ARTIFICIAL NEURAL NETWORK AND RESPONSE SURFACE METHODOLOGY APPROACHES FOR OPTIMIZATION OF PARAMETERS FOR AN AIR-JET SUNFLOWER SEEDS REMOVER MACHINE
DOI:10.30825/5.EJPAU.185.2020.23.2

Amir Hossein Mirzabe¹, Gholam Reza Chegini², Jafar Massah², Ali Mansouri², Javad Khazaei^2
1 Department of Mechanical Engineering of Biosystems, College of Agriculture & Natural Resources, University of Tehran, Tehran, Iran
² Department of Mechanical Engineering of Biosystems, College of Aboureihan, University of Tehran, Tehran, Iran

ABSTRACT

An impingement jet method was employed for extracting of sunflower seeds from sunflower heads (SHs). The method was based on holding SHs with a rotating plate and extracting the sunflower seeds with the help of pressurized air-jets. Artificial neural networks (ANNs) and response surface methodology (RSM) were used to model the effects of operational parameters of impingement air-jet on performance of preliminary model of the remover machine. The operational parameters were diameter of nozzle (ND), angle of impingement (AI), distance between nozzle outlet and sunflower head (DBNS), air pressure (A_P) and rotational velocity of sunflower head (RV). The final ANN model, 3-5-1, successfully modeled the relationship between three operational parameters, ND, AI and RV with removing performance of machine (RP_AJSSR) with R²of 0.98 and T value of 0.96. The RSM method was applied for three different locations of SHs at the optimum AP of 7 bar. The maximum value of RP_AJSSR, (57%) was obtained for ND of 8 mm, AI of 30°, DBNS of 20 mm and RV of 10 rpm at side region of SH (SR_SH). Also, the minimum value (4.49%) belonged to ND of 4 mm, AI of 30°, DBNS of 20 mm and RV of 15 rpm for central region of SH (CR_SH).

Key words: Sunflower, Jet impingement, Neural networks, Response surface methodology.

Abbreviations
AI	Angle of impingement [deg]	MS	Mean squares
AJSSR	Air-jet impingement sunflower seeds remover	MSE	Mean squares error
ANN	Artificial neural network	ND	Nozzle diameter, mm
AP	Air pressure [bar]	RE	Relative error
CRSH	Central region of sunflower head	RSM	Response surface methodology
D	Sunflower head diameter [mm]	RV	Rotational velocity of SH, rpm
Da	Arithmetic mean diameter of SH [mm]	RPAJSSR	Removing performance of machine
DBNS	Distance between nozzle and SH [mm]	RPM	Measured removing performance
DF	Degree of freedom	RPP	Predicted removing performance
Dg	Geometric mean diameter of SH [mm]	SH	Sunflower head
EAP	Predicted extracted area	SRSH	Side region of sunflower head
EAM	Measured extracted area	SPH	Sphericity of SH, %
M	Mass of whole seeds of SH [g]	SS	Sum of squares
MLP	Multi-layer perceptron	STD	Standard deviation
MRSH	Middle region of sunflower head	SV	Source of variations
		T	Thickness of sunflower head, mm

INTRODUCTION

Sunflower (Helianthus annuus L.) oil is one of the most important seed crops product which can also be a valuable source of protein. In post-harvest process of sunflowers, the seeds should be removed from sunflower head (SH). Nowadays, the removing of seeds from SHs is done using traditional manual or mechanized methods.

In traditional methods, SHs are rubbed to each other or over a rigid surface such as: brick, stone, a piece of metal, wood, rubber, etc. Performance and experience of the laborers are the principle factors affecting the efficiency of seeds’ removing [14]. The mechanisms of mechanical machines have been designed based on beating and friction theories. The mechanical machines are categorized into two types: combined harvesters and stationary thresher machines. Pedal operated, wire mesh type and peg tooth and/or rasp-bar drum threshers are the developed industrial threshing tools available in stationary which are combined harvester machines [27, 30].

The most important problems of these machines are high energy consumption and high amount of damaged seeds and losses. Furthermore, the mentioned tools usually are labor intensive and have high mechanical adjustments. Also, in some areas of the world (in many developing countries), these machines cannot be used due to traditional planting of sunflowers, low level of farmers’ know-how of using the machines, and complex settings of combined harvester machines.

To overcome these problems, an air-jet impingement method was developed. The air and water-jet impingement extraction techniques were successfully used to remove pomegranate fruits arils [16, 25, 26] and juice sacs [2, 15, 17, 23]. Literature reported that the air-jet and water-jet impingement methods can sufficiently improve the performance of extraction and reduce the mechanical damage of fruits sacs and arils. When using an air-jet impingement method to remove sunflower seeds from the head, each impingement jet parameter may vary in the degree of effect on machine performance. Each of these parameters must be considered while designing machines.

Optimization is an act, process, or methodology of making something as fully perfect, functional, or effective as possible. The aim of optimization of the removing seeds process is to find the specific conditions which the removing process would have the best performance efficiency. Artificial neural networks (ANNs) and response surface methodology (RSM) are the applicable and the two of the most popular methods to model complex nonlinear relationships between independent variables and responses of a system.

ANNs use different sophisticate mathematical functions to process and interpret various sets of erratic data [24]. Therefore, it is being used as an alternative method for the modeling of complex processes. ANNs have been used in various filed, function approximation, pattern recognition, noise reduction, generation of new meaningful pattern and multivariate data analysis [6, 18].

The response surface methodology (RSM) is the collection of mathematical functions that define the relationship between various process parameters and responses with the various desired criteria [11]. It is a sequential experimentation approach to improve efficiency of a process and design of products [5].

The aim of present study was to develop the optimized model of preliminarily air-jet sunflower seed remover (AJSSR) parameters in order to design and construct an efficient version of AJSSR machine. ANNs and RSM methods were used to develop a model to optimize effects of nozzle diameter (ND), distance between nozzle and SH (DBNS), angle of impingement (AI), rotational velocity of SH (RV), and air pressure (AP) on AJSSR machine removing performance (RP_AJSSR). These are the fundamental steps towards the development of an automated intelligence version of AJSSR machine for industrial usage.

MATERIALS AND METHODS

1. Sample preparation
In the present study, the mature heads of Songhori variety were used. The Songhori variety is native of Iran and is planted in large amounts in many provinces of the country. The sample used in the current work was planted on May 22th, 2012 in local farms of Foodan plain, located on Shahreza, Isfahan province, Iran (longitude of 31.59° N, latitude of 51.50° E, average annual Precipitation 135 mm from 2000 to 2010, height above the sea level of 1845 m, average annual temperature of 14.7°C from 1993 to 2010). The SHs were harvested manually in late October, after they were completely matured.

Since the size of sunflower heads had a great effect on removing force of sunflower seeds [21], in order to avoid experimental errors, the diameter of each SH was measured in 4 directions by a ruler, then mean diameters were calculated. After that, only samples with diameters ranging between 200 and 230 mm were chosen for the tests.

In order to determine the effects of air-jet optional parameters on percent of removed seeds from the different areas of SH, the surface of selected SHs were divided into three regions namely central region (CR_SH), middle region (MR_SH) and side region (SR_SH), as shown in Figure 1.

Fig. 1. Three regions of SH, Central region (CRSH), Middle region (MRSH) and Side region (SRSH)

In order to determine the performance of machine, the weight of removed seeds and residual for each SH were recorded. The weights of the seeds were measured for one rotation of SH with an accurate digital balance (Kern, EMB600-2, 600 g ± 0.01 g, Philippines).

2. Physical properties of sunflower heads
The diameter of each SH was measured in 4 directions, then, the mean diameter (D) was calculated. Also, in order to measure the total mass of seeds (M), at the first, all seeds of sunflower heads were extracted manually. After that, the mass of the whole seeds of each SH was measured by an accurate digital balance (Kern, EMB600-2, 600 g ± 0.01 g, Philippines).

Moreover, in order to calculate arithmetic mean diameter (Da), geometric mean diameter (Dg), and sphericity (Φ) of sunflower heads, the following equations were used [4, 19]:

 3. Experimental setup
The air-jet parameter levels were reported in Table 1. All parameters were set at three levels. A schematic diagram of the preliminarily model of air-jet impingement sunflower seed remover machine is shown in Figure 2.

Table 1. The adjusted levels of AJSSR parameters

Parameter	Unit	Levels
Diameter of nozzle (ND)	mm	4	6
Angle of impingement (AI)	degree	30	60
Nozzle distance (DBNS)	mm	10	20
rotational velocity (RV)	rpm or rev min-1	10	15
Air pressure (PA)	bar	5	6

Fig. 2. The schematic view of preliminarily model of AJSSR machine. (1) frame, (2) mechanical jack, (3) gear box, (4) electromotor, (5) sloped plate, (6) holder disc, (7), two nozzles, (8) changing angle mechanisms, (9) changing horizontal position of nozzles, (10) valve, (11) pneumatic hoses [20]

The main parts of the machine consisted of a frame, an electromotor, a worm gear box, a holder disc (Fig. 3a), variable angle of impingement mechanism (Fig. 3b), horizontal position of nozzle mechanism, distance between nozzle outlet and sunflower head (DBNS) change mechanism (mechanical jack), a switching valve, an inclined metallic plate, and transparent polycarbonate talcs.

A) B)

Fig. 3. Schematics of mechanisms used to set the holding disk, A, and angle of impingement, B. [20]

The power required for rotating the sunflower head was provided by electrical motor (Hummer, MS632-4, Iran). In order to reduce the rotational velocity of the electro motor and change the rotational direction, a worm gear box (Taizhou Jiaoxing Transmission Equipment Co., Ltd, RV30, China) was used.

On order to hold sunflower heads during the tests, a holder disc was designed and constructed. The holder disc is capable of holding SHs with the diameters of 6 to 36 cm (Fig. 3a). The rotational velocity of the holder disc was adjusted with an inverter (LS industrial systems, SV 015ic5-1F, China) in the range of 10 to 20 rpm.

The air pressure was increased in piston compressor; the high pressure flow of the air was transferred to the nozzle using pneumatic hoses. To connect and disconnect the air flow, a switching valve was used. To avoid seeds’ throwing out from the machine space, the transparent polycarbonate talc was used. The seeds, after removing from the heads, fall down on inclined plate; then the sunflower seeds were transferred out of the machine.

4. Artificial neural network model development
Nowadays, there are lots of applications of new technologies and new methods in agricultural science. Signal processing techniques, image processing techniques, and computer vision are some of the most important new technologies [3]; neural networks, fuzzy logic, and response surface methodology are some of the most important new methods used for studying and optimizing parameters and processes.

Using the experimental data, multi-layer perceptron (MLP) ANNs with one and two hidden layers were developed to model correlation between RPAJSSR and the three air-jet operational parameters, namely: ND, AI and RV (Fig. 4). Preliminary trials showed that one hidden layer ANNs performed better than two-hidden-layer ones to learn and predict the input/output parameters correlation. There were a total of 108 patterns (3 levels of ND × 3 levels of AI × 3 levels of RV × 4 replication), each with 4 components, (x1, x2, x3, Y), three of which were the input parameters, while the Y was the single output parameter. The input data were divided into three parts, 60% for training process, 15% for validation phase and 25% for test phase. In order to acquire the same significance for all ranges of data during training, the inputs and outputs were normalized between -1 and 1 value using the MATLAB subroutine mapminmax. The best ANN model was selected on the basis of the lowest error on the training, validation and testing of developed ANNs. The performance of the ANNs was compared by mean square error (MSE), correlation coefficient, R², and T statistics [12]. ANNs were trained three times and the best values of parameters were recorded. A laptop computer (SONY VAIO VPCEG34F/W, Intel Core i5 processor, China) equipped with MATLAB R2013b software package was used to develop the ANN topologies.

Fig. 4. Topology of multi-layer ANN for predicting the removing performance of AJSSR machine

5. Response surface methodology (RSM)
The optimization by means of RSM approach was performed in six stages: (1) selection of independent variables and possible responses, (2) selection of the experimental design strategy, (3) execution of experiments and obtaining results, (4) fitting the model equation to experimental data, (5) obtaining response graphs and verification of the model, (6) determination of optimal conditions [28].

The mass flow rate of gas, weight ratio of removed seeds to all seeds for each SH and time of the one rotation of holding disk were considered as response of the tests; because, the optimum conditions for machine can be obtained when the weight ratio of removed seeds, mass flow rate of gas and time of the test are maximum, minimum and minimum, respectively. The mass flow rate of gas was calculated as reported by [13].

The experimental strategy was developed with Box-Behnken design method [8]. These designs are more efficient and economical than their corresponding factorial (3^k) designs, especially for large numbers of variables [7]. In Box-Behnken design, the number of experiments needed to be done are lower than number of factorial designs; it can be calculated according to following Equation [8–10]:

Where: x is the number of factors and y is the number of the central points; 2) all factor levels have to be adjusted only at three levels with equally spaced intervals between these levels.

In the factorial design, 3⁴ = 81 tests were carried out while response surface method was used and tests were carried out based on Eq. (1), 4² + (2×4) + 5 = 29 (see Table 2).

Table 2. Designed tests based on response surface methodology for each region of sunflower head

Test number	ND [mm]	DBNS [mm]	AI [deg]
1	4	10	60
2	8	10	60
3	4	30	60
4	8	30	60
5	6	20	30
6	6	20	90
7	6	20	30
8	6	20	90
9	4	20	60
10	8	20	60
11	4	20	60
12	8	20	60
13	6	10	30
14	6	30	30
15	6	10	90
16	6	30	90
17	4	20	30
18	8	20	30
19	4	20	90
20	8	20	90
21	6	10	60
22	6	30	60
23	6	10	60
24	6	30	60
25	6	20	60
26	6	20	60
27	6	20	60
28	6	20	60
29	6	20	60

A second order quadratic model (Eq. 1) was used to describe the effect of independent variables in terms of linear, quadratic and interactions. The result in an empirical model related to the response (ρ) is:

where: ρ is the predicted response (density), b₀ is the interception coefficient, b_i, b_ii, and b_ij are the linear, quadratic, and interaction terms, θ is the random error and x_i is the independent variables studied [29]. Data were analyzed using the response surface regression procedure and fitted to the second-order quadratic equation. The Design Expert 10 software package (Stat-Ease, Inc., Minneapolis, MN) was used for regression and graphical analysis of the data. The significance of the statistical model was evaluated by the F-test analysis of variation (ANOVA).

RESULTS

Effects of ND, AI, DBNS, and RV parameters on RPAJSSR were examined with application of ANN and RSM methods. To illustrate of removed areas of sunflower heads and performance of AJSSR machine, three random sunflower heads were chosen which are shown in Figure 5. It is clear from this figure, that the area of separated seeds in B (side region) is more than A (middle region). It is due to higher jet force effect at MR of SHs relative to SR (Fig. 5). Furthermore, increasing of ND leads to an increase in removed area of SHs (Fig. 5 B and C).

Fig. 5. Some removed regions of SHs due to impingement of air-jet, A, middle region of SHs (ND = 6 mm), B, side region of SHs (ND = 6 mm), C, side region of SHs (ND = 8 mm)

1. Physical properties of sunflower heads
Table 3 summarizes the values of some physical properties of sunflower heads that they were measured. The total mass of seeds (M), SH thickness (T) and SH diameter (D) ranged from 344 to 110 gr, 83 to 42 mm and 345 to 170 mm, respectively. The mean values of arithmetic mean diameter (D_a), geometric mean diameter (D_g), and sphericity (SPH) were equal to 165.61 mm, 140.65 mm and 62.90%, respectively.

Table 3. Some physical properties of sunflower head of Songhori variety

Parameter	Maximum	Minimum	Mean ± STD	Skewness
M [g]	344	110	192.83 ± 44.61	1.27
T [mm]	83	42	57.84 ± 8.12	0.82
D [mm]	345	170	223.73 ± 33.52	1.44
Da [mm]	253.33	125.67	165.61 ± 24.94	1.36
Dg [mm]	209.89	106.83	140.65 ± 20.64	1.24
SPH [%]	67.87	58.17	62.90 ± 1.71	-0.08
M: total mass of the seeds; T: head thickness; D: head diameter; Da: arithmetic mean diameter; Dg: geometric mean diameter and SPH: sphericity

2. Artificial neural network model
Various feed-forward ANNs were trained and tested with back propagation algorithms for predicting the maximum RPAJSSR based on the three operational parameters of ND, AI and RV. The performance of different error minimization algorithms, transfer functions, nodes number and training iterations on ANN performance were investigated (Table 4). Among the different ANN topologies, the model with the best performance was produced by three-layer ANN structure, 3-5-1, one-step secant (oss) error minimization algorithm and log-sigmoid activation function. This model produced the smallest MSE in training, 0.0018, validation, 0.0023 and testing, 0.0035. The optimal ANN structure and its parameters are presented in Table 5. The range of ANN parameters tried was: number of neurons (from 1 to 10), activation function (log-sigmoid, hyperbolic tangent sigmoid, and linear), and number of epochs (100–1000). Pretests with other ANN topologies indicated that one hidden layer networks produced better results than two-hidden-layer one. The number of hidden nodes required for proper generalization was determined by trial and errors. The value of MSE in both training and test processes for each training ANN models is reported in Figure 6. Initially, the train and test subset errors decreased with increasing the number of neurons in hidden layer (Fig. 6). As the network loses its ability to generalize the test data, the errors increased (see after 5 neuron numbers in Fig. 6). Although the MSE on the test data may not follow a smooth path, the onset of a major increase in the error was considered to represent the optimal number of hidden neurons for that final network architecture [6]. Too few numbers of hidden neurons will hinder the learning process. In contrast, too many hidden neurons will depress prediction abilities through overtraining [1]. The final ANN architecture is obtained at the onset of the increase (5 neurons) in test data errors (Fig. 6).

Table 4. Variation of training and validation MSE for different configurations of the learning algorithms and transfer functions

Transfer function	Learning algorithm	4 Neuron/ 100 epoch		6 Neuron/ 400 epoch		8 Neuron/ 700 epoch		10 Neurons/ 1000 epoch
		Training	Validation	Training	Validation	Training	Validation	Training
logsig	gda	0.0045	0.060	0.0038	0.062	0.0043	0.043	0.0097
	gdm	0.0041	0.051	0.0043	0.041	0.0051	0.062	0.0080
	lm	0.0034	0.029	0.0030	0.031	0.0069	0.046	0.0088
	oss	0.0025	0.014	0.0027	0.029	0.0040	0.068	0.0037
tansig	gda	0.0043	0.040	0.0041	0.041	0.0077	0.045	0.0054
	gdm	0.0040	0.060	0.0042	0.030	0.0073	0.039	0.0048
	lm	0.0039	0.038	0.0032	0.042	0.0058	0.065	0.0042
	oss	0.0037	0.047	0.0037	0.034	0.0059	0.043	0.0065

|

Table 5. The optimum values of the ANN model used to predict RPAJSSR

Optimum		Transfer function	Mean values			T value
ANN structure	Algorithm		MSE train	MSE validation	MSE test
3-5-1	oss*	logsig	0.0018	0.0023	0.0035	0.96
* One-step secant backpropagation algorithm

Fig. 6. The performance of training and testing processes for various ANN models as a function of hidden neurons.

Figure 7 shows the convergence of training, validation and testing errors as function of epochs for final ANN structure, 3-5-1. The errors on training, validation and testing subsets data generally decreased with increasing iteration numbers. An initial drop down in MSE values shows the satisfactory learning of input/output relationship by oss error minimization algorithm. Approximately after 60 epochs, MSE arrived a stable phase and no increasing due to memorization and overtraining of trained ANN is observed (Fig. 7). The errors on training, validation and testing sets were in the acceptable range (MSE ≤ 0.0035).

Fig. 7. Convergence of the MSE during training, validation and testing processes of the final ANN structure, 3-5-1.

Figure 8 shows the 18 predicted values of RP_AJSSR versus the same set of measured data for final ANN with 1000 epochs.

Fig. 8. Correlation between the measured and predicted removing performance of AJSSR machine using the ANN mode

It is clear that the network sufficiently learned the correlation between input and output variables. The linear adjustment between the predicted and measured values gave a slop close to 1 (RPP = 1.107×RPM – 9.292). The resulting MSE, correlation of determination and T values were 0.0035, 0.98 and 0.96 for linear regression between measured and predicted values, respectively (Table 5 and Fig. 8). T value represents the distribution of data around line (1:1) and the practically values equal to 1 are desirable [12]. The distribution pattern of training, validation and testing relative errors versus corresponding predicted values are illustrated in Figure 9. It is obvious that the residuals were well scattered on both side of the horizontal line (relative error = 0) without any systematic tendencies and clear pattern (RE = 0.0007×RP_P - 0.0503, R² = 0.049).

Fig. 9. Relative error distribution of the ANN model for the prediction of removing performance

If the residual plots indicate a clear pattern, the model could not be accepted [12]. These results prove the sufficient predicting performance of final network for the whole range of data.

3. Response surface methodology
According to the pretest results, performance of machine at AP of 7 bar was better than other AP values, consequently the RSM optimization was performed to determine the interaction effects of ND, AI, DBNS, and RV parameters at AP of 7 bar. The interaction effects of ND, AI, DBNS, and RV on AJSSR machine performance for SR_SH are illustrated in Figure 10. Results showed that maximum and minimum values of RPAJSSR were equal to 57.66 and 23.38%, respectively. Table 6 indicates the results of the statistical analysis carried out to examine the effects of AJSSR machine parameters on percentage of the removed seeds for SR_SH. ANOVA indicated that ND, AI, DBNS and RV, had a significant effect on machine performance (P≤0.001). For all cases of SR_SH, no significant effects were observed for combination of AJSSR parameters.

Table 6. ANOVA results for percentage of the removed seeds from the SH on SRSH

SV	DF	SS	MS	F-value
Model	14	2471.432	176.531	303.912**
Diameter of nozzle (ND)	1	641.379	641.379	1104.187**
Nozzle distance (DBNS)	1	54.955	54.955	94.610**
Angle of impingement (AI)	1	1006.318	1006.318	1732.458**
Rotational velocity (RV)	1	544.053	544.053	936.632**
ND × DBNS	1	0.931	0.931	1.603
ND × AI	1	1.891	1.891	3.255
ND × RV	1	5.929	5.929	10.208
DBNS × AI	1	1.199	1.199	2.064
DBNS× RV	1	0.689	0.689	1.186
AI × RV	1	13.432	13.432	23.125
Residual	14	8.132	0.581
Lack of Fit	10	8.113	0.811	169.018 s
R-squared	0.997
Adjusted R-squared	0.993
Pred R-squared	0.981
s and ** show  significant and significant difference at probability of % 1, respectively. Numbers have been rounded to three decimal places.

Fig. 10. Interaction effects between operating parameters of air-jet impingement on percentage of removed seeds for SRSH

The interaction effects between independent parameters of AJSSR machine on removing sunflower seeds for MR_SH are shown in Figure 11. The RSM results indicated that maximum and minimum percentages of the removed seeds were equal to 44.55 and 18.46%, respectively. Table 7 represents the results of the statistical analysis carried out to examine the effects of AJSSR parameters on percentage of the removed seeds for MR_SH. ANOVA indicated that all air-jet operational parameters had significant effects on machine performance (P≤0.001). No significant effects were determined for combination of parameters in MR_SH (Table 7).

Table 7. Effects of AJSSR parameter percentage of the removed seeds for MRSH

SV	DF	SS	MS	F-value
Model	14	1147.752	81.9823	149.267**
Diameter of nozzle (ND)	1	551.173	551.173	1003.531**
Nozzle distance (DBNS)	1	14.070	14.070	25.618*
Angle of impingement (AI)	1	10.498	10.498	19.114
Rotational velocity (RV)	1	206.579	206.579	376.121**
ND × DBNS	1	0.343	0.343	0.625
ND × AI	1	0.261	0.261	0.475
ND × RV	1	5.394	5.394	9.821
DBNS × AI	1	0.006	0.006	0.010
DBNS× RV	1	0.166	0.166	0.302
AI × RV	1	0.080	0.080	0.145
Residual	14	7.689	0.549
Lack of Fit	10	7.563	0.756	23.879 s
R-squared	0.993
Adjusted R-squared	0.987
Pred R-squared	0.962
s, *, and ** show  significant and significant difference at probability of 5% and  1%, respectively. Numbers have been rounded to three decimal places.

Fig. 11. Interaction effects between operating parameters of air-jet impingement on percentage of removed seeds for MRSH

The interaction effects of air-jet impingement parameters on removing sunflower seeds for CR_SH are illustrated in Figure 12. The RSM results of CR_SH indicated that the maximum and minimum percentages of the removed seeds were equal to 9.24 and 4.49%, respectively. The statistical analyses of AJSSR parameters for MR_SH are given in Table 7. Results indicated that all parameters had significant effects on machine performance (P≤0.001). No significant effects were observed for combination of parameters in MR_SH (Table 7). Table 8 represents the results of the statistical analysis carried out to examine the effect of AJSSR parameters on percentage of the removed seeds for CR_SH. The ND and RV, parameters had significant effects on machine performance (P≤0.001). No significant effects were observed for combination of parameters in CR_SH (Table 8).

Table 8. ANOVA results for percentage of the removed seeds from the SH on SRSH

SV	DF	SS	MS	F-value
Model	14	39.396	2.814	110.894**
Diameter of nozzle (ND)	1	15.350	15.350	604.906**
Nozzle distance (DBNS)	1	0.000	0.000	0.000
Angle of impingement (AI)	1	0.002	0.002	0.089
Rotational velocity (RV)	1	4.222	4.222	166.386**
ND × DBNS	1	0.000	0.000	0.000
ND × AI	1	0.000	0.000	0.000
ND × RV	1	0.059	0.059	2.337
DBNS × AI	1	0.000	0.000	0.003
DBNS× RV	1	0.000	0.000	0.001
AI × RV	1	0.000	0.000	0.001
Residual	14	0.355	0.025
Lack of Fit	10	0.351	0.035	35.851 s
R-squared	0.991
Adjusted R-squared	0.982
Pred R-squared	0.949
s and ** show  significant and significant difference at probability of % 1, respectively. Numbers have been rounded to three decimal places.

Fig. 12. Interaction effects between operating parameters of air-jet impingement on percentage of removed seeds for CRSH

DISCUSSIONS

In SR_SH, the maximum percentage of the removed seeds was obtained when the values of ND, AI, DBNS and RV were equal to 8 mm, 30°, 20 mm and 10 rpm respectively. In contrast, the minimum percentage of the removed seeds with AJSSR machine belonged to ND of 4 mm, AI of 90°, DBNS of 30 mm and RV of 20 rev min^-1. According to RSM optimization results, the optimal condition of machine operation in SR_SH belonged to ND of 4.19 mm, AI of 30°, DBNS of 17.83 mm, and RV of 16.48 rev min^-1. Additionally, results showed that in all cases, percentage of the removed seeds increased with increasing ND and decreasing AI and RV values. 

In MR_SH, the maximum performance of AJSSR machine was provided by ND of 8 mm, AI of 60°, DBNS of 20 mm and RV of 10 rpm, respectively. Corresponding values for the minimum percentage of the removed seeds were equal to 4 mm, 90°, 30 mm and 20 rev min^-1. Results indicated that in all cases, percentage of the removed seeds increased with increasing of ND and decreasing of RV. Consequently, the optimum condition of machine operation for SR_SH belonged to ND of 8 mm, AI of 62.42°, DBNS of 15.65 mm and RV of 17.59 rpm (Fig. 11).

In CR_SH, the maximum and minimum percent of the removed seeds belonged to ND of 8 mm and 4 mm, AI of 60° and 30°, DBNS of 20 mm and RV of 10 rpm and 15 rpm, respectively. Results indicated that in all cases, percentage of the removed seeds increased with increasing ND and decreasing RV. As well, the optimum condition of machine operation for CR_SH were obtained with ND of 8 mm,  AI of 62.08°, DBNS of 18.74 mm and RV of 20 rev min^-1.

Statistical F-test was performed to determine the effects of three different regions of SHs (SR_SH, MR_SH and CR_SH) on performance of AJSSR machine. The quadratic equations were used to model the percentage of the removed seeds through its lower number of coefficients in compare with third, fourth-degree, etc. equations. The coefficients of developed quadratic model for SR_SH, MR_SH and CR_SH are presented in Table 9.

Table 9. The constant coefficients of the quadratic model to predict percentage of removed seeds for SRSH, MRSH and CRSH

SV	Side region	Middle region	Central region
Nozzle diameter (ND)	3.55417	-0.53213	0.62850
Nozzle distance (DBNS)	1.66825	1.39584	0.33780
Angle of impingement (AI)	-0.56425	0.75973	0.20836
Rotational velocity (RV)	-0.22367	1.48828	-0.01293
ND × DBNS	-0.02413	-0.01465	-0.00001
ND × AI	-0.01146	-0.00426	0.00000
ND × RV	-0.12175	-0.11613	-0.01218
DBNS × AI	0.00183	0.00013	0.00001
DBNS× RV	0.00830	0.00408	0.00005
AI × RV	0.01222	0.00094	0.00002
ND 2	0.25813	0.51759	0.00999
DBNS2	-0.04929	-0.03712	-0.00848
AI 2	0.00090	-0.00652	-0.00174
RV 2	-0.04305	-0.05865	-0.00115
Constant coefficient	39.93125	-17.99817	-5.19375
Numbers have been rounded to five decimal places.

According to the experimental results, in the same operation condition, the highest performance belonged to the SR_SH and the least performance belonged to the CR_SH. This is due to more maturity of SR_SH seeds toward the CR_SH seeds. Physiological maturity of sunflower seeds starts from the SR_SH to the CR_SH, so, when the sunflower head is matured, there are immature seeds in central region which are still absorbing nutrition from the plant; therefore, in most cases, in the CR_SH, maturity has not occurred completely. Accordingly, a higher picking force for CR_SH seeds compared to the SR_SH and MR_SH is required [22].

With increasing ND and A_P, jet momentum and jet force increases; therefore, with increasing ND, percentage of removed seeds increases; the experimental results confirmed this expectation. Furthermore, with increasing DBNS, the covered area by jet increases, but the experimental results showed that percentages of the removed seeds had a decreasing and increasing phases from 10 to 30 mm. Furthermore, with increasing AI from 30° to 90°, the covered area by the jet decreases. This issue was confirmed by experimental results of SR_SH compared to MR_SH and CR_SH. Arrangement of the seeds on SH and distance between adjacent seeds are the principle factors for adjustment of AI parameter.

Results of [16] showed that the air pressure, nozzle diameter, the number of passes, and the route of the nozzle over the surface of halved pomegranate or citrus fruit significantly affected the percentage of removed arils of pomegranate, citrus juice and juice sacs. Their studies also indicated that with increasing reservoir pressure and nozzle diameter, the percentage of removed arils of pomegranate, citrus juice and juice sacs increased. Furthermore, [16] showed that air pressure and nozzle diameter significantly influenced the percentage of damaged pomegranate arils.

Effects of operational parameters of impingement air-jet on damaged pomegranate arils and citrus sacs were investigated by [16] and [17], respectively. They cited that nozzle diameter and reservoir pressure had a significant influence on the percentage of damaged pomegranate arils and citrus sacs, while in the present work, in all tests, no seeds have been observed to have been damaged due to air-jet impinging, and this is one of the most important advantages of the AJSSR machine.

CONCLUSIONS

Effects of different air-jet impingement operational parameters including diameter of nozzle (ND), angle of impingement (AI), distance between nozzle outlet and sunflower heads (DBNS), air pressure (A_P), and rotational velocity of the SH (RV) on removing performance of machine (RP_AJSSR) for three different areas of the SH were examined.

In order to investigate the mentioned effects and parameters, a preliminarily model machine was designed and constructed based on air-jet impingement method. To model and optimize the effects of the mentioned parameters, artificial neural network and response surface methodology were used. The final selected artificial neural network (ANN) model, 3-5-1, successfully modeled the relationship between three air-jet impingement parameters (ND, AI and RV), and output parameters (RP_AJSSR). The final trained ANN model was able to predict simultaneously the removing performance with MSE of 0.0035, R² of 0.98 and T value of 0.96.

Results of RSM optimization indicated that in the same conditions, the maximum percentage of the removed seeds belonged to the SRSH and the minimum percentage belonged to the CRSH. For all regions of SHs, with increasing ND and decreasing RV, performance of machine increased. Furthermore, when DBNS values ranged between 15 and 19 mm and RV values ranged from of 17 to 20 rev min-1, the optimum performance of power consumption and removed seeds were obtained. It is interesting to note that in all cases, no damaged seeds during AJSSR machine operation were observed. It can be a significant advantage of jet-impingement method over mechanical methods for removing sunflower seeds from SHs.

For instance, removing success of air jet system for sunflower seeds in sunflower heads is 57% at most, in the present study, which is not low compared to the mechanical systems. But in the present study we used just one nozzle; by increasing number of nozzles we will able to remove high present of sunflower seeds from their heads. Also, on the other hand, most of the seeds that are located on the central region of the SHs are too small and their quality and their marketability are low, and a lot of them are hollow. So, with good regulation of machine’s operating parameters, the machine is capable to sort out low quality seeds by not removing them.

Acknowledgements

The authors would like to thank the University of Tehran for providing technical support for this work. We also want to sincerely thank Dr. Mohammad Hassan Torabi, Mr. Asghar Mirzabe Mr. Fazlollah Ansari, deceased Mr. Feizollah Aghasi, Mr. Ali Kharaji, Mr. Ehsan Afshari and Eng. Javad Yousefi for his technical help and supervision while writing the paper.

REFERENCES

1. Agatonovic-Kustrin S., Beresford R., 2000. Basic concepts of artificial neural network (ANN) modeling and its application in pharmaceutical research. J. Pharm. Biomed. Anal., 22(5): 717–727.
2. Ando T., Suzuki T., Ishii K., Omura H., Yamazaki J., 1988. Apparatus for separating juice sacs of citrus fruits. U.S. Patent 4,738,194.
3. Arribas J.I., Sánchez-Ferrero G.V., Ruiz-Ruiz G., Gómez-Gil J., 2011. Leaf classification in sunflower crops by computer vision and neural networks. Comput. Electron. Agric., 78(1): 9–18.
4. Aydin C., 2003. Physical properties of almond nut and kernel. J. Food Eng., 60(3): 315–320.
5. Bas D., Boyaci I.H., 2007. Modeling and optimization II: comparison of estimation capabilities of response surface methodology with artificial neural networks in a biochemical reaction. J. Food Eng., 78(3): 846–854.
6. Basheer I.A., Hajmeer M., 2000. Artificial neural networks: fundamentals, computing, design, and application. J. Microbiol. Methods, 43(1): 3–31.
7. Bezerra M.A., Santelli R.E., Oliveira E.P., Villar L.S., Escaleira L.A., 2008. Response surface methodology (RSM) as a tool for optimization in analytical chemistry. Talanta, 76(5): 965–977.
8. Box G.E.P., Behnken D.W., 1960. Some new three level designs for the study of quantitative variables. Technometrics, 2(4): 455–475.
9. Box G.E.P., Draper N.R., 1987. Empirical model-building and response surfaces. John Wiley & Sons, New York.
10. Bradley N., 2007. The response surface methodology. PhD dissertation, Indiana University South Bend.
11. Chamoli S., 2015. ANN and RSM approach for modeling and optimization of designing parameters for a V down perforated baffle roughened rectangular channel. Alexandria Eng. J., 54(3): 429–446.
12. Chegini G.R., Khazaei J., Ghobadian B., Goudarzi A.M., 2008. Prediction of process and product parameters in an orange juice spray dryer using artificial neural networks. J. Food Eng., 84(4): 534–543.
13. Fay J.A, Sonwalkar N., 1996. A fluid mechanics hypercourse. A Fluid Mech. Hypercourse, by James A. Fay Nishikant Sonwalkar. ISBN 0-262-56103-4. Cambridge, Massachusetts, USA MIT Press. May 1996 (Paper).
14. Goel A.K, Behera D., Swain S., Behera B.K., 2009. Performance evaluation of a low-cost manual sunflower thresher. Indian J. Agric. Res., 43(1): 37–41.
15. Hayashi M., Ifuku Y., Uchiyama H., Kaga Y., Nakamori A., 1981. Apparatus for extracting pulp from citrus fruits. U.S. Patent 4,300,448.
16. Khazaei J., Ekrami-Rad N., Safa M., Nosrati S.-Z., 2008. Effect of air-jet impingement parameters on the extraction of pomegranate arils. Biosyst. Eng., 100(2): 214–226.
17. Khazaei J., Massah J., Mansouri G.H., 2008. Effect of some parameters of air-jet on pneumatic extraction of citrus juice and juice sacs. J. Food Eng., 88(3): 388–398.
18. Menlik T., Özdemir M.B., Kirmaci V., 2010. Determination of freeze-drying behaviors of apples by artificial neural network. Expert Syst Appl, 37(12): 7669–7677.
19. Milani E., Seyed M., Razavi A., Koocheki A., Nikzadeh V., Vahedi N., MoeinFard M., GholamhosseinPour A., 2007. Moisture dependent physical properties of cucurbit seeds. Int. Agrophysics, 21(2): 157.
20. Mirzabe A.H., Chegini G.R., Khazaei J., Massah J., 2014. Design, construction and evaluation of preliminarily machine for removing sunflower seeds from the head using air-jet impingement. Agric. Eng. Int. CIGR J., 16(1): 294–302.
21. Mirzabe A.H., Chegini G.R., 2015. Measuring picking force of sunflower seeds and prediction of reasonable range of air-jet parameters to remove sunflower seeds from the head. Agric. Eng. Int. CIGR J., 17(3), 415–429.
22. Mirzabe A.H., Khazaei J., Chegini G.R., 2012. Physical properties and modeling for sunflower seeds. Agric. Eng. Int. CIGR J., 14(3): 190–202.
23. Nahir D., Ronen B., 1992. Apparatus for removing pulp from fruit. U.S. Patent 5,088,393.
24. Prasad V.S.S., Gupta S.D., 2008. Applications and potentials of artificial neural networks in plant tissue culture. Plan Tissue Cult. Eng. Springer, 47–67.
25. Sarig Y., Regev Y., Grosz F., 1985. Apparatus for separating pomegranate seeds, scanning apparatus and techniques useful in connection therewith and storage and packaging techniques for separated seeds. U.S. Patent 4,530,278.
26. Schmilovitch Z., Sarig Y., Daskal A., Weinberg E., Grosz F., Ronen B., Hoffman A., Egozi H., 2014. Apparatus and method for extracting pomegranate seeds from pomegranates. U.S. Patent 7,968,136.
27. Sudajan S., Salokhe V.M., Triratanasirichai K., 2002. PM – Power and machinery: effect of type of drum, drum speed and feed rate on sunflower threshing. Biosyst. Eng., 83(4): 413–421.
28. Witek-Krowiak A., Chojnacka K., Podstawczyk D., Dawiec A., Pokomeda K., 2014. Application of response surface methodology and artificial neural network methods in modelling and optimization of biosorption process. Bioresour. Technol., 160, 150–160.
29. Yeh W.-C., 2003. A MCS-RSM approach for network reliability to minimise the total cost. Int. J. Adv. Manuf. Technol., 22(9–10), 681–688.
30. Zakiuddin K.S., Sondawale H.V., Modak J.P., Ceccarelli M., 2012. History of Human Powered Threshing Machines: a Literature Review. Explor. Hist. Mach. Mech. Springer, 431–445.

Received: 29.12.2019
Reviewed: 3.04.2020
Accepted: 27.04.2020

---

Amir Hossein Mirzabe
Department of Mechanical Engineering of Biosystems, College of Agriculture & Natural Resources, University of Tehran, Tehran, Iran
Telephone: 098 3153239185
Cell phone: 0989399442161
a_h_mirzabe@yahoo.com
email: a_h_mirzabe@alumni.ut.ac.ir

Gholam Reza Chegini
Department of Mechanical Engineering of Biosystems, College of Aboureihan, University of Tehran, Tehran, Iran
Telephone: 098 21 360 406 14
Cell phone: 0989126356329
email: chegini@ut.ac.ir

Jafar Massah
Department of Mechanical Engineering of Biosystems, College of Aboureihan, University of Tehran, Tehran, Iran
Telephone: 098 21 360 406 14
Cell phone: 0989198028454
email: jmassah@ut.ac.ir

Ali Mansouri
Department of Mechanical Engineering of Biosystems, College of Aboureihan, University of Tehran, Tehran, Iran
Telephone: 098 3153239185
Cell phone: 0989171837151
email: ali.mansouri@ut.ac.ir

Javad Khazaei
Department of Mechanical Engineering of Biosystems, College of Aboureihan, University of Tehran, Tehran, Iran
Telephone: 098 21 360 406 14
Cell phone: 0989123880128
email: jkhazaei@ut.ac.ir

---

Responses to this article, comments are invited and should be submitted within three months of the publication of the article. If accepted for publication, they will be published in the chapter headed 'Discussions' and hyperlinked to the article.

---

Main  -  Issues  -  How to Submit  -  From the Publisher  -  Search  -  Subscription