Evaluation of Artificial Neural Network to Model Performance Attributes of a Mechanization Unit (Tractor-Chisel Plow) under Different Working Variables

Abstract

The focal objective of the current research is to apply artificial neural network (ANN) and multiple linear regression (MLR) methods for modeling the performance attributes of a mechanization unit (tractor-chisel plow) during the plowing process under both different soil types and working variables. Two different parameters to represent working conditions and soil type were considered as potential input parameters. The first parameter represented soil type by calculating soil texture index as a combination of clay, silt, and sand. The second one was constructed into one dimensionless parameter, namely the field working index. This index linked most working variables such as plowing speed, plow width, soil moisture content, soil bulk density, tractor power, and plowing depth. The performance of the created ANN and MLR models was appraised by computing mean-absolute-error criterion for the testing dataset. The mean absolute error values for draft force, effective field capacity, fuel consumption, drawbar power, overall energy efficiency, rate of plowed soil volume, and loading factor, were 1.44 kN, 0.03 ha/h, 1.17 L/h, 2.28 kW, 0.68%, 73.97 m³/h, and 0.08 (decimal), respectively, when the ANN model was applied. In addition, coefficient of determination (R²) acted as a criterion for judging the performance of the developed models, and their values when ANN was applied were 0.569, 0.384, 0.516, 0.454, 0.486, 0.777, and 0.730 for the same performance attributes, respectively. When the MLR model was applied, the corresponding values of R² were 0.239, 0.358, 0.352, 0.429, 0.511, 0.482, and 0.422, respectively, for the same attributes. The current study adds to the standing literature by contributing data and information regarding the performance attributes of a tractor-chisel plow unit under specific working variables and soil types. In addition, the models developed for plowing operations in different soil texture and under the field working index are recommended for use in budgeting for diesel consumption, in calculating operation cost, in matching mechanization units of tractor-chisel plows, in estimating energy requirements of tractor-chisel plow combinations, etc.

1. Introduction

For tillage or soil preparation and other agricultural working tasks, the matching of tractors and farm implements in one mechanization unit is highly important [1] because this such unit is the main player and one of the largest cost factors in agricultural production [2]. The chisel plow is a widely used agricultural tool for primary tillage and for initially working the soil. The tractor is interfaced with farm implement to deliver power and tractive work to move and control the chisel plow through the field, and together they form a tractor-implement unit [3]. Thus, the information related with the performance of the tractor-implement unit is necessary for both implement manufacture and field-operation management [4]. As previously stated by Abbaspour-Gilandeh et al. [4], field study of factors affecting the field performance of a tractor-implement unit is accordingly expensive and time consuming, so empirical mathematical models or computer simulations could be useful for agricultural engineers to determine the effects of relative factors. The factors include specifications of the tractor and implement, soil conditions (firm, tilled, or soft) and operation conditions (tillage depth and speed), etc. [5].

The prediction of tractor-implement unit performance is a major objective for reducing the cost of agricultural production [6]. It is vital to assess the correct combination of a tractor and a farm implement unit to determine the best utilization of available power of a tractor [7]. In addition, as farmers extensively use chisel plows as a tool for primary tillage, gathering performance data for its operation is necessary to decrease the cost of agricultural production and optimize performance [8]. Normally, wheel slip, fuel consumption, draft force, tractive performance, drawbar power, overall energy efficiency, and rate of plowed soil volume are considered to represent performance attributes of a mechanization unit [9,10,11]. Numerous research papers have established practical mathematical equations to predict performance attributes of a tractor-chisel plow unit, to enhance and realize the activities of the main parameters of draft force requirements [12] and fuel consumption [13,14], which are in fact better indicators of the energy requirement for each implement [15].

Recent studies have appraised the prediction behavior of soft-computing approaches (neural networks, adaptive-network fuzzy-inference systems, and genetic algorithms) for the examination of tractive properties of an agricultural tractor under various soil conditions [16]. However, the artificial neural network (ANN) technique is well-known among soft-computing approaches. ANNs are nonlinear processes based on computer algorithms that represent the behavior of complex tasks [17] and can identify complex relationships in data [18]. Roul et al. [19] appraised an ANN structure to expect the draft requirements of numerous tillage implements, particularly in a soil having a sandy clay loam texture. The created ANN structure expected the draft requirements of disk harrow, moldboard plow, and cultivator with an experimental error less than 6.5% when compared to the measured draft values. Ajdadi and Abbaspour-Gilandeh [20] applied the ANN method to forecast fuel consumption using data that were available from the tractor testing by staff of Nebraska Tractor Test Laboratory. The results revealed that both stepwise regression and ANN models presented very little variation in R² values of 0.986 and of 0.973, respectively. Taghavifar et al. [21] applied a 3-9-1 feedforward with a back-propagation learning algorithm as the modeling structure for predicting the power requirements for the driving wheels of off-road vehicles. The inputs were wheel load, velocity, and slip. The computed R² for the training and testing phases of the best artificial neural network–genetic algorithm model was obtained at 0.9696 and 0.9672, respectively. Almaliki et al. [3] applied different ANN structures with inputs of tillage depth, engine speed, tractor forward speed, tire inflation pressure, soil moisture content, and soil cone index to estimate performance attributes of a tractor-moldboard plow unit. The ANN with 6-7-1 structure and training algorithm of Bayesian regulation was the best ANN model for drawbar power prediction, with R² of 0.995 and mean square error (MSE) of 0.00024. In addition, applying the ANN model with 6-7-1 topology under the training algorithm of Levenberg–Marquardt had the greatest performance for estimating temporal fuel consumption, with R² of 0.969 and MSE of 0.13427. The 6-8-1 structure showed the top ANN model for area-specific fuel-consumption prediction, with MSE of 0.01348 and R² of 0.885. In addition, the 6-10-1 topology produced the finest performance for specific fuel-consumption estimation, with R² of 0.935 and MSE of 0.012756. They stated that the created ANN topologies could be clever to acquire the correlation between the input factors and performance attributes of a tractor-moldboard plow unit to give accurate results. Borges et al. [22] estimated the tractor fuel consumption in the course of soil tillage using an ANN model. They decided that the ANN model performed very well and accurately in estimating the rate of tractor fuel consumption during the soil-tillage process. Shafaei et al. [23] applied an ANN model to estimate the rate of fuel consumption during the soil-plowing process and the results supported that the ANN can reliably acquire the correlation between the input factors and fuel consumption. Abbaspour-Gilandeh et al. [24] employed the ANN method to estimate the horizontal force of a chisel cultivator with a rigid tine. The inputs were soil cone index, plowing speed, plowing depth, and soil moisture content. The results displayed that in comparison to the ANN model, the MLR model had both worse accuracy and a lower correlation coefficient. Çarman et al. [25] employed MLR and ANN approaches to forecast the horizontal force and the area of plowed soil by operation of a chisel tine. The study inputs were blade (share) width, tillage speed, and working depth. In estimation of required horizontal force and plowed soil area according to root-mean-square error criterion, the better values with the ANN model were 0.010 kN and 0.016 cm², respectively, which were better performed than the MLR model. Küçüksarıyıldız et al. [26] predicted the specific fuel consumption of a 60 HP tractor by an ANN model. The study inputs were drive wheel pressure, dynamic rear axle load, and drawbar force. The root-mean-square error value obtained for the estimation of specific fuel consumption was 0.007551 g/kWh.

Understanding and quantifying the features of a tractor’s performance, while executing tasks in tillage operations, is a new and valuable asset to agricultural machinery management. The performance knowledge can be used to assess potential machine efficiencies when the tractor is coupled with varying implements [14]. Additionally, the presentation of soft-computing approaches, involving ANN and adaptive-network fuzzy-inference systems, has provided positive results for the analysis of tractive performance of an agricultural tractor under various soil conditions [16]. Therefore, the objective of the current research was to assess the ANN and MLR models’ capability to predict performance attributes of a tractor-chisel plow unit.

2. Materials and Methods

2.1. Plowing-Field Experiment

The plowing-field experiments were conducted in varying soil types. The first soil type had a texture of silty clay loam with 18% silt, 52% sand, and 30% clay located in a field belonging to Alexandria Governorate, Egypt (longitude: 29°58′34.6″ E and latitude: 31°130′09.3″ N) and was inside the Abies area. The second soil type had a texture of clay soil with 15% sand, 40.6% silt, and 44.4% clay that was part of a field inside El Gemmaiza Research Station, El Gharbia Governorate, Egypt (longitude: 31°07′31.8″ E and latitude: 30°47′19.8″ N). The third soil type had a texture of clay soil with 28.5% sand, 17.7% silt, and 53.7% clay and was located in a field belonging to Kafer El Sheikh Governorate, Egypt (longitude: 30°51′17.6″ E and latitude: 31°06′59.3″ N) inside the Rice Mechanization Center, Meet El Deeba Province. The overall energy efficiency, effective field capacity, draft force, fuel consumption, rate of plowed soil volume, and loading factor were determined using different tractor powers under different levels of plowing speeds by selecting different gears, different levels of tillage depths, and different soil moisture content levels. The dataset contained 70 instances. The levels of plowing speed, tractor power, soil bulk density, tillage depth, and soil moisture content at each experimental location are shown in

Table 1. Number of data points and levels of plowing depth, tractor power, soil bulk density, plowing speed, and soil moisture content in the experimental sites.

Sites	Tractor Power	Plowing Depth	Plowing Speed	Soil Moisture Content	Soil Bulk Density	Number of Data Points
	(kW)	(cm)	(km/h)	(%db)	(g/cm3)	(-)
Abies area	67	14.0	2.5,3.8,4.8	18.2	1.28	3
El Gemmaiza Research Station	82	10,12,14,15	3.5,4.8,5.7	18.2	1.40	12
		10,12,15	4.8	18.2	1.40	3
Meet El Deeba Province	82	10,13,15,16	2.5,3.4,4.8	17.4	1.35	12
		10,14,17,18	2.4,3.5,4.6,5.1	17.6	1.30	12
		10,13,14,16	2.5,3.2,5.1	20.1	1.38	12
		9,11,14,16	3.2,3.7,4.7,6.9	19.8	1.36	16

. For each soil type, soil samples were picked up using a cylindrical core tool from five unplanned spots from a 30 cm topsoil layer. In an electric oven, the collected soil samples were dried at 105 °C for 24 h to determine levels of the soil moisture content and soil bulk density.

A locally manufactured chisel plow (Type RAU, Behera Company) was used in all field experiments. The plow weighed 460 kg (4.51 kN) and had a width of 1.75 m. It had seven shanks distributed in two rows. A different tractor make was utilized to hitch the chisel plow in each site, and the auxiliary tractors were different makes. The horizontal force (draft) was acquired in Abies area using a locally manufactured strain-gauge pull meter [27,28] with a 10 kN capacity that was joined between the two tractors. However, an Arduino Uno board arrangement was developed to acquire draft signals [29]. In addition, a fabricated locally made hydraulic pull meter, which consisted of a cylinder-piston system, connected to a Bourdon tube gage [30] was used to obtain draft data during field experiments at Meet El Deeba Province and El Gemmaiza Research Station sites. One pass was considered as a data point accurate throughout the tillage experiments.

2.2. Formulas Used

In this research, the dataset covered vast levels of working variables under different soil types. It was formed from field-experiment data and a considerable writing of literature. The raw data included the levels of draft force of chisel plows with several widths under different levels of soil moisture contents, soil bulk densities, plowing depths, plowing speeds, and tractor powers as well as different percentages of sand, clay, and silt in the soils. Thus, two parameters were adaptable from these variables. The first adaptable parameter, namely the soil texture index, was denoted by STI (dimensionless), which linked all the soil percentages in one value and was offered by Oskoui and Harvey [31] as follows:

$$\mathrm{STI}=\frac{\log (\mathrm{Sa}+{\mathrm{Ca}}^{\mathrm{Si}})}{100}$$

(1)

Here, Sa denotes the sand percentage (%) in the soil. Ca and Si indicate the percentages of clay and silt in the soil type, respectively. Oskoui and Harvey [31] verified that the STI fluctuates for different arrangements of sand, silt, and clay. The second adaptable parameter was denoted by FWI (dimensionless), which is well-defined as follows [32]:

$$\mathrm{FWI}=\frac{\mathrm{MC}\times \mathrm{TP}}{\mathrm{BD}\times {\mathrm{d}}^2\times \mathrm{V}\times \mathrm{W}}$$

(2)

Here, FWI is called field working index, BD is the soil bulk density (kN/m³), TP is the tractor power (kW), MC is the soil moisture content (decimal, db), W is the plow width (m), d is the plowing depth (m), and V is the plowing speed (m/s).

The drawbar power (DBP) was calculated as follows:

$$\mathrm{DBP}(\mathrm{kW})=\mathrm{DF}\times \mathrm{V}$$

(3)

where DF is draft force (kN).

Effective field capacity (EFC) was determined for field experiments as the ratio of area plowed by the plow to the summation of nonproductive and productive times, shown as follows:

$$\mathrm{E}\mathrm{F}\mathrm{C} (\mathrm{h}\mathrm{a}/\mathrm{h})=\frac{\mathrm{A}}{\mathrm{T}}\times 0.36$$

(4)

where A is plowed area (m²), T is total time required to finish the plowed area (sec), and 0.36 is conversion factor.

The overall energy efficiency (OEE) was determined by dividing the drawbar power by fuel power, according to the formula stated by Bowers [33] as follows:

$$\mathrm{O}\mathrm{E}\mathrm{E}(\%)=\frac{\mathrm{D}\mathrm{B}\mathrm{P}}{\mathrm{P}\mathrm{F}}\times 100$$

(5)

where PF is fuel power (kW) and can be determined as follows:

$$\mathrm{P}\mathrm{F}(\mathrm{k}\mathrm{W})=\frac{\mathrm{F}\mathrm{C}(\mathrm{L}/\mathrm{h})\times \mathrm{H}\mathrm{V}(\mathrm{k}\mathrm{J}/\mathrm{k}\mathrm{g})\times \mathsf{\rho }(\mathrm{k}\mathrm{g}/\mathrm{L})}{3600}=10.21\times \mathrm{F}\mathrm{C}(\mathrm{L}/\mathrm{h})$$

(6)

where $\mathsf{\rho }$ is fuel density, and for diesel fuel, it was expected to be 0.835 kg/L; HV is heating value, and for diesel fuel, it expected to be 44,000 kJ/kg as reported in Srivastava et al. [34], 3600 is conversion factor, 10.21 is constant; and FC is fuel consumption (L/h).

The rate of plowed soil volume, denoted by PSV, is well-defined as the volume of plowed soil during tillage per unit time. It was calculated according to Legahri et al. [35] as follows:

$$\mathrm{P}\mathrm{S}\mathrm{V} ({\mathrm{m}}^3/\mathrm{h})=\mathrm{E}\mathrm{F}\mathrm{C}\times \mathrm{d}\times 10,000$$

(7)

where 10,000 is conversion factor.

2.3. Literature Data Collection

To create the ANN and MLR models for estimating performance attributes of the tractor-chisel mechanization unit, previous available datasets were acquired for the tractor-chisel plow mechanization unit that are directly related to the research subject [36,37,38,39,40,41,42,43,44,45]. The dataset was clustered based on actual field experiments in which different chisel plow configurations were used (only one pass over the soil) in different soil types with different soil moisture content levels, levels of soil bulk densities, and with different changeable working variables of plowing depths and speeds and tractor powers. The dataset enclosed 174 instances, each with draft force corresponding to the variables of plow width, tractor power, soil percentages of silt, sand, and clay, initial soil bulk density, plowing depth, initial soil moisture content, and plowing speed. Furthermore, the collected data had no values for effective field capacity of the investigated mechanization unit. The field efficiency was assumed to be 75%, as stated by Lar et al. [46] for chisel plow, where the typical range of field efficiency [47] ranged from 75 to 90%. The collected datasets have no fuel-consumption data, but actual fuel consumption should be measured in the field. Moreover, if no opportunity exists to carry this out, fuel consumption can be estimated when the machine application is known [48]. Thus, for collecting data from previous studies in this research, the following steps (from Equation (8) to Equation (13)) were employed to acquire the fuel-consumption data:

Define the loading factor using the following form [49]:

$$\mathrm{L}\mathrm{F} (\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{l})=\frac{\mathrm{E}\mathrm{P}\mathrm{T}\mathrm{O}}{\mathrm{A}\mathrm{P}\mathrm{T}\mathrm{O}}$$

(8)

where LF is loading factor, EPTO is the equivalent power of power take-off for an implement (kW), and APTO is the tractor available take-off power (kW). However, EPTO can be calculated using the drawbar power (DBP) and tractive efficiency (TE). Still, the tractive efficiency was dependent on tractor make and soil characteristics, and in this research for data collected, average TE was supposed as 0.65 [50]. Using the following form [49,51], EPTO can be determined as follows:

$$\mathrm{E}\mathrm{P}\mathrm{T}\mathrm{O} (\mathrm{k}\mathrm{W})=\frac{\mathrm{D}\mathrm{B}\mathrm{P}}{0.96\times \mathrm{T}\mathrm{E}}$$

(9)

where 0.96 is constant.

Calculate APTO based on the formula presented in Zoz and Grisso [52] as follows:

$$\mathrm{A}\mathrm{P}\mathrm{T}\mathrm{O} (\mathrm{k}\mathrm{W})=\mathrm{T}\mathrm{P}\times 0.83$$

(10)

where 0.83 is constant.

Calculate loading factor (LF) as follows:

$$\mathrm{L}\mathrm{F} (\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{l})=\frac{\mathrm{D}\mathrm{B}\mathrm{P}}{0.96\times \mathrm{T}\mathrm{E}\times \mathrm{T}\mathrm{P}\times 0.83}$$

(11)

Calculate specific fuel consumption for tillage process (SFC, L/kWh) as mentioned in ASABE [50] for diesel fuel using the following form:

$$\mathrm{S}\mathrm{F}\mathrm{C}=2.64\times \mathrm{L}\mathrm{F}+3.91-0.203\sqrt{738\times \mathrm{L}\mathrm{F}+173}$$

(12)

Calculate fuel consumption (FC) for tillage process as follows:

$$\mathrm{F}\mathrm{C} (\mathrm{L}/\mathrm{h})=\mathrm{S}\mathrm{F}\mathrm{C}\times \mathrm{E}\mathrm{P}\mathrm{T}\mathrm{O}$$

(13)

To check the steps (from Equation (8) to Equation (13)) for describing fuel-consumption rates, the amount of specific fuel consumption for data collected from soil-tillage field experiments was determined using the steps (from Equation (8) to Equation (13)), and the average value was 0.487 L/kWh. However, for diesel engines, the benchmark range of specific fuel consumption is usually from 0.244 to 0.57 L/kWh [53], and for diesel engines, the specific fuel consumption is affected by loading percentage on the engine [53].

Table 2. Statistical criteria of the working variables for literature and experimental work.

Working Variables	Statistical Criteria
	Minimum	Maximum	Average	Kurtosis	Skewness	Coefficient of Variation (%)
Tractor power (kW)	33.55	104.38	54.99	−0.50	0.88	33.71
Plow width (m)	1.68	2.63	1.74	142.16	10.48	3.82
Plowing speed (km/h)	2.30	6.92	4.12	1.16	0.18	19.29
Soil bulk density (g/cm3)	1.11	1.6	1.37	−0.83	0.42	8.88
Soil moisture content (%, db)	4.60	20.05	16.92	3.24	−1.93	19.99
Plowing depth (cm)	10.00	25.00	15.95	0.65	1.08	23.37
No. of data points	244	244	244	244	244	244

displays statistical criteria of the working variables collected from literature and experimental work.

Table 3. Statistical criteria of the input variables and performance attributes (outputs) used in ANN and MLR models.

Input and Output Attributes	Statistical Criteria
	Minimum	Maximum	Average	Kurtosis	Skewness	Coefficient of Variation (%)
STI (input) (dimensionless)	0.03	0.84	0.54	−0.27	−1.10	46.95
FWI (input) (dimensionless)	1.07	101.17	17.87	10.36	2.77	85.42
Draft (output) (kN)	5.32	21.55	15.13	−0.69	0.19	21.27
Fuel consumption (output) (L/h)	5.31	18.17	10.11	1.43	0.87	21.02
Effective field capacity(output) (ha/h)	0.32	0.95	0.60	4.08	−1.38	15.23
Drawbar power (output) (kW)	6.38	28.85	17.25	0.16	0.002	25.94
Overall energy efficiency (output) (%)	7.88	20.81	16.55	3.82	−1.86	10.82
Rate of plowed soil volume (output) (m3/h)	315.00	1814.40	950.51	0.94	0.61	27.15
Loading factor (output) (decimal)	0.13	0.99	0.67	−0.66	−0.70	35.73
No. of data points	244	244	244	244	244	244

displays statistical criteria of the input variables and performance attributes (outputs) used in ANN and MLR models.

2.4. Building the ANN Model

ANNs are computational modeling tools motivated by biological neural networks that try to simulate structures or functions. In this research, feedforward ANN topology and an error-back-propagation algorithm were used, which are commonly handy in the field of agricultural engineering [23]. The ANN in this investigation was comprised of input layer, one or more hidden layers and output layer. The input layer (i) is linked to the hidden layer (j), which is in turn connected to the output layer (k) through the connection weights (W) and biases (B). The weights are employed to modify the parameters of the throughput and the varying links to the neurons (n). The biases are associated to all nodes in the hidden and output layers and are employed to maintain the universal estimate of the ANN. A neuron (processing element) comprises two parts in the hidden layer. The first part aggregates the weighted inputs adding up to a quantity 1. The second part is the transfer/activation function that assists the translation of the input attributes (the activation values of the nodes) into the desired output parameter. The precise mathematical expression and explanation of the ANN are as follows [54]: The output-layer neuron (Y_k) can be expressed as follows:

$${\mathrm{Y}}_{\mathrm{k}}={\mathrm{f}}_{\mathrm{o}}\left({\mathrm{n}}_{\mathrm{kj}}\right)$$

(14)

where n_kj is the input to the k-th output neuron and can be expected using the formula:

$${\mathrm{n}}_{\mathrm{kj}}={\displaystyle \sum }_{\mathrm{j}=1}^{{\mathrm{N}}_{\mathrm{j}}}{\left({\mathrm{W}}_2\right)}_{\mathrm{kj}}{\mathrm{h}}_{\mathrm{j}}+{\left({\mathrm{B}}_2\right)}_{\mathrm{k}}$$

(15)

Therefore,

$${\mathrm{Y}}_{\mathrm{k}}={\mathrm{f}}_{\mathrm{o}}\left({\displaystyle \sum }_{\mathrm{j}=1}^{{\mathrm{N}}_{\mathrm{j}}}{\left({\mathrm{W}}_2\right)}_{\mathrm{kj}}{\mathrm{h}}_{\mathrm{j}}+{\left({\mathrm{B}}_2\right)}_{\mathrm{k}}\right)$$

(16)

Further, the neuron’s activation value (h_j) in the hidden layer is mathematically expressed using the following formula:

$${\mathrm{h}}_{\mathrm{j}}=\mathrm{f}\left({\mathrm{n}}_{\mathrm{ji}}\right)$$

(17)

where n_ji is the input of the j-th neuron in the hidden layer, which is calculated from:

$${\mathrm{n}}_{\mathrm{ji}}={\displaystyle \sum }_{\mathrm{i}=1}^{{\mathrm{N}}_{\mathrm{i}}}{\left({\mathrm{W}}_1\right)}_{\mathrm{ji}}{\mathrm{X}}_{\mathrm{i}}+{\left({\mathrm{B}}_1\right)}_{\mathrm{j}}$$

(18)

Accordingly, the h_j can be written as:

$${\mathrm{h}}_{\mathrm{j}}={\mathrm{f}}_{\mathrm{h}}\left({\displaystyle \sum }_{\mathrm{i}=1}^{{\mathrm{N}}_{\mathrm{i}}}{\left({\mathrm{W}}_1\right)}_{\mathrm{ji}}{\mathrm{X}}_{\mathrm{i}}+{\left({\mathrm{B}}_1\right)}_{\mathrm{j}}\right)$$

(19)

By substituting Equation (10) in Equation (6), Y_k can be calculated as follows:

$${\mathrm{Y}}_{\mathrm{k}}={\mathrm{f}}_{\mathrm{o}}({\displaystyle \sum }_{\mathrm{j}=1}^{{\mathrm{N}}_{\mathrm{j}}}{\left({\mathrm{W}}_2\right)}_{\mathrm{kj}}({\mathrm{f}}_{\mathrm{h}}({\displaystyle \sum }_{\mathrm{i}=1}^{{\mathrm{N}}_{\mathrm{i}}}{\left({\mathrm{W}}_1\right)}_{\mathrm{ji}}{\mathrm{X}}_{\mathrm{i}}+{\left({\mathrm{B}}_1\right)}_{\mathrm{i}}))+{\left({\mathrm{B}}_2\right)}_{\mathrm{k}})$$

(20)

with rearrangement, Equation (15) might be written as:

$${\mathrm{Y}}_{\mathrm{k}}={\mathrm{f}}_{\mathrm{o}}\left[{\left({\mathrm{B}}_2\right)}_{\mathrm{k}}+{\displaystyle \sum }_{\mathrm{j}=1}^{{\mathrm{N}}_{\mathrm{j}}}\right[{\left({\mathrm{W}}_2\right)}_{\mathrm{kj}}{\mathrm{f}}_{\mathrm{h}}\left({\left({\mathrm{B}}_1\right)}_{\mathrm{j}}+{\displaystyle \sum }_{\mathrm{i}=1}^{{\mathrm{N}}_{\mathrm{i}}}{\left({\mathrm{W}}_1\right)}_{\mathrm{ji}}{\mathrm{X}}_{\mathrm{i}}\right)]]$$

(21)

where X_i are input parameters; N_i is the number of input nodes; N_j is the number of output neurons; (W₁)_ji are the developed weights through the input layer to the hidden layer; (W₂)_kj are the weights transferred from the hidden layer to the output layer; (B₁)_j are the biases in the hidden layer; (B₂)_k are the biases in the output layer; f_h is the activation function (transfer function) in the hidden layer; and f_o is the transfer function in the output layer.

From the development of ANN model, the developer can face some difficulties for a specific application, such as how many hidden layers and their neurons can be used, as well as how many training patterns should be used in training phase [55]. In addition, the difficulties may be related to training algorithm and which neural-network topology should be used [55]. On the other hand, the hidden nodes can affect the error on the neurons to which their output is connected [56]. The reliability of an ANN mode model is assessed by error value. The least error reveals better reliability, and higher error value redirects worst reliability, and the excessive hidden neurons will cause overfitting [56]. In ANN application in agricultural engineering field, a single hidden layer may be ideal [3,21,57] due to there is no rule of thumb to find out how many hidden layers we need.

In this investigation, three different forms of activation functions were tested: hyperbolic tangent, sigmoid, and hyperbolic secant in each of the hidden and output layers to learn the established ANN. Moreover, the ANN model was established as three layers (i.e., input, hidden, and output). STI and FWI acted as input parameters in the input layer. The output layer included seven attributes, namely the draft force, drawbar power, effective field capacity, overall energy efficiency, fuel consumption, rate of plowed soil volume, and loading factor. A flowchart labeling the different stages used to build the ANN model with the commercially available Qnet2000 package (the publisher is Vesta Services Inc., Winnetka, IL, USA, version 1.0, https://qnetv2kt.software.informer.com/1.0/, accessed on 27 April 2022) [58] is presented in a former study by Marey et al. [59]. Howevere, Qnet2000 is a multi-layer perceptron whose training is achieved using a back-propagation algorithm. This program permits the definition of up to eight intermediate layers of neurons and the choice of different activation functions

The dataset was fully treated to be appropriate for the ANN simulation setting. Thus, the input and output values were normalized to be in the range of 0.15–0.85 by using the following equation [58]:

$$\mathrm{T}=\frac{(\mathrm{t}-{\mathrm{t}}_{\mathrm{m}\mathrm{i}\mathrm{n}})}{({\mathrm{t}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{t}}_{\mathrm{m}\mathrm{i}\mathrm{n}})}\times (0.85-0.15)+0.15$$

(22)

where T is the normalized value, t is the actual values of input or output parameters, t_max is the highest value of the input or output parameters, and t_min is the lowest value of the input or the output parameters.

In addition, the used of the commercially package [58] was employed to randomly hand-pick 48 points as test data, which suggest about 20% of the total dataset. Several topologies of ANN were created using a trial-and-error methodology to realize the topology with the highest predictive ability. Different structure parameters for developing the ANN model, such as the number of training iterations, activation functions, the number of nodes in the hidden layer, and the number of hidden layers, were tested through several simulations using the commercially package [58]. The best topology of the top ANN model used in this research was formed by 2–15–7. The training style used was a standard back-propagation algorithm with fully connected structures that used around 196 patterns as training points. The final learning rate was 0.010604. The output and hidden layers had a sigmoid activation function. The training and testing errors were 0.071999 and 0.082041, respectively. The momentum was 0.8 and the desired performance was terminated at iteration of 109,880.

2.5. Building MLR Model

Linear regression is the oldest statistical method in modeling in different fields and can be considered a benchmark for new modeling techniques [60]. However, MLR is a linear statistical method that attempts to establish the best correlation between a dependent parameter and several independent parameters by adjusting regression coefficients through the training data. The MLR model can be formulated as follows:

$$\mathrm{Y}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{X}}_1+{\mathsf{\beta }}_2{\mathrm{X}}_2+\dots \mathrm{\dots }+{\mathsf{\beta }}_{\mathrm{K}}{\mathrm{X}}_{\mathrm{K}}+\mathsf{\epsilon }$$

(23)

where Y is the dependent parameter or response, k is the number of independent parameters, X_j is the independent parameter, β is the regression coefficient, j = 0, 1, 2, …, k, and ε is a term that covers the influences of unmodeled sources of variability that impact the dependent parameter. The SPSS software (IBM SPSS Statistics 20; Chicago, IL, USA) was utilized to express the values of the regression coefficients for different outputs, and the training dataset used for building ANN model was employed as learning dataset for building the MLR model.

2.6. Statistical Criteria for Models Evaluation

By linking the predictions to the actual data in the testing dataset, ANN or MLR model trained on the training dataset can be assessed using statistical criteria. These criteria include the root mean square error (RMSE) and the mean absolute error (MAE) [61].

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{\mathrm{n}}{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{P}}_{\mathrm{i}}-{\mathrm{O}}_{\mathrm{i}}\right)}^2}}$$

(24)

$$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{\mathrm{n}}{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{P}}_{\mathrm{i}}-{\mathrm{O}}_{\mathrm{i}}\right|}$$

(25)

Here, P_i and O_i are the predicted and actual values, respectively, and n is the total number of points in each dataset. Moreover, scatter graphs are presented for visual correlation of the actual and predicted values.

3. Results

3.1. Data Analysis

The correlation between the FWI and the performance attributes of effective field capacity, draft, and fuel consumption of the mechanization unit during the soil-tillage process is described in Figure 1.

Noticeably, increasing the FWI decreases the draft, effective field capacity, and fuel consumption with a low correlation. This is owing to the explanation of the FWI parameter, which is in the range of 1.07 to 101.17 for the whole dataset. These findings specify that increasing the plowing speed, soil bulk density, plowing depth, and plow width decreases the FWI value at both specific tractor power and soil moisture content. For the FWI creation, increasing tractor power and the soil moisture content at a constant plowing speed, plowing depth, and soil bulk density led to an increase in FWI value. However, Ahmed [62] observed that the draft force of the disk plow working in a soil having clay texture was 11.72 and 12 kN when the tractor power was 60 and 67.5 kW, respectively, with the other variables remaining constant. The correlation between the FWI and the performance attributes of drawbar power, overall energy efficiency, rate of plowed soil volume, and loading factor of the mechanization unit during the soil-tillage process is presented in Figure 2. It is clear that increasing the FWI decreases the drawbar power, overall energy efficiency, rate of plowed soil volume, and loading factor of the mechanization unit during the soil-tillage process.

The correlation between the STI, which is in the range of 0.03 to 0.81 for the entire dataset, and the performance attributes of fuel consumption, draft, and effective field capacity of the tractor-chisel plow unit during the soil-tillage process is presented in Figure 3.

In addition, the correlation between the STI and the performance attributes of OEE, drawbar power, rate of plowed soil volume, and loading factor is shown in Figure 4.

Different researchers around the world have also determined the performance attributes. Thus, to validate the data, the present study was compared to prior researches. At first, the OEE includes the load matching of tractor and tillage implement [63] and the normal range for it is 10–20% [33]. A tractor- tillage implement arrangement with an OEE below 10% shows poor load matching and/or low tractive performance. A value of OEE above 20% specifies a respectable load match and/or high tractive performance. The rate of plowed soil volume is a critical attribute used to conclude the evaluation of the performance of tillage implement [64].

The noted values of the horizontal (draft) force in this investigation were within the range of 5.32 to 21.55 kN (Table 3) which are similar to those reported in previous studies. Increasing the depth from 7 to 47 cm increased the total drawbar power from 1.35 to 14.11 kW for the chisel plow in a soil texture of silty loam when the plowing speed of 2.5 km/h was applied [65]. For a chisel plow working in clay soil, the mean draft force was 17.62 kN [66]. The draft forces were 8.56 and 8.97 kN at a tillage depth of 20 cm when using a two-row chisel plow for plowing clay soil having different levels of soil moisture contents [67]. For clay soil under different operation and soil conditions, the forces were 9.97–13.03 kN [68]. In a clay type of soil, for chisel plow, the high value draft was 10.95 kN, while the lowest draft of 9.14 kN was measured when the plow had different arrangement of its shanks [69].

The amount of fuel consumption in this research ranged from 5.31 to 18.17 L/h (Table 3). For chisel plow working in clay soil, the mean fuel consumption was 15.5 L/h in a study by Khadr [66]. Meselhy [70] found that the fuel consumption rate fluctuated between 9.2 and 11.1 L/h, with a plowing speed ranging from 3.5 to 7.5 km/h for different chisel plow arrangements under many plowing speeds in a soil having sandy clay loam texture. In a similar study conducted in clay loam soil by Karparvarfard and Rahmanian-Koushkaki [71], the average fuel consumption rate was 14.03 L/h at a soil moisture content of 8.4% db (dry basis), a tillage depth of 20 cm, and plowing speed of 5 km/h, with the use of a chisel implement with a total width of 2.25 m. Jebur and Alsayyah [72] performed a tillage experiment with a chisel implement having a 2 m width in light soil texture under a tillage depth of 20 cm, levels of soil moisture content of 10–11.96%, and a plowing speed range of 5–5.2 km/h. Under these working variables, their study informed that the amount of fuel consumption ranged between 8.16 and 11.63 L/h. Mostly, the findings of the amount fuel consumption of our research agreed with the values described in the previous studies.

The noted effective field capacity in this research was ranged from 0.315 to 0.945 ha/h (Table 3). These results were similar to values obtained by previous researches [70,73] and are therefore considered acceptable. Meselhy [70] presented that field efficiency and the effective field capacity were 64% and 0.53 ha/h, respectively, of a customary chisel plow with seven tools, which was working in soil texture of sandy clay at a plowing speed of 5.5 km/h. In a study conducted by Osman et al. [73], a conventional chisel plow was operated in clayey sand soil. It was a fully tractor mounted with 7-rigid tools prepared in two rows, with three tools in the front row and four tools in the rear row. The effective field capacity was 0.83 ha/h where the plowing speed was 6.66 km/h and the effective field capacity increased to 0.953 ha/h when the speed was 7.2 km/h. When operating a chisel plow in a clay type of soil, the mean of noted effective field capacity was 0.36 ha/h and 0.62 ha/h as the plow varied the arrangement of its tools [69].

In the present study, drawbar power ranged from 6.38 kW to 28.85 kW (Table 3). Khadr [66] found that for a chisel plow working in clay soil, the mean drawbar power was 22.12 kW. There was a variance of the average power requirement for the chisel plow operation between 12.3 kW and 21.32 kW, respectively, when the plow was applied on clay soil at two different plowing speeds 4.5 km/h and 7.5 km/h [69].

The present OEE ranged from 7.88 to 20.81% (Table 3). These discoveries agree with the results stated by Ranjbarian et al. [74]. López-Vázquez et al. [63] reported that OEE was relatively low when employing a chisel plow (6.88%). Similar findings were observed in other studies, where the OEE values ranged from 11 to 20.08% for different tillage equipment [66,75]. Bowers [33] reported that higher OEE points to a worthy load match and/or a high tractive performance. For a chisel plow working in clay soil, the mean OEE was 14.13% [66]. The rate of plowed soil volume in this research ranged from 315 to 1814.4 m³/h (Table 3). These numbers were related to the values described in previous studies achieved under similar settings [10,76].

In this study, the loading factor, also called the fraction of equivalent PTO power available, had average value of 0.67, however, its maximum and minimum values were 0.99 and 0.13, respectively (Table 3). The loading factor, or working load estimation, is becoming avital variable at predicting fuel consumption and power take-off power requirements for specific tillage processes [77]. Lee et al. [78] used workload as one of the variables required to predict fuel consumption along with tractor ballast, tire inflation pressure, transmission gear, and engine speed.

3.2. Evaluation of Investigated ANN and MLR Models for Prediction Performance Attributes

The MLR equations that ensued from the regression analysis are as follows:

Draft

(kN) = 14.99 + 4.262 × STI − 0.118 × FWI R² = 0.249

(26)

Fuel consumption

(L/h) = 8.949 + 4.631 × STI − 0.075 × FWI R² = 0.345

(27)

Effective field capacity

(ha/h) = 0.622 + 0.068 × STI − 0.004 × FWI R² = 0.298

(28)

Drawbar power

(kW) = 15.859 + 8.79 × STI − 0.190 × FWI R² = 0.376

(29)

Overall energy efficiency

(%) = 17.459 + 0.855 × STI − 0.079 × FWI R² = 0.346

(30)

Rate of plowed soil volume

(m³/h) = 1262.905 − 170.478 × STI − 12.306 × FWI R² = 0.604

(31)

Loading factor

(decimal) = 0.841 + 0.067 × STI − 0.012 × FWI R² = 0.460

(32)

Taking into consideration the above parameters, these established simple equations are not available in literature. Hence, the present attempt therefore is important for first identifying the parameters affecting the performance attributes of a tractor-chisel unit and for decision making in selecting such arrangements. However, the predictability and efficiency of the established MLR and ANN models were evaluated using the statistical criteria of the MAE, RMSE, and R².

Table 4. Statistical criteria of the established ANN model for prediction of the performance attributes of the tractor-chisel unit during soil-tillage process in the case of the testing and training datasets.

Performance Attributes	Statistical Criteria
	MAE		R2		RMSE
	Training	Testing	Training	Testing	Training	Testing
Draft (kN)	1.22	1.44	0.734	0.569	1.61	2.06
Fuel consumption (L/h)	0.92	1.17	0.602	0.486	1.15	1.64
Effective field capacity (ha/h)	0.04	0.03	0.467	0.384	0.03	0.07
Drawbar power (kW)	1.99	2.28	0.633	0.516	2.07	3.03
Overall energy efficiency (%)	0.76	0.68	0.510	0.454	0.89	1.13
Rate of plowed soil volume (m3/h)	70.22	73.97	0.835	0.777	85.29	110.21
Loading factor (decimal)	0.07	0.08	0.853	0.730	0.08	0.12

shows the MAE, RMSE, and R² for different performance attributes using the ANN model.

Meanwhile,

Table 5. Statistical criteria of the established MLR models for prediction of the performance attributes for tractor-chisel unit during soil-tillage process in the case of the testing and training datasets.

Performance Attributes	Statistical Criteria
	MAE		R2		RMSE
	Training	Testing	Training	Testing	Training	Testing
Draft (kN)	2.42	2.41	0.249	0.239	2.81	2.72
Fuel consumption (L/h)	1.28	1.35	0.345	0.352	1.68	1.85
Effective field capacity (ha/h)	0.05	0.04	0.298	0.358	0.08	0.07
Drawbar power (kW)	2.78	2.58	0.370	0.429	3.57	3.27
Overall energy efficiency (%)	0.93	0.74	0.346	0.511	1.51	1.02
Rate of plowed soil volume (m3/h)	125.13	134.10	0.604	0.482	165.64	179.72
Loading factor (decimal)	0.14	0.14	0.460	0.422	0.18	0.18

shows the same statistical criteria for different performance attributes using MLR models.

Using the testing dataset as a base for the two models’ comparison, we found that for draft predictions, the ANN and MLR models have MAEs of 1.44 kN (Table 4) and 2.41 kN (Table 5), respectively. For fuel consumption, the ANN and MLR models have MAEs of 1.17 L/h (Table 4) and 1.35 L/h (Table 5), respectively. For the effective field capacity, the ANN- and MLR-developed models have MAEs of 0.03 ha/h (Table 4) and 0.04 ha/h (Table 5), respectively. For drawbar power, the ANN- and MLR-established models have MAEs of 2.28 kW (Table 4) and 2.58 kW (Table 5), respectively. For overall energy efficiency, the ANN- and MLR-created models have MAEs of 0.68% (Table 4) and 0.74% (Table 5), respectively. For rate of plowed soil volume, the ANN- and MLR-established models have MAEs of 73.97 m³/h (Table 4) and 134.10 m³/h (Table 5), respectively. For loading factor, the ANN- and MLR-constructed models have MAEs of 0.08 (decimal) (Table 4) and 0.14 (decimal) (Table 5), respectively. These findings show that the ANN-created model yields more accurate values than the MLR model since the ANN model has high robustness. The results specified that the ANN-created model was adequate for predicting the performance attributes of the tractor-chisel plow unit during plowing operations at different FWIs and STIs.

In the case of employing the testing dataset, correlations between the noted and predicted performance attributes of the horizontal (draft) force, effective field capacity, fuel consumption, drawbar power, overall energy efficiency, rate of plowed soil volume, and loading factor are clarified in Figure 5, Figure 6, Figure 7, Figure 8, Figure 9, Figure 10 and Figure 11.

Moderate correlation between the noted and ANN-predicted numbers is observed, as represented by the values of R² in the range of 0.384 to 0.777 for all performance attributes (Table 4). However, R² can be realized as the proportion of the difference in the predicted values that is clarified by the discrepancy in the noted values.

3.3. Contribution Percentage Analysis

A vital feature of the used the commercially package [58] is that it permits quantification of the contribution percentage of each input node to the computed output number. Thus, it is promising to detect the most related variables affecting the performance attributes of a tractor-chisel plow unit. The individual contribution percentages of the inputs are indicated in Figure 12 and Figure 13.

The STI makes average contribution percentages of 46.36%, 35.97%, and 23.42%, as indicated in Figure 12, during prediction of draft force, fuel consumption, and effective field capacity, respectively, using the ANN model. Furthermore, STI makes average contribution percentages of 33.53%, 16.54%, 30.99%, and 32.81, as indicated in Figure 13, during prediction of drawbar power, rate of plowed soil volume, overall energy efficiency, and loading factor, respectively, using the ANN model. The FWI makes average contribution percentages of 53.64%, 64.03%, and 76.58%, as indicated in Figure 12, during prediction of draft force, fuel consumption, effective field capacity, respectively, using the ANN model. Moreover, the FWI makes average contribution percentages of 66.47%, 83.46%, 69.01%, and 67.19%, as indicated in Figure 12, during prediction of drawbar power, rate of plowed soil volume, overall energy efficiency, and loading factor, respectively, using the ANN model. However, the matching and performance of a mechanization unit depend upon some variables, such as the type of tractor tires and implement characteristics, soil conditions, etc., which are essential to a mechanization unit and cannot be changed or controlled. In addition, there are other variables that influence the performance of the tractor tillage implement system such as hitching type (semimounted, mounted, and trailed), working variables (depth and speed of plowing), types of plowing conditions (primary or secondary) tillage, etc. These variables can be adjusted for achieving the maximum performance. All of these variables, which are manageable, cover a wide variation of options on which choices have to be based, such as the appropriate size of the tillage implement for the tractor [79]. Moreover, the prediction of tractor drawbar performance can lead to simulation and optimization of tractor performance, allowing for the optimum setting of different parameters, as well as guiding a manufacturer in decision making for design of new tractors [80].

4. Discussion

An important feature of the use of ANN models in realizing deterministic issues is the selection of a suitable network structure. A difficult point of testing altered ANN topologies permitted us to specify the most appropriate type of network for the issues offered in this paper. Finally, the work uses an artificial neural network model with two inputs, fifteen neurons in the hidden layer, and seven neurons in the output layer (2–15–7) to develop an operational model of the performance attributes of a mechanization unit (tractor-chisel plow) during the plowing process. The draft force, effective field capacity, fuel consumption, drawbar power, overall energy efficiency, rate of plowed soil volume, and loading factor were addressed as the performance attributes. The model presented is simple in operation and faster in application, which makes it suitable in predicting different performance attributes of a mechanization unit (tractor-chisel plow) during the plowing process for farm machinery management. Moreover, this model was verified by comparing its response with the actual performance attributes measured and calculated for several tractor-chisel plow units. Our study presented a new technique for predicting draft, fuel consumption, etc., for a chisel plow based on using two different parameters to represent working conditions (FWI) and soil type (STI), and all the previous studies in this field directly considered tillage speed, tillage depth, soil moisture content, etc., as potential input parameters influencing draft, fuel consumption, etc. Thus, to make comparisons of previous ANN models with our ANN model, assessments of R² values were considered. Overall, our ANN-based models had R² in the range of 0.384 to 0.777 for the investigated attributes. These results were found to be less satisfactory to the main results of previous studies using ANNs in tillage research: draft force of a moldboard plow (R² = 0.9996) [81], fuel consumption of a moldboard plow (R² = 0.9996) [81], draft force of a chisel cultivator (correlation coefficient = 0.9445) [24], and tractor axle torque estimation (R² value ranging from 0.857 to 0.904) [82]. The variations in correlation values may be because our data were acquired from different field experiments. Additionally, the variations are assumed to be due to the low input dimensions and high model complexity of the prediction model [82], which stated that higher input dimensions in an ANN model could make the model more fit to training data. As a result, the generalization performance of the ANN model reflected poor behavior. Nevertheless, it revealed the moderate performance gained in the training dataset, which revealed that the model’s ability was adequately fit for the data; thus, it can be realized that the ANN-established model can be considered to be appropriate for estimation of performance attributes of a mechanization unit (tractor-chisel plow) during the plowing process under both different soil types and working conditions. The sensitivity analysis performed for the described ANN model indicated that the field working index (FWI), which combined six variables, had the greatest contribution for all investigated performance attributes compared to the soil texture index (STI). This can be used to select appropriate tractor power, soil moisture content, soil bulk density, tillage depth, plow width, and tillage speed for achieving the maximum performance of a mechanization unit.

5. Conclusions

Two variables, specifically FWI and STI, were recognized as input nodes in an ANN model to estimate the performance attributes of the tractor-chisel plow unit during plowing operations. The established ANN model had 2 nodes for the input layer, 15 neurons for the first hidden layer, and 7 nodes for the output layer, i.e., a topology of 2–15–7. The results indicated that standard backpropagation as a training algorithm was adequate for estimating fuel consumption, draft force, drawbar power, effective field capacity, overall energy efficiency, rate of plowed soil volume, and loading factor. The prospects of the developed ANN and MLR models for prediction of each performance attribute was evaluated: For draft, the RMSEs of the ANN and MLR models were 2.06 and 2.72 kN, respectively; for fuel consumption, they were 1.64 and 1.85 lit/h, respectively; for effective field capacity, they were 0.07 and 0.07 ha/h, respectively; for drawbar power, they were 3.03 and 3.27 kW, respectively; for overall energy efficiency, they were 1.13 and 1.02%, respectively. For rate of plowed soil volume, the RMSEs of the ANN and MLR models were 110.21 and 179.72 m³/h, respectively; and for loading factor, they were 0.12 and 0.18 (decimal), respectively. The ANN model is a dominant tool that is easy to apply for nonlinear issues. The established ANN model can be steadfastly used for the prediction of the performance attributes of a tractor-chisel plow unit during plowing operations over a large range of FWIs (1.07–101.17) and STIs (0.03–0.84).