equivalent bi-blade consisting of the balsa core and the bottom skin, induced a mass loss of 85% during 2150 s, and the mass loss rate was equal to 0.03 g s<sup>−</sup>1. The remaining mass was about 10% of the initial mass of the test sample. Figure 5 presented that the evolution of the mass loss as a function of exposure time to fire of the moisture saturation sandwich sample. The degradation mechanism in the moisture saturation sandwich sample occurred slowly, but in the dry sandwich sample it occurred rapidly. This is because the water saturation sandwich sample contains a high water quantity which might significantly slow down the material pyrolysis. Thus, the complete deterioration due to fire of the water saturation sample was achieved in the 2250 s.

**Figure 6.** Cross-sectional view of the during fire exposure of the dry specimens (**a**) and the moisture saturation specimens (**b**).

#### *3.3. Post-Hygro-Thermo-Mechanical Properties*

Following mass loss measurements, the residual mechanical properties were determined to understand the hygro–thermo–mechanical response of these materials. Three-point bending measurements were performed on dry and water saturation samples at different times of the fire exposure. This analysis allows examining the residual mechanical strength of the sandwich structured composite after fire exposure.

Thus, the first step is to measure the ability of the sandwich structure to transmit a load up to the elastic limit. Figure 7 illustrated the evolution of the force as a function of displacement for dry and water saturation sandwich samples that did not undergo heat treatment. The mechanical behaviour consists of three periods: the first (period I) is similarly a behaviour of the porous materials. The displacement increases while the force is constant. This is supposed to be due to a change of the free volume of the top composite skins that were directly subjected to the action force. Thus, the residual deformation is maintained after discharge. A linear elastic behaviour of the sandwich structured composite material was observed. The force increases linearly with the displacement. Therefore, this linear portion of the curve was selected to make definite the Young's modulus. The final portion of the curve in period I corresponds to a nonlinear behaviour of the material until the abrupt rupture of the specimen. At the beginning of this phase, the force increases which corresponds to a small displacement. This clearly demonstrates that the flexural stiffness of the sandwich structure is improved by the shear stiffness of the balsa core. The end of this phase highlights a failure mode of the balsa core. The core is major enforced to shear, and failure takes place as the critical value (shear strength) of the core material is attained by the maximum shear stress. In the case of the moisture saturation specimen, after reaching critical load, the specimen did not break completely. The large amount of water contained in the core made it more elastic, so there was then a period of internal structure reorganization with constant force (period II). Finally, the force increases linearly with the increase of displacement (period III). The end of this period corresponds to a nonlinear behaviour of the moisture material until the rupture of the specimen. The critical load of the moisture specimen is lower than this one of the dry specimen, but the elastic deformation of the moisture specimen is larger by about 1.5 times. Figure 7b shows the failure mode of dry and water saturation sandwich specimens.

**Figure 7.** Three-point bending test: Force—displacement curve of the dry and water saturation sandwich specimens (**a**), failure mode of dry sandwich specimen (**b**) and water saturation sandwich specimen (**c**).

The second step was to measure the hygro–thermo–mechanical properties of the sandwich composite beam. The failure mode of the sandwich beam is characterized by two parts. The first is the core's shear failure and the second failure is due to global buckling of the skins. The displacement obtained by the three-point bending test is therefore composed of two parts: the deflection due to global buckling and the displacement caused by shear deformation of the core. The total displacement is expressed as follows [13,22,24]:

$$\frac{\delta}{FL} = \frac{L^2}{48D} + \frac{1}{4S} \tag{2}$$

where δ is the displacement (mm) measured at mid-span under load *F* (N). *L* is the span length (mm), *<sup>D</sup>* <sup>=</sup> *IE* is the flexural stiffness of the sandwich beam (N·mm2) within *<sup>E</sup>* (MPa) is the Young's modulus I is the inertia moment, and *S* = *Gel* is the shear stiffness of the sandwich beam (N) with G being the shear modulus (MPa). The coefficients *l* and *e* are, respectively, the width and the thickness of the sandwich specimen. Equation (2) is used to determine the values of the Young's modulus *E* and the shear modulus *G* of the specimen by measuring three-point bending for five sandwich beam specimens with the different values of *L*. The corresponding curve is represented in the Figure 8 as well as the equation of the line obtained by linear regression analysis on the experimental points. The calculated values of the Young's modulus is *E* = 7.8 <sup>×</sup> 103 MPa and of the shear modulus, it is *G* = 85.9 MPa for the dry balsa core sandwich beam specimen. Results are coincident with one found in the research in the particular case for pure balsa wood or E-glass-reinforced polyester resin [2,3,6,7,12].

It is supposed that the parameter (1/(4*S*)) of the Equation (2) is corresponding to zero (Ignoring the contribution of core to the flexural stiffness). The Equation (2) is also rewritten to identify the Young's modulus *E* as follows:

$$E = \frac{L^3}{4lc^3} \times \frac{F}{\delta} \tag{3}$$

Identification of the elastic modulus *E* is performed by using the analytical resolving method, given by the Equation (3) and the experimental results *F* and δ, obtained by three-point bending test.

Finally, the residual flexural mechanical responses were investigated on the dry and moisture saturation sandwich beam specimens. The normalized evolutions of the maximum force and the Young's modulus as a function of the time exposed to fire are shown in Figure 9. These plots show the thermo–mechanical responses as the maximum force (*Fm*) and the flexural modulus (*E*) decreases followed exponential law. This law had also been observed for other composite materials [2–4,9,27]. The dry and water saturation sandwich structured composite materials lost their mechanical properties during 100 s of exposure time to fire. This time characteristic corresponds to the thermal degradation of the top skin. When the top skin is degraded to weaken the sandwich structure, the maximum force and the flexural modulus decreases like the exponential law. The dependence of the thermo–mechanical properties on the moisture content is visible in Figure 9. These obtained experimental results indicate that the sandwich structured composite specimen and the thermo–mechanical responses significantly related to the degradation of the top skin, and they are not assured of the moisture content in the balsa core. The moisture concentration only strongly influences the duration of the fire exposure. The normalized force curve expresses a light disparity between dry and moisture materials and the normalized Young's modulus indicated the similar trend. Thus, it allows us to express that the relative elasticity of the sandwich material is independent of the internal water concentration, but it is significantly influenced by the duration of the fire exposure time. Figure 10 represents the cross-section of the sandwich structured composite beam after 100 s exposed to fire for the dry and the moisture saturation specimens to confirm well the weakened sandwich structure due to the degradation of its top skin.

**Figure 8.** The curve of the δ/(*FL*) = *L*<sup>2</sup> to determine the Young's modulus (*E*) and the shear modulus (*G*).

**Figure 9.** Post-residual flexural mechanical responses as a function of fire exposure time: normalized residual force (**a**) and normalized residual modulus (**b**).

**Figure 10.** Cross-section of the sandwich structured composite beam after 100 s exposed to fire for the dry specimen (**a**) and the moisture saturation specimen (**b**).

#### **4. Conclusions**

The diffusive behaviour of the sandwich specimen and the bi-blade specimen (one skin and core) were investigated to understand the moisture impervious barrier significance of the skin. Thermal degradation rates under fire were identified for each skin and bi-blade sandwich specimens. The residual flexural mechanical responses as a function of the fire exposure time under a heat flux of 50 kW m−<sup>2</sup> and 750 ◦C were analysed for the dry and moisture saturation sandwich specimens. The obtained results show that the sandwich structure undergoes a rapid thermo–mechanical degradation in the first 100 s of fire exposure, and this degradation is strongly influenced by the degradation of the top skin. Although the relative elasticity modulus of the sandwich structure composite material is independent of the moisture content, it is strongly influenced by the fire exposure time.

**Author Contributions:** Conceptualization, L.T.; Data curation, V.L.; Methodology, Software and Resources: P.C.; Original Draft Writing: L.T. and R.S.; Review & Editing: R.S.; Supervision: P.L.S. and A.B.; Project Administration: P.L.S. and C.-H.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Acknowledgments:** Luan TranVan is indebted to the Ministry of Education and Training (MOET) for funding under grant number KYTH-74 (B2017.DNA.11). We would like to thank Didier Andeler and Philippe Frétaud (IUT Saint-Nazaire, France) for technical support.

**Conflicts of Interest:** The authors declare no conflict of interest in any choice of research project funding sponsors; design of the study; in the collection, analyses or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Multivariate Analysis and Machine Learning for Ripeness Classification of Cape Gooseberry Fruits**

**Miguel De-la-Torre 1, Omar Zatarain 1, Himer Avila-George 1,\*, Mirna Muñoz 2, Jimy Oblitas 3, Russel Lozada 4, Jezreel Mejía <sup>2</sup> and Wilson Castro 3,5**


Received: 1 November 2019; Accepted: 3 December 2019; Published: 5 December 2019

**Abstract:** This paper explores five multivariate techniques for information fusion on sorting the visual ripeness of Cape gooseberry fruits (principal component analysis, linear discriminant analysis, independent component analysis, eigenvector centrality feature selection, and multi-cluster feature selection.) These techniques are applied to the concatenated channels corresponding to red, green, and blue (RGB), hue, saturation, value (HSV), and lightness, red/green value, and blue/yellow value (L\*a\*b) color spaces (9 features in total). Machine learning techniques have been reported for sorting the Cape gooseberry fruits' ripeness. Classifiers such as neural networks, support vector machines, and nearest neighbors discriminate on fruit samples using different color spaces. Despite the color spaces being equivalent up to a transformation, a few classifiers enable better performances due to differences in the pixel distribution of samples. Experimental results show that selection and combination of color channels allow classifiers to reach similar levels of accuracy; however, combination methods still require higher computational complexity. The highest level of accuracy was obtained using the seven-dimensional principal component analysis feature space.

**Keywords:** Cape gooseberry; color space selection; color space combination; food engineering

#### **1. Introduction**

In the advent of the fourth industrial revolution, the growing tendency of automation of human activities encourages the use of robotic systems in the food industry [1]. In this context, the automation of food packing processes is essential to accelerating the production rate, and reducing human contact and possible contamination of food products. Moreover, machine vision techniques allow robotic systems to retrieve information from food products, using different sensors that depend on the particular characteristics to be measured, and each sensor represents an additional cost to construct an information retrieval system. For instance, an application that requires such automation systems is the classification of Cape gooseberry (*Physalis peruviana* L.) fruits according to their level of ripeness. Current industry practices address this repetitive task through visual inspection of color, size, and

shape parameters [2]. While automated sorting systems based on computer vision techniques have been proposed to improve production methods and provide high-quality products, their operation relies on classification algorithms that consider either different color spaces or a combination of them [3,4].

The most common representation of color images employed by computer vision systems is a combination of the three primary colors: Red, green, and blue (RGB). The triplet with the values for each primary color is typically considered as a coordinate system with either Euclidean, Mahalanobis, Hamming, or a different metric of distance. In such a three-dimensional coordinate system, each point (e.g., 3D vector) represents a different color in the visible spectrum. Other color spaces different than RGB are commonly employed, providing different three-dimensional representations, and can be classified into three categories according to [5]: Hardware-orientated spaces, human-orientated spaces, and instrumental spaces. In the first category, hardware-orientated spaces (e.g., RGB, YIQ, and CMYK ) are defined based on the properties of the hardware devices used to display images. On the other hand, human-orientated spaces (e.g., HSI, HSL, HSV, and HSB ) are based on hue and saturation, following the principles of an artist and based on inherent color characteristics. Finally, instrumental spaces (e.g., XYZ, L\*a\*b\*, and L\*u\*v\* ) are those used for color instruments, where the color coordinates of an instrumental space are the same on all output media.

As will be considered in Section 2, the color spaces that are most commonly employed in the classification of fruits are RGB, L\*a\*b\*, and HSV. However, the accuracy of the same classifier on the same dataset may vary from one color space to the other. Some authors have investigated these differences in classification accuracy due to the variation of the distribution of pixels in distinct color spaces or the use of different segmentation techniques. According to [6,7], the practice of color measurement in food engineering, the L\*a\*b\* color space, is the most commonly used. The main reasons are related to the uniform distribution of colors and because the L\*a\*b\* is perceptually uniform (i.e., equal changes in data are visually perceived as equal changes in the color space). However, it is known that color spaces like RGB, L\*a\*b\*, and HSV are equivalent up to a transformation.

Regardless of the color space used for classification, the objective of classifiers applied to fruit sorting consists of finding a criterion to separate samples from one or other ripeness levels in the so-called feature space. The goal is to establish a decision boundary that may be applied as a fixed borderline between categories. Supervised classifiers employ labeled samples to learn a model that is used to predict a category in new, never seen unlabeled samples. Supervised classifiers commonly employed in the food industry include support vector machines (SVM), k-nearest neighbors (KNN), artificial neural networks (ANN), and decision tree (DT) [8,9].

In practice, any pattern classifier may be employed, presenting a trade-off between accuracy and complexity. While the equivalence between color spaces is well-known [10], it has been found that different color spaces allow the same classifier to produce distinct classification rates, due to variations in the distribution of color samples [3,11]. Moreover, the combination of color spaces using multivariate analysis may provide a feature space where an increase in classification accuracy is possible. For instance, in [3], a methodology to compare different combinations of machine learning techniques and color spaces (RGB, HSV, and L\*a\*b\*) was proposed in order to evaluate their ability to classify Cape gooseberry fruits. The results showed that the classification of Cape gooseberry fruits by their ripeness level was sensitive to both the color space and the classification technique used. The models based on the L\*a\*b\* color space and the support vector machine (SVM) classifier showed the highest performance regardless of the color space. An improvement was obtained by employing principal component analysis (PCA) for the combination of the three-color spaces at the expense of increased complexity. An extension of such a study was proposed in [4], where a supervised multivariate analysis method was compared with previous results (linear discriminant analysis, LDA).

In this paper, an extension of previous work described in [3,4] is proposed to compare multivariate analysis methods and machine learning techniques for ripeness classification. The color channels from RGB, HSV, and L\*a\*b\* color spaces were concatenated to spam a nine-dimensional feature space. The five multivariate methods employed to combine information from the nine color channels include PCA, LDA, independent component analysis (ICA), multi-cluster feature selection (MCFS), and eigenvector centrality feature selection (ECFS). In the last case, selection methods applied to find the most relevant features for classification were MCFS and ECFS. The main contribution of this paper is the use of multivariate techniques to find an appropriate feature space for classification.

The manuscript is organized as follows. Section 2 summarizes the most recent works published on ripeness classification, including diverse approaches and methodologies. Some of the most popular methods were selected for this comparison, and Section 3 describes the material and methods employed in experiments to compare the distinct approaches. Section 4 presents the results and a discussion on the relevant findings. Finally, Section 5 presents conclusions and future work.

#### **2. Ripeness Classification**

As reported in the literature, different color spaces have been used to create automated fruit classification systems, presenting different levels of accuracy that are related to both, the color space and the sorting algorithm. Table 1 shows common methods and color spaces reported in the literature used to classify distinct fruits according to their ripeness level.

**Table 1.** Color spaces and classification approaches for fruit classification in literature. NA stands for non-available information MDA stands for multiple discriminant analysis, QDA for quadratic discriminant analysis, PLSR for partial least squares regression, RF for the random forest, and CNN for the convolutional neural network. The table was taken from [3] and updated with new findings.


According to Table 1, the most common color space used for classification is RGB, with 50% of the works, followed by L\*a\*b\* with 32%, and HSV with 14%. Similarly, the most common classifiers are ANN and SVM, with 32% of the experiments reporting results using color spaces that include RGB, L\*a\*b\*, and HSV. The accuracy obtained by each approach depends on the experimental settings and are not comparable at this point. However, reported accuracy ranges between 73 and 100 percent.

#### *2.1. Methods for Color Selection and Extraction*

The distribution of samples in the feature space depends on the measurements obtained from sensors, and in this case, the color channels for the distinct color spaces. The search for the color channels that are most relevant for classification is important to help classifiers to find the decision frontier between classes. Features that are noisy or not relevant may difficult classification problems and may conduce to a low performance even by the most sophisticated classifiers. Finding a subset of the *k* most relevant features, either by selecting them or applying feature extraction techniques, favors the reduction of redundant and irrelevant information. The so obtained *k*-dimensional feature space employed for classification instead of the original *d*-dimensional feature space is suitable to facilitate finding a separation criterion between classes. Whereas feature extraction algorithms find a mapping to a new feature space, feature selection methods aim to select a subset of vectors that spans a feature subspace that facilitates classification.

Feature extraction approaches can be categorized according to the use of data labels in supervised and unsupervised. Unsupervised feature extraction techniques consider the underlying distribution of data solely, and aim to find a mapping to a new feature space with desired characteristics. An example of unsupervised methods is PCA, which employs the Eigenvectors of the covariance matrix of samples to maximizes their spread in each new axis. Additionally, supervised approaches employ the information from class-labels to find the mapping. A representative supervised approach is the linear discriminant analysis (LDA), that aims to maximize the spread of samples distinct classes, and minimize the within-class spread.

Analogously, supervised feature selection considers class labels to find the most relevant features, and unsupervised feature selection strategies are based exclusively on the underlying distribution of samples. The selection of the subset of the most relevant features is a computational expensive combinatorial problem. The optimally of an algorithm to find a good enough feature subset may depend on the strategy followed for ranking or selection of features.

In feature selection and extraction, the problem can be stated by considering a set of points (sample tuples or feature vectors) *<sup>X</sup>* = {*x*1, *<sup>x</sup>*2, ..., *xN*}, where *xi* ∈ *<sup>R</sup>d*. The algorithms for feature extraction and selection, find a new set *X* = {*x* <sup>1</sup>, *x* <sup>2</sup>, ..., *x <sup>k</sup>*}, where *x <sup>i</sup>* ∈ *<sup>R</sup>k*, and *<sup>k</sup>* ≤ *<sup>d</sup>* is the new dimension of the feature vectors.

#### *2.2. Principal Component Analysis (PCA)*

The PCA method is applied to find a linear transformation that finds the directions of maximum variance data. Sample patterns are projected onto a new feature space, and the axes with more explained variance provide a distribution that facilitates the separation between classes. The algorithm is shown in the Figure 1 depicts the general procedure to transform data samples from X to their new representation in the *k*-dimensional feature space X'. The new *k*-dimensional feature space corresponds to the *k* eigenvectors of the covariance matrix *C*. The axis with the highest eigenvalues expresses a higher explained variance.


**Figure 1.** The procedure followed by principal component analysis (PCA) to map the input data samples to the new *k*-dimensional feature space.

#### *2.3. Linear Discriminant Analysis (LDA)*

The LDA method allows obtaining and applying a linear transformation that finds the directions of maximum variance input data samples. The main difference with PCA is that LDA aims to minimize intraclass variability, whereas it maximizes interclass variability employing class labels. The main limitation is that the number of classes bounds the number of features in the new *k*-dimensional space (e.g., 1 < *k* < *c*, where *c* is the number of classes). This limitation makes this approach advantageous only with a high number of classes, and unpractical for data with a few classes (e.g., *c* << *d*). The procedure followed in computing the mapping and transforming the data is shown in Figure 2.

**Figure 2.** The procedure followed by linear discriminant analysis (LDA) to map the data samples to the new *k*-dimensional feature space.

#### *2.4. Independent Component Analysis (ICA)*

The ICA method finds underlying components from multivariate statistical data, where data is decomposed into components that are maximally independent in an appropriate sense (e.g., kurtosis and negentropy). The difference between PCA and LDA is that low-dimensional signals do not necessarily correspond to the directions of maximum variance; rather, the ICA components have maximal statistical independence and are nongaussian. In practice, ICA can be used to find disjoint underlying trends in multi-dimensional data [35].

The algorithm is shown in the Figure 3 depicts the procedure followed by the FastICA algorithm to obtain the independent components from *X*, using kurtosis as a measure of non-gaussianity. In this case, dimensionality reduction is not obvious, given that there is no measure of how important a particular independent component is. The relevance of the individual features obtained with PCA and LDA is given by the algorithms shown in Figures 1 and 2, respectively. In the case of ICA, feature selection techniques may be employed to provide a relevance level for each of the features that result from the transformation, as described in Sections 2.5 and 2.6.


**Figure 3.** The procedure followed by FastICA to map data samples in *X* to the new feature space that respects nongaussianity using kurtosis.

#### *2.5. Eigenvector Centrality Feature Selection (ECFS)*

The feature selection via eigenvector centrality is a supervised method that ranks features by identifying the most important ones. It maps the selection problem to an affinity graph with features as nodes and assesses the rank features according to the eigenvector centrality (EC) [36].

The algorithm shown in the Figure 4 presents the method to rank and select the most relevant features from the data samples *X*. While this does not constitute a proper transformation in terms of linear algebra, every sample is represented in a new *k*-dimensional feature space with the highest-ranked features.


**Figure 4.** The procedure followed by eigenvector centrality feature selection to select the variables that constitute the new feature space.

#### *2.6. Multi Cluster Feature Selection (MCFS)*

Multi-class feature selection (MCFS) is an unsupervised technique that aims to find those features that preserve the multi-cluster underlying structure of the samples used for training [37]. Given that the number of clusters is unknown a priori, it is a good practice to explore distinct values to find a good feature subspace. The most relevant features are found following the procedure shown in the algorithm shown in the Figure 5.

**Figure 5.** The procedure followed by eigenvector centrality feature selection to select the variables that constitute the new feature space.

While the simplest method to choose W was presented in Step 1, other methods exist that range between accuracy and complexity (See [37]). According to the authors, the default number of nearest neighbors is *p* = 5, and the default number of eigenfunctions is *K* = 5. This last parameter *K* usually influences the accuracy of the algorithm and should be optimized before usage.

#### *2.7. Classification for Fruit Sorting*

According to Table 1, some of the most popular supervised classifiers in fruit sorting are the artificial neural networks (ANN), decision trees (DT), support vector machines (SVM), and k-nearest neighbor (KNN). These classifiers were used in this paper for the experimental settings. While these techniques have been present in the literature for many years now, see [3,38], their usage in practice increased due to their capacity to address diverse real-world problems.

ANN is a non-linear classifier that simulates biological neural networks. A common implementation of ANN corresponds to the probabilistic ANN, which produces an estimated posterior probability for each input sample to belong to any of the classes, and the max function allows to select the most likely class. In this research, the Matlab's Neural Network Toolbox was used to implement the probabilistic ANN classifier, byways of the newpnn function, tunned to optimize hyperparameters using linear search.

DT is a tree-based example of the knowledge used to represent the classification rules. Internal nodes are representations of tests of an attribute; each branch represents the outcome of the test, and leaf nodes represent class labels. In this paper, the Matlab's Machine Learning Toolbox (MLT) was used the train and simulate DTs, using the Classification & Regression Trees (CART) algorithm to create decision trees, with the fitctree and predict functions. The function fitctree employes standard classification and regression trees algorithm to create DTs.

SVM is a non-parametric statistical learning classifier that constructs a separating hyperplane (or a set of hyperplanes) in a high-dimensional feature space. Some versions use the so-called kernel trick to map data to higher dimensional feature space and find the separating hyperplane there. The functions fitcecoc and predict functions were used for simulations, both implemented in Matlab's MLT. The fitcecoc function was tunned to use a linear kernel and Bayesian hyperparameter optimization.

KNN is a non-parametric classifier that keeps all training samples, and prediction is based on the number of closest neighbors belonging to a class. Given an input sample, the distance to all stored samples is computed and compared to all pre-stored samples, presenting a high computational complexity at prediction. For simulations, the fitcknn and predict functions from Matlab's MLT were used. This function employs Bayesian hyperparameter optimization.

#### **3. Materials and Methods**

For experiments, a set of 925 samples of Cape gooseberry fruits were collected from a plantation located in the village of El Faro, Celendin Province, Cajamarca, Peru (UTM: −6.906469, −78.257071). Fruit samples were homogeneously disposed on a conveyor belt used in a production line (160 × 25 cm,

and 80 cm high). A Halion-HA-411 VGA webcam was used for image capture, which provides RGB raw images in JPG format. The resolution of the resulting images is 1280 × 1720 pixels. The camera was fixed 35 cm over the conveyor belt, and the captured scene was covered with black matte panels to reduce variations in light, as implemented by Pedreschi et al. [39]. The illumination system included two long fluorescent tubes (Philips TL-D Super, cold daylight, 80 cm, 36 W) that were placed on both sides of the conveyor belt. Additionally, a circular fluorescent tube (Philips GX23 PH-T9, cold daylight, 21.6 cm, 22 W) was located at the top of the setting. Images captured with the camera were stored on a portable computer running Matlab to control image acquisition and data analysis.

Seven levels of ripeness were employed for visual classification, following the standard for Cape gooseberry, and the visual scale proposed in [40] and shown in Figure 6. The process followed for evaluation is depicted in Figure 7. Images captured from the conveyor belt (step 1) were employed to find the regions of interest corresponding to Cape gooseberry fruits in the image, employing standard segmentation techniques (steps 2 and 3); the size of resulting regions depends on the size of the actual fruit. Color versions of segmented fruits were labeled by five experts according to the categorization proposed by Fischer et al. in [40] (step 4). One-color sample was selected for each fruit region in each of the RGB, HSV and L\*a\*b\* color spaces, by computing the average for each color channel; and a nine-dimensional feature vector was built through concatenation: *x* = [*R*, *G*, *B*, *H*, *S*, *V*, *L*∗, *a*∗, *b*∗] *T* (step 5). Then, multivariate analysis techniques for feature extraction/selection were applied to the set of feature vectors (step 6), and the resulting samples were organized for five-fold cross-validation. Four standard classifiers were trained (step 7) and performance evaluation computed (step 8).

**Figure 7.** Experimental process followed to evaluate the system with distinct feature extraction/selection methods and different classifiers.

The performance of the seven-class classifiers was evaluated using the F-measure, as defined in [3]. First, the confusion matrix is computed according to the responses of each classifier, and true positives (*TPi*), false positives (*FPi*), true negatives (*TNi*), and false negatives (*FNi*) are obtained for each class *i*, using the elements *Nij* of the confusion matrix. Class-specific precision and recall are computed using Equations (1) and (2), respectively.

$$Precision\_{i} = \frac{TP\_{i}}{TP\_{i} + FP\_{i}} \tag{1}$$

$$Recall\_i = \frac{TP\_i}{TP\_i + FN\_i} \tag{2}$$

Finally, the multiclass F-measure was used for comparison along with the experimental results, due to its representativeness of the classification performance on target classes (Equation (3)).

$$F-measure\_i = 2 \times \frac{Precision\_i \times Recall\_i}{Precision\_i + Recall\_i} \tag{3}$$

The three analyses followed to characterize the performance of the system started by fixing the classifier (e.g., SVM). First, the *k* (number of clusters) was explored in order to find the *k* that allows the highest classification performance. Then, the size of the feature space was explored in terms of average F-measure. The last analysis explores the performance using the parameters found in previous steps, and the four classifiers presented in Section 2 ANN, SVM, DT, and KNN.

#### **4. Experimental Results**

As explained in Section 2.6, MCFS needs a search to find the number of clusters that maximizes the classification performance. The number of characteristics was set to seven, to make it comparable with previous results using PCA [3].

Figure 8 shows the box plots that summarize the distribution of performance for the SVM classifier trained with seven color channels (features) selected with the MCFS algorithm. The parameter that controlled the number of clusters was moved from 1 to 9 (i.e., the maximum number of possible features). In most cases, the median of the F-measure was maintained around 71.75, and only two cases were different: 2 and 9. Using nine clusters appears to provide lower performance related to the creation of an excessive number of clusters. On the other hand, using only two clusters for feature selection seems to provide a higher level of accuracy. However, and regardless of the median accuracy, the variability between cases shows a difference that makes no significant difference in using a different number of clusters. Therefore, in the following experiments, the number of clusters is fixed at 2, and it explored other variables.

**Figure 8.** Boxplots corresponding to the F-measure for nine distinct values of the parameter controlling the number of clusters in multi-cluster feature selection (MCFS). The number of characteristics was fixed at 7, and the experimentation follows a five-fold cross-validation strategy.

#### *4.1. Analysis of Feature Spaces*

Table 2 shows the average F-measure and standard deviation corresponding to the outputs generated by the SVM classifier after training on feature spaces selected or extracted with the different methods explained in Section 2. The feature space was varied from *d* ∈ {1...9} features, generating nine *d*-dimensional feature spaces for classification. The performance was estimated using a five-fold cross-validation strategy to obtain a measure of dispersion.

**Table 2.** Average F-measure of the support vector machine (SVM) classifier applied to distinct feature spaces obtained with the four methods for feature extraction/selection. IC stands for the independent component, bold numbers symbolize the highest F-measure obtained for each method, and numbers in parenthesis symbolize standard deviation.


Results showed in Table 2 evidence that was using seven features provide a level of performance that is similar either using MCFS or PCA. On the other hand, ECFS and LDA present the highest level of performance using five and six features, respectively, with a slightly lower average performance compared to PCA and MCFS. Moreover, in all cases, using more than three features allows classifiers to obtain a significantly higher performance with a lower standard deviation. In that sense, when a feature space with more color channels—or features—is employed, the SVM classifier presents a higher and more stable classification performance, at the expense of the evident increase in computational complexity. This is evident either if features are selected (e.g., MCFS, ECFS) or extracted (e.g., PCA, LDA). Different behavior is presented when ICA is employed for feature extraction, due to the strategy to find the independent components instead of the axis of maximum spread.

In the hypothetic case that only three-color channels were allowed, and these channels could be arbitrarily chosen from the nine provided by our three-color spaces, in this case, a selection method should be used. Then, a quick look at the 3D column of Table 2 evidences that the MCFS provided a better channel selection, achieving the highest performance level with a feature space composed of channels [*L*∗, *V*, *H*].

#### *4.2. Performance across Classifiers*

The comparative of performance in terms of F-measure, between the ANN, DT, SVM, and KNN classifiers, evaluated on the best d-dimensional feature space found in Section 4.1, is presented in Figure 9. The distinct feature spaces provided a different optimal number of characteristics, and those features were employed in each case. In particular, seven features were selected for PCA and MCFS, six features for LDA, and five features for ECFS. In the case of ICA, all nine features were employed to obtain the highest level of performance.

**Figure 9.** Boxplots representing the distribution of F-measure performance for the feature selection/extraction approaches, using four different classifiers.

As shown in Figure 9, the SVM classifier F-measure outperforms the rest of the approaches, and only ANN performance is close to SVM performance on the six-dimensional LDA feature space. The highest level of F-measure achieved by ANN is shown in the space extracted with LDA. In general, in these settings, the performance of all classifiers presents its highest level on the seven-dimensional PCA space. The settings suggest that PCA provides a feature space that facilitates the work of a classifier after combining information from multiple color spaces. On the other hand, focusing on the two feature selection methods, a similar level of performance is provided by all classifiers, without significant difference. The only case where KNN presents a lower performance is on the nine-dimensional ICA feature space. However, if a minimum number of features is required for a

given application (e.g., to reduce computational complexity and cost), a feature selection method may provide the means to select a few color channels (sensors), at expenses of a reduction in performance.

#### **5. Conclusions and Future Work**

In this paper, an extension of a food packaging process was proposed for Cape gooseberry fruit sorting, according to its ripeness level. As a difference from previous works, five techniques from the multivariate analysis were employed to find the feature spaces that facilitate classification. Whereas PCA, LDA, and ICA provided mapping to a new feature space, the two selection methods (MCFS and ECFS) provided the most relevant features for classification. The configuration of the experiments provided a realistic scenario, including accommodating Cape gooseberry fruits with distinct levels of ripeness on a conveyor belt, that were captured with a VGA camera located on top. Segmentation and manual sorting were performed before feature extraction/selection and classification. Four classifiers commonly employed in literature for ripeness classification were compared, including ANN, SVM, DT, and kNN.

Results reveal that selection and extraction methods allow classifiers to reach similar levels of accuracy, but feature extraction methods require an increased computational complexity. This evidence must be considered in a final implementation, and the real-time performance of the whole system should be observed once running with the distinct algorithms on the selected computational platform.

Considering the classifiers, the SVM classifier outperformed the rest in terms of F-measure regardless of the feature space. This may respond to the organization of samples in the feature space, and the capacity of the Bayesian optimization on SVM to find a good separating hyperplane. Moreover, the four classifiers employed in the test presented the highest level of accuracy on the seven-dimensional PCA feature space. This combination of 7-D PCA feature space and the SVM classifier should be considered when a final implementation. However, the hyperparameters for this (and other classifiers) were fixed before training, and some optimization may allow finding a higher level of performance for the distinct classifiers used in experiments.

On the other hand, the lowest level of accuracy was achieved on the one-dimensional feature space, employing the ICA feature extraction technique without a feature selection method. This evidences the need for a feature selection method when ICA is employed for finding new feature spaces with independent spanning vectors. On the opposite, the highest level of accuracy for a one-dimensional feature space was obtained with the MCFS channel selection, obtaining an F-measure of 64.74 (0.07), with a single feature (L\*) using the SVM classifier. This result suggests that the L\* color channel from the L\*a\*b\* perceptually uniform color space is the feature that provides the highest degree of separation between classes. The L\* feature, combined with the other six features (R, G, B, b\*, H, and V), allows it to obtain a performance that is similar to PCA with the same seven features.

The results obtained in the experiments suggest some paths for further research. Future work may include the use of distinct and more sophisticated algorithms for feature selection and extraction that may be explored and combined to find the best combination for a particular application. Similarly, other algorithms for classification may be tested in this configuration, such as those employing deep learning and large data sets. Additionally, optimization techniques like particle swarm optimization or evolutionary algorithms may be employed to find the best hyperparameters for the application.

On the other hand, information fusion techniques like classifier combination strategies may also enhance the establishment of the decision borderlines between classes, with the inherent performance increase. Finally, another kind of problem may benefit from feature extraction and selection techniques in food engineering, like using multi- or hyperspectral sensors to measure the level of ripeness of Cape gooseberry or any different type of fruit.

**Author Contributions:** Individual contributions for the authors are as follows: conceptualization, W.C., H.A.-G. and M.D.-l.-T.; methodology, J.O., M.M. and J.M.; software, M.D.-l.-T. and J.O.; validation, W.C. and H.A.-G.; formal analysis, M.D.-l.-T. and O.Z.; resources, R.L., M.M. and J.M.; data curation, H.A.-G. and M.D.-l.-T.; writing–original draft preparation, M.D.-l.-T.; writing–review and editing, O.Z., H.A.-G. and W.C.; visualization, M.D.-l.-T. and H.A.-G.; supervision, M.M. and J.M.; project administration, H.A.-G.; funding acquisition, W.C.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflicts of interest.

#### **References**


c 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Recovery of Protein from Dairy Milk Waste Product Using Alcohol-Salt Liquid Biphasic Flotation**

#### **Pei En Tham 1, Yan Jer Ng 1, Revathy Sankaran 2, Kuan Shiong Khoo 1, Kit Wayne Chew 3,\*, Yee Jiun Yap 3, Masnindah Malahubban 4, Fitri Abdul Aziz Zakry <sup>4</sup> and Pau Loke Show 1,\***


Received: 25 October 2019; Accepted: 19 November 2019; Published: 21 November 2019

**Abstract:** Expired dairy products are often disposed of due to the potential health hazard they pose to living organisms. Lack of methods to recover valuable components from them are also a reason for manufactures to dispose of the expired dairy products. Milk encompasses several different components with their own functional properties that can be applied in production of food and non-food technical products. This study aims to investigate the novel approach of using liquid biphasic flotation (LBF) method for protein extraction from expired milk products and obtaining the optimal operating conditions for protein extraction. The optimized conditions were found at 80% concentration ethanol as top phase, 150 g/L dipotassium hydrogen phosphate along with 10% (*w*/*v*) milk as bottom phase, and a flotation time of 7.5 min. The protein recovery yield and separation efficiency after optimization were 94.97% and 86.289%, respectively. The experiment has been scaled up by 40 times to ensure it can be commercialized, and the protein recovery yield and separation efficiency were found to be 78.92% and 85.62%, respectively. This novel approach gives a chance for expired milk products to be changed from waste to raw materials which is beneficial for the environment and the economy.

**Keywords:** milk; protein; liquid biphasic flotation; dairy waste; recovery

#### **1. Introduction**

A large quantity of dairy waste is produced per annum in every country. Taking UK as an example, a total of 330,000 tons of milk waste is produced annually with approximately 90% of the total waste produced from homes. This is equivalent to 490 million pints nationwide or 18.5 pints per household. Milk should be kept at the right temperature to prevent it from spoiling before the expiry date. However, the typical household UK fridge operates at a temperature that is 2 ◦C warmer than the recommended storage temperature of milk, which is between 0 and 5 ◦C [1]. This amount of milk waste creates an environmental problem as it creates greenhouse gas emissions equivalent to approximately 20,000 cars annually [2].

Milk contains approximately 87.4% of water and 12.6% of milk solids. Fats makes up 3.7% of the 12.6% milk solids while the remaining 8.9% is 3.4% of proteins, 4.8% lactose, and 0.7% minerals. Of the proteins in milk, 80% is Casein and the remaining 20% is Whey protein [3]. Casein is chiefly phosphate-conjugated and mainly consists of calcium phosphate-micelle complex. Whey protein is a collection of a globular proteins with a high level of α-helix structure and the acidic-basic and hydrophobic-hydrophilic amino acids are distributed in a fairly balanced form. Whey proteins have substantial levels of secondary, tertiary, and quaternary structure. They are heat-labile stabilizing their protein structure through intermolecular disulfide linkage [4]. The proteins in milk are considered to be complete as they contain all types of essential amino acids in amounts that match the amino acid requirements. They are used as a standard reference for proteins to compare with other food proteins due to their high quality. Branched-chain amino acids contents such as valine, isoleucine, and leucine in milk are also higher than many other foods [4].

Since the conventional technique for extracting bioactive compounds need longer extraction time yet cost-consuming with complex scale-up, the liquid biphasic flotation (LBF) method was proposed [5]. LBF system is an integration of the adsorptive bubbles floatation system, where the biphasic system is supported with air bubbling to transport the biomolecules from one phase to another. The surface-active compound of biomolecules present will be absorbed onto the surface of ascending gas bubble and be brought from the bottom phase to the top organic phase [6]. LBF is formed by combining an immiscible polymer and a salt solution. Addition of salt to water will cause segregation of ions into their preferred water structuring [7]. Aqueous biphasic systems will occur when certain solutes cause an aqueous solution to fully separate into two aqueous phases. The basic aqueous two-phase system (ATPS) phasing strategy is based on the separation of proteins into one phase with the contaminants being present in the other phase. The smaller biomolecules will be present mostly at the bottom phase which also can be known as the salt-rich phase. Whereas proteins will be brought up to the top phase [8]. Polarity is believed to play a role in the separation; molecules with lower polarity will be partitioned to the top phase, while molecules with higher polarity will be partitioned to the bottom phase. This aqueous liquid–liquid two-phase system is more widely used in the extractive separation of labile biomolecules such as proteins. This system operates under mild conditions due to the low interfacial tension between the two phases, achieving small droplet size, large interfacial areas, and efficient mixing under very gentle stirring and rapid partition.

LBF is a well-known method for the separation, concentration, and purification of biological material, particularly for protein, enzyme, and DNA [9]. Extraction using LBF is based on the separation of biomolecules between the two aqueous phases [10]. Much work has been done by using LBF to exploit and study the behavior of the aqueous rich phase and driving forces which will affect the partitioning of biomolecules in the separation process. These systems were based on aqueous mixtures of two incompatible polymers, such as polyethylene glycol (PEG), dextran, and/or maltodextrin [11]. Since then, many immiscible aqueous systems were found by using hydrophilic solutions. However, other types of LBF, with components of different phase, had been focused on to increase the mass transfer rates and the selectivity of certain biomolecules. Ionic liquids [12], inorganic salts, and carbohydrates are three examples of solutes used in LBF [13]. These molecules were applied in the separation or purification of a wide range of compounds, including proteins, enzymes, antibiotics, organic acids, and many other bio- or synthetic molecules [13].

LBF is a very promising method, and it indicates a great potential for a wider usage in partitioning, concentrating, and purification of labile biology products from natural sources or fermentation broths, as well as in enzyme technology during industrial or laboratory production of enzymes. LBF is an integration of the principles of ATPS but with additional bubbling action to enhance the separation of biomolecules. This integration will utilize the adsorptive gas bubble separation technique in which the biomolecules with surface-active sites in the bottom aqueous phase are selectively adsorbed onto the surface of the ascending gas bubbles which are then collected in the immiscible top aqueous phase. With this, water soluble biomolecules can be separated from their crude aqueous extracts [9].

A detailed study was made with aims to obtain optimal operating conditions for the extraction of protein from expired milk using alcohol/salt LBF. The effect of milk concentration, type of salt, type of alcohol, concentration of salt solution, concentration of alcohol solution, pH, flotation time along with a scaled-up LBF system were studied. Up to current date, no study has been made on recovery of protein from expired milk using alcohol/salt LBF. Partitioning of the milk protein into the alcohol phase through LBF using low-cost and recyclable phase forming components would lead to a cost-efficient protein recovery process. Additionally, alcohol-salt LBF has a low viscosity, easier constituent for recovery and short settling time. As such, this approach would enable the production of milk protein to be economically feasible and sustainable. This study has led to a novel discovery of liquid biphasic flotation application for protein extraction from milk waste with economic processes that will be beneficial at the industrial scale.

#### **2. Materials and Methods**

#### *2.1. Materials*

Food grade alcohols of ethanol, 1-propanol, 2-propanol (R&M Chemicals, Selangor, Malaysia) were used as the extraction solvents of proteins. Salts for the bottom phase that are utilized in this study were ammonium sulphate [(NH4)2SO4], di-potassium hydrogen phosphate (K2HPO4), sodium sulphate (Na2SO4), di-sodium hydrogen phosphate (Na2HPO4) and magnesium sulphate (MgSO4) purchased from R&M Chemicals (Selangor, Malaysia). Bradford reagent was used to quantify protein concentration in the two solutions (top phase and bottom phase) after the flotation. Bradford reagent is also purchased from R&M Chemicals (Selangor, Malaysia).

#### *2.2. Apparatus*

Liquid biphasic flotation unit of 50 mL volume capacity was used as the separation system, and it was obtained from Donewell Resources (Puchong, Selangor, Malaysia). A 50 mL glass tube was connected from the bottom to a gas compressor. The bottom of the glass tube was drilled and fitted with a rubber tube to be connected to the gas compressor. A sintered glass disk (Grade 4 (G4) porosity) was fitted at the bottom of the glass tube so that air bubbles will be produced when compressed air is passed through it. The flowrate of air supplied to the LBF system is controlled by using a flowmeter (model: RMA-26-SSV) with a range of 50 to 200 cc/min (Dwyer, Michigan, IN, USA). The air compressor is powered by plugging it into a wall socket. Figure 1 shows the schematic diagram of the LBF system.

**Figure 1.** Figure illustrating schematic diagram of LBF system.

#### *2.3. Preparation of Milk Samples*

Expired milk was supplied by local producers (Dutch Lady) and was stored under room temperature. The concentration of proteins in the milk was tested to be 13.64 mg/mL. The milk was stored into a refrigerator at a temperature of 2 ◦C to reduce the effects of bacteria activity.

#### *2.4. Protein Assay*

The protein is determined by using Bradford Reagent. The dilution has been prepared with 10× dilution with adding 2 mL reagent with 0.2 mL of either top or bottom solution and was incubated for 10 min before the reading was tested using UV-Vis spectrophotometer at a reading wavelength of 595 nm. The absorbance of the protein concentration was based on the calibration between BSA concentration and OD595.

#### *2.5. Protein Extraction Using LBF*

A mixture of salt solution and expired milk was mixed and top up to 15 mL to be used as the bottom phase of the experiment. 15 mL of pure alcohol was used as the top phase of the experiment. After pouring the two solutions into the LBF system, the mixture was allowed to settle for 30 s so that two layers of liquid can be formed inside the system. The flowmeter was set to 25 cc/min to allow compressed air to flow into the system. Air was passed through the system for 10 min before closing the flowmeter and allowing the system to settle for 5 min again. The top and bottom layer was pipetted out from the glass tube and tested for their respective protein concentrations.

#### *2.6. Optimization of LBF Operating Parameters*

The operating parameters of LBF such as type of salt/alcohol, concentration of salt/alcohol/milk, pH of the bottom phase, and the flotation time were investigated by one factor at a time (OFAT) approach to maximize protein extraction and recovery. The optimization started with the initial operating conditions which is 100% of 2-propanol, 20 g/L salt solution, 15% (*w*/*v*) milk solution, flotation time of 10 min and the initial pH of the solution. The initial volume for both top phase and bottom phase was kept at 15 mL each, and the experiment was carried out at room temperature. Table 1 shows the parameters and variables tested for this experiment.



#### *2.7. Calculations of Recovery Yield and Separation E*ffi*ciency*

Recovery yield (*R*) of proteins in the hydrophilic organic solvent phase was measured using the following equation:

$$R = \frac{C\_T V\_T}{M} \times 100\%$$

where,

*R* is the recovery yield

*CT* is the concentration of proteins in the hydrophilic organic solvent phase

*VT* is the volume of the hydrophilic organic solvent phase

*M* is the total mass of proteins in the initial milk used

The separation efficiency from the milk to the top phase after LBF is calculated by the following equation:

$$E = \frac{\mathcal{C}\_T V\_T}{\mathcal{C}\_T V\_T + \mathcal{C}\_B V\_B} \times 100\%$$

where,

*E* is the separation efficiency

*CT* is the concentration of proteins in the hydrophilic organic solvent phase

*VT* is the volume of the hydrophilic organic solvent phase

*CB* is the concentration of proteins at the bottom phase

*VB* is the volume of the bottom phase

#### **3. Results and Discussion**

#### *3.1. E*ff*ects of Di*ff*erent Types of Inorganic Salt in Protein Recovery Using LBF*

The type of salts used for LBF is a key for protein extraction in this system as different salts induce different interactions with the protein, causing the separation efficiency of the proteins to alter. This is because salt solutions at the bottom phase are responsible of manipulating the surface tension of water thus changing the hydrophobic interactions between proteins and water at bottom phase [14]. As a result, when the protein solubility is reduced, proteins will start to migrate to the top phase when aided by flotation. It is reported that the Gibbs free energy of hydration of salt was the key to the formation of a biphasic system between salt and alcohol solution [15].

This experiment was conducted by firstly determining the most suitable type of salt to be used to obtain the best results and then changing the type of alcohol used. The types of salts used include ammonium sulphate, dipotassium hydrogen phosphate, disodium hydrogen phosphate, sodium sulphate, and magnesium sulphate. The results are illustrated in Figure 2. Dipotassium hydrogen phosphate and disodium hydrogen phosphate show a higher efficiency followed by ammonium sulphate and sodium sulphate in the descending order of 85.18%, 68.99%, 58.85%, and 53.81%, respectively. Magnesium sulphate showed the lowest separation efficiency which is 50.10%. For disodium hydrogen phosphate, the bottom phase forms salt crystals after flotation has been completed. This is due to the volume of the bottom phase being reduced; thus, the salt is unable to be fully dissolved in the remaining solution. Higher maintenance cost of the system is required if disodium hydrogen phosphate were to be used as a medium to create the biphasic conditions of this system.

As for the protein recovery yield, dipotassium hydrogen phosphate exhibits a higher protein recovery yield than all other salts, with a recovery yield of 29.997%. The lowest recovery yield of all salt tested was ammonium sulphate, which has a recovery yield of 7.74%. The other salts, which are disodium hydrogen phosphate, magnesium sulphate, and sodium sulphate, each has a recovery yield of 20.33%, 16.04%, and 10.81% respectively. The recovery yield for dipotassium hydrogen phosphate is significantly higher than that of disodium hydrogen phosphate. Taking into consideration the fact that when disodium hydrogen phosphate is used, the bottom phase will form salt crystals after flotation, and the two results obtained, dipotassium hydrogen phosphate was chosen to be the inorganic salt used in the following tests.

**Figure 2.** Figure showing the effect of different types of salts on the protein recovery yield and separation efficiency using LBF system.

#### *3.2. E*ff*ect of Di*ff*erent Types of Alcohols for Protein Recovery*

The type of alcohol used for LBF plays an important role in the system as different types of alcohol have different levels of interactions with the proteins, which will determine how much protein can the system extract. The type of alcohol used also affects the formation of biphasic system with salt solution. Many proteins are found to be non-compatible with the alcohol-rich top phase in the LBF process [16]. Some alcohols such as methanol will form triphasic solutions rather than biphasic solutions when mixed with salt solution. The selection of the type of alcohol used is very crucial as it will affect the overall performance of the system. In this study, ethanol, 1-propanol and 2-propanol of 100% were selected to form a biphasic system with dipotassium hydrogen phosphate at a concentration of 20 g/L.

All three alcohols were found to successfully form a biphasic system with the dipotassium hydrogen phosphate solution and the protein recovery yield for ethanol, 1-propanol and 2-propanol found to be 42.00%, 13.23%, and 23.35%, respectively. Based on Figure 3, ethanol has outperformed both 1-propanol and 2-propanol in terms of recovery yield, being almost two times the yield of 2-propanol and almost three times the yield of 1-propanol. In terms of separation efficiency, ethanol also outperforms both 1-propanol and 2-propanol. The separation efficiency of ethanol, 1-propanol, and 2-propanol is found to be 92.20%, 87.54%, and 90.56%, respectively. Alcohols usually contain a carbon chain and a functional group (-OH); the difference between ethanol and the other two alcohols is that ethanol has a shorter carbon chain, resulting in it having more ethanol molecules at the same volume. This can be proven by dividing the density of the respective alcohol with its molar mass. For example, the density of ethanol is 0.789 g/cm3 at 20 ◦C [17] and ethanol has a molar mass of 46.07 g/mol [18], this will result in ethanol a volume of 0.01715 mol/cm3. This is higher than 1-propanol and 2-propanol as they have volumes of 0.01336 mol/cm3 and 0.01306 mol/cm3, respectively. The density of 1-propanol and 2-propanol is 0.803 g/cm3 [19] and 0.785 g/cm3 [20], respectively at 20 ◦C while their molar mass is 60.096 g/mol [21] and 60.1 g/mol [22], respectively. This difference will result in the protein molecules being able to interact more with the alcohol molecules and not settle back to the bottom phase after flotation. Additionally, the high recovery yield and separation efficiency in ethanol could be because of high polarity of the alcohol compared to the other two alcohols used. High

hydroxyl group in ethanol could allow more protein to be accumulated at the top phase thus, giving high recovery yield [23]. Due to the above reasons, ethanol is chosen to carry out the following tests.

**Figure 3.** Effect of various types of alcohols on protein recovery yield and separation efficiency using LBF system.

#### *3.3. E*ff*ect of Di*ff*erent Concentration of K2HPO4 Salt on the Recovery of Proteins*

The concentration of salt used for the bottom phase is also optimized in this study. Varying salt concentrations have been used in the separation of proteins. When dipotassium hydrogen phosphate concentration is increased from 150 g/L to 350 g/L, the volume of top phase showed a decreasing trend while the volume of bottom phase showed an increasing trend. More proteins are retained in the lower phase.

From Figure 4, the highest protein recovery yield is exhibited by a salt concentration of 150 g/L, with a yield of 46.83%, while the lowest yield was obtained from a salt concentration of 350 g/L, with a yield of 37.14%. At salt concentration of 200 g/L, 250 g/L, and 300 g/L, the recovery yield of proteins is 45.01%, 39.36%, and 44.96% respectively. When increasing the salt concentration, the yield shows a decreasing trend due to more proteins being denatured when exposed to higher concentrations of salts. Thus, the lowest concentration of salt to form a biphasic solution should be obtained [24]. As for the separation efficiency when salt concentration is altered, the separation efficiency shows a decreasing trend when the salt concentration is increased. This is also due to the proteins in the milk being denatured by the salts when the solution is mixed together. Due to the above reasons, a salt concentration of 150 g/L was used for the following tests.

**Figure 4.** Figure illustrating effect of various K2HPO4 concentrations on protein recovery yield and separation efficiency using LBF system.

#### *3.4. E*ff*ect of Di*ff*erent Concentrations of Ethanol on the Recovery of Proteins*

The concentration of alcohol used will also affect the overall performance of the LBF system. Thus, the next parameter to be optimized is the concentration of alcohol. Various concentrations from 60% to 100% of ethanol were tested by using 15% (g/L) dipotassium hydrogen phosphate. As shown in Figure 5, 80% (*W*/*V*) shows the best recovery yield of 77.30%, followed by 70% of ethanol concentration which shows a yield of 54.50%. Concentrations of 60%, 90%, and 100% show 45.88%, 46.30%, and 41.48% yield, respectively. The protein recovery yield increases when the ethanol concentration increases, however when the alcohol concentration exceeds a certain point, the protein recovery yield starts to reduce. This is because the formation of the biphasic layers is weak when the concentration of alcohol is low. This result follows the trend in the previous study on protein recovery of wet microalgae using LBF where the highest protein recovery was obtained at 60% of 1-propanol, and the recovery yield decreased when the concentration of alcohol reduced below 40% [25]. These phenomena are due to the concentration of alcohol decreasing; more hydrophilic proteins can be dissolved into the top phase when the proteins are brought up by flotation air bubbles. However, when the concentration of alcohol decreases, the water at the top phase tends to migrate down to the bottom phase of the LBF system [16].

In terms of separation efficiency, ethanol solution with 80% concentration also has the highest separation efficiency, being at a value of 93.82%. The separation efficiency is followed by a pure ethanol solution which has a value of 90.59%. Values for 60%, 70%, and 90% are 69.16%, 54.17%, and 84.60%, respectively. The high value obtained at 80% of ethanol could be contributed due to cluster formation of ethanol. The number and size of the clusters strongly depend on the number of hydrogen bonds, and at higher concentration of ethanol, the cluster size is higher, which contributed to higher recovery of protein. Generally, ethanol has maximum viscosity of 75% to 80%, thus, this supports the high recovery of protein and separation efficiency. This indicates that at 80% of ethanol concentration, more proteins can be separated from the bottom phase which is the main point of this study. Therefore, 80% concentration ethanol solutions will be used in the following optimizations.

**Figure 5.** Figure showing effect of various concentrations of ethanol on protein recovery yield and separation efficiency using LBF system.

#### *3.5. E*ff*ect of Various Concentrations of Milk*

In this section, the effect of milk concentration of the bottom phase of the LBF system was studied. The concentration of milk may pose a potential effect on protein extraction by affecting the formation of the biphasic system, indicating that the concentration of milk used for extraction will have an impact on the yield of proteins recovered [26]. Milk with concentrations of 5% (*w*/*v*), 10% (*w*/*v*), 15% (*w*/*v*), 20% (*w*/*v*), and 25% (*w*/*v*) mixed along with dipotassium hydrogen phosphate solution with 150 g/L concentration were tested to investigate the effects of milk concentration on protein recovery. The highest protein yield obtained is at 10% (*w*/*v*), closely followed by 5% (*w*/*v*), which has the values of 93.96% and 93.75%, respectively. At 25% (*w*/*v*), the recovery yield of proteins is the lowest, having a value of 50.98%. At 15% (*w*/*v*) and 20% (*w*/*v*), it has values of 89.77% and 64.74%, respectively. According to Figure 6, it was observed that the concentration of milk has increased from 5% (*w*/*v*) to 10% (*w*/*v*), the recovery yield has only increased slightly; however, when the concentration is further increased, the yield starts to drop significantly, especially from 15% (*w*/*v*) to 20% (*w*/*v*), with a total drop of more than 25% protein yield. This is due to the fact that when a high concentration of milk is used, the salt solution mixture tends to form a liquid of which its high viscosity will result in the formation of flotation bubbles to be too difficult to control. By increasing the concentration of milk used, the performance of the LBF would be reduced as the level of impurities within the solution will also increase. The overall composition of the bottom phase will be altered as there is a lot of impurities in the milk [16,27]. In terms of separation efficiency, however, the highest value obtained is 94.02%, which is achieved by 25% (*w*/*v*). The separation efficiency gradually increases as the concentration is increased, starting from 5% (*w*/*v*), the separation efficiency was found to be 72.62%, 87.14%, 90.96%, and 92.59%, respectively. Due to the higher separation efficiency, 10% (*w*/*v*) of milk concentration was selected to carry on the following tests.

**Figure 6.** The effect of different concentrations of milk on the protein recovery yield and separation efficiency utilizing LBF system.

#### *3.6. E*ff*ect of pH on the Recovery of Proteins*

Impact of pH partitioning of proteins and enzymes to the phases in the LBF system depends on their isoelectric points. The pH of the system, however, affects the charge of target protein molecules and ionic composition, as well as introduces differential partitioning into the two phases. Most of the biomolecules, especially proteins and enzymes, are stable at neutral pH, which is a favorable condition to conduct the LBF partitioning. However, an increase in pH of the LBF from 7.0 to 9.0 reduced the protein recovery and activity recovered. Enzyme stability is slightly reduced in the acidic area except at the lowest pH and was dramatically lost at pH above 9.0. This dependence on pH for optimal protein recovery can be explained in terms of the charge in the protein. The protein in the LBF is predominantly casein. From the literature, isoelectric point of casein is 4.6, and since the pH of milk is 6.6, casein molecules are positively charged due to the protons provided by the milk medium. Given that the formula for pH is *pH* = <sup>−</sup>*log*[*H*+], where *<sup>H</sup>*<sup>+</sup> is concentration of hydrogen ions, a pH of 7 to 8 in the system (see Figure 7 corresponds to a negatively charged medium. With a positive charge, the casein molecules are thus hydrophobic, making them less soluble in water. Given that bubbles are used to push the particles up to the top phase and that the casein molecules are positively charged, the surface charge on the bubbles plays a vital role in the protein extraction efficiency. In particular, the charge of the casein and the surface charge of the bubble will be responsible for the adsorption of the protein molecule to the bubble surface. In order for the protein molecules to be attached to the bubble surface, the charge of the bubble surface must therefore be negative.

**Figure 7.** Figure showing the effect of pH system on protein recovery yield and separation efficiency.

To understand the mechanics behind the adsorption of the bubbles, we have developed a mathematical model with the use of partial differential equations (PDEs).

The region of interaction in our model is constructed using the same approach as the Lamm equations (Lamm O., 1929), which is to divide the volume of the container into sector-shaped cells. With reference to Figure 8 consider a region *R* in a sector-shaped cell within the chamber of the flotation system. Let *M*, *Min*, and *Mout* be the mass of solute inside *R*, the mass flow into *R*, and the mass flow out of *R*, respectively. By the principle of conservation of mass,

$$\frac{\partial M}{\partial t} = \frac{\partial M\_{in}}{\partial t} - \frac{\partial M\_{out}}{\partial t} \tag{1}$$

**Figure 8.** Region of collision of the protein particles against the bubbles, *R*.

Flux *j* is defined as the number of bubbles passing though an area *A* per unit time. Assuming no diffusion, i.e., only convection,

$$
\mathfrak{j} = \sigma \mathfrak{s} \tag{2}
$$

where σ and *s* are the density and velocity of the bubbles respectively.

Putting Equation (2) into (1) gives

$$\frac{\partial \sigma}{\partial t} = -\nabla \cdot \underline{\mathcal{J}}.\tag{3}$$

*Processes* **2019**, *7*, 875

The bubbles are assumed to move into the region *R*, resulting in a negative ∇·*j*, and the negative sign is to make it positive.

Integrating,

$$\frac{\partial}{\partial t} \int\_{h\_0}^{h\_1} \sigma(h, t) A(\Omega\_h) dh = -\int\_{h\_0}^{h\_1} \frac{\partial}{\partial h} [j(h, t) A(\Omega\_h)] dh \tag{4}$$

where σ(*h*, *t*) and *j*(*h*, *t*) are density and magnitude of flux of bubbles respectively, *A*(Ω*h*) is the area in the top and bottom surfaces of the region *R* (see Figure 1), and cylindrical coordinates are used: *r* = *x*<sup>2</sup> + *y*2, φ = *<sup>y</sup> <sup>x</sup>* , *h* = *z*.

From Sminov and Berry (Smirnov et al., 2015), the velocity of the bubble is given by

$$s = \frac{2dgr\_{\nu}2}{\theta^{\nu}}\tag{5}$$

where *g* is the free fall acceleration, *d* is the density difference for liquid and air inside bubbles or the liquid density, *rb* is the radius of bubble, ν is the liquid viscosity.

From previous work (Lin YK, 2015), the number of bubbles that can be adsorbed to the surface of a bubble, *Np*, is given by *Np* = π *rb rp* 2 where *rp* is the radius of the particle.

For maximum adsorption, the number of particles in the region *R* must be π *rb rp* 2 times the number of bubbles in the same region *R*, and assuming a very small *R*, all bubbles that go into the *R* will collide with all particles that go into *R*.

From Equation (4), substituting *A*(Ω*r*) = *h*φ*r* and rearranging gives

$$\int\_{h\_0}^{h\_1} h\phi r \frac{\partial \sigma}{\partial t} + \frac{\partial \sigma}{\partial r} (h\phi r j) dh = 0 \tag{6}$$

From Equation (4),

$$\int\_{z\_0}^{z\_1} \frac{\partial}{\partial t} [\sigma(z, t) A(\Omega\_{\bar{z}})] + \frac{\partial}{\partial z} [j(z, t) A(\Omega\_{\bar{z}})] dz = 0. \tag{7}$$

Putting *Area* = *r*φ(*r*<sup>1</sup> − *r*0) into Equation (7) and integrated w.r.t. *z* gives

$$r\phi(r\_1 - r\_0)\frac{\partial}{\partial t}\sigma(z, t) + r\phi(r\_1 - r\_0)\frac{\partial}{\partial z}j(z, t) = 0$$

Rearranging, we have

$$\frac{\partial}{\partial t}\sigma(z,t) + \frac{\partial}{\partial z}j(z,t) = 0\tag{8}$$

Putting *j*(*z*, *t*) = σ*s* into Equation (8) yields

$$\frac{\partial}{\partial t}\sigma(z,t) + \frac{\partial}{\partial z}(\sigma s) = 0$$

$$\frac{\partial}{\partial t}\sigma(z,t) + s\frac{\partial}{\partial z}\sigma(z,t) + \sigma(z,t)\frac{\partial}{\partial z}s = 0\tag{9}$$

Given that *s* = <sup>2</sup>*dgrb* 2 <sup>9</sup><sup>ν</sup> , we can take *d* ≈ ρ. We know that

$$P = \rho gh\tag{10}$$

where *P* is pressure of fluid, which is equal to the pressure in the bubble, ρ is density of fluid, *g* is acceleration due to gravity, and *h* is height. From Sminov and Berry,

$$P = P\_{\varepsilon x} + \frac{2\alpha}{r\_b} \tag{11}$$

where *Pex* is exteral pressure acting on liquid, α is the surface tension and is a constant, and *rb* is the radius of the bubble. Combining Equations (10) and (11) gives:

$$
\sigma\_b = \frac{2\alpha}{\rho g h - P\_{cx}} \tag{12}
$$

where *Pex* is the pressure of air in the room (most likely atmospheric pressure, 1 atm).

Putting Equation (12) into Equation (6) gives:

$$s = \frac{2\rho g \left(\frac{2a}{\rho g h - P\_{\rm ex}}\right)^2}{g\_V} = \frac{8\rho g}{g\_V} \left(\frac{a}{\rho g z - P\_{\rm ex}}\right)^2 = f(z) \tag{13}$$

where, *h* = *z*.

Putting (13) into (9) gives:

$$\frac{\partial}{\partial t}\sigma(z,t) + \frac{8\rho g}{9\nu} \left(\frac{a}{\rho gz - P\_{cx}}\right)^2 \frac{\partial}{\partial z}\sigma(z,t) - \frac{16(\rho g)^2 a^2}{9\nu (\rho gz - P\_{cx})} \sigma(z,t) = 0\tag{14}$$

$$\frac{\partial}{\partial t}\sigma(z,t) + f(z)\frac{\partial}{\partial z}\sigma(z,t) + \sigma(z,t)\frac{d}{dz}f(z) = 0\tag{15}$$

$$f(z)\sigma(z,t)\_z + \sigma(z,t)\_t + F(z)\sigma(z,t) = 0\tag{16}$$

where, *<sup>F</sup>*(*z*) <sup>≡</sup> *<sup>d</sup> dz f*(*z*).

To solve Equation (16), Laplace transform is applied to Equation (16), giving

$$f(z)\underline{\sigma}\_z(z,s) + s\underline{\sigma}(z,s) - \sigma(z,0) + F(z)\underline{\sigma}(z,s) = 0$$

$$f(z)\underline{\sigma}\_z(z,s) + [s + F(z)]\underline{\sigma}(z,s) - \sigma(z,0) = 0\tag{17}$$

Since σ(*z*, 0) = 0, Equation (17) becomes

$$f(z)\underline{\sigma}\_z(z,s) + [s + F(z)]\underline{\sigma}(z,s) = 0\tag{18}$$

To solve Equation (18), we divide Equation (18) by *f*(*z*) to give

$$
\underline{\underline{\sigma\_z}}(z,s) + \frac{s + F(z)}{f(z)} \underline{\underline{\sigma}}(z,s) = 0 \tag{19}
$$

Multiplying Equation (19) with integrating factor *e <sup>s</sup>*+*F*(*z*) *<sup>f</sup>*(*z*) *dz* yields

$$\frac{d}{dz}[e^{q(z,s)}\underline{\sigma}(z,s)] = 0\tag{20}$$

where,

$$q(z,s) = \frac{9s\nu}{8\rho g a^2} \left(\frac{\rho^2 g^2 z^3}{3} - \rho g z^2 P\_{cx} + (P\_{cx})^2 z\right) + \ln \ln \left|\frac{8\rho g a^2}{9\nu \left(\rho g z - P\_{cx}\right)^2}\right| \tag{21}$$

Integrating Equation (20) gives

$$\underline{\sigma}\_{z}(z,s) = \mathbb{C}e^{Q(z,s)} \tag{22}$$

where *C* is a constant, and *Q*(*z*,*s*) = *q*(*z*,*s*) −1 . Transforming Equation (22) from the *s* domain back to the *t* domain yields

$$\begin{split} L^{-1} \left| \underline{\sigma\_z} (z, s) \right| &= \\ &= \delta \left| t - \frac{g\_V}{8 \rho g a^2} \left( \rho g z^2 P\_{cx} - \left( P\_{cx} \right)^2 z - \frac{\rho^2 g^2 z^3}{3} \right) - \ln \ln \left| \frac{8 \rho g a^2}{9 \nu \left( \rho g z - P\_{cx} \right)^2} \right| \right| \end{split} \tag{23}$$

From Equation (23), for the bubbles to hit all protein molecules, a sufficient amount of time is required. In particular, the Dirac delta function in Equation (23) indicates that for maximum protein extraction, *t* ∼ *O z*2 - . Thus, with a uniform distribution of bubble holes at the bottom of the chamber, the time it takes to collect all molecules increases as the square of the height of the chamber. This explains the increasing amount of time required for higher protein collection in Figure 9, which shows both the theoretical and experimental data. It can be seen from this figure that the experimental data is in agreement with the theoretical data for flotation time less than 8 min. The reduction in yield after 8 min can be attributed to the lack of protein in the system after prolonged extraction.

**Figure 9.** Figure illustrating the effect of flotation time on protein recovery yield and separation efficiency.

Besides allowing for a better understanding of the mechanism behind the system, this model also serves to provide some insights on design of efficient flotation systems at industrial scale. In particular, since it has been mathematically shown that an increase in height of the chamber drastically increases the flotation time, future chambers of flotation should preferably be as low as industrially viable.

The effect of pH has been studied by altering the pH of the bottom phase by using 1M hydrochloric acid. The initial pH of the bottom phase is 9.15, the tested pH is 6.5, 7.0, 7.5, and 8.0. The acid is added drop by drop until the bottom phase reached the desired pH with the aid of a pH meter. The highest recovery yield of proteins is obtained by the solution to which hydrochloric acid has not been added, which is the solution with a pH of 9.15. The recovery yield of the solution with this pH is at 93.96%, surpassing the second highest, which is a pH of 8.0 with a value of 65.71%, by more than 28%. The other three pH, being 6.5, 7.0, and 7.5, each has a recovery yield of 61.45%, 60.37%, and 62.33%, respectively. The big difference in yield is caused by the hydrochloric acid used to alter the pH having denatured the proteins, thus greatly reducing the yield of the LBF [24]. As the acid was added drop by drop, each time a drop of acid hits the surface of the bottom phase, the extreme pH of the acid will denature some of the proteins at the bottom phase before being diluted by the rest of the solution at this phase. This results in the big difference in the recovery yield when comparing between a bottom phase of pH 9.15 and 8.0; this is also why the difference in recovery yield between pH 8 and pH 6.5 does not show a big difference when compared with 9.15 and 8.0. The separation efficiency increases when the pH is changed from neutral to 6.5, which is from 85.59% to 86.21%. At a pH of 7.5, the separation efficiency is the highest which is at 97.18%. The separation efficiency of pH 8.0 and 9.15 is 94.33% and 87.14%, respectively. Due to the above reasons, the following test was conducted without altering the pH system.

#### *3.7. E*ff*ect of Flotation Time on the Recovery of Proteins*

The duration of the flotation process being conducted is very important as it could cause a major impact on the area of air–water interface per unit volume of aqueous solution in time [27]. Flotation times of 5, 7.5, 10, 12.5, and 15 min are tested for this study. The highest recovery yield of proteins is obtained when the system is run for 7.5 min, having a yield of 94.97%, closely followed by a flotation time of 10 min having a yield of 93.96%. With a flotation time of 5, 12 and 15 min, the yield is 85.03%, 63.76%, and 63.24%, respectively. The yield of proteins obtained shows an increasing trend when the flotation time is increased; however, it starts to decrease when the flotation time exceeds 10 min. This is because molecules other than proteins is being blown to the top phase, causing the overall concentration of the proteins at the top phase to be reduced. This is proven as the volume of the bottom phase shows a decreasing trend when the flotation time is increased, decreasing from 4 mL for 5 min flotation time to 2.5 mL for 15 min flotation time. In terms of separation efficiency, the values show a general increasing trend, but the increase in efficiency is only by a small amount, starting from 5 min flotation time, the separation efficiency is 85.58%, 86.29%, 87.20%, 88.64%, and 88.46%, respectively. Due to the reasons stated above, a flotation time of 7.5 min is taken to be optimum for this study. In general, the recovery of protein from the beginning of the experiment is low, however; it increases as the optimization is carried out. The low in recovery may be due to high contamination of bacteria that could affect the recovery of protein.

#### *3.8. E*ff*ect of Scaling up LBF for Industrial Application Purposes*

For industrial reasons, this experiment was tested again under large scale conditions. The experiment was scaled up by 40 times (from 30 mL to 1.2 L), and the results show that this experiment is suitable to be scaled up, with a recovery yield of more than 70% and a separation efficiency of more than 80%. This indicates that LBF is suitable to be commercialized as a method to separate proteins from expired milk wastes.

In this study, several parameters and their impacts on the LBF system have been studied by using the one-factor-at-a-time approach. Given that there might be a possibility that there will be interaction effects between several parameters, further studies on interaction factors such as milk concentration and flotation time could have a high chance of further optimizing this process. There is also a need to further improve this method so that protein recovery rate may be increased. An air compressor that can achieve higher and more accurate flowrates can be used in future LBF experiments. Furthermore, different types of gases can be used in replacement of atmospheric air for flotation. Different types of gases may help to bring up proteins or have other interactions with proteins that can improve the protein recovery yield. Thus, the effects of different types of gases such as pure oxygen or pure nitrogen can be tested to improve the recovery yield of the protein. Besides, the liquid used as top phase can be changed to different materials as using large amount of alcohol in industrial scale is a safety hazard. Alternative materials such as other organic solvents can be considered as an alternative to alcohol. Another aspect that is worth mentioning is the brand of milk used. As each brand of milk has a different formula for the milk they produce, changing the brand of milk used may also improve the protein recovery yield as there might be some components in milk from other brands that help in protein separation. Moreover, milk of different expiry dates can also be tested as well, as the level of microorganisms inside the milk may differ as time progresses. Further studies can be carried out by comparing milk of different expiry dates and some milk that is close to their expiry date so that the effects are clear.

#### **4. Conclusions**

The parameters for protein extraction from expired dairy products were optimized in this study. The effects of the type of inorganic salt used, the type of alcohol used, the concentration of salt used, the concentration of alcohol used, the concentration of raw material (milk) used, pH of the bottom phase, and the flotation time of the LBF system were discussed. The optimum conditions for protein extraction from dairy wastes tested in this study were found to be 150 g/L dipotassium hydrogen phosphate, 80% of ethanol, 10% (*w*/*v*) milk, a pH system of 9.15 (initial pH), and a flotation time of 7.5 min. The final protein recovery yield and separation efficiency after optimization were 94.97% and 86.29%, respectively. A scaling up of the LBF system was also performed at a factor of 40 times, and the protein recovery yield and separation efficiency for this test were 78.92% and 85.62% respectively. This study showed that proteins can be extracted from dairy waste effectively. The advantages of this novel approach include providing a use for expired milk products, reducing wastes being thrown away and benefiting the environment, and turning waste that once needed money to be disposed of into a raw material that can provide profit. Additionally, the utilization of high concentration of alcohol and salt will help inhibit further contamination by bacteria as they cannot survive in high alcohol and salt concentration environments. The concern with environmental impact due to high salt and alcohol concentration can be avoided by studying the recycling ability of the phase components. Studies have shown that there is a great potential to reuse recycling phase components in the subsequent extraction of LBF. Future studies on the interaction of parameters and methods for recycling the top and bottom phases after separation will provide a great opportunity for future industries to apply this method as a waste treatment process.

**Author Contributions:** Conceptualization, K.W.C.; data curation, P.E.T., Y.J.N., and K.W.C.; formal analysis, P.E.T. and Y.J.N.; funding acquisition, P.L.S.; methodology, R.S. and K.W.C.; supervision, K.W.C. and P.L.S.; visualization, K.S.K.; writing—original draft, P.E.T. and Y.J.N.; writing—review and editing, R.S., K.S.K., Y.J.Y., M.M., F.A.A.Z., and P.L.S.

**Funding:** This research was supported financially by SATU Joint Research Scheme (ST014-2018, ST022-2019), Geran Penyelidikan Universiti Malaya (UMRG Programme)—SUS (Sustainability Science) (RP025B-18SUS) and Fundamental Research Grant Scheme (Malaysia FRGS/1/2019/STG05/UNIM/02/2).

**Acknowledgments:** This study is supported by the Fundamental Research Grant Scheme (Malaysia, FRGS/1/2019/STG05/UNIM/02/2).

**Conflicts of Interest:** The authors declare that they have no conflicts of interests.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
