Optimized Support Vector Machines Combined with Evolutionary Random Forest for Prediction of Back-Break Caused by Blasting Operation

Abstract

Back-break is an adverse event in blasting works that causes the instability of mine walls, equipment collapsing, and reduction in effectiveness of drilling. Therefore, it boosts the total cost of mining operations. This investigation intends to develop optimized support vector machine models to forecast back-break caused by blasting. The Support Vector Machine (SVM) model was optimized using two advanced metaheuristic algorithms, including whale optimization algorithm (WOA) and moth–flame optimization (MFO). Before the models’ development, an evolutionary random forest (ERF) technique was used for input selection. This model selected five inputs out of 10 candidate inputs to be used to predict the back break. These two optimized SVM models were evaluated using various performance criteria. The performance of these two models was also compared with other hybridized SVM models. In addition, a sensitivity evaluation was made to find how the selected inputs influence the back-break magnitude. The outcomes of this study demonstrated that both the SVM–MFO and SVM–WOA improved the performance of the standard SVM. Additionally, the SVM–MFO showed a better performance than the SVM–WOA and other hybridized SVM models. The outcomes of this research recommend that the SVM–MFO can be considered as a powerful model to forecast the back-break induced by blasting.

1. Introduction

Back-break (BB) is an undesired outcome of blasting in mining operations. This phenomenon refers to the breaking into pieces of rocks exceeding the thresholds of the rear row of holes in a blast design [1]. Some undesirable impacts of the BB include instability of rock mine wall, fallings, and increment of the overall cost of blasting [2,3]. There are three main categories of parameters which affect the BB; these categories include (1) parameters related to blast design, (2) explosive material characteristics, and (3) the rock mass traits and breaks. While the first and second types of variables are regarded as manageable, the third group is viewed as uncontrollable blasting variables [4,5,6,7,8]. The factors that may affect the BB include low stiffness ratio, extreme burden, over stemming of the hole, and geological structure [1,9,10]. As many parameters affect BB, suitable appraisal and prediction of this environmental consequence are extremely challenging. The mentioned challenge can be solved using different, well-designed models for blasting pattern parameters. In addition, geological conditions should be observed and considered before blasting operations [11]. Many previous studies endeavored to forecast BB through numerous machine learning (ML) techniques such as support vector machines (SVM), artificial neural networks (ANN), and so on [3,12,13,14,15,16]. Monjezi, Rezaei and Yazdian [14] developed two multiple-based techniques, namely multiple regression (MR) and fuzzy inference system (FIS) to forecast BB induced by blasting. They used burden, the charge per delay, hole depth, specific drilling, stemming spacing, powder factor, and rock density as input variables. They also discovered the FIS model outperforms the MR model. The ANN, neuro-fuzzy and MR models were used by Esmaeili, Osanloo, Rashidinejad, Bazzazi and Taji [3] to forecast the BB employing data of Sangan iron mine, Iran. Compared to other models, they demonstrated that the neuro-fuzzy model receives a superior performance. Using 10 input parameters, for the same problem, Monjezi, Ahmadi, Varjani and Khandelwal [13] developed an ANN model using a database comprising of 97 data samples. In addition to the ANN model, they suggested an MR equation for the BB prediction. They successfully showed that the ANN model outperforms the MR model. Another study, which was conducted by Mohammadnejad, Gholami, Sereshki and Jamshidi [2], showed the applicability and practicability of the SVM for prediction of BB and concluded that the SVM technique is a reliable and precise tool for the BB prediction. Some other ML techniques, such as random forest and its optimized approaches, optimized fuzzy-rule techniques with rock engineering systems, and genetic-based models have been suggested in literature for forecasting BB resulting from blasting [17,18,19]. The details of some recent studies on BB prediction using ML and artificial intelligence (AI) techniques are described in

Table 1. Some recent ML and AI studies on the BB prediction.

Method	Study	Input	Dataset Size
FIS	Monjezi, Rezaei and Yazdian [14]	Burden, charge per delay, hole depth, powder factor, rock density, spacing, specific drilling, stemming	-
ANN–GA	Monjezi et al. [42]	Burden, charge per delay, hole diameter, hole length, powder factor, rock mass rating, spacing, specific drilling	195
ANN	Monjezi, Ahmadi, Varjani and Khandelwal [13]	Burden, charge per delay, hole depth, hole diameter, powder factor, spacing, specific drilling, stemming, uniaxial compressive strength water content	97
ANN	Sayadi et al. [43]	Burden, hole depth, spacing, specific charge, specific drilling	103
SVM	Mohammadnejad, Gholami, Sereshki and Jamshidi [2]	Burden, hole depth, powder factor, spacing, specific drilling, stemming	193
SVM	Khandelwal and Monjezi [12]	Burden, hole length, powder factor, spacing, specific drilling, stemming	234
ANN	Monjezi et al. [44]	Burden, delay per burden, number of rows, powder factor rock factor, spacing, specific drilling, stemming	-
ANN, neuro-fuzzy	Esmaeili, Osanloo, Rashidinejad, Bazzazi and Taji [3]	Charge last row, number of rows, specific charge, stemming	42
ANN, ABC	Ebrahimi et al. [45]	Burden, hole depth, powder factor, spacing, stemming length	34
GP	Faradonbeh et al. [46]	Burden, powder factor, spacing, stemming, stiffness ratio	175
fuzzy RES–GA, fuzzy RES–ICA	Hasanipanah and Bakhshandeh Amnieh [17]	Burden, blast-hole inclination, burden to hole diameter ratio, charge per delay, hole diameter, spacing to burden ratio, stemming to burden ratio, velocity of detonation	62
RF	Kumar et al. [47].	Spacing to burden ratio, P-wave, hole length to stemming ratio, density of explosive	140
SCA–RF, HHO–RF	Zhou, Dai, Khandelwal, Monjezi, Yu and Qiu [18]	Burden, hole length, powder factor, spacing, specific drilling, stemming	234
ANN, ACO	Saghatforoush et al. [48]	Burden, hole length, powder factor, spacing, stemming length	97

GP: genetic programing, ACO: ant colony optimization, GA: genetic algorithm, ICA: imperialism competitive algorithm, RES: rock engineering system, RF: random forest, HHO: Harris hawks optimizer, SCA: sine cosine algorithm.

. In this table, input variables of BB predictive models together with the size of the database used by various researchers are presented. It is worth stating that these models have been used by other investigators in fields of blasting, rock mechanics, geotechnics, and civil and mining engineering [20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41].

The victorious utilizations of SVM and its associated combination variants in determining different geotechnical difficulties were informed by many scholars [49,50,51,52]. In the domain of blasting and its environmental effects, many scholars established SVM models to forecast ground vibration, fly-rock distance, rock fragmentation, and blast-caused rock movement [24,53,54,55]. The SVM is regarded as a robust method that can confidently determine geotechnical-related difficulties. Thus, this research concluded to utilize various hybrid SVM models to explain the BB issue. It is necessary to remark that to the best of the authors’ knowledge, the SVM models were implemented and offered in the domain of BB forecast. Nevertheless, the application of innovative combined predictive models based on the idea of an SVM model optimized by any robust optimization methods are neglected in this domain. Therefore, the central role of this study is to utilize and propose the novel combination of SVM models in forecasting BB extent. To accomplish this, the authors selected to use two renowned, robust, and fit optimization methods, including moth–flame optimization (MFO) and whale optimization algorithm (WOA), in integrated SVM models of this research. Consequently, two SVM-centered models, such as SVM–MFO and SVM–WOA, are developed in this study for BB magnitude forecasts. The objective of the MFO and WOA systems is to optimize the SVM parameters, including ‘C’ and ‘ε’, to observe more eminent performance capabilities for forecast views. Tuning and optimizing SVM parameters by optimization algorithms has been used by researchers for improving the performance of the base model (SVM).

The following explains the remainder of this paper. First, the case study of this research and the methods of the data collection are presented. Second, the structure of the SVM model and its optimizers (MFO and WOA) is presented. Third, the results of this study are presented, evaluated, and discussed. Finally, a summary of this paper is presented in the Section 7.

2. Field Observation and Measurement

The data of this present research were gathered from a blasting operation in Gol-E-Gohar Iron mine. This mine is situated in Kerman province, Iran (Figure 1). During the blasting operation, the height of the blast holes and diameter were 17 m and 203 mm, respectively. The lag time between the first and second row was 80 ms. The lag time for the other rows was set as 50 ms. The stemming material used was drilling cuttings. In this mine, the BB phenomenon has worsened and reached up to 20 m because of its inappropriate blasting patterns. The problem of the BB (with 20 m) induced by the blasting operation in the investigated mine is shown in Figure 2.

Table 2. Summary of gathered data with their symbol, unit and range.

Parameter	Class	Unit	Acronym	Min	Max
Powder factor	Input	kg/ton	PF	1.90	0.30
Burden	Input	m	B	5.00	6.50
Spacing to burden ratio	Input	---	S/B	1.18	1.30
Number of rows	Input	---	No. row	2.00	5.00
Charge per delay	Input	Kg	CPD	72.00	455.96
Last row charge to total charge ratio	Input	---	LRC/TC	0.04	0.55
Stemming to burden ratio	Input	---	St/B	0.80	1.60
Joint condition	Input	---	JC	1.00	4.00
Uniaxial compressive strength	Input	MPa	UCS	55.00	90.00
Water height to burden ratio	Input	---	W/B	0.00	2.00
Back-break	Target	m	BB	3.00	20.00

demonstrates the scale, unit and symbol of the assessed parameters in the study site together with more information of the BB parameter. In total, 85 blasting events were considered and their data were used in this research for the BB forecast. In the following, the process of selecting the most important parameters among these 10 measured parameters on the BB will be discussed and then the modeling procedure using tree-based models will be described in detail.

3. Methods

3.1. SVM

The SVM is considered as a supervised ML technique that is successfully implemented in the domain of geotechnical and tunnelling engineering [56]. The linear function of SVM can be explained as follow:

$$f\left(a\right)=w.a+d$$

(1)

where $a$ denotes the input variable, w denotes the weight vector, and d points to model error values. SVM strives to decrease the disparity between the real and predicted values. Therefore, SVM predicts according to reducing the objective function, which is an error indicator. Below is the optimization procedure [57]:

$$min\frac{1}{2}\Vert w{\Vert }^2+c{\displaystyle \sum }_{i=1}^k\left({\xi }_i^{-}-{\xi }_i^{+}\right)$$

(2)

$$Subject\left(to\right)\left(w{x}_i+d\right)-b_i<\epsilon +{\xi }_i^{+}$$

(3)

$$b_i-\left(w_i{a}_i+d\right)\le \left(\epsilon +{\epsilon }_i^{-}\right)$$

(4)

where C denotes the coefficient of penalty, k means the number of training data, ${\xi }_i^{-}$ and ${\xi }_i^{+}$ signify the data violations whose various values are higher than $\xi$ the allowed range with observable values, and $w_i$, $a_i$, and $b_i$ refer to the variables’ weight, the input variable, and the target observation. Equations (2) and (3) are used to estimate the values of w and d and are then replaced into Equation (1). In SVM, to represent the input data points to a high-dimensional feature space, the kernel function can be employed. The kernels are able to resolve the issues with numerous dimensions. A total of four well-known SVM kernels are used, including sigmoid, linear, polynomial, and radial basis function (RBF). In this study, the RBF kernel was used since this kernel was proved to possess a desirable generalization capability for various types of datasets. Thus, Equation (1) is regarded as follows:

$$f\left(a\right)=w.H\left(a,a_i\right)+d$$

(5)

$$H(a,a_i)=\exp (-\frac{a-a_i}{2{\gamma }^2})$$

(6)

where $H\left(a,a_i\right)$ stands for the kernel function, and γ signifies the kernel function’s parameter. SVM parameters such as C and ε have unknown values and are assumed as decision variables. Thus, they should enter the optimization process. The aim of hybrid SVM, MFO, and WOA is to determine the precise value of the parameters mentioned above and predict BB by SVM. Figure 3 indicates the optimization process employed by each optimization method, including WOA and MFO, as well as their functions in two combined models of SVM–MFO and SVM–WOA to forecast BB values.

3.2. MFO

MFO is one of the most effective optimization algorithms that mimics the fly styles of moths in the darkness. Typically, moths try to keep a fixed position to the moon for shuttling at nighttime [58]. They engage a method defined as a transverse orientation to navigate. Nevertheless, sometimes this method is useless, and especially so for straight movement if the light origin is extremely far away. If moths find irregular lighting, they endeavor to preserve an analogous form with the brightness to pass it in a straight way. Notwithstanding that this bright origin is closer to the moths than the moon, holding a comparable angle to the origin of light creates an incompetent or killing spiral fly-path for moths. This sort of killing flow, while the origin of light is close, is used to determine the optimization problems in the actual practice. In this method, the probable answers are moths, and variables are moths’ coordinate vectors in the exploration space. The MFO equation used for optimizing SVM is presented in Equation (7).

$$Flame number=round(N-m\ast \frac{N-1}{Q})$$

(7)

where m indicates the current number of repetitions, the highest quantity of flames is represented by N, and Q implies the largest number of repetitions.

3.3. WOA

The WOA is an optimization approach that imitates the humpback whale’s social practice in following their hunt in seas [59]. This algorithm utilizes a different method following the particular bubble net and the feeding habits of the whale. The WOA method comprises three principal activities, including encircling prey, bubble-net attacking (Equation (8)) and prey exploring (Equation (9)). Here is the mathematical presentation of these activities.

$$X^{s+1}=\{\begin{array}{c}{X}_{gbest}^s-F.\left|M.X_{gbest}^s-X^s\right|,p<0.5\\ X_{gbest}^s+Q.e^{bl}.\cos \left(2\pi l\right), p\ge 0.5\end{array}$$

(8)

$$X^{s+1}=X_{rand}^s-F.\left|M.X_{rand}^s-X^s\right|$$

(9)

where s refers to the present repetition, $X^s$ is the existing position vector, $X^{s+1}$ is the new position, $X_{gbest}^s$ represents the existing position of the best solution achieved, F and M are the coefficient vectors, l stands for an arbitrary number within [−1,1], b signifies a constant for limiting the form of the logarithmic spiral, p signifies an arbitrary number within [0,1], $X_{rand}^s$ denotes a position vector of a whale individual at random selected from the existing inhabitants, and $Q=\left|X_{gbest}^s-X_s\right|$ stands for the space between the whale and the target.

Every humpback whale signifies an individual, and the location of every individual in the search space describes a solution. The whale can know the position of the prey and circle the prey within the echolocation (encircling prey). The whale strikes the target by spiraling up and constantly narrowing the circling (bubble-net attacking). If the coefficient vector |F| > 1, it indicates that the whale moves beyond the lessening encircling circle. At this point in time, the humpback whale explores arbitrarily based on every other’s position.

4. SVM Optimized Models

For developing the optimized SVM models, the aim of employing WOA and MFO algorithms is to optimize the SVM hyperparameters ‘ε’ and ‘C’. The values of these parameters were set within the following ranges:

• C: 0.01–100;
• ε: 0.01–50.

The key procedure of optimizing SVM parameters utilizing WOA and MFO optimization methods is indicated in Figure 4. The first step involved is preparing the data and dividing it into training and testing sets. The second step involved is assigning appropriate values to the WOA and MFO parameters. In the third step, RMSE was used as an indicator for assessing the models’ fitness. The parameters were revised based on the outcomes of each repetition in the fourth step. In the final step, the most suitable values for the parameters were achieved, and the stopping conditions were met.

5. Results and Discussions

5.1. Feature Selection

An evolutionary random forest (ERF) technique was employed to select the most significant factors for BB prediction. This model was applied to ten candidate inputs, including PF, B, S/B, No. row, CPD, LRC/TC, St/B, JC, and UCS. The ERF selected five important inputs, including PF, No. row, CPD, LRC/TC, and St/B. These inputs were used to develop the SVM model and its optimized variants. This model was ran using the parameters shown in Figure 5. Figure 6 demonstrates the frequency distribution of each selected input and BB. The coefficient of determination (R²)-linear regression of this model was obtained as 0.909.

5.2. Models’ Development and Evaluation

Once the models are developed, it is vital to assess the performance of them. To this end, four performance criteria, including R², the mean absolute error (MAE), the root mean squared error (RMSE), and the variance accounted for (VAF), were employed. The formulas for estimating these criteria are shown in Equations (10)–(13). In these equations, BB_i is the real value, $\widehat{{\mathit{BB}}_i}$ stands for the forecasted value, $\overline{{\mathit{BB}}_i}$ implies the mean of the real values, and N signifies the number of samples in the training or testing phases. It is important to mention that these performance criteria have been used in many published works (e.g., [60,61,62,63,64,65]).

$$RMSE=\sqrt{\frac{1}{N}{\displaystyle \sum }_{i=1}^N{(\widehat{B{B}_i}-B{B}_i)}^2}$$

(10)

$$R^2=1-\frac{{\sum }_{i=1}^N{(B{B}_i-\widehat{B{B}_i})}^2}{{\sum }_{i=1}^N{(B{B}_i-\overline{B{B}_i})}^2}$$

(11)

$$MAE=\frac{1}{N}{\displaystyle \sum }_{i=1}^N\left|\widehat{B{B}_i-}B{B}_i\right|$$

(12)

$$VAF=\left[1-\frac{var(B{B}_i-\widehat{B{B}_i})}{var\left(B{B}_i\right)}\right]\times 100$$

(13)

Figure 7 shows the association between the predicted and real values of BB. The outcomes indicate that the testing and training outcomes of these algorithms are great. The training and test values are scattered adjacent to the best fitting line. Concerning the performance indicators, the forecast performance of the SVM–MFO is somewhat greater than the other model. For the training phase, the R² (linear regression), RMSE, MAE, and VAF values were 0.992, 0.364, 0.044, and 99.150, correspondingly. For the testing phase, the R², RMSE, MAE, and VAF values were 0.985, 0.629, 0.332, and 98.371, respectively. It is worth noticing that the SVM–MFO is a more reliable model than the other optimized model for both the training and testing phases. Expectedly, these two hybrid models are capable of considerably improving the performance capability of a single SVM model in predicting BB, as displayed in

Table 3. Performance of the models developed.

Model	Training				Testing
	R2 (rank)	RMSE (rank)	VAF (rank)	MAE (rank)	R2 (rank)	RMSE (rank)	VAF (rank)	MAE (rank)
SVM	0.885 (1)	1.714 (1)	88.404 (1)	1.135 (1)	0.844 (1)	1.949 (1)	82.391 (1)	1.553 (1)
SVM–MFO	0.992 (3)	0.364 (3)	99.150 (3)	0.044 (3)	0.985 (3)	0.629 (3)	98.371 (3)	0.332 (3)
SVM–WOA	0.981 (2)	0.559 (2)	98.064 (2)	0.125 (2)	0.974 (2)	0.805 (2)	97.168 (2)	0.391 (2)

. For example, the RMSE value can be lessened from 1.714 to below 0.6 (in the training phase) by optimizing the SVM models.

Table 3 and

Table 4. Total ranking of the models developed.

	Training Rank	Testing Rank	Total Rank
SVM	4	4	8
SVM–MFO	12	12	24
SVM–WOA	8	8	16

review the performance indicator outcomes (R², RMSE, MAE, and VAF) and overall ranking outcomes of the standard SVM and two hybrid models in forecasting BB. The joining outcomes of the training and testing sets are that the total ranking of SVM–MFO is superior. This explains that the SVM–MFO gives greater precision and robustness in forecasting BB compared to the SVM–WOA.

The performance of the two models established in this research was also compared with two more well-known SVM optimized models, including SVM–particle swarm optimization (PSO) and SVM–cuckoo optimization algorithm (COA). The outcomes of this comparison are displayed in Figure 8. The parallel graph outcomes explain that the forecast performance of the SVM models established in this research is more precise than the other algorithms. Amongst all the models, the SVM–MFO is more reliable. According to the findings of this study, it is evident that the SVM–MFO has excellent learning and forecast capacities. Hence, this present study suggests employing the developed SVM–MFO model for BB prediction.

The data used in this study were also used by Monjezi and Dehghani [66]. They did not employ a feature selection technique and used seven inputs for developing neural network models. The best values of R² and RMSE that they achieved were 0.972 and 0.643, respectively. For both training and testing, the R² of SVM–MFO is better than those presented by Monjezi and Dehghani [60]. In addition, these data were used by Monjezi, Rezaei and Yazdian [14], and they applied fuzzy set theory and MR models to eight inputs. Their best R² value was 0.954. This comparison shows that despite the usage of fewer inputs for developing the SVM–MFO, this model showed better performance than the models developed in the above studies. It is important to mention that one of the shortcomings and disadvantages of ML and AI models is their limited practical application. We as engineers should always try to make them as simple as possible in practice for other researchers and designers. In this way, one of the possible options is related to the number of inputs that we need to give to the system. The level of complexity can be decreased by reducing the number of input parameters. Another point is related to the fact that if a lower number of inputs are needed to be collected, the process of data collection would be easier and faster compared with the situation in which we need to collect and have all inputs.

6. Analysis of Sensitivity

In mechanical tunnel engineering, the forecast of a BB is the solution under specific rock circumstances. Different determinants of BB should be systematically examined to forecast the BB precisely and decrease the great cost and danger of tunnel building. It can be identified that whole inputs, namely PF, No. row, CPD, LRC/TC, and St/B, contribute to the BB forecast. Nevertheless, the sensitivity of every input is ambiguous and requires further investigation.

In this part, the mutual information (MI) test method [67] is employed for examining the significance of BB factors and their sensitivity. MI is a filtering technique that obtains the arbitrary connection between inputs and the target. MI tests the dependence among variables and shows the intensity of the association among them. The MI magnitude among variables is estimated by means of the information gain:

$$Gain\left(A, B\right)=Ent\left(A\right)-{\displaystyle \sum }_{s-1}^S\frac{\left|A^s\right|}{\left|A\right|}Ent\left(A^s\right)$$

(14)

where s denotes the number of all probable values of B, As is the set of A when B takes the value Bs, and Ent(A) signifies the information entropy. The larger the value of gain (A, B), the better the relationship between B and A.

Finally, based on the variable score in the MI examination, the significance intensity of the input that forecasts BB was ascertained. The results of this analysis are indicated in Figure 9. The most crucial variables for forecasting BB were CPD, PF, and St/B. Their significance scores were 59.52, 15.09, and 11.92, respectively. The lowest score belonged to LRC/TC (significance score = 2.94). Nevertheless, it should be mentioned that inputs, such as LRC/TC and No. row, still have a deep influence on BB.

7. Conclusions

This study attempts to hybridize the SVM algorithm with well-known and efficient optimization algorithms in the domain of BB prediction. To achieve this goal, two renowned optimization algorithms, including WOA and MFO, which were effectively investigated by previous scientists, were chosen and integrated with SVM, and later, SVM–MFO and SVM–WOA were built for forecasting intentions. The models were built employing ten inputs and one target, which was BB. Before the models’ development, an ERF was used as the feature selection method to lessen the data dimensionality and identify the most relevant inputs for BB prediction. The inputs selected by this technique were PF, No. row, CPD, LRC/TC, and St/B. To appraise the performance of the developed models, several measures were employed, including R², RMSE, VAF, and MAE. Additionally, for the purpose of comparison, the authors have forecasted BB developing different models, including standard SVM, SVM-PSO, and SVM-COA. Finally, following the evaluation of the performance of the entire implemented and built models, it was discovered that the SVM–MFO had an R² of (0.992 and 0.985), RMSE of (0.364 and 0.629), VAF of (99.150 and 98.371), and MAE of (0.044 and 0.332), correspondingly, for training and test phases, which are better results than those of other employed projecting methods. Consequently, the model offered in this research can be employed in different schemes relating BBs for forecasting their accomplishments. By administering the sensitivity investigation, the relevance score of all inputs was achieved utilizing the MI method. This method was executed among those five variables that were identified by the ERF technique. The importance scores of LRC/TC, No. row, St/B, PF, and CPD were 2.94, 10.52, 11.92, 15.09, and 59.52. These results proved that CPD, PF, and St/B variables are regarded as greatly sensitive determinants on BB. Nevertheless, it should be mentioned that additional data and examination are required to analyze the BB under different, severe circumstances. Hence, the employment of the integrated model offered in this article is solely advised under comparable circumstances and in a consistent range of data. Future studies should employ datasets with more data and inputs to enhance the predictive capability of the model. Furthermore, AI-based systems cannot entirely substitute conventional practical techniques. Regarding geotechnical engineering, the potential advancement path of AI technology is a mixed method, which simply evolves in the direction of decision support instruments. Prominently, the smart systems employed are just suggested to be implemented under comparable circumstances in this research. The principal weakness of such methods in the geotechnical domain can be regarded as site-specific data employed in developing AI models. The geotechnical data and measurements can vary from location to location, and because of this issue, the generalization of the developed AI models is a challenging task.