*Article* **Magnetic Bioreactor for Magneto-, Mechano- and Electroactive Tissue Engineering Strategies**

**Nelson Castro 1,\*, Margarida M. Fernandes 2,3, Clarisse Ribeiro 2,3, Vítor Correia 4, Rikardo Minguez 5,\* and Senentxu Lanceros-Méndez 1,6**


Received: 12 May 2020; Accepted: 9 June 2020; Published: 12 June 2020

**Abstract:** Biomimetic bioreactor systems are increasingly being developed for tissue engineering applications, due to their ability to recreate the native cell/tissue microenvironment. Regarding bone-related diseases and considering the piezoelectric nature of bone, piezoelectric scaffolds electromechanically stimulated by a bioreactor, providing the stimuli to the cells, allows a biomimetic approach and thus, mimicking the required microenvironment for effective growth and differentiation of bone cells. In this work, a bioreactor has been designed and built allowing to magnetically stimulate magnetoelectric scaffolds and therefore provide mechanical and electrical stimuli to the cells through magnetomechanical or magnetoelectrical effects, depending on the piezoelectric nature of the scaffold. While mechanical bioreactors need direct application of the stimuli on the scaffolds, the herein proposed magnetic bioreactors allow for a remote stimulation without direct contact with the material. Thus, the stimuli application (23 mT at a frequency of 0.3 Hz) to cells seeded on the magnetoelectric, leads to an increase in cell viability of almost 30% with respect to cell culture under static conditions. This could be valuable to mimic what occurs in the human body and for application in immobilized patients. Thus, special emphasis has been placed on the control, design and modeling parameters governing the bioreactor as well as its functional mechanism.

**Keywords:** magnetic bioreactor; magnetoactive scaffolds; tissue engineering; magnetic actuator; magnetoelectric stimulation

#### **1. Introduction**

Fundamental biological studies and therapeutic applications rely on tissue engineering (TE) techniques, which aim to mimic the physicochemical and bioactive characteristics of natural cellular matrices [1,2], in order to achieve the replacement and/or regeneration of damaged tissues or organs [3,4]. When building a new tissue culture, three tools are mainly used: cells, scaffolds and stimuli. The cells are the building blocks for tissue culture as they contain the pre-programmed information that allows tissue regeneration. Cells are thus placed in a scaffold, which acts as the cell culture support, where the necessary environment is present, mainly in terms of biochemical stimuli, through the inclusion of growth factors and biophysical stimuli, by using a bioreactor [5]. Bioreactors allow us to introduce

different chemical or physical stimuli on tissue culture, depending on both bioreactor technology and on how scaffold structures respond to those stimuli, in order to create a synergistic environment, thus stimulating cell response. Biomaterials used as scaffolds can be tailored, allowing them to be passively tolerated by the organism or actively providing the most appropriate and specific cell responses [6,7].

In the context of physically active TE by using bioreactors, different approaches have been implemented, including mechanoactive and electroactive scaffolds, among others [1,8]. In particular, electroactive materials are gaining increased attention due to the possibility of regulating different cell functions by providing electrical signals to the tissue culture [9–11]. Examples of such materials are piezoelectric scaffolds, which provide the necessary stimuli for the effective regeneration of bone [7,9,12], neural tissue [13,14], muscle [15–17], among others. The underlying mechanism relies on the possibility for mechanoelectrical transduction from materials to the cells [12], but lack of appropriate bioreactors able to stimulate those materials and take full advantages of their smart and multifunctional nature. In this work, a bioreactor is presented able to apply magnetic, mechanical and electrical stimuli to the cells in culture, based on the application of a magnetic field to a magnetically responsive scaffold containing magnetostrictive nano/microparticles embedded in a specific matrix [18,19] which can be electrically responsive (e.g., piezoelectric [20]) or not. These stimuli can be important for different types of tissue and, in particular, for bone TE due to the piezoelectric characteristics of bone [21], allowing to address novel TE strategies [22]. In particular, the performance of bone tissue along with its development and functional characteristics is strongly influenced by the voltage variation generated by mechanical stress to which the bone is subjected and, therefore, piezoelectric stimuli must be considered for proper regeneration strategies [23].

In fact, the electric sensitivity of osteoblasts has been regarded as an important tool for enhancing the ossification and healing through electric stimulation, as proven by piezoelectric scaffolds stimulated by a mechanical bioreactor, thus providing a proper electroactive environment to the cells [24,25]. In a similar approach, conductive composites have been proven to deliver exogenous electric currents to cells and increase their function [26], while evidence of electric stimulation influence on the ossification has been indeed observed [27].

In a novel approach, magnetoelectric (ME) scaffolds were used in order to provide support for cell seeding and proliferation while taking advantage of the scaffolds material, which enables electromechanical stimuli to the cells, generated by a magnetic field [19,28,29]. These composites are composed of magnetostrictive and piezoelectric layers working synergistically to produce the ME effect [30]. The ME effect can be described as a transduction from a magnetic field to an electrical field once the vibration of the magnetostrictive phase generated by an alternated magnetic field results in an electrical charge variation at the piezoelectric phase terminals at room temperature [31,32]. Further, if the support is not piezoelectric, just a mechanical vibration will be induced, providing such stimuli to the neighboring cells [33].

The flexibility, versatility and biocompatibility of these materials [8,34] can take advantage of in-vitro dynamic cultures through the support of a remote magnetic field [33]. Thus, materials with ME properties are therefore regarded as breakthrough platforms for TE applications that allow for remote generation of these physical stimuli, resulting in a controled influence on the surrounding tissue [9]. This effect has already been proven to induce a magnetomechanical [20] or a local magnetomechanical and magnetoelectrical effect [19], on the cells thus triggering improved cell proliferation and differentiation effect.

It is important to notice that the application of the magnetic field by itself, without further magnetomechanical or magnetoelectrical transduction, is also interesting for biomedical applications. This enables stimulation of cellular functions and cell manipulation to create cellular clusters, enabling more complex tissue structures than conventional strategies based on static scaffolds [35]. As a recent example, the use of superparamagnetic iron oxide nanoparticles proved to be a promising bioactive additive for scaffold fabrication [36], the scaffold enhancing the performance of human dental pulp

stem cells yielding a higher count of phosphatase activity, higher osteogenic marker gene expression and improved cell-synthesized bone minerals. Other methods include the marking of C2C12 cells with magnetite cationic liposomes, mixed in a collagen solution, and seeded in a cell culture space of a hollow-fiber bioreactor [37]. The results demonstrated that high cell-density and viable tissue constructs containing myotubes were successfully obtained. Magnetic stimuli through permanent magnetic displacement were also proposed [38]. Rotation of permanent magnets was also employed in order to induce cellular growth proving that the variation of the magnetic field between 7 and 10 Hz increased the growth of neurite on chromaffin cells [39]. These devices can thus give an important contribution to the field, in order to overcome the issues related to the traditional cell culture conditions, improving the cellular distribution and accelerating cellular growth [40].

Besides biocompatibility and sterility, the design of mechanical bioreactors requires accurate control of the applied stimulus in order to get accurate data and to replicate results within the same parameters. In order to comply with these requirements and build a user-friendly device, which enables the user to apply controlled stimuli while providing control over temperature, culture stimuli active and resting time schedule, as well as total time, a modular magnetic bioreactor with an interchangeable magnets table was designed and developed. The interchangeable magnets table module has been designed to be used with 24-multiwell standard plates that can be easily operated and calibrated by the user. Magnetic stimulation has been previously reported [19,20,41], however, the present work reports on the development and validation of a novel bioreactor for magnetic stimulation of cells and/or scaffolds, comprehending cells-scaffold-stimuli relation, electromechanical study for actuation, control user interface and cell culture validation.

In terms of practical applications, the herein developed bioreactor can be a valuable tool for novel and more efficient tissue engineering strategies, including (i) to perform cell culture assays in vitro to validate the use of magneto-active materials for tissue engineering, thus avoiding extensive animal testing; (ii) to grow in vitro cellular tissue to be further implanted in a patient, after detaching the tissue from the surface of a magneto-active material and (iii) to be used in vitro to grow cellular tissue that is further implanted in a patient with the magneto-active material, which would allow for a remote stimulation of the material in the body, an important tool for immobilized patients.

#### **2. Bioreactor Design**

A bioreactor is an important tool for TE purposes since it allows us to mimic essential elements of the tissue environment and thus evaluate the influence of the stimuli on cell proliferation and differentiation. In order to study the influence of the electromechanical stimulation in bone tissue cells remotely [19,20], magnetoelectric scaffolds have been developed to provide the required stimuli [19], where an alternated magnetic field was applied to supply the necessary magnetic stimulation to the scaffolds. In this way, the herein developed device was designed to fulfill these characteristics and to meet experimental requirements to which the bioreactor will be subjected, which includes an incubation chamber with controlled temperature and humidity (37 ◦C and 95%, respectively). Thus, increasing the scaffolds temperature while applying the magnetic field is a critical design parameter, excluding the use of electromagnets, which requires high current flow, consequently resulting in radial heat from windings [42]. On the other hand, the displacement of permanent magnets as magnetic actuators on magnetoelectric scaffolds were used instead, to avoid a bulkier system and further heat. A schematic representation of the designed system is presented in Figure 1. The actuation system is composed of a permanent magnetic table that is displaced at a controlled frequency until certain limits, in order to get the required alternated magnetic field at the culture plate. For that, a mechanical structure comprising a motor in a ball screw assembly was installed to obtain an electromechanical actuation system. For mechanical protection, limit switches and precision sensors were applied to obtain electronic control of table position through a linear sensor and magnetic encoder for speed, as main operation components.

**Figure 1.** Magnetoelectric bioreactor operating principle through the use of electrical and mechanical controls to produce an alternated magnetic field and thus stimulate the magnetoelectric scaffolds and, consequently, the cells.

Furthermore, the system required a stable power supply in order to handle all digital and power electronic components, as well as remote control through wireless communication such as Bluetooth®® and respective user-machine interface through buttons and light and/or display feedback. To facilitate the use of this bioreactor in TE, the design included the application of a commercially available culture plate on the top of the system, where the ME scaffolds are easily placed and tested for cell culture as represented in Figure 1. The starting position of the magnets is user-defined as well as its displacement and motion frequency. It is important to note that the displacement will influence the magnetic variation over time in a specific place of the culture wells. The magnetic field peak values are directly dependent on magnets grade and distance to the scaffold, which can be calibrated mechanically by adding extra layers below the magnet table. The required magnetic field for the stimulation of the ME samples must reach values within a range of 20 to 50 mT [20]. The selected magnets are nickel-plated neodymium disks S−15–03-N52N from Supermagnete. The distance between the culture bottom and the N52 grade neodymium permanent magnets influences the magnetic field intensity, as analyzed through simulation with ANSYS®® Software. The magnetic field generated by a permanent magnet is easily stronger at a given z distance from the source, by comparison with fair current amplitude within a reasonable size coil for this system. It can be calculated for a cylinder type of permanent magnet, using Equation (1).

$$B = \frac{B\_r}{2} \left( \frac{D+z}{\sqrt{R^2 + (D+z)^2}} - \frac{z}{\sqrt{R^2 + z^2}} \right). \tag{1}$$

This way, magnetic flux density can be calculated at a certain distance, where *Br* is the remnant field, independent of the magnet's geometry, *z* the distance from a pole face on the symmetrical axis, *D* the thickness (or height) of the cylinder and R the semi-diameter (radius) of the cylinder [43]. Figure 2 displays the more appropriate distance according to each culture wells plates setup, which is different according to the number of magnets and radius used, resulting in a distance of 10 mm, where a field intensity of 30 mT was achieved at culture plate bottom (Figure 2a). In Figure 2b, it is possible to observe the side views with a larger range of the magnetic field variation, due to the substantial higher field at the N52 permanent magnets core. This fact enables higher magnetic fields at the culture bottom by reducing distance through magnetic table mechanical elevation in the same mechanism represented in Figure 2c. Furthermore, tailoring/exchanging the permanent magnets table with different magnetic grades, sizes and geometries, enables the mechanics to fit with the number of wells and geometries that a culture plate may present.

**Figure 2.** (**a**) Magnetic field intensity distribution at the bottom of 24-wells cell culture plates, (**b**) magnetic field force lines simulation in frontal and side planes and (**c**) rendered model of the mechanical permanent magnets table ball-screw assembly.

As schematized in Figure 1, the DC motor is controlled by a custom electrical system built and designed for the purpose of this application (control system). For this role, the system takes advantage of the sensors (limit switches, encoder and linear) in order to perform a close loop control of the magnets table positioning and displacement frequency. Furthermore, the system handles a user firmware interface that allows us to select the culture cycles for active and resting times. Regarding the firmware tasks, they are divided into two main control units: one unit controls the user inputs while the other controls the magnets positioning, sensors and display feedback. The local user interface is composed of an ILI9341 LCD providing 240 × 320 resolution with 262 k color and a side capacitive touch wheel with one button and a sliding circular panel for selection and option confirmation. The setup allows the user to locally stop, start or change the culture control parameters. A remote interface was also installed by Bluetooth, which allows us to monitor the culture status and the sensor reads from outside the incubator, using a mobile terminal. Since cell culture experiments require aseptic environments, requiring a sterilization process before each cell culture experiment, the creation of a waterproof enclosure was necessary. Nevertheless, such enclosure should withstand the temperature without overheating the bioreactor thus damaging the cell culture. The three-dimensional project of the device complying with such requirements is presented in Figure 3.

**Figure 3.** (**a**) Representation of the bioreactor assembled with a cell culture plate; (**b**) Bioreactor prototype built mechanism with every component; and schematic representation of (**c**) all disassembled main electric and mechanical components and (**d**) of the mechanical component represented as a transversal cut.

The magnetostriction of the scaffolds is obtained as a consequence of the magnetic field application, achieved by the movement of the magnets along the horizontal axis of the magnets table, below the cell culture plate. The selected construction material was nylon in order to avoid magnetic field interference. The system was designed for 24 and 48 well culture plates, although several types of commercial plates

can be used with a different number of culture wells. Further, it was designed to be modular, thus the permanent magnets table can be replaced with a higher or lower number of magnets to fit the culture plate and apply an even magnetic field to all scaffolds (Figure 3). The result of the assembled system is simple, compact and a sealed design (Figure 3a). It is worth noting that the herein used permanent magnets were selected according to the magnetic field level required for inducing an electroactive environment on the ME scaffolds. Thus, when subjected to a magnetic field, a magneto-mechanical and magneto-electric stimulation is induced on the scaffold due to the incorporated magnetostrictive particles [19,44]. The magnetostrictive particles deform and generate the mechanical stimulus to the piezoelectric polymer within the scaffold which, in turn, develops an electrical charge that passes to the cells. The movement of magnets along the horizontal axis is achieved by using two side supports and a central motor shaft coupled with a DC motor, a component that can be observed in the device sectional detail in Figure 3d. The mechanical setup is implemented with a 25GA370 DC motor with 400 rpm and is controlled through an H-Bridge at 20 kHz pulse width modulation (PWM) pulses. This was selected over a stepper motor due to the possibility of applying lower currents and avoid heating in order to keep the system temperature low to protect the cell culture. The magnet table position is controlled by a linear sensor 9615R5.1KL2.0 (from BEI Sensors, Attleboro, MA, USA), together with an ADC resolution of 12 bits, which results in a linear resolution of 0.01 mm. The bioreactor was designed with an IP68 waterproof rating system to withstand the sterilization process, through waterproof power connector, rubber protection in the joints and capacitive touch avoiding mechanical buttons and leaks. Figure 3 displays a photo of the prototype device after the design and development phase.

#### *2.1. Power and Control Circuitry*

The main system communication and control between modules is illustrated in the block diagram of Figure 1. Thus, an electrical system was developed according to the mechanical design and user interface operational requirements. Further, the designed circuits were implemented using commercial electronic components. The electrical circuits designed are represented in the schematic of Figure 4. These circuits can be divided between power conversions (A), user interface (B), sensors (C), main control of operation (D), actuation (E) and system wireless communications (F). Power conversion is required in order to comply with three different voltage levels, whereas the motor operates with 12 V, the LED lighting and sensors operate with 5 V and logical CMOS level of 3.3 V. The user interface was designed in order to be intuitive and waterproof with no leaking fissures to the interior of the device. Thus, the capacitive interface was selected with a dedicated microcontroller STM32F091CBT6 (IC2) ARM®® 32-bit Cortex®® M0 CPU frequency up to 48 MHz for both RGB led lighting (D1~D8) and capacitive detection. The main control of the operation was performed by STM32F303RET6 (IC1) ARM®® Cortex®® M4 32-bit CPU with 72 MHz FPU, which controlled the overall firmware architecture. Therefore, IC1 was able to control the operation of the motor with the aid of an H-Bridge DRV8872-Q1 (IC3) with a high range of operation up to 3.6 A and 45 V. In addition, it made the interface communication with IC2 and Bluetooth HM−10 (IC7) through UART, LCD screen design through SPI, sensors input through ADC channels and system-timings for culture operation. In order to measure the temperature, an LMT85 sensor (IC4) and magnetic field (AD22151) (IC5) in the most adequate position an extra PCB was designed in order to house both sensors and wire communicate with the main board for a closed-loop system response. A linear position sensor 9610R3.4KL2.0 (IC6) was used in order to access magnets platform position to control permanent magnets positioning and resulting magnetic field, was easily integrated with an ADC channel input. All integrated circuits were designed with their respective decoupling capacitors in order to avoid high-frequency power transitions to dwell in the rest of the circuit power lines. The 5 V regulator LM2596 (IC9) was a switching step-down with 80% efficiency, being important for this application by comparison with linear regulators, during long periods of use the amount of heat produced is considerably less for this level of power required. In order to power the logic circuits at 3.3 V, an LD1117 voltage regulator (IC10) was employed.

**Figure 4.** Main circuits used for the power conversion (**A**), user interface (**B**), sensors (**C**), system control (**D**), actuators (**E**) and wireless communications (**F**).

#### *2.2. Firmware and Interface Design*

The developed circuit was based on microcontrollers and digital communication with every component on the control boards. Regarding magnetic output, it must be noted that the peak amplitude was topped by the permanent magnets magnetization grade, thus the electromechanical component will control solely how much amplitude will reach the culture plate through the Hall sensor, as well as the displacement and frequency. Thus, the resulting mechanism control followed an approach by calculated displacement steps that will fit a distance at a given displacement frequency in order to output at the culture plate an approximate alternated magnetic field wave. However, motor sliding was a problem, using firmware breaking mechanisms with the aid of the IC3 H-Bridge, together with linear sensor for displacement error, more precise control was enabled over the mechanical response, and consequently the magnetic field. The interface, communication and each machine state (Figure 5) was controlled by firmware developed for each microprocessor according to their own respective tasks. The different machine states allowed the user to interface with each system functionality and take advantage of its calibration, variable settings and running processes.

**Figure 5.** State machine control nodes.

The device works through four main states: (i) a menu state where every state returns to and consists of the core control of the device; (ii) a calibration state where the user sets the starting position; (iii) a program calibration state where the user sets the culture variables; (iv) a running state where the machine performs the programmed culture by the user. Through the capacitive interface, the developed system allows us to navigate a menu to adjust variables (Table 1) such as whole culture duration; active and resting cycles duration which work through two temporal levels (short and long cycles); table position calibration and setting starting point. It is possible to store up to three culture programs through emulated electric erasable programmable read-only memory (EEPROM) in the program memory. Mechanical characteristics of the developed system allows us to move the magnets up to 25 mm at a speed of up to 2.5 mm per second. The firmware provides these limitations in order to protect the cells and hardware from human error by calculating hardware limits according to distance and operation frequency. The user can define various stimulation parameters such as critical temperature, displacement between 1 and 25 mm, resulting in operating frequencies between 0.1 and 2 Hz and different stimuli cycle timings to adapt the culture to the cell's native environments.



#### **3. Bioreactor Evaluation**

The performance of the herein developed bioreactor was evaluated using magneto-active materials based on composites comprising the piezoelectric poly(vinylidene fluoride) (PVDF) and the magnetostrictive Terfenol-D (TD) particles (Etrema Products) with approximately 1 μm diameter, as described in References [20,45]. This material was selected due to its magnetoelectrical properties, i.e., actively responding to the magnetic field provided by the magnetic bioreactor. Due to their magnetostrictive component (TD), the material senses the magnetic field, inducing a mechanical stimulation on PVDF, which due to its piezoelectric properties further induce an electrical polarization variation, creating the electrically active microenvironment that is translated to the cells [46].

MC3T3-E1 pre-osteoblast cells (Riken Bank) were used for the cell proliferation assays, as a proof of concept for bone regeneration studies. Previous to the cell culture studies, the cells were maintained in Dulbecco's modified Eagle's medium (DMEM from Gibco, ThermoFisher, Loughborough, UK) containing 1 g L−<sup>1</sup> glucose, 10% fetal bovine serum (FBS from Biochrom, Cambridge, UK), and 1% penicillin/streptomycin (P/S, Biochrom) in a controlled atmosphere at 37 ◦C and 5% CO2. The culture medium was replaced every 2 days, and at pre-confluence, cells were harvested using trypsin−ethylenediaminetetraacetic acid (EDTA)(Biochrom). Non-poled (non-charged) films were used to study the effect of the mechanical stimuli provided by the magnetostrictive particles in cell proliferation while poled films were used to study the influence of the mechano-electrical stimulus provided by the combination of electroactive PVDF and TD particles. ME films with a diameter of 1.3 cm were sterilized using UV for 1 h each side and placed in a 24-well tissue culture polystyrene plate. Then, 3 <sup>×</sup> 104 cells·mL−<sup>1</sup> in DMEM were seeded on each well and incubated for 24 h. For this, a drop method was used, in which approximately thirty-five microliter of DMEM containing 15,000 cells was first placed on the surface of the material for 30 min to allow the cell adhesion, and then 250 μL DMEM was added to the well. After 24 h incubation time, one plate was used for the static cell culture (without any stimulation) and the other was transferred onto the bioreactor for 48 h at 37 ◦C in a 95% humidified air containing 5% CO2, totalizing two cycles of magnetic stimulation. The dynamic stimuli provided by the magnetic bioreactor was achieved through the following procedure: an active time of 16 h under magnetic stimuli, which was divided into intervals of 5 min active stimuli and 25 min of resting followed by a period of complete inactivity of 8 h (Figure 6a). After 48 h, the 3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2 H-tetrazolium (MTS, Promega) assay was used in order to determine the cell viability at the defined time-points. MTS assay is a coloring method that allows determining the cell viability and is based on the NADPH or NADP-assisted bioreduction in living cells. For this assay, the samples were transferred to a new 48-well plate and further incubated with an MTS solution (in a 1:5 ratio) at 37 ◦C and 5% CO2. After 2 h, 100 μL of each well was transferred to a 96-well plate, and the optical density (OD) of each well was measured at 490 nm using a spectrophotometric plate reader (Synergy HT from BioTek, Colmar Cedex, France).

The selected conditions were employed in order to resemble the human body's daily mechanical conditions divided by 16 h of activity and 8 h of sleep, also considering the fact that bone is piezoelectric itself [21] and the magnetoelectrical scaffold is able to mimic the electroactive microenvironment upon magnetic stimulation. Those short bursts of stimuli for a duration of 5 min were performed by displacing 25 mm the magnetic table at a frequency of 0.3 Hz, resulting in a magnetic field variation of up to 23 mT within the cell culture wells. For every studied condition, three samples were assayed, and growing cell viability was determined through the MTS assay. For this assay, three main variables were considered: (i) under static conditions, i.e., without magnetic stimulus and bearing in mind the single effect of the different morphologies related to pore size differences, (ii) under dynamic conditions considering magnetic stimuli effect, and (iii) the relative effect between the material surface charge and magnetic stimuli.

**Figure 6.** (**a**) Stimuli schedule timing programmed in the bioreactor for pre-osteoblast tissue culture assays using either static or dynamic conditions and (**b**) cell viability after 48 h of cell culture on TD/PVDF films with and without magnetic stimuli. The cell viability was calculated regarding the cells growing on the non-poled ME film at static conditions presented as % of growth. In each study, three samples were assayed per studied condition.

The bioreactor system was found to be completely biocompatible and suitable for cell culture. After 48 h of cell culture, MTT results show that cells grow and proliferate independently of the condition applied, showing more than 100% of cell growth under all conditions. All the components are thus biocompatible and the system working properly to avoid the increase in the temperature, one of the main concerns related to this device.

Different conditions applied to the scaffolds further induce different effects in terms of cell proliferation. The application of magnetic stimuli brings an unequivocal increase of proliferation rate in all samples, indicating a clear response of the ME films to the magnetic field, thus demonstrating that the bioreactor provides a suitable microenvironment to the pre-osteoblast cells especially in positively charge TD/PVDF film (Figure 6b). The clear increase in cell viability upon application of the stimuli indicates that a mechanoelectrical effect occurs on non-poled samples while a magnetoelectric effect occurs on poled samples, being the later more beneficial for cell growth. On all tested magneto-responsive materials, statistically significant differences in proliferation rate were observed on the growing cells.

#### **4. Conclusions**

There is an ever-increasing need for more efficient strategies in TE applications. This fact is becoming a driving force in the R&D efforts to develop a new class of materials, smart materials that respond to stimuli that further triggers appropriate cellular response through the creation of a proper microenvironment. The complexity of these microenvironments where cells are able to optimize cell growth and control cellular functions, make it difficult to recapitulate in vitro. Thereby, the need for these materials and devices capable of recreating in vivo conditions are key elements for the next developments in TE applications.

The magnetic bioreactor developed in this study represents an advance on the state of the art in TE and provides, together with specific magnetoelectric scaffolds, the electrically active microenvironments necessary for cell tissue regeneration. The bioreactor was designed and constructed taking into consideration the envisaged operating principles, by using biocompatible materials, conventional mechanics and digital electronics. It also has the potential to integrate other electronic modules that support digital communication for synchronization. This study further proved that tissue cultures may be performed with this system since a boosting effect on the proliferation rate was observed upon application of the stimuli and no signs of toxicity were found. Simultaneously, this experiment demonstrates the suitability of magneto-responsive scaffolds for adhesion and proliferation of pre-osteoblasts, availing itself from the mechanical and electrical microenvironment conceived in the material. It was possible to conclude that the magnetic module of this bioreactor was able to provide an important contribution to building the proper microenvironment as a device, whereas a scaffold which provides the proper cues to cells with the physical environment, mechanical and electrical stimuli that can be used synergistically with the system.

Therefore, this work has provided the development of a novel bioreactor based on magnetic stimulation that has proven that the developed bioreactor is biocompatible and that may be used for advanced tissue engineering applications, allowing for advanced tissue engineering strategies. It will certainly act as a valuable tool for mimicking in vitro the human stimulations provided by the electrically active tissues that are present in the body. It could also be important for growing well-formed cellular tissues in vitro in a more effective and rapid way, which could be further implanted in the human body without the material. In the case that magnetoactive materials to be implanted in the human body, it would provide a suitable platform to evaluate the remote stimulation and thus for effective growth and differentiation of cells in immobilized patients. In fact, mimicking cell microenvironments is thus a key issue to recapitulate in vitro what occurs in vivo and this bioreactor holds great promise to fulfill such requirements.

**Author Contributions:** Conceptualization, S.L.-M., V.C. and C.R.; methodology V.C. and C.R.; hardware N.C. and V.C.; software N.C.; validation M.M.F., C.R. and N.C.; formal analysis N.C., V.C., R.M., S.L.-M., C.R. and M.M.F.; investigation C.R. and M.M.F.; writing—original draft preparation N.C.; writing—review and editing S.L.-M., R.M., M.M.F. and N.C.; supervision, S.L.-M. and R.M.; project administration, S.L.-M.; funding acquisition, S.L.-M. All authors have read and agreed to the published version of the manuscript.

**Funding:** FCT—Fundação para a Ciência e Tecnologia: UID/FIS/04650/2020; PTDC/BTM-MAT/28237/2017; PTDC/EMD-EMD/28159/2017 and SFRH/BPD/121464/2016. Spanish Ministry of Economy and Competitiveness (MINECO): MAT2016–76039-C4–3-R (AEI/FEDER, UE). Basque Government Industry and Education Department: ELKARTEK, PIB and PIBA (PIBA−2018–06) programs, respectively.

**Acknowledgments:** The authors thank the FCT—Fundação para a Ciência e Tecnologia—for financial support under Strategic Funding UID/FIS/04650/2020 and projects PTDC/BTM-MAT/28237/2017 and PTDC/EMD-EMD/28159/2017. MMF thank FCT for the SFR H/BPD/121464/2016 grant. The authors acknowledge funding by the Spanish Ministry of Economy and Competitiveness (MINECO) through the project MAT2016–76039-C4–3-R (AEI/FEDER, UE) and from the Basque Government Industry and Education Department under the ELKARTEK, HAZITEK and PIBA (PIBA−2018–06) programs, respectively.

**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/).

## *Letter* **Component Analysis of Gas Mixture Based on One-Dimensional Convolutional Neural Network**

**Canjian Zhan, Jiafeng He \*, Mingjin Pan and Dehan Luo**

School of Information Engineering, Guangdong University of Technology, Guangzhou 510006, China; zcj@mail2.gdut.edu.cn (C.Z.); illusionnol66@gmail.com (M.P.); dehanluo@gdut.edu.cn (D.L.) **\*** Correspondence: jfhe@gdut.edu.cn; Tel.:+86-1814-891-8377

**Abstract:** Indoor harmful gases are a considerable threat to the health of residents. In order to improve the accuracy of indoor harmful gas component identification, we propose an indoor toxic gas component analysis method that is based on the combination of bionic olfactory and convolutional neural network. This method uses the convolutional neural network's ability to extract nonlinear features and identify each component of bionic oflactory respense signal. A comparison with the results of other methods verifies the improvement of recognition rate while with the same level of time cost, which proved the effectiveness of the proposed model. The experimental results showed that the recognition rate of different types and concentrations of harmful gas components reached 90.96% and it solved the problem of mutual interference between gases.

**Keywords:** electronic nose; convolutional neural network; component analysis

#### **1. Introduction**

Because low-concentration indoor harmful gases are invisible and tasteless, they are difficult for people to distinguish. We can detect low-concentration indoor toxic gases through physical and chemical identification methods. Still, they are cumbersome and complicated operations, and to use the instrument needs to be professionally trained. It is difficult to promote in the market.

Many different methods are applied in indoor air environment monitoring for the quantitative analysis of harmful gases, including the non-dispersive infrared method [1], gas chromatography [2], nessler's reagent colorimetry [3], and ion-selective electrode method [4]. The methods, as mentioned above, are relatively complicated and they cannot perform real-time on-site air quality testing. With the rapid development of information science and sensor technology, the bionic olfactory system has been applied in medical, food processing, and environmental detection fields, with its advantages of simplicity and economy.

However, when machine olfactory technology is used in the quantitative analysis of substance odor in an open environment, it is easily affected by interfering gases and environmental temperature and humidity, which causes the problem of reduced recognition accuracy.

The machine learning algorithms and their optimization methods were applied to the quantitative analysis of machine olfactory rapidly. For example, Xianjiang Li et al. proposed an optimization model for mixed gas quantitative detection that combines an adaptive genetic algorithm and a traditional BP neural network. This algorithm can overcome the shortcomings of the slow search rate and easily fall into local minimum. The BP neural network can obtain better initial weights and early stage thresholds through the adaptive genetic algorithm. Subsequently, the experiment uses this algorithm to quantitatively analyze the machine olfactory odor data of a set of five-element gas mixtures. The experiment shows that the accuracy of the algorithm for identifying the gas concentration of the gas mixture is higher than that of the traditional BP neural network [5]. Shurui

**Citation:** Zhan, C.; He, J.; Pan, M.; Luo, D. Component Analysis of Gas Mixture Based on One-Dimensional Convolutional Neural Network. *Sensors* **2021**, *21*, 347. https:// doi.org/10.3390/s21020347

Received: 16 November 2020 Accepted: 2 January 2021 Published: 6 January 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 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 (https:// creativecommons.org/licenses/by/ 4.0/).

Fan et al. used principal component analysis and random forest modeling, which can qualitatively identify methane and ethylene gases. Subsequently, the optimal regression model is constructed through support vector regression and particle swarm optimization for quantitative analysis of the two types of gases. The experimental results show that the average recognition rate of principal component analysis, combined with random forest, is 97% higher than logistic regression and the support vector machine. The fitting effect of support vector regression is optimized by the particle swarm algorithm, and better fitting results are obtained than support vector regression [6].

In practical applications, when using machine olfaction to quantitatively analyze the odor of substances, it is often accompanied by the influence of interfering gas. Therefore, nonlinear or linear inseparable phenomenon, or even gas shielding, will often appear in the machine olfactory system. The phenomenon of shielding between gases means that the response of the electronic nose sensor when measuring mixed gas is lower than that of measuring pure gas of the same concentration [7]. However, the above-mentioned literature does not fully consider the influence of these phenomena on the quantitative analysis of machine olfaction [8], so some improved algorithms are proposed.

Yu Lu et al. combined the artificial neural network with the basic concepts of analytical chemistry, designed an error function based on analytical chemistry, and applied the error function to the neural network. This method can be used to control alcohol, petroleum gas, and water. The experimental results show that the method predicts a gas concentration error less than 10% [9]. Tang K TZ et al. proposed a Locally Weighted Nearest Neighbor (LWNN) algorithm that is based on the K-Nearest Neighbor classifier (KNN) algorithm to determine the odor components, and then combined with the types of odor components, while using the weighted and constrained least squares (Weighted and Constrained Least-Squares, WCLS) gas concentration estimation method that is based on the least squares method measures the target gas concentration in the mixed gas [10].

The above analysis shows that, in addition to investigate the feasibility of applying machine odor perception to quantitative gas analysis, great progress has been made in the component analysis methods of machine odor perception; however, there are still various shortcomings. The methods are not universally applicable and they do not take the mutual influence between different gas concentrations and the reversal of the response curves into consideration. Therefore, none of the above methods can be used for the component analysis of indoor pollutant gases while using the odor perception engine. It is necessary to find appropriate methods that are based on the properties of the pollutant gas odor data for indoor spaces in order to extract the component properties of the target gas (for example, formaldehyde) among other indoor gases. Based on the background values of indoor air pollution, this paper proposes a component analysis method for indoor pollutant gases when considering the influence of interfering gases.

#### **2. Gas Data Preprocessing**

According to the requirement of the research objectives and content of this article, a machine olfactory system is required for collecting the odor information of indoor harmful gases. This article contains the odor data of multiple indoor toxic gas samples through the PEN3 electronic nose system in order to ensure the reliability of the data. The PEN3 electronic nose is the third generation of the PEN series developed by AIRSENSE, Germany. It is built with 10 cross-sensitive metal oxide gas sensors. Table 1 shows the characteristics of the sensors in PEN3 electronic nose.


**Table 1.** The characteristics of the sensors in PEN3 electronic nose.

This article selected the four most common pollution gases in daily life based on the needs of the research objectives and content of this article. They are formaldehyde, ammonia, benzene, and methanol, produced by Henan Testing Center. Additionally, the multi-channel gas mixing system performs the ratio of indoor harmful gases in order to ensure the objectivity and reliability of the data. Appropriate experimental materials and exhaust gas treatment equipment should be selected when designing the practical plan in order to ensure the safety of the experiment. Figure 1 shows the process of harmful indoor gas collection based on machine olfactory. In order to obtain accurate gas concentration sample data, the MT-500X dynamic gas mixing system (gas mixing instrument) is used in this paper to carry out the precise ratio of indoor harmful gas samples, to ensure the objectivity and accuracy of the data that were collected in this paper. The system is equipped with a high-precision mass flow controller, which can meet the requirements of stable, reliable, and high-precision gas distribution.

**Figure 1.** The experimental procedure of indoor hazardous gas collection based on machine olfaction.

Before starting the experiment, the indoor air conditioner and humidifier should be turned on, so that the temperature should be controlled within the range of 25 ± 1 ◦C and the humidity should be controlled at 75 ± 1%. Table 2 shows the other experimental parameters. First, we prepare the gas samples that are required for the experiment with the standard gas through the gas mixing instrument, transport the prepared gas samples to the gas testing chamber, and then collect the odor data through the PEN3 electronic nose. Finally, the residual test gas in the gas test box is passed through the tail gas treatment device for harmless treatment.


**Table 2.** The parameters of the gas data acquisition experiment.

According to different interference groups, the response curve of the electronic nose to different indoor harmful gases in the same interference group can be drawn through the data set. The concentrations of formaldehyde gas containing 0.01 mg/m3, 0.05 mg/m3, 0.09 mg/m3, 0.13 mg/m3, 0.17 mg/m3, and 0.21 mg/m3 are drawn, respectively, as shown in Figures 2 and 3.

**Figure 2.** The response curve of weak interference group.

**Figure 3.** The response curve of moderate interference group.

By observing Figures 2 and 3, it can be found that, in the case of the same interference group, each response curve of the sensor is very similar in the overall listing, and it is still necessary to use the radar chart to supplement the observation. The radar chart data still uses the data with a sampling interval of 30 s to 40 s in order to calculate the average value, and then the average value is converted into a radar chart, as shown in Figure 4.

**Figure 4.** Radar diagram for each set of data.

The response values of sensors numbered S2, S6, and S8 are quite different from other sensors, especially for S2. The response of the sample is more sensitive. However, with the exception of the S2, S6, and S8 sensors, the difference in the response values of other sensors is small enough to determine that "high dimensionality, redundant information, and non-linearity" are the characteristics of indoor harmful gas data that are collected by machine olfactory.

The above data show that we can infer that indoor harmful gas samples containing different formaldehyde gas have differences in data that are based on the difference between the response curve of the sensor and the radar chart. According to their differences, suitable identification methods can be selected in order to train the computer to identify the level of formaldehyde pollution in indoor harmful gases.

In this experiment, Table 3 shows the 60 × 10 odor data matrix generated by PEN3. Each column represents a different sensor, and each row represents the response of the same sensor at different sampling times.


**Table 3.** Raw data format.

According to the original data format of the PEN3 electronic nose in Table 3, the data matrix of 60 × 10 is first transposed to the data matrix of 10 × 60, and then the data matrix is represented by a brand new row vector. The specific method is as follows. Based on the corresponding sensor that generates the response, the original matrix is divided into 10 row vectors of 60, and is then sequentially connected by the sensor number to form a new row vector, and stored in the .csv format. If the sampling time of the acquisition experiment is 60 s to collect m samples, then the converted data set of the electronic nose data file is a .csv file with m 600-dimensional features. In this paper, 1040 gas samples are collected through experiments, so the data in this paper are concentrated.

In this paper, a total of 1040 data samples were collected through the collection experiment. According to the indoor air quality standard (GB/T18883-2002), the data samples are divided into three types of data with different pollution levels. Among them, 320 data samples containing formaldehyde gas from the concentration of 0.01 mg/m3 to 0.08 mg/m<sup>3</sup> (without 0.08 mg/m3) are qualified (Normal), and 320 data samples with the formaldehyde concentration from 0.08 mg/m<sup>3</sup> to 0.16 mg/m<sup>3</sup> were classified as mildly polluted (Mild), and 400 data samples containing formaldehyde concentrations that were greater than 0.16 mg/m<sup>3</sup> were classified as severely polluted (Serious). Select 70% of the data samples of each type of pollution level as the training set and the remaining 30% data samples as the test set of the experiment. The data set used for model training has 728 data samples, and the data set used for testing has 312 data samples.

Therefore, we construct an indoor hazardous gas odor information collection system. A number of indoor harmful gas samples with different concentrations were prepared by a dynamic mixing gas distributor, and the PEN3 electronic noses was used in order to collect odor data from these samples, thereby obtaining a data set of the response of the indoor harmful gases. The exhaust gas of the experiment was treated in a harmless manner in order to eliminate the impact of harmful gases on the health of the experimenters. According to the characteristics of the electronic nose response data, a component analysis method of indoor harmful gases based on machine olfaction and global average pooling convolutional neural network model is proposed.

#### **3. One-Dimensional GAP-CNN**

CNN is a neural network with deep structure and a classic algorithm that is widely used in deep learning [11]. Nowadays, many typical and widely used CNN models have been proposed, such as LeNet-5 [12], AlexNet [13], and GoogleLeNet [14]. They have been successfully appled in face detection [15], role recognition [16], pedestrian detection [17], and robot navigation area [18]. Because CNN has the structural characteristics of local connection, weight sharing, and down-sampling, the model is sample-invariant to translation, scaling, and distortion, and, thus, has strong robustness [19]. This feature makes convolutional neural networks a great success in the field of image processing. The main difference between CNN and the traditional BP (Back Propagation, BP) neural network lies in the two aspects of weight sharing and local connection. Weight sharing makes the convolutional neural network more suitable for the structure of biological neural networks. The local connection of convolutional neural network is not like a traditional neural network. Each neuron in the first layer is connected to all neurons in the first layer, but the neurons in the first layer are partially connected to the neurons in the first layer. The role of these two characteristics makes the model have lower model complexity and fewer weights than traditional BP neural networks.

The convolution layer performs convolution processing on the input data through multiple convolution kernels and extracts the convolutional features, which is the feature map. A corresponding type of feature is extracted through a convolution kernel. Because the operation of the same convolution kernel has the characteristics of local connection, parameter sharing, and multiple convolution kernels, when compared with the fully connected layer, the convolution layer can propose more features with fewer parameters when extracting data features. Because the convolution structure is not affected by the input dimensions and the training depth structure is simple, it can effectively extract features from complex and high-latitude inputs. The convolution formula of the convolution layer is:

$$\log(i) = \sum\_{x=1}^{m} \sum\_{y=1}^{n} \sum\_{z=1}^{p} \alpha\_{x,y,z} \times w\_{x,y,z}^{i} + \beta^{i}, i = 1, 2, \dots, q \tag{1}$$

where: *i* is the *i*-th convolution kernel, *g*(*i*) is the feature map that is extracted by the *i*-th convolution kernel; *α* is the input data; *β* is the bias of the convolution kernel; *x*, *y*, *z* represent three different dimensions of data. After completing the convolution of the data, it is necessary to use a nonlinear activation function in order to perform nonlinear

conversion on the data. The commonly used activation function in CNN is generally ReLU, and its formula is:

$$y(i) = f(\lg(i)) = \max\{0, g(i)\}, i = 1, 2, \dots, q \tag{2}$$

$$p^{l(i,j)} = \max\_{(j-1)w < t < jw} \left\{ a^{l(i,t)} \right\}, j = 1, 2, \dots, q \tag{3}$$

$$p^{l(i,j)} = \underset{(j-1): w < t < jw}{\text{avg}} \left\{ a^{l(i,t)} \right\}, j = 1, 2, \dots, q \tag{4}$$

where: *αl*(*i*,*t*) is the *t*-th neuron of the *i*-th feature map in the *l*-th layer, *w* is the width of the convolution kernel, and *j* is the *j*-th pooling kernel [20].

In the convolutional neural network, the convolutional layer and the pooling layer both perform feature extraction on the data, and the operation of data classification is performed in the fully connected layer based on the feature. The fully connected layer can integrate features through the feature maps output by the convolutional layer, thereby obtaining classification information with high-level meaning, and then classify and output according to these. As the output of the CNN model, the output of the fully connected layer is a fixed-length feature vector that is obtained by transforming the feature map input from the layer. The feature information of all combinations of the original data will be integrated by this feature vector. Although this fixed-length feature vector does not have the location information of the original data, the feature information used to effectively complete the data classification task has been fully extracted [21].

In the traditional convolutional neural network, the parameter ratio of the fully connected layer is almost 80% of the entire neural network model [22]. In order to achieve the purpose of reducing network parameters, there are generally two common methods. The first method is to replace the fully connected layer in the model with a convolutional layer. If the convolution kernel of the same size as the feature map in the full connection layer is used as the input of the output layer, this often makes the model output result far inferior to using the fully connected layer as the input of the output layer accurate. The second approach is to reduce the feature output dimensions of all the layers in the network, but the disadvantage of this approach is that the neural network lacks sufficient features for network training, which results in the final output feature information that contains less feature information than the original dimension, which reduces the model accuracy of recognition.

In this article, we refer to the idea of using global pooling in GoogleNet,and use the global pooling layer to replace the fully connected layer in order to extract features [23]. There are generally two methods for global pooling, global average pooling (GAP) and global maxing pooling (GMP). This article takes the global average pooling as an example to illustrate the difference between global pooling layer and full connection. The global average pooling layer, like the fully connected layer, has the function of connecting global feature information, so as to ensure the accuracy of the output result of the recognition model. However, because the global pooling layer does not contain any parameters, the phenomenon where the fully connected layer is easy to overfit is avoided during the model training process.

#### **4. Result**

In order to verify that GAP-CNN can reduce the weight parameters while ensuring the accuracy of identifying formaldehyde pollution in indoor harmful gases, a CNN model with the same depth and the same convolution kernel size as GAP-CNN will be used as the control group. Both network models use the same convolution layer and convolution kernel with the same parameter settings. However, in the CNN model, the number of parameters that need to be trained is 14,963, which is much higher than the number of 4259 parameters shown in the GAP-CNN model. The reason is that in the CNN model, the number of parameters in the first dense layer, which is the fully connected layer, accounts for 70% of the total network model parameters. This also proves that global pooling can effectively compress the volume of the convolutional neural network model. The parameters of the prediction models of the two convolutional neural networks are set as: the number of training data is epoch = 400, the size of the training set data and the test set data is randomly selected each time batchsize = 32, and the learning rate *α* is 0.01. Figures 5 and 6 show the training curves of the two models, respectively.

**Figure 5.** Training curve of global average pooling-CNN (GAP-CNN) model.

**Figure 6.** Training curve of CNN model.

Among them, the abscissa represents the number of epochs of model training, the ordinate represents the accuracy and the size of the loss function value, train acc represents the recognition rate of the model to the training set, train loss represents the value of the loss function of the model to the training set, the val acc model represents the recognition rate of the set, and val loss represents the value of the loss function of the model to the training set. In Figure 5, the values of train loss and train acc of GAP-CNN reach a stable state after the number of training is 260. The GAP-CNN model has been trained. The value of train acc of the model is 0.9684, and the value of train loss is 0.0954. The correct recognition rate val acc of the model on the training set is 0.9140. From Figure 6, the train loss and train acc values of the CNN model reach a stable state after 230 training times. The train acc value of the CNN model is 0.9615 and the train loss value is 0.1083. The correct recognition rate val acc of the CNN model on the training set is 0.9071.

When comparing the training curves of the two models, we can find that, under the same training parameters, although the convergence speed of the GAP-CNN network model is slightly slower than that of the classic CNN, the recognition of GAP-CNN on the

test set of indoor harmful gases. The accuracy is slightly better than the classic CNN model, which also proves that the GAP-CNN network model can reduce the model parameters while reducing the occurrence of overfitting, and it can also ensure a certain model accuracy. By saving the GAP-CNN model to identify the test set again, it is found that there are three main reasons why the classification accuracy rate cannot reach 100%. The first is that GAP-CNN will make a small part of the formaldehyde concentration in the strong interference group close to 0.08 mg/m. The Normal class is incorrectly identified as the Mild class. The second reason is that GAP-CNN will incorrectly classify a small part of the Mild and Serious, especially gas samples whose concentration is at the classification boundary. For example, gas samples containing Mild's 0.14 mg/m<sup>3</sup> formaldehyde will be classified as Serious. The Serious class' 0.18 mg/m3 will be classified into the Mild class. The third reason is that there is a high concentration of interfering gas, indoor harmful gas samples containing low concentration of formaldehyde gas, the model can easily identify it as a higher level of formaldehyde pollution level.

In order to further verify the effectiveness of the GAP-CNN algorithm,in Table 4, PLS model, PCA+LDA model, CNN model, t-SNE+RF model, and GAP-CNN model are used in order to calculate the average of the classification results while using five-fold cross-validation.

**Table 4.** Cross validation results of 5 models.


Through the analysis and research of the above experimental results,it can be known that the traditional PCA+LDA machine olfactory recognition method cannot effectively extract the effective features for the classification of formaldehyde pollution in indoor harmful gases. The PLS model can extract the effective features of formaldehyde pollution classification in indoor harmful gas to a certain extent, but, in the presence of other indoor harmful gas interference, the accuracy of algorithm recognition is low. The t-SNE+RF algorithm is less accurate than the classification that is based on the CNN neural network. In the CNN and GAP-CNN models, although the classification accuracy of the Normal level of formaldehyde pollution is not as good as t-SNE+RF, as compared with the t-SNE+RF algorithm, it is found that the advantage of the CNN neural network formaldehyde pollution classification is that CNN When the neural network recognizes the level of formaldehyde pollution of Mild and Serious, it has higher classification accuracy. At the same time, GAP-CNN has better generalization performance than t-SNE+RF.

By observing the results of cross-validation experiments, it can be found that CNN, t-SNE+RF, and GAP-CNN have an ideal effect on identifying the level of formaldehyde pollution in indoor harmful gases. Therefore, this paper uses the aboved algorithm to compare and analyze the total computing time of data preprocessing, feature extraction, model training, and output results in order to further compare the characteristics and performance of the three machine learning algorithms to identify machine olfactory odor data. The operation time presented in Table 5 shows that GAP-CNN algorithm needs to go through multi-layer convolution operation, and the convergence speed is slow, which leads to greater time consumption of GAP-CNN algorithm.

**Table 5.** Computational time comparison of discrimination algorithm.


#### **5. Discussion**

This paper proposes an indoor harmful gas component analysis algorithm that is based on the combination of CNN and bionic olfactory. This method uses one dimensional convolutional neural network weight sharing and adding a global maximum pooling layer, so that the neural network has a higher recognition rate for indoor harmful gases when the number of training parameters is small. The research of this algorithm is of great significance to the solution of the subsequent concentration estimation problem of the bionic olfactory system. The algorithm that is proposed in this paper has not been well verified in the concentration regression experiment. In the process of the experiment, the influence of external factors on the experimental results has not been considered. This will be the direction of follow-up research.

#### **6. Conclusions**

In this work, a novel component analysis strategy was proposed for formaldehyde pollution in harmful indoor gases based on machine olfaction. While using a portable electronic nose system, indoor harmful gas samples' odor information is collected and processed to achieve an real-time monitoring. By converting the smell information of the machine olfaction into a one-dimensional time series, the machine olfactory analysis method GAP-CNN that is based on global average pooling and one-dimensional convolutional neural network is innovatively proposed. Additionally, through the comparison of the experimental results, in the case of compressing the model volume, the recognition accuracy of the model is guaranteed, which proves the performance of the GAP-CNN model. At the same time, this paper uses the GAP-CNN model and the classic CNN model to build an indoor hazardous gas formaldehyde pollution classification model. Through comparing the experimental results, it is found that the accuracy of the two classification models is greater than 90%, and the weight parameters of the GAP-CNN model are much lower.

**Author Contributions:** The work described in this article is the collaborative development of all authors. C.Z. and J.H. contributed to the idea of data processing and designed the algorithm. D.L. and M.P. made contributions to data measurement and analysis. C.Z. and M.P. participated in the writing of the paper. All authors have read and agreed to the published version of the manuscript.

**Funding:** The authors acknowledge the financial support of the General Program of the National Natural Science Fund (Grant No. 61571140), Guangdong Provincial Project (Grant No. 2016A020226018), Guangdong Provincial Administration of traditional Chinese Medicine Project (Grant No. 20161152), Guangdong University Project (Grant No. 51348000).

**Data Availability Statement:** All data generated or appeared in this study are available upon request by contact with the corresponding author. Furthermore, the models and code used during the study cannot be shared at this time as the data also forms part of an ongoing study.

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

#### **Abbreviations**

The following abbreviations are used in this manuscript:


#### **References**

