Application of Cascade Forward Backpropagation Neural Networks for Selecting Mining Methods

Abstract

Decision-making is very important in many fields, such as mining engineering. In addition, there has been a growth of computer applications in all fields, especially mining operations. One of these application fields is mine design and the selection of suitable mining methods, and computer applications can help mine engineers to decide upon and choose more satisfactory methods. The selection of mining methods depends on the rock-layer specification. All rock characteristics should be classified in terms of technical and economic concerns related to mining rock specifications, such as mechanical and physical properties, and evaluated according to their weights and ratings. Methodologically, in this study, the criteria considered in the University of British Columbia (UBC) method were used as references to establish general criteria. These criteria consist of general shape, ore thickness, ore plunge, and grade distribution, in addition to the rock quality designation (ore zone, hanging wall, and foot wall) and rock substance strength (ore zone, hanging wall, and foot wall). The technique for order of preference by similarity to ideal solution (TOPSIS) was adopted, and an improved TOPSIS method was developed based on experimental testing and checked by means of the application of cascade forward backpropagation neural networks in mining method selection. The results provide indicators that decision makers can use to choose between different mining methods based on the total points given to all ore properties. The best mining method is cut and fill stopping, with a rank of 0.70, and the second is top slicing, with a rank of 0.67.

1. Introduction

Mining method selection (MMS) is a time-consuming and difficult task that requires outstanding knowledge and experience. Thus, it is a difficult assignment for mining engineers and managers. For correct and powerful evaluation, the decision maker may also need to investigate a huge set of data and to consider many factors. This selection is vital for mine control because of the operational cost, as well as being an essential part of mine planning and design. Most importantly, the optimal mining approach increases the protection of personnel and productivity [1]. Extensive research has been performed to discover an appropriate procedure for MMS. This is a procedure for choosing an extraction approach for a specific deposit. It entails correct planning, research, and knowledgeable decisions by professionals within the mining industry. Ooriad et al. stated that MMS is complicated and difficult because of the numerous factors that shape a part of the decision process. In addition, because the character of an orebody is unique, it may no longer be a practical method to simply undertake a mining approach without considering the requirements of a particular orebody [2].

Previously developed methods have been used for all commodities. Some were successful at implementation. However, others have had to evolve to be effective [3]. One is the University of British Columbia (UBC) method. The UBC method was advanced in 1995 by Miller-Tait [4] as an amendment to the Nicholas approach [5,6]. The scoring domain of the Nicholas method, which is between the maximum and minimum, was extended. This emphasizes the stopping method rather than mass-mining techniques. This is because it was designed to represent the typical Canadian practice, which is a limitation for use outside Canada [7]. The selection process is similar to that of the Nicholas method. The rankings and characteristics, except for grade distribution and plunge, are different. The ranking in the UBC method ranges from zero to six. Six is given to the characteristics of the most suitable mining method. Additionally, −10 was introduced to the method to discount a method strongly without fully eliminating it. There is also an improvement in the rock mechanics ratings because the internationally recognized rock mass rating is used [8]. Mahrous et al. focused on updating the UBC method to apply the technique for order of preference by the similarity to ideal solution (TOPSIS) technique for selecting mining methods [9]. They targeted increasing the burden for every criterion, achieving scoring normalization for every criterion, and estimating the geometric distance between every opportunity and the right opportunity, to achieve the best value for every criterion. In recent years, scientists worldwide have introduced a number of new theories and procedures for selecting underground mining methods, which generally involve gray correlation and decision making using multiple criteria (AHP, FAHP, TOPSIS, PROMETHEE, ELECTRE, and VIKOR). Multiple-criterion decision-making (MCDM) methods have been demonstrated as useful problem-solving tools in various engineering fields [10,11,12].

There is much research on MMS using MCDM methods. Some of these studies did not take into consideration the uncertainty of the parameters. This uncertainty may be conquered with an artificial neural network (ANN). Science and technology have improved, and accessible data are increasing. Meaningful knowledge discovery through big data collection has gained importance. As the demand for significant information increases, the popularity of such data-processing fields as data mining, big data, machine learning, and artificial intelligence has increased.

An artificial neural network (ANN) mimics the mechanisms of the studying and problem-solving capabilities of the human mind. It is flexible, quite parallel, robust, and fault-tolerant [13]. In the implementation of ANNs, knowledge is represented as numeric weights, which are used to gather the relationships within data that are difficult to relate analytically, and this iteratively adjusts the network parameters to minimize the sum of squared approximation errors. Many researchers have benefited from ANNs for fixing mining problems. Adeli and Wu anticipated the financial results of mining activity [14], and Leu et al. [15] modeled the reaction of the support device to stress, while Ambrozic and Turk [16], Hu [17], and Liu and Li [18] used ANNs for alternatives in methane concentration. Khandelwal and Singh [19] and Singh et al. [20] applied them to blasting and its environmental outputs. Cheng et al. optimized an airflow system in an underground mining facility [21]. Ozyurt (2018) stated that ANNs, which are computer programs that offer answers for comparable or specific cases (regardless of the shortage of information) by learning from reason and impact relationships in pattern cases, can overcome the abovementioned problems [22]. Yang and Zhang (1997) and Lv and Zhang (2014) employed ANNs to choose the most suitable mining technique for a mine deposit [23,24]. Ozyurt and Karadogan (2020) advanced six exceptional ANN models that can examine the geometry, rock mass properties, environmental factors, and air conditions of underground mine deposits to determine underground mining techniques and programs that fulfill an underground protection situation. Among the underground mining techniques determined using ANNs, the optimal underground mining technique was decided by means of the ultimatum games, wherein a compromise between protection and monetary situations was simulated. Using a mixture of ANN models and ultimatum games, a new model primarily based on ANNs and game theory for the choice of an underground mining technique was advanced. This model could make predictions despite a loss of facts by means of the following technological traits and new findings received in scientific/sectoral research if learning is continuous. Moreover, the model can examine all choice standards and offer primarily literature-based solutions [25]. Lawal and Musa applied an artificial neural network (ANN)-based mathematical model for the prediction of blast-induced ground vibrations [26]. Lawal et al. (2021) focused on using sine cosine algorithm optimized artificial neural network (SCA-ANN) models for predicting the blast-initiated ground vibration in five granite quarries [27].

In this study, a trainable cascade forward backpropagation network was used. This is because, as soon as a mining approach is selected, extrusion is almost impossible because of the excessive charges and losses entailed. Thus, it is vital to re-examine a choice earlier [28,29]. The method that decision-makers generally use is a sensitivity assessment of the final decision. The mining method requirements influence every mining operation and are important for estimating the capital and operational costs of alternatives in such a manner that economic returns are maximized. Considering the requirements is also important in mine management because of their effect on operational costs, as well as in mine planning and design. Most importantly, using the optimal mining method increases the safety of employees and ensures steady production. [30]. The focal point of this article is to achieve consistency with the proposed solution of TOPSIS with cascade forward backpropagation neural networks.

The remainder of this manuscript is organized as follows. In Section 2, the site investigation and location are described. In Section 3, the methodology, data collection, constraints, problem formulation, raw material requirements, and implementation are presented according to the analysis methods, including the relevant mathematical formalism. The results are presented, and solutions are discussed in relation to the previous methodology in Section 4. Finally, the conclusions are drawn in Section 5.

2. Site Investigation and Data Collection

The geotechnical characterization of the Boleo copper mine (Mexico) was conducted to compare its diverse geological structural functions and depositional environments. The mineral-bearing zones of the study area are bedded clay seams with a moderate dip, known regionally as “mantos”, and an overlying brecciated zone. One of the authors of this article is a group manager of technical service for a Korean corporation in Mexico, and the investigations were conducted between 2017 and 2020.

The three mines at this site are M303, M303S, and M303C. To evaluate the location of Manto 3, step mining was used to attain the ore frame after excavation through its top interburden. Severe abrasions and pillar damage were triggered, while conglomerate and repeated grading added to the lateral strain of the mine. The crack displacements and cement injections were measured. In addition, the water did not penetrate through cracks during the wet season, as shown in Figure 1. For quick wall mining, the primary gateway was excavated within the side of the Manto 3 layer. This mine has panels. One phase of panel SW1 had a width of 80 m and was 2.4 m high. Currently, it is approximately 90 m long, and therefore produces approximately 17,280 m³ of extracted ore.

Table 1. Characterization of rock collected from the mine site.

	Ore Property
General shape	T, Tabular
Ore thickness, m	N, narrow	Description
	15	Magnitude
Ore plunge, degree	70°
Grade distribution	G, Gradational	Description
	G, Gradational	Magnitude
Depth below surface, m	SH, Shallow	Description
	100, m	Magnitude
Ore zone	M, Moderate	Description
	60	Magnitude
Hanging wall	VW, very weak	Description
	35	Magnitude
Foot wall	VW, very weak	Description
	35	Magnitude
Ore zone	VW, very weak
	35
Hanging wall	W, weak
	25
Foot wall	W, weak
	35

lists the properties of the rock within the studied location, consisting of information on ore thickness, shape, ore plunge, grade distribution, depth, and rock mass classification.

3. Methodology

3.1. University of British Columbia (UBC) Method

The UBC method was developed in 1995 by Miller-Tait as a modification of the Nicholas method [4]. The scoring domain of the Nicholas method, which is between the maximum and minimum, was extended. This emphasizes the stopping method rather than mass-mining techniques. This is because it was designed to represent the typical Canadian practice, which is a limitation for its use outside Canada [7]. In addition, the importance of the criteria was not considered [31]. The selection process is similar to that of the Nicholas method. The rankings and characteristics, except for grade distribution and plunge, are different. The ranking in the UBC method ranges from zero to six. Six is given to the characteristics of the most suitable mining method. In addition, –10 was introduced to the method to discount a method strongly without fully eliminating it. There is also an improvement in the rock mechanics ratings because the internationally recognized rock mass rating is used [32]. Rock characterizations, particularly the mechanical values, are also modified. Tomich provided more details, as shown in Figure 2 [4].

The essential properties are categorized qualitatively into specific categories, and weights are assigned in line with the precise usage of mining methods. In this study, the focus was on converting all parameters of the UBC approach for ease of use and to use new strategies without problems and with faster calculation. To avoid the use of poor parameters (minus umber in UBC method) for weight and charge, the poor parameters of the UBC approach were dispensed with. The UBC standards were transformed to weight and charge, and the load relies upon the mining methods (the highest weight is given benefits and lowest given to is risks, similar to the mining methods), and the charge relies on the actual geotechnical geometric and financial elements associated with a minefield investigation, as presented in

Table 2. Assignment of the weights of the various parameters according to mining methods.

	Parameter/Mining Methods		Open Pit	Block Caving	Sublevel Stopping	Sublevel Caving	Longwall	Room and Pillar	Shrinkage Stopping	Cut and Fill Stopping	Top Slicing	Square Set Stopping
1	General shape		0.8	0.8	0.6	0.6	0	0	0	0.2	0.2	0
2	Ore thickness, m		0.8	0.8	0.6	0.8	0	0	0	0	0.2	0
3	Grade distribution		0.6	0.6	0.8	0.6	0.8	0.8	0.6	0.4	0.4	0
4	Depth below surface, m		0.8	0.4	0.6	0.6	0.4	0.6	0.6	0.4	0.4	0.2
5	Plunge		0.2	0.8	0.8	0.8	0	0	0.8	0.8	0	0.4
6	RQD	Ore zone	0.6	0	0.8	0	0.4	0.9	0.6	0.6	0	0
7		Hanging wall	0.8	0.4	0.8	0.4	0.6	0.9	0.8	0.6	0.6	0
8		Foot wall	0.8	0.4	0.6	0.6	0	0	0.6	0.4	0.4	0
9	RSS	Ore zone	0.6	0	0.8	0.4	0.2	0.9	0.8	0.6	0	0
10		Hanging wall	0.8	0	0.6	0.25	0.4	0.9	0.8	0.4	0.4	0
11		Foot wall	0.8	0.2	0.6	0.4	0	0	0.6	0.4	0.25	0

and

Table 3. Ratings assigned to the different ranges of parameters.

Parameter		Rating
		0	0.4	0.6	0.8	1
General shape		Irregular	Platy/tabular with depth not exceeding 35 m	Platy/tabular with depth not exceeding 30 m	Platy/tabular with depth not exceeding 25 m	Equal dimensions
		Very low	Low	Medium	High	Very high
Ore thickness, m		<3	3–10	10–30	30–100	>100
		Very narrow	Narrow	Intermediate	Thick	Very thick
Plunge		>55	45–55	35–45	20–35	<20
		Steep	Semi-steep	Intermediate	Semi-flat	Flat
Depth below surface, m		>600	300–600	150–300	70–150	<70
		Deep	Semi-deep	Intermediate	Semi-shallow	Shallow
Grade distribution, %		<50	50–70	70–80	80–99	100
		Erratic	Ununiform	Medium	Semi-uniform	Uniform
RQD, %	Ore zone	<20	30–40	40–60	60–80	80–100
		Very weak	Weak	Moderate	Strong	Very strong
RSS, %	Ore zone	<5	5–10	10–15	15–20	20
		Very weak	Weak	Moderate	Strong	Very strong

.

3.2. Cascade forward Backpropagation Neural Network (CFBNN)

The cascade forward backpropagation model is similar to feed-forward networks, as shown in Figure 3 [33]; however, two-layer feed-forward networks can be used to study any input–output relationship, and feed-forward networks with extra layers can be used to study complicated relationships more quickly. The cascade forward backpropagation ANN model is similar to a feed-forward backpropagation neural network in the use of the backpropagation algorithm for weight updating. However, a fundamental characteristic of this network is that every layer of neurons is associated with all preceding layers of neurons [34]. In CFBNN, as in other feed-forward networks, there are one or multiple interconnected hidden layers and activation functions. Neurons have their own biases, and the connections have specified weights. In ANN modeling, a set of adjusted weights must be found, such that the error of the model prediction falls to a minimum acceptable level [35].

A CFBNN is similar to a feed-forward backpropagation neural network in the use of the backpropagation algorithm to update weights. However, the fundamental characteristic is that each layer of neurons is interrelated to all preceding layers of neurons [36].

The tan-sigmoid transfer, log-sigmoid transfer, and pure linear threshold features are calculated to optimize the response of the CFBNN, as shown in Figure 4. The mean square error (MSE), which is given in Equation (1), the root means square error (RMSE) in Equation (2), and R² are calculated by Equation (3) to indicate the success of the algorithms.

$$MSE=\left[{\displaystyle \sum }_1^n{((Qexp-Qcal)|n)}^2\right]$$

(1)

$$RMSE=\sqrt{\frac{1}{2}{\displaystyle \sum }_1^n[{((Qexp-Qcal)|n)}^2}]$$

(2)

$$R^2=\left[{\displaystyle \sum }_1^n{((Qexp-Qcal)|n)}^2\right]$$

(3)

where Qexp is the value of the measurement, Qcal is the calculated value, and n denotes the observation number.

Generally, the cascade-forward back-propagation network models are similar to feed-forward networks, but they include a weight connection from the input to each layer and from each layer to the successive layers. Filik and Kurban [37] found that the cascade-forward backpropagation method can be more efficient than the feed-forward backpropagation method in some cases.

$$f\left(ne{t}_j\right)=\frac{1}{\left(1+e^{-netj }\right)}$$

(4)

$$f(ne{t}_j)=net$$

(5)

The schematic of a trainable cascade-forward back-propagation is used to predict the yield of the selected mining methods. As mentioned above, the training stage is a crucial stage. Generally, backpropagation and quick-propagation training methods are the most widely used methods. Therefore, in the present study, the backpropagation method was applied in the training stage according to the approach of Lashkarbolooki et al. and its unique advantages. In more detail, the Levenberg–Marquardt backpropagation method was used because it is one of the quickest and most accurate training methods. Therefore, the proposed ANN model was trained using the Levenberg–Marquardt algorithm [38]. In the next stage, the number of hidden layers was optimized. Only one hidden layer was considered for the proposed ANN model because Cybenko [39] reported that a network with only one hidden layer can approximate almost any type of nonlinear relation. The second important parameter is the selection of the optimal number of neurons in the hidden layer. The selection of the number of neurons in the hidden layer is of great concern because having a small number of neurons produces a network with low precision, and a higher number leads to overfitting and poor quality of interpolation, because as the number of neurons increases, the risk of overtraining increases [40].

3.3. Using Excel for Selecting of Mining Methods via ANN

This work describes the use of Microsoft Excel for the sequential solving of mining method selection problems according to the cascade-forward backpropagation model. Solver is a much simpler, intuitive, and easily available tool for MCDM managers who usually learn how to use it in the first years of their degree in informatics or similar subjects. Thus, it is not necessary to spend extra time in teaching how to use the software to save time that can be devoted to other tasks. The following steps are used for converting the proposed method into an Excel sheet.

Step 1: Weight (W) and bias (b) values of the ANN, which are shown in Figure 4, are calculated via trainable cascade-forward backpropagation using the MATLAB toolbox and then placed in an Excel sheet, as shown in

Table 4. Weight (W) and bias (b) values according to MATLAB toolbox.

Properties/Input	w1	w21 w2	b1	b2
General shape	5.4438	3.4113	8.803278	6.128579
Ore thickness	12.4184	0.0339
Grade distribution	−12.5975	−5.4184
Depth below surface	−0.1821	−1.4329
plunge	11.1132	0.301
RQD	6.8993	2.6115
RSS	−22.7358	0.7069
		−2.78847

.

Step 2: The output of the first layer (K) in the Excel sheet function is calculated, using Equation (6), and all the results will be presented in the following section.

K1 = 2/(1 + EXP(−2 × (SUM(O1) + b1))) − 1

(6)

Step 3: The output of the second layer (rank of selection method) in the Excel sheet is calculated using Equation (7).

R = SUM (Column O2) + K × W21W1 + b2

(7)

Step 4: The name of the method is selected, and it is viewed in an Excel sheet by converting approximation values of rank into integer values using V function.

In this report, the use of Microsoft Excel and a cascade-forward backpropagation ANN for the sequential solving of mining method selection problems according to the cascade-forward backpropagation model is described. Solver is a much simpler, intuitive, and widely available tool for MCDM managers, who usually learn how to use it in the first years of their degree in informatics or similar subjects. Thus, it is not necessary to spend extra time in the teaching of the software—time that can be devoted to other tasks.

4. Results and Discussion

4.1. Results

The TOPSIS technique has been employed as a multi-criterion decision-making method for sustainable selection among different mining methods, considering the layer specifications and its properties. In this study, according to weight values estimated based on questionnaires and interviews with experts, the weights for the material properties that impact the mining method performance were determined. The UBC method was modified by considering various economic, technical, sociocultural, and environmental factors. Each of these factors was weighted according to its impact on the application and purpose of the study.

Table 5. Criteria for conversion to new approach based on weight and rate.

Weights/Rates	0.1	0.4	0.6	0.8	1	0.3	0.4
	General Shape	Ore Thickness, m	Grade Distribution	Depth Below surface, m	Plunge	RQD	RSS
Open pit mining	0.4	0.8	0.6	0.6	0.6	0.1	0.1
Block caving	0.8	0.8	0.6	0.8	0.1	0.1	0.1
Sublevel stopping	0.6	0.6	0.8	0.6	0.8	0.8	0.6
Sublevel caving	0.8	0.4	0.6	0.6	0.4	0.6	0.6
Longwall	0.2	0.8	0.8	0.8	0.1	0.1	0.8
Room and pillar	0.6	0.1	0.8	0.1	0.4	0.9	0.6
Shrinkage stopping	0.1	0.4	0.8	0.4	0.6	0.9	0.8
Cut and fill stopping	0.8	0.4	0.6	0.6	0.1	0.1	0.6
Top slicing	0.6	0.1	0.8	0.4	0.2	0.9	0.8
Square set stopping	0.8	0.1	0.6	0.25	0.4	0.9	0.8

presents the criteria for conversion of the UBC criteria to the new approach, whereby all properties were weighted to approximately 1.

Table 6. Normalized matrix.

	General Shape	Ore Thickness, m	Grade Distribution	Depth Below Surface, m	Plunge	RQD	RSS
Open pit mining	0.203858	0.478947	0.26832	0.34009	0.434144	0.048336	0.0498
Block caving	0.407717	0.478947	0.26832	0.45346	0.072357	0.048336	0.0498
Sublevel stopping	0.305788	0.359210	0.35777	0.34009	0.578859	0.386694	0.2992
Sublevel caving	0.407717	0.2394737	0.26832	0.34009	0.289429	0.290020	0.2995
Longwall	0.101929	0.4789474	0.35777	0.45346	0.072357	0.048336	0.3990
Room and pillar	0.305788	0.0598684	0.35777	0.05668	0.289429	0.435031	0.2992
Shrinkage stopping	0.057259	0.3255153	0.38669	0.27521	0.483493	0.436051	0.4
Cut and fill stopping	0.40771775	0.23947374	0.26832	0.34009	0.0723574	0.0483368	0.29925
Top slicing	0.30578831	0.05986843	0.35777	0.22673	0.1447149	0.4350314	0.39900
Square set stopping	0.4077177	0.05986843	0.2683282	0.14171	0.28942984	0.4350314	0.39900

presents the normalized matrix, which was calculated using Equation (1).

Table 7. Normalized matrix multiplied by rate for every property.

	General Shape	Ore Thickness, m	Grade Distribution	Depth Below Surface, m	Plunge	RQD	RSS
Open pit mining	0.02039	0.19158	0.161	0.27	0.4341	0.0145	0.02
Block caving	0.04077	0.19158	0.161	0.36	0.0724	0.0145	0.02
Sublevel stopping	0.03058	0.14368	0.2147	0.27	0.5789	0.116	0.1197
Sublevel caving	0.04077	0.09579	0.161	0.27	0.2894	0.087	0.1197
Longwall	0.01019	0.19158	0.2147	0.36	0.0724	0.0145	0.1596
Room and pillar	0.03058	0.02395	0.2147	0.05	0.2894	0.1305	0.1197
Shrinkage stopping	0.00573	0.13021	0.232	0.22	0.4835	0.1308	0.16
Cut and fill stopping	0.04077	0.09579	0.161	0.27	0.0724	0.0145	0.1197
Top slicing	0.03058	0.02395	0.2147	0.18	0.1447	0.1305	0.1596
Square set stopping	0.04077	0.02395	0.161	0.11	0.2894	0.1305	0.1596

summarizes the normalized matrix multiplied by the rate for each property.

Table 8. Positive ideal and negative ideal solutions.

V+	0.00573	0.02395	0.161	0.05	0.0724	0.0145	0.02
V−	0.04077	0.19158	0.232	0.36	0.5789	0.1308	0.16

illustrates the positive and negative ideal solutions, and

Table 9. Euclidean distance from the ideal and worst solutions and the ranking.

Si+	Si−	Pi	Rank
0.4589	0.2603	0.36	8	Open pit mining
0.3607	0.5429	0.60	4	Block caving
0.5883	0.113	0.16	10	Sublevel stopping
0.3466	0.3313	0.49	7	Sublevel caving
0.3889	0.5209	0.57	6	longwall
0.2721	0.4633	0.63	3	room and pillar
0.4991	0.1855	0.27	9	Shrinkage stopping
0.2603	0.6	0.70	1	Cut and fill stopping
0.2454	0.4999	0.67	2	Top slicing
0.2931	0.4232	0.59	5	Square set stopping

presents the results according to the Euclidean distance from the ideal worst and ranking.

4.2. Discussion

According to the above results which reflect the integration between the method requirements and the actual layer set, the following layer properties were used: a plunge of 70°, a depth below the surface of 100 m, and rock quality designations of 60 (moderate) and 35 in the hanging wall (very weak). The suitable method for all previous layer specifications led to cut and fill stopping being much better than other methods. Figure 5 illustrates the main design for the cut and fill stopping mining method, which is considered as the optimal mining method.

An example of using Microsoft Excel for the sequential solving of mining method selection problems according to the cascade-forward backpropagation model, including the output of the proposed method, is shown in

Table 10. Example for estimation MMS via CFBNN using Excel spreadsheet.

Input	w1	w21 w2	b1	b2	O1	O2
0.8	5.44	3.41	8.8	6.13	4.4	2.7					N	Method Name
0.4	12.4	0.0339			4.9	0.01					1	Open pit mining
0.6	−12.59	−5.4184			−7.6	−3.25		Method 8			2	Block caving
0.6	−0.182	−1.4329			−0.11	−0.85		Cut and fill stopping			3	Sublevel stopping
0.1	11.11	0.301			1.11	0.03					4	Sublevel caving
0.11	6.899	2.6115			0.7589	0.287					5	Longwall
0.6	−22.73	0.7069			−13.6	0.424					6	Room and Pillar
		−2.78847			−1.313	7.9142					7	Shrinkage stopping
					−0.865	2.729					8	Cut and fill stopping
											9	Top slicing
											10	Square set stopping

. The results provide indicators which decision makers can use to choose between different mining methods based on the total points given to all ore properties.

Figure 6 shows that the rank of the mining method of the CFBNN is identical in order to evaluate the errors, and then the values of rank Max error less than 1%. The selected method given by Excel via CFBNN is identical to the base selected method, as shown in Table 10.

5. Conclusions

Selecting a mining method depends on many criteria, all of which are related to safety and economic considerations. The modification of the UBC method to TOPSIS focused on linking all parameters related to all criteria in a simple manner and obtaining accurate results.

• The CFBNN method was applied in this research to select suitable mining methods under different conditions. Comparing the two methods showed that the results were identical. In addition, the CFBNN method is more accurate and easier to apply than TOPSIS.
• The results provide indicators which decision makers can use to choose between different mining methods based on the total points given to all ore properties.
• The best mining method is cut and fill stopping, with a rank equal to 0.70, and the second-best mining method is top slicing, with a rank of 0.67. The pattern continues, as shown in Table 8. This modified method was applied to other cases, and good results were obtained; furthermore, it is easy to input and output all data and solutions.