Strip mine slope stress prediction method and system based on VMD-DBO optimized GRU

Abstract

The invention discloses a strip mine slope stress prediction method and system based on VMD-DBO optimized GRU, and relates to the technical field of mine engineering safety monitoring. According to the method, VMD (variational mode decomposition), a dung beetle optimization algorithm (DBO) and an improved gating cycle unit (GRU-A) are combined, and decomposition and noise reduction are performed on stress data through the VMD so as to extract key features; dBO is utilized to optimize and improve hyper-parameters of a gating cycle unit GRU-A, and the learning ability of the model is enhanced; and in combination with a self-attention mechanism, key features of the stress time sequence are dynamically captured. According to the strip mine slope stress prediction method and system based on VMD-DBO optimization GRU, the prediction result is more accurate compared with a traditional method, and an efficient and reliable solution is provided for surface mine slope stress prediction and landslide disaster early warning.

Description

Technical Field

[0001] The present invention belongs to the technical field of mine engineering safety monitoring, and particularly relates to an open-pit mine slope stress prediction method and system based on VMD-DBO optimized GRU. Background Art

[0002] The prediction of the stability of open-pit mine slopes is of great significance in mining safety monitoring, and its accuracy directly affects the production progress and the safety of personnel and property. With the continuous expansion of open-pit mine development activities and the increasing demand for fuel resources, the disaster risks of open-pit coal mines are also continuously intensifying. The stability of open-pit mine slopes is affected by various complex factors, including slope displacement, rainfall, stress, earthquakes, and artificial mining activities, etc. Its evolution process usually shows significant non-linear characteristics and is further complicated by the dynamic changes of human activities and triggering factors.

[0003] Traditional slope prediction methods mainly rely on mathematical modeling techniques, such as numerical simulation methods and quantitative analysis based on geological parameters. These methods can meticulously simulate the evolution process of the stress field, but generally have problems such as high requirements for geological parameters, large calculation costs, and strong sensitivity of results to parameters. For example, landslide monitoring based on synthetic aperture radar (InSAR) data can identify the slope acceleration stage, quantify the deformation rate and time series changes before landslides, but it is difficult to capture all landslide precursors due to limitations in image quality and time resolution. Nie et al. proposed a medium- and short-term prediction method based on the polynomial (MsTPLP) model for slope monitoring of Fushun West Open-pit Mine, while Wang et al. improved the landslide susceptibility assessment model and verified its prediction ability in open-pit mines by introducing the rock quality designation (RQD) and numerical simulation (NS), combined with the gradient boosting decision tree (GBDT), but its dependence on factor data and high-precision monitoring equipment limits its wider application. In addition, Li et al. used MatDEM to model and simulate mining-induced rock mass fractures, revealing the evolution law of overlying rock mass fractures after the transition from surface mining to underground mining, and evaluating the risks of open-pit mine slopes. Ding et al. proposed a landslide monitoring and warning model based on ground-based real aperture radar (RAR).

[0004] Although the above methods perform well in specific scenarios, due to the complexity and diversity of the influencing factors of open-pit mine slope stability, these methods are often based on certain assumptions or linear relationships and do not fully consider the non-linear characteristics and uncertain factors in the slope instability process. In recent years, machine learning (ML) has gradually become an important tool for open-pit mine slope stability analysis because of its advantage in reflecting complex non-linear relationships. Mohd Maneeb Masood et al. developed an automated algorithm based on monitored displacement data for predicting slope failure; Song et al. used the gradient boosting regression tree (GBRT) model to predict the factor of safety (FOS) of slopes under heavy rainfall conditions. To further improve the prediction performance of traditional models, some studies have introduced optimization algorithms and multi-factor coupling models. For example, Bui et al. improved the M5Rules-GA model through the genetic algorithm (GA), improving the accuracy of slope stability prediction. Ruan et al. proposed an intelligent crack detection algorithm based on improved Mask R-CNN, emphasizing the key role of crack detection in open-pit mine safety management; Lin et al. constructed an ML model for slope stability assessment, meeting the requirements of high precision and rapidity in slope engineering; Li et al. used ensemble learning techniques for slope stability prediction driven by multi-source monitoring data; Du and Chen used support vector machines (SVM) to construct a slope monitoring model, further improving the accuracy of slope stability assessment.

[0005] However, existing research has mostly focused on the prediction of variables such as slope displacement and rainfall, while the research on slope stress prediction is relatively scarce. As an important influencing factor of landslide disasters, the change of slope stress can effectively characterize the trend of landslide occurrence. Liu et al. analyzed the relationship between stress and resistance through the limit equilibrium method and Morgenstern method and calculated the stability of potential slip surfaces, indicating that stress prediction is of great significance for landslide monitoring and early warning. However, the current prediction methods for slope stress still have great room for improvement in terms of model accuracy, robustness, and generalization ability. Summary of the Invention

[0006] Aiming at the deficiencies of the existing technology, the present invention proposes an open-pit mine slope stress prediction method and system based on VMD-DBO optimized GRU, which combines variational mode decomposition (VMD), dung beetle optimization algorithm (DBO), and improved gated recurrent unit (GRU-A) to solve the problems of scarcity and inaccuracy in the existing prediction of slope stress.

[0007] The first aspect of the present invention provides an open-pit mine slope stress prediction method based on VMD-DBO optimized GRU, including the following steps:

[0008] Step 1: Construct a basic stress database for open-pit mine slopes;

[0009] The open-pit mine slope foundation stress database includes several stress history data, and the stress history data is a time series, including the open-pit mine slope stress data at different time points;

[0010] Step 2: Preprocess the stress history data in the open-pit mine slope foundation stress database to obtain the preprocessed stress history data; the preprocessing is missing value imputation;

[0011] Step 3: For each preprocessed stress history data, use the variational mode decomposition method to decompose it to obtain several intrinsic mode function components and normalize them, and then obtain several normalized intrinsic mode function components corresponding to each stress history data;

[0012] Step 4: According to the normalized intrinsic mode function components obtained in Step 3, construct a stress dataset and divide it according to a set ratio to obtain a training set and a test set; the stress dataset includes several samples, and each sample includes several normalized intrinsic mode function components corresponding to a stress history data;

[0013] Step 5: Construct a stress prediction model; the stress prediction model is an improved gated recurrent unit GRU-A;

[0014] The input of the improved gated recurrent unit GRU-A is a normalized intrinsic mode function component, and the output is the prediction result of the intrinsic mode function component. The improved gated recurrent unit GRU-A adds a self-attention mechanism layer between the hidden layer and the fully connected layer of the gated recurrent unit GRU. The self-attention mechanism layer uses the self-attention mechanism to weight the time series data output by the hidden layer to generate a context vector;

[0015] Step 6: Use the dung beetle optimization algorithm to determine the hyperparameters of the stress prediction model;

[0016] Step 6.1: Initialize the dung beetle swarm;

[0017] The dung beetle swarm includes several dung beetle individuals, and all dung beetle individuals are divided into four roles according to a ratio, namely the ball-rolling dung beetle, the brood ball, the juvenile dung beetle, and the stealing dung beetle. Each dung beetle individual represents a set of hyperparameter combinations of the stress prediction model;

[0018] Step 6.2: Set the maximum number of iterations;

[0019] Step 6.3: Set the current iteration number q to 0;

[0020] Step 6.4: Calculate the fitness values of all dung beetle individuals, and take the dung beetle individual with the minimum fitness value as the current optimal solution; the fitness value is the error index after a set of hyperparameters corresponding to the dung beetle individual are applied to the improved gated recurrent unit GRU-A.

[0021] Step 6.5: Determine whether the position of the dung beetle individual exceeds the set optimization range. If it exceeds, adjust the position of the dung beetle individual to within the optimization range.

[0022] Step 6.6: Update the position of the dung beetle individual according to its role.

[0023] Specifically: The ball-rolling dung beetle uses the ball-rolling strategy to update the position of the dung beetle individual; the brood ball uses the reproduction strategy to update the position of the dung beetle individual; the juvenile dung beetle uses the foraging strategy to update the position of the dung beetle individual; the kleptoparasitic dung beetle uses the stealing strategy to update the position of the dung beetle individual.

[0024] The ball-rolling strategy is as follows: Obtain a random number rand, and determine whether the random number is less than the set threshold. If so, execute the obstacle-free mode to update the position of the dung beetle individual, otherwise execute the obstacle mode to update the position of the dung beetle individual; the value range of the random number is from 0 to 1.

[0025] The method for updating the position of the dung beetle individual in the obstacle-free mode is:

[0026] x l (q + 1)=x l (q)+α×m×x l (q - 1)+b×Δx (15)

[0027] Δx = |x l (q)-X w | (16)

[0028] In the formula, x l (q) represents the position of the l-th dung beetle individual at the q-th iteration, x l (q + 1) represents the updated position of the dung beetle individual, x l (q - 1) represents the position of the l-th dung beetle individual at the (q - 1)-th iteration, α is the natural coefficient, and its value is -1 or 1; m is the deflection coefficient with a value of (0.0, 0.2), b is a constant of (0, 1), Δx is used to simulate the change in light intensity, and X w represents the global worst position.

[0029] The method for updating the position of the dung beetle individual in the obstacle mode is:

[0030] x l (q + 1)=x l (q)+tan(θ)|x l (q)-xl (q - 1)| (17)

[0031] In the formula, θ represents the deflection angle, and its value range is [0, π]. When θ is equal to 0, or π, the position of the dung beetle individual is not updated;

[0032] The reproduction strategy is as follows:

[0033] First, define the spawning area:

[0034] Lb * = max(X * × (1 - R), Lb) (18)

[0035] Ub * = min(X * × (1 + R), Ub) (19)

[0036] In the formula, X * represents the current optimal position, Lb * and Ub * are the lower and upper limits of the spawning area respectively. Define R = 1 - q / T max , where T max is the maximum number of iterations, and Lb and Ub represent the lower and upper limits in the optimization problem respectively;

[0037] The method for updating the position of the brood ball according to the spawning area is as follows:

[0038] B l (q + 1)= X * + b1×(B l (q)- Lb * )+ b2×(B l (q)- Ub * ) (20)

[0039] In the formula, B l (q) represents the position of the l-th brood ball in the q-th iteration, and B l (q + 1) represents the updated position of the brood ball. b1 and b2 are irrelevant random vectors representing the dimensions of the optimization problem;

[0040] The foraging strategy is as follows:

[0041] Define the best foraging area:

[0042] Lb b = max(X b × (1 - R), Lb) (21)

[0043] Ub b = min(X b×(1 + R), Ub) (22)

[0044] In the formula, X b represents the global optimal position, Lb b and Ub b represent the lower and upper limits of the best foraging area respectively;

[0045] Update the position of the juvenile dung beetle according to the best foraging area:

[0046] x l (q + 1) = x l (q) + C1 × (x l (q) - Lb b ) + C2 × (x l (q) - Ub b ) (23)

[0047] where C1 is a random number subject to a normal distribution, and C2 is a random vector belonging to the interval (0, 1);

[0048] The theft strategy is as follows:

[0049] The method for updating the position of the thieving dung beetle is:

[0050] x l (q + 1) = X b + S × g × |x l (q) - X * | + |x l (q) - X b | (24)

[0051] In the formula, g is a random vector of size 1 × D, subject to a normal distribution, D is the space dimension, and S is a constant value;

[0052] Step 6.7: Increment the current iteration count by 1;

[0053] Step 6.8: Calculate the fitness value of each dung beetle individual, and take the dung beetle individual with the minimum fitness value as the current optimal solution. Compare the current optimal solution with the optimal solution of the previous iteration, and take the one with the smaller fitness value as the global optimal solution;

[0054] Step 6.9: Determine whether the current iteration count has reached the maximum iteration count. If so, output the current global optimal solution and its fitness value; otherwise, return to Step 6.5;

[0055] Step 7: Apply the hyperparameters of the stress prediction model determined in Step 6 to the stress prediction model, and train it using the training set to obtain the trained stress prediction model;

[0056] Step 8: Input the test set into the trained stress prediction model to obtain the prediction results of the stress;

[0057] Specifically, input the normalized intrinsic mode function components corresponding to a stress history data into the trained stress prediction model respectively to obtain the prediction results of each intrinsic mode function component, and add the prediction results of the intrinsic mode function components according to the time points to obtain the prediction results of the stress;

[0058] Step 9: Denormalize the prediction results of the stress to obtain the final prediction results of the stress.

[0059] The second aspect of the present invention provides an open-pit mine slope stress prediction system based on VMD-DBO optimized GRU for implementing the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU, including:

[0060] A stress history data acquisition module for acquiring stress history data;

[0061] A data preprocessing module for preprocessing the stress history data to obtain the preprocessed stress history data;

[0062] A variational mode decomposition and normalization module for decomposing the preprocessed stress history data by using the variational mode decomposition method to obtain a number of intrinsic mode function components and normalizing them, so as to obtain a number of normalized intrinsic mode function components corresponding to each stress history data;

[0063] A dung beetle optimization module for determining the hyperparameters of the stress prediction model;

[0064] A stress prediction model, which is an improved gated recurrent unit GRU-A, applies the hyperparameters determined by the dung beetle optimization module, inputs a normalized intrinsic mode function component, and outputs the prediction results of the intrinsic mode function component;

[0065] A summation module for adding the prediction results of all the obtained intrinsic mode function components according to the time points to obtain the prediction results of the stress;

[0066] A denormalization module for denormalizing the prediction results of the stress to obtain the final prediction results of the stress.

[0067] The third aspect of the present invention provides an electronic device, including: a processor, a memory and a bus, the memory stores machine-readable instructions executable by the processor, when the electronic device runs, the processor communicates with the memory through the bus, and when the machine-readable instructions are executed by the processor, the steps of the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU are executed.

[0068] In the fourth aspect of the present invention, a computer-readable storage medium is provided, in which a computer program is stored, and when the computer program is run by a processor, it executes the steps of the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU as described above.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] The present invention combines variational mode decomposition (VMD), dung beetle optimization algorithm (DBO) and improved gated recurrent unit (GRU-A). Through VMD, stress data is decomposed and denoised to extract key features; the DBO is used to optimize the hyperparameters of the improved gated recurrent unit GRU-A to enhance the learning ability of the model; and the self-attention mechanism is combined to dynamically capture the key features of the stress time series. The prediction results of the open-pit mine slope stress prediction method and system based on VMD-DBO optimized GRU proposed by the present invention are more accurate than traditional methods, providing an efficient and reliable solution for open-pit mine slope stress prediction and landslide disaster warning. Description of the Drawings

[0071] Figure 1 It is a flowchart of the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU in an embodiment of the present invention;

[0072] Figure 2 It is a schematic diagram of slope stress data acquisition in an embodiment of the present invention;

[0073] Figure 3 It is the VMD decomposition waveform in an embodiment of the present invention;

[0074] Figure 4 It is a diagram of the division of the training set and the data set in an embodiment of the present invention;

[0075] Figure 5 It is a structural diagram of the improved gated recurrent unit GRU-A in an embodiment of the present invention;

[0076] Figure 6 It is a structural diagram of the GRU unit in an embodiment of the present invention;

[0077] Figure 7 It is a schematic diagram of the self-attention mechanism in an embodiment of the present invention;

[0078] Figure 8 It is an iterative effect diagram of the DBO algorithm in an embodiment of the present invention;

[0079] Figure 9 It is a diagram of the true value and predicted value results in an embodiment of the present invention;

[0080] Figure 10The result graph of the true value and predicted value of the prediction model in the embodiment of the present invention;

[0081] Figure 11 The comparison between the true value and predicted value of the ablation model in the embodiment of the present invention. Specific implementation manners

[0082] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0083] The open-pit mine slope stress prediction method and system based on VMD-DBO optimized GRU, as Figure 1 shown, include the following specific steps:

[0084] Step 1: Construct a basic stress database for the open-pit mine slope to provide high-quality data support for subsequent model construction;

[0085] The basic stress database for the open-pit mine slope includes several stress historical data, and the stress historical data is a time series, including the stress data of the open-pit mine slope at different time points, showing certain trend changes and volatility characteristics, and at the same time including long-term trends, short-term fluctuations, and high-frequency noise;

[0086] Before a landslide occurs in a rock mass, the internal stress will change prior to the landslide. As the stress accumulates, when the downward sliding force of the sliding mass along the sliding surface is greater than the anti-sliding force, the sliding mass will slide along the sliding surface, ultimately triggering a landslide accident. The slope stress data acquisition system used in this embodiment consists of three parts: an anchoring system, a stress monitoring and communication system, and an information processing system, as Figure 2 shown. The slope stress data acquisition system first collects the measurement data of the vibrating wire sensor at the top of the anchor cable, and transmits the data to the monitoring center in a wireless transmission manner. The monitoring center receives and stores the monitoring data. The slope stress data acquisition system adopts an active monitoring method, making up for the deficiencies of GPS displacement monitoring, and having the characteristics of being forward-looking, accurate, efficient, and not affected by weather.

[0087] The data acquisition location is an open-pit mine in Northeast China. The monitoring data of the stress acquisition device with node number 8 was mainly collected. This device recorded the stress changes from February 4, 2010 to November 25, 2015, with a total of 9,821 data. Since the data was collected at multiple random time points every day, the data distribution was uneven. To solve this problem, the data was processed, and the average stress of each day was calculated to form a new data set, with a total of 1,976 records.

[0088] Step 2: Preprocess the stress historical data in the basic stress database for the open-pit mine slope to obtain the preprocessed stress historical data; the preprocessing is missing value imputation;

[0089] Step 3: For each piece of preprocessed stress history data, use the variational mode decomposition method for decomposition to obtain several intrinsic mode function (IMF) components and perform normalization, thereby obtaining several normalized intrinsic mode function (IMF) components corresponding to each stress history data;

[0090] To better analyze the stress history data, the variational mode decomposition (VMD) method is used to process several stress history data of open-pit mines after missing value imputation. VMD decomposes the original stress history data into various components such as long-term trend mode, periodic mode, and high-frequency mode by constructing a variational optimization problem, thereby providing clearer input features for the prediction model; avoiding the common mode mixing problem in EMD / EEMD and enhancing the ability to suppress high-frequency noise. In addition, the non-recursive decomposition method of VMD ensures high computational efficiency and decomposition accuracy, and can provide higher-quality input data for subsequent modeling.

[0091] VMD is a Wiener filter-based adaptive signal decomposition method that decomposes a signal into a specified number of intrinsic mode function (IMF) components based on the central frequency by solving a variational problem while minimizing the bandwidth. The variational problem is defined by formulas (1) and (2):

[0092]

[0093] where u k is the intrinsic mode function (IMF) component obtained by decomposition, k is the index, w k is the central frequency of the intrinsic mode function (IMF) component, t represents the time variable, represents the partial derivative operation with respect to the time variable t, δ(t) is the unit impulse signal, j is the imaginary unit, and f(t) represents the original stress history data;

[0094] By adding a quadratic penalty term α and a Lagrange multiplier λ to the constrained variational problem, this problem can be transformed into an unconstrained form. To achieve this transformation, an extended Lagrangian method is introduced, so that the constraint conditions can be handled by introducing the penalty term and the Lagrange multiplier. The resulting augmented Lagrangian function expression is:

[0095]

[0096] where L({u k},{w k},λ) is the augmented Lagrangian function, K is the number of intrinsic mode function (IMF) components, α is the quadratic penalty term, and λ is the Lagrange multiplier;

[0097] Using the alternating direction multiplier method, by successively updating and to solve for the optimal solution in Equation (1):

[0098]

[0099] wherein, represents the intrinsic mode function (IMF) component after the (n + 1)-th step update and is evaluated according to the central frequency, w represents the frequency (or a related variable in the frequency domain), especially the frequency component, x(w) is a function representing the stress value or other characteristics related to the variable w, represents the result of the frequency-domain transformation of the signal after Fourier transform, i represents the index, u i (w) is the stress value or predicted value related to the index i, representing the stress of different nodes or variables, represents u i (w) in the frequency domain, is the representation of λ(w) in the frequency domain, represents w after the (n + 1)-th step update k , represents the frequency domain of the k-th intrinsic mode (IMF) component, represents λ(w) after the (n + 1)-th step update, τ represents a scalar step size or learning rate, controlling the update amplitude of the parameters in each iteration, represents the representation of the original signal f(t) in the frequency domain, represents the Hilbert transform;

[0100] After iterating until the stop condition is met, the variational mode decomposition process ends, and K IMF components are obtained;

[0101] The stop condition for variational mode decomposition (VMD) is:

[0102]

[0103] wherein, ε is a set threshold. When the sum of the changes of all modal components is less than this threshold, it indicates convergence, and the decomposition process can stop;

[0104] In this embodiment, when performing VMD decomposition, the selectable parameters are as follows: The selection of the modulus K is determined by the central frequency of the corresponding spectrum. When the value of K is small, under-decomposition may occur, and the main information cannot be fully decomposed into IMF components; while when the value of K is large, over-decomposition is likely to occur, that is, multiple IMF components may have similar central frequencies. The bandwidth constraint parameter (quadratic penalty term) α has an important impact on the bandwidth of the IMF. When the value of α is small, modal aliasing is likely to occur, and when the value of α is too large, effective decomposition may not be possible. In addition, the fidelity coefficient τ and the convergence stop condition ε are also key parameters, and their values are shown in Table 1.

[0105] Values of VMD parameters in Table 1

[0106]

[0107] The original stress history data of the open-pit mine slope contains obvious fluctuations and non-stationary characteristics, and there are also high-frequency noises and complex trend components. Directly analyzing this data may be affected by noises and multimodal characteristics, making it difficult to accurately extract specific features from it.

[0108] Through VMD decomposition, the original stress history data is decomposed into 6 Intrinsic Mode Function (IMF) components (IMF1 to IMF6), corresponding to characteristics in different frequency ranges respectively, such as Figure 3 shown, IMF1 mainly contains high-frequency components, extracting the rapid fluctuations and noises in the original data; IMF2 and IMF3 capture medium-high frequency and medium-frequency characteristics respectively, reflecting the secondary fluctuations and local changes in the data; IMF4 gradually transitions to low frequency, showing the medium- and long-term trend fluctuation characteristics; IMF5 and IMF6 mainly contain low-frequency components, among which IMF6 is highly consistent with the overall trend change of the original data, reflecting the long-term stable trend pattern. This multi-level decomposition effectively separates the high-frequency noises, local fluctuations and long-term trends of the original data, providing clearer characteristic components for subsequent analysis and modeling.

[0109] Step 4: According to the normalized Intrinsic Mode Function (IMF) components obtained in Step 3, construct a stress data set and divide it according to a set ratio to obtain a training set and a test set;

[0110] The stress data set includes several samples, and each sample includes several normalized Intrinsic Mode Function (IMF) components corresponding to a stress history data;

[0111] In this embodiment, the ratio of the training set to the test set is 8:2, and the specific data is as Figure 4 shown;

[0112] Step 5: Construct a stress prediction model; the stress prediction model is an improved Gated Recurrent Unit GRU-A;

[0113] As Figure 5 shown, the input of the improved Gated Recurrent Unit GRU-A is a normalized Intrinsic Mode Function component, and the output is the prediction result of this Intrinsic Mode Function component. The improved Gated Recurrent Unit GRU-A adds a self-attention mechanism layer between the hidden layer and the fully connected layer of the Gated Recurrent Unit GRU. The self-attention mechanism layer uses the self-attention mechanism to weight the sequential data output by the hidden layer to obtain a weight distribution focusing on key features;

[0114] AsFigure 6 As shown in Figure 6 , the gated recurrent unit (GRU) is an improved recurrent neural network (RNN) model based on the long short-term memory network (LSTM), which is used to solve the problems of traditional RNNs in processing long sequences and long-term dependencies. Compared with LSTM, GRU improves the gating structure, combines the input gate and forget gate in LSTM into an update gate, and changes the output gate to a reset gate, enabling it to capture dependencies in time series and extract their inherent features, further reducing the complexity of the model, reducing a certain number of training parameters, and improving the model running efficiency;

[0115] The calculation method of the hidden state of GRU at time t is as follows:

[0116] z t = σ(W z X t + U z h t-1 ) (8)

[0117] r t = σ(W r X t + U r h t-1 ) (9)

[0118]

[0119] where h t is the final hidden state, h t-1 is the hidden state at the previous moment, z t and r t are the update gate and reset gate respectively, which determine the update degree of h t and the retention degree of h t-1 , is obtained by fusing h t and h t-1 according to the weight of the update gate. σ is the sigmoid function, ⊙ is the element-wise multiplication, W z and U z are the weight matrices of the update gate network state, W r and U r are the weight matrices of the update gate network state, W h and U h are the weight matrices of the hidden layer network state, X t is the current input;

[0120] The self-attention mechanism obtains the weighted weights of the input features through a learning process, and then realizes the weighted average of multiple inputs, thereby reflecting the attention degree of the model to different inputs in the time series. Its core goal is to extract key information from the input features to help the model capture and describe the dynamic features in the data stream. The key advantage of the self-attention mechanism is its ability to achieve selective attention for specific features, thus effectively alleviating the bottleneck problem in the information processing process. In addition, by establishing a direct connection between the encoder (before the self-attention mechanism) and the decoder (after the attention mechanism), the self-attention mechanism significantly alleviates the vanishing gradient phenomenon. Figure 7 It shows the structure and working process of the self-attention mechanism, including the processing of input data, the calculation and distribution of attention scores, and the generation process of the final output. Specifically, the self-attention mechanism first receives the input data, extracts important features and focuses on key information, and finally generates the required output result.

[0121] The self-attention mechanism consists of a series of steps, including calculating attention scores, priority weights, and the distribution of context vectors. These steps are carried out in sequence: first, calculate the attention scores, then determine the weights, and finally generate the context vector. Through the encoder hidden state h x and the previous decoder output s g-1 to calculate the score e g,x , which measures the degree of fit between the components of the input sequence and the current output at position g. The specific calculation is shown in Equation (12). The attention score is obtained through the function a(.), which is usually implemented by a feed-forward neural network. Then, by applying the softmax operator to the attention scores (as shown in Equation (13)), the weights a g,x can be obtained (as shown in Equation (13)). The context vector c g is unique and is obtained by summing the products of all weights and the corresponding encoder hidden states, and is input into the decoder at each time step. The calculation process is shown in Equation (14).

[0122] The design of the self-attention mechanism effectively reduces the consumption of computing memory, while reducing the interference of redundant and noisy data, thereby improving the efficiency and accuracy of the model.

[0123] e g,x = a(s g-1 , h x ) (12)

[0124] a g,x = softmax(e g,x ) (13)

[0125]

[0126] Among them, x represents the index or position of a specific element, T represents the total number of elements, and a x represents the attention weight of each element;

[0127] Step 6: Use the dung beetle optimization algorithm to determine the hyperparameters of the stress prediction model;

[0128] Although GRU-A effectively simplifies the network structure and improves the computational speed, the performance of GRU-A largely depends on the adjustment of its hyperparameters, which can be a complex and resource-intensive process. Therefore, integrating DBO to find the optimal set of hyperparameters can significantly improve the performance of GRU-A in predicting stress changes based on stress history data. In this study, the number of hidden neurons and the learning rate in GRU-A are selected as the parameters to be optimized.

[0129] In terms of model hyperparameter optimization, this paper introduces the dung beetle optimization algorithm (DBO). As a new type of swarm intelligence optimization algorithm, it simulates the foraging and reproduction behaviors of dung beetles and has strong global search ability and fast convergence characteristics. Compared with traditional optimization algorithms (such as particle swarm optimization and genetic algorithms), DBO shows higher efficiency and robustness in complex optimization problems and can effectively avoid the local optimum problem. In addition, according to the characteristics of different modal data after VMD decomposition, the DBO algorithm can dynamically optimize the hyperparameter combination of the GRU-A model, thereby improving the prediction accuracy and stability of the model.

[0130] The DBO algorithm is a swarm intelligence optimization algorithm proposed by XUE et al. in 2022, inspired by the behaviors of dung beetles such as rolling balls, dancing, foraging, stealing, and reproducing. The DBO algorithm is inspired by the behaviors of dung beetles such as foraging, rolling balls, dancing, reproducing, and stealing, and has the characteristics of fast convergence speed and high accuracy. The dung beetle optimization algorithm realizes parameter optimization by performing corresponding operations on different types of dung beetles.

[0131] Step 6.1: Initialize the dung beetle swarm;

[0132] The dung beetle swarm includes several dung beetle individuals. All dung beetle individuals are divided into four roles according to a certain proportion, namely ball-rolling dung beetles, brood balls, juvenile dung beetles, and stealing dung beetles. Each dung beetle individual represents a set of hyperparameter combinations, such as hyperparameters like the number of hidden layer neurons and the learning rate in the GRU-A network. These dung beetle individuals perform global search by simulating the foraging and reproduction behaviors of dung beetles to optimize the hyperparameter combination of the model, and ultimately achieve the goal of minimizing the prediction error or improving the model performance;

[0133] Step 6.2: Set the maximum number of iterations;

[0134] Step 6.3: Set the current iteration number q to 0;

[0135] Step 6.4: Calculate the fitness values of all dung beetle individuals; the fitness value is the error index (such as RMSE, MAE or MAPE) after a set of hyperparameters corresponding to the dung beetle individual are applied to the improved gated recurrent unit GRU-A. The smaller the error, the smaller the fitness value, indicating that the hyperparameter combination of the dung beetle individual has better performance and higher survival advantage. The dung beetle individual with the smallest fitness value is taken as the current optimal solution;

[0136] Step 6.5: Determine whether the position of the dung beetle individual exceeds the set optimization range. If it exceeds, adjust the position of the dung beetle individual to within the optimization range. The specific method can be to set boundary conditions to limit the position of the individual. For example, if the position of the individual exceeds the minimum or maximum boundary, set its position to the boundary value; if it does not exceed, continue to update the position of the individual according to the optimization rule so that it can perform effective optimization within the search space;

[0137] Step 6.6: Update the position of the dung beetle individual according to its role;

[0138] Specifically: The ball-rolling dung beetle uses the ball-rolling strategy to update the position of the dung beetle individual; the brood ball uses the reproduction strategy to update the position of the dung beetle individual; the juvenile dung beetle uses the foraging strategy to update the position of the dung beetle individual; the kleptoparasite dung beetle uses the stealing strategy to update the position of the dung beetle individual;

[0139] The ball-rolling strategy is: Obtain a random number rand, and determine whether the random number is less than the set threshold. If so, execute the obstacle-free mode to update the position of the dung beetle individual, otherwise execute the obstacle mode to update the position of the dung beetle individual; the value range of the random number is from 0 to 1;

[0140] The method for updating the position of the dung beetle individual in the obstacle-free mode is: Dung beetles usually use the sun as a guide to ensure that they roll the dung ball on a straight path, but natural factors such as light intensity and wind will affect the dung beetle's rolling during movement. The position of the dung beetle when rolling the ball is updated according to equations (15) and (16).

[0141] x l (q + 1) = x l (q) + α × m × x l (q - 1) + b × Δx (15)

[0142] Δx = |x l (q) - X w | (16)

[0143] In the formula, x l (q) represents the position of the l-th dung beetle individual at the q-th iteration, x l (q + 1) represents the updated position of the dung beetle individual, x l(q - 1) represents the position of the l-th dung beetle individual at the (q - 1)-th iteration, α is the natural coefficient, taking values of -1 or 1; m is the deflection coefficient with values in (0.0, 0.2), b is a constant in (0, 1), Δx is used to simulate the change in light intensity, and X w represents the globally worst position;

[0144] The method for updating the position of the dung beetle individual in the obstacle mode is as follows:

[0145] When a dung beetle individual encounters an obstacle that hinders its progress, it adjusts itself by dancing to find a new route. Once the dung beetle determines a new direction, it continues to roll the ball backward, and its position is updated as follows:

[0146] x l (q + 1) = x l (q) + tan(θ)|x l (q) - x l (q - 1)| (17)

[0147] In the formula, θ represents the deflection angle, taking values in [0, π]. When θ is equal to 0, or π, the position of the dung beetle individual is not updated;

[0148] The reproduction strategy is as follows:

[0149] In nature, a dung beetle individual will roll the dung ball to a safe location and hide it to ensure the safety of its offspring and provide a suitable environment for them. Selecting a suitable egg-laying site is crucial for female dung beetles. Inspired by this, a boundary selection strategy is proposed to simulate the area where female dung beetles lay eggs, which is defined as:

[0150] Lb * = max(X * ×(1 - R), Lb) (18)

[0151] Ub * = min(X * ×(1 + R), Ub) (19)

[0152] In the formula, X * represents the current optimal position, Lb * and Ub * are the lower and upper limits of the egg-laying area respectively. Define R = 1 - q / T max , where T max is the maximum number of iterations, and Lb and Ub represent the lower and upper limits in the optimization problem respectively;

[0153] Once the oviposition area is determined, the female dung beetle will select a brood ball within this area for egg-laying. It should be noted that in the DBO algorithm, each female dung beetle lays only one egg per iteration. Additionally, it can be clearly seen from Equation (20) that the boundary range of the oviposition area changes dynamically, mainly determined by the value of R. Therefore, the position of the brood ball also changes dynamically during the iteration process, and it is defined as:

[0154] B l (q + 1) = X * + b1×(B l (q) - Lb * ) + b2×(B l (q) - Ub * ) (20)

[0155] In the formula, B l (q) represents the position of the l-th brood ball in the q-th iteration, B l (q + 1) represents the updated position of the brood ball, and b1 and b2 are irrelevant random vectors representing the dimensions of the optimization problem;

[0156] The foraging strategy is as follows:

[0157] The eggs laid by the female dung beetle will gradually hatch and grow. Meanwhile, some mature dung beetle larvae will also emerge from the ground and start looking for food. To simulate the optimal foraging area of juvenile dung beetles, the optimal foraging area is defined by Equations (21) and (22);

[0158] Lb b = max(X b ×(1 - R), Lb) (21)

[0159] Ub b = min(X b ×(1 + R), Ub) (22)

[0160] In the formula, X b represents the global optimal position, Lb b and Ub b represent the lower and upper limits of the optimal foraging area respectively;

[0161] The position update of juvenile dung beetles is as follows:

[0162] x l (q + 1) = x l (q) + C1×(x l (q) - Lb b ) + C2×(x l (q) - Ub b ) (23)

[0163] Among them, C1 is a random number that follows a normal distribution, and C2 is a random vector belonging to the interval (0, 1);

[0164] The stealing strategy is as follows:

[0165] Some dung beetles exhibit the behavior of stealing the dung balls of other dung beetles. The position update process of the stealing dung beetles can be described as follows:

[0166] x l (q + 1) = X b + S × g × |x l (q) - X * | + |x l (q) - X b | (24)

[0167] In the formula, g is a random vector of size 1×D that follows a normal distribution, D is the spatial dimension, and S is a constant value;

[0168] Step 6.7: Increment the current iteration count by 1;

[0169] Step 6.8: Calculate the fitness value of each dung beetle individual, and use the dung beetle individual with the smallest fitness value as the current optimal solution. Compare the current optimal solution with the optimal solution of the previous iteration, and use the one with the smaller fitness value as the global optimal solution;

[0170] Step 6.9: Determine whether the current iteration count has reached the maximum iteration count. If so, output the current global optimal solution and its fitness value; otherwise, return to Step 6.5;

[0171] The software environment in this embodiment is based on the Pytorch framework, with a Xeon 4210 CPU, an RTX 3090 GPU, an Ubuntu 18.04 version 64-bit operating system, and 32GB of running memory. The number of dung beetle populations in DBO is 4, and the population ratios of rolling dung beetles, brooding balls, juvenile dung beetles, and stealing dung beetles are set to 1:2:1:1. The learning rate is randomly selected within the range of [0.001, 0.01], and the training iteration count is 10 times. The iteration effect diagram is as Figure 8 shown, and the search range of the hidden layer neurons for the upper and lower boundaries is [10, 100]. The GRU-A model uses L2 regularization, and the number of iteration rounds is set to 50. Under the stress data set, the best learning rate output by the GRU-A model after the DBO algorithm iteration is 0.0026, and the number of hidden neuron layers is 46.

[0172] Step 7: Apply the hyperparameters of the stress prediction model determined in Step 6 to the stress prediction model, and use the training set to train it to obtain the trained stress prediction model;

[0173] Step 8: Input the test set into the trained stress prediction model to obtain the predicted stress results;

[0174] Specifically, input the normalized intrinsic mode function components corresponding to a stress history data into the trained stress prediction model respectively to obtain the predicted results of each intrinsic mode function component, and add the predicted results of the intrinsic mode function components by time point to obtain the predicted stress results;

[0175] Step 9: Denormalize the predicted stress results to obtain the final predicted stress results;

[0176] Figure 9 The comparison between the predicted results and the true values of the present invention on the test set of the open-pit mine slope stress prediction data set is shown. It can be seen from the figure that the model can accurately fit the stress change trend of the test set. Whether it is the overall trend or the local detail fluctuations, the predicted values are highly consistent with the true values. Especially in capturing the rapidly changing regions and long-term trends, the model shows excellent robustness and adaptability. This indicates that GRU-A effectively combines signal decomposition, hyperparameter optimization, and dynamic adjustment capabilities, and can provide high-precision prediction results in complex time series prediction tasks, providing reliable technical support for mine landslide monitoring and risk assessment.

[0177] In this embodiment, the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and regression coefficient (R 2 ) are used as four evaluation indicators to evaluate the performance of the improved prediction model, and their mathematical formulas are as follows:

[0178]

[0179] In the formula: p y is the true value; is the model predicted value; is the average value of the true values; N represents the number of samples in the data set;

[0180] In this study, five models, namely LSTM, CNN-LSTM, DBO-LSTM, SVM, and XGB, are selected as comparison objects. These models cover different methods and characteristics of time series analysis. For example, LSTM can capture long-term time dependencies and is suitable for dealing with complex time series problems; CNN-LSTM combines convolutional feature extraction and sequence modeling capabilities; DBO-LSTM improves performance by optimizing hyperparameters; SVM, as a traditional machine learning method, is suitable for dealing with small sample data; XGB, as a gradient boosting tree algorithm, has strong generalization ability.

[0181] From the experimental results, the VMD-DBO-GRU-A model shows significant advantages in various performance indicators. In terms of the RMSE indicator, compared with LSTM and SVM, it is reduced by 77% and 84% respectively, significantly reducing the overall prediction error; in terms of the MAE indicator, compared with CNN-LSTM and DBO-LSTM, it is reduced by 74% and 77% respectively, indicating higher quality of the predicted data; in terms of the MAPE indicator, compared with SVM and XGB, it is reduced by 86% and 67% respectively, reflecting higher prediction accuracy; in the R 2 indicator, the VMD-DBO-GRU-A model is improved by 3% and 10% compared with LSTM and SVM respectively, indicating better fitting effect. The specific results are shown in Table 2.

[0182] In addition, the VMD-DBO-GRU-A model not only effectively decomposes and denoises the input data through the VMD method, but also significantly improves the hyperparameter adjustment efficiency through the DBO optimization algorithm. Compared with other models, this model integrates multi-level functions of data preprocessing, parameter optimization and feature extraction, greatly improving the overall modeling effect and prediction ability.

[0183] Table 2 Comparison results of prediction models

[0184]

[0185] From Figure 10 the results of the true value and predicted value of the prediction model, although SVM can handle some non-linear features, it performs poorly in complex time-dependent modeling, and the overall prediction deviation is obvious; XGB has certain advantages in non-linear feature extraction, but its ability to capture the dynamic characteristics and fluctuations of time series is insufficient; although LSTM and CNN-LSTM can model the long-term dependence of time series, their accuracy in capturing local fluctuations is limited, and lags or errors are likely to occur; DBO-LSTM improves the model performance by optimizing hyperparameters, but its performance under multi-modal data is still limited, and the detail capture ability is not as expected. The VMD-DBO-GRU-A model performs the best in time series prediction. Compared with other models, this model can capture the overall trend and local details of time series more accurately, especially in the fast-changing region.

[0186] To verify the contribution of each module to the performance of the VMD-DBO-GRU-A model, five models were designed for comparative analysis: the basic GRU model, which only processes the original time series data; the DBO-GRU model, which introduces the dung beetle optimization algorithm to optimize hyperparameters; the GRU-A model, which adds an attention mechanism to enhance the dynamic change response ability; the DBO-GRU-A model, a GRU model that combines DBO optimization and the attention mechanism; and the complete VMD-DBO-GRU-A model, which adds variational mode decomposition on the basis of DBO-GRU-A to separate data modes and denoise. By gradually adding modules in this way, the improvement of each module on the model prediction performance was comprehensively evaluated.

[0187] As shown in Table 3, GRU, as a benchmark model for comparison, only uses basic time series modeling capabilities and does not include any optimization mechanisms. Its performance is the worst, indicating that the complex characteristics of stress time series data are difficult to capture by a single model. After introducing DBO, through global search and fast convergence, the parameter adaptation ability of the model was effectively enhanced. Compared with the basic GRU model, the RMSE decreased by 16.22%, the MAE decreased by 17.04%, and the MAPE decreased by 16.88%, indicating that DBO's contribution to hyperparameter optimization is significant. After adding the attention mechanism, the model can dynamically focus on key information points in the time series, and its ability to capture complex features is further improved. Compared with the basic GRU model, the RMSE decreased by 22.73%, the MAE decreased by 24.60%, and the MAPE decreased by 24.83%, reflecting the important role of the attention mechanism in enhancing the model's response ability to dynamic changes. After combining DBO and the attention mechanism, the model showed significant improvement in overall trend prediction and local feature capture. Compared with the DBO-GRU model, the RMSE decreased by 19.52%, the MAE decreased by 21.67%, and the MAPE decreased by 21.78%, indicating that the collaboration between the two enhanced the model performance. After adding VMD to the DBO-GRU-A model, the model performance was further improved, especially in removing high-frequency noise and separating multimodal data. The introduction of VMD enabled the model to extract key information more clearly, and finally it was comprehensively superior to other models in terms of indicators such as RMSE, MAE, MAPE, and R 2 and other indicators, verifying the complementarity of the functions between modules.

[0188] Further analysis reveals that the combination between modules does not produce an obvious redundancy effect. On the contrary, the functions of each module are complementary. VMD effectively removes high-frequency noise and separates multi-modal data. DBO significantly enhances the parameter adaptation ability of the model through global optimization. The attention mechanism dynamically focuses on key information points in the time series, enabling the model to capture complex features more accurately. In addition, in module collaboration, VMD and the attention mechanism contribute relatively more to the overall performance. In the future, the weight allocation of the modules can be further optimized to improve the prediction ability of the model in specific scenarios.

[0189] Table 3 Results of ablation experiments for each module

[0190]

[0191]

[0192] Figure 11 is the comparison between the true value and the predicted value of each ablation model on the test set. It can be seen from Figure 10 that the prediction curve of the basic GRU model fits the true value curve worst, and it is difficult to capture the dynamic characteristics especially in the rapidly changing area, which is the same as the situation reflected by the evaluation index; the DBO-GRU model significantly improves the adaptation ability of hyperparameters, and the overall trend prediction is more accurate; the GRU-A model enhances the response to local rapid fluctuations, but is slightly inferior to the DBO-GRU in capturing the overall trend; the combined DBO-GRU-A model performs excellently in both trend and detail capture, showing the advantages of synergistic effects; finally, the complete VMD-DBO-GRU-A model achieves the best effect in capturing the overall trend and local details. Its prediction curve is highly consistent with the true value, showing extremely high accuracy both in the stable area and the rapidly fluctuating area. This situation further illustrates the effectiveness and reliability of the model designed in this paper for predicting the stress of mine landslides.

[0193] This embodiment provides an open-pit mine slope stress prediction system based on VMD-DBO optimized GRU for implementing the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU, including:

[0194] A stress historical data acquisition module for acquiring stress historical data;

[0195] A data preprocessing module for preprocessing the stress historical data to obtain the preprocessed stress historical data;

[0196] A variational mode decomposition and normalization module, which is used to decompose the preprocessed stress history data by using the variational mode decomposition method, obtain several intrinsic mode function components and normalize them, and then obtain several normalized intrinsic mode function components corresponding to each stress history data;

[0197] A dung beetle optimization module, which is used to determine the hyperparameters of the stress prediction model;

[0198] A stress prediction model, which is an improved gated recurrent unit GRU-A. Applying the hyperparameters determined by the dung beetle optimization module, it inputs a normalized intrinsic mode function component and outputs the prediction result of this intrinsic mode function component;

[0199] A summation module, which adds up the prediction results of all the intrinsic mode function components obtained according to time points to obtain the prediction result of the stress;

[0200] An inverse normalization module, which is used to perform inverse normalization on the prediction result of the stress to obtain the final prediction result of the stress.

[0201] This embodiment provides an electronic device, including: a processor, a memory and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are executed by the processor, the steps of the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU are executed.

[0202] This embodiment provides a computer-readable storage medium, in which a computer program is stored. When the computer program is run by a processor, the steps of the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU are executed.

Claims

1. A method for predicting slope stress in open-pit mines based on VMD-DBO optimized GRU, characterized in that: The following steps are involved: Step 1: Construct open-pit mine slope foundation stress database; Step 2: preprocess the stress history data in the open pit mine slope foundation stress database to obtain the preprocessed stress history data; Step 3: For each preprocessed stress history data, the variational mode decomposition method is used to decompose it, and several intrinsic mode function components are obtained and normalized, thereby obtaining several normalized intrinsic mode function components corresponding to each stress history data; Step 4: Based on the normalized intrinsic mode function components obtained in step 3, a stress data set is constructed and divided according to a set ratio to obtain a training set and a test set; Step 5: Construct stress prediction model; Step 6: Determine the hyperparameters of the stress prediction model using the dung beetle optimization algorithm; Step 7: Apply the hyperparameters of the stress prediction model determined in step 6 to the stress prediction model, and train it using the training set to obtain a trained stress prediction model; Step 8: Input the test set into the trained stress prediction model to obtain the stress prediction result; Step 9: Denormalize the stress prediction results to obtain the final stress prediction results.

2. The open-pit mine slope stress prediction method based on VMD-DBO optimized GRU according to claim 1 is characterized in that: The open-pit mine slope foundation stress database in step 1 includes a number of stress history data, and the stress history data is a time series, including open-pit mine slope stress data at different time points.

3. The open-pit mine slope stress prediction method based on VMD-DBO optimized GRU according to claim 1 is characterized in that: The preprocessing described in step 2 is missing value interpolation.

4. The open-pit mine slope stress prediction method based on VMD-DBO optimized GRU according to claim 1, characterized in that: The stress data set in step 4 includes a plurality of samples, and each sample includes a plurality of normalized intrinsic mode function components corresponding to a piece of stress history data.

5. The open-pit mine slope stress prediction method based on VMD-DBO optimized GRU according to claim 1, characterized in that: The stress prediction model in step 5 is an improved gated recurrent unit GRU-A; The improved gated recurrent unit GRU-A inputs a normalized intrinsic mode function component, and outputs a prediction result of the intrinsic mode function component. The improved gated recurrent unit GRU-A adds a self-attention mechanism layer between the hidden layer and the fully connected layer of the gated recurrent unit GRU. The self-attention mechanism layer uses the self-attention mechanism to perform weighted processing on the time series data output by the hidden layer to generate a context vector.

6. The open-pit mine slope stress prediction method based on VMD-DBO optimized GRU according to claim 1, characterized in that: Step 6 specifically includes: Step 6.1: Initialize the dung beetle swarm; The dung beetle group includes a number of dung beetle individuals, and all dung beetle individuals are divided into four roles according to proportion, namely, rolling ball dung beetles, brooding balls, young dung beetles and stealing dung beetles, and each dung beetle individual represents a set of hyperparameter combinations of the stress prediction model; Step 6.2: Set the maximum number of iterations; Step 6.3: Set the current number of iterations q to 0; Step 6.4: Calculate the fitness values ​​of all dung beetle individuals, and take the dung beetle individual with the smallest fitness value as the current optimal solution; the fitness value is an error index after a set of hyperparameters corresponding to the dung beetle individual is applied to the improved gated recurrent unit GRU-A; Step 6.5: Determine whether the position of the dung beetle individual exceeds the set optimization range. If so, adjust the position of the dung beetle individual to within the optimization range; Step 6.6: Update the position of the dung beetle according to its role; Specifically, the rolling ball dung beetle uses the rolling ball strategy to update the position of the individual dung beetle; the brood ball uses the breeding strategy to update the position of the individual dung beetle; the young dung beetle uses the foraging strategy to update the position of the individual dung beetle; the stealing dung beetle uses the stealing strategy to update the position of the individual dung beetle; The rolling ball strategy is: obtain a random number rand, determine whether the random number is less than a set threshold, if so, execute the barrier-free mode to update the position of the dung beetle individual, otherwise execute the barrier mode to update the position of the dung beetle individual; the value range of the random number is 0 to 1; The method for updating the position of the individual dung beetle in the barrier-free mode is: x l (q+1)=x l (q)+α×m×x l (q-1)+b×Δx (15) Δx=|x l (q)-X w | (16) In the formula, x l (q) represents the position of the lth dung beetle individual at the qth iteration, x l (q+1) represents the updated position of the dung beetle individual, x l (q-1) represents the position of the lth dung beetle individual at the q-1th iteration, α is the natural coefficient, which takes the value of -1 or 1; m is the deflection coefficient with the value of (0.0, 0.2), b is a constant (0, 1), Δx is used to simulate the change of light intensity, X w represents the global worst position; The method for updating the position of the dung beetle individual in the obstacle mode is: x l (q+1)=x l (q)+tan(θ)|x l (q)-x l (q-1)| (17) Where θ represents the deflection angle, which takes values ​​of [0, π]. When θ is equal to 0, or π, the position of the dung beetle individual is not updated; The breeding strategy is: First define the spawning area: Lb * =max(X * ×(1-R),Lb) (18) Ub * =min(X * ×(1+R),Ub) (19) In the formula, X * Indicates the current optimal position, Lb * and Ub * are the lower and upper limits of the spawning area, respectively, and R = 1-q / T is defined max , where T max is the maximum number of iterations, Lb and Ub represent the lower and upper bounds of the optimization problem, respectively; The method to update the position of the brooding ball according to the spawning area is: B l (q+1)=X * +b1×(B l (q)-Lb * )+b2×(B l (q)-Ub * ) (20) In the formula, B l (q) represents the position of the lth brooding ball in the qth iteration, B l (q+1) represents the updated position of the brooding ball, b1 and b2 are unrelated random vectors representing the dimensions of the optimization problem; The foraging strategy is: Defining optimal foraging areas: Lb b =max(X b ×(1-R),Lb) (21) Ub b =min(X b ×(1+R),Ub) (22) In the formula, X b represents the global optimal position, Lb b and Ub b represent the lower and upper limits of the optimal foraging area, respectively; Update the location of the juvenile dung beetles based on the best foraging areas: x l (q+1)=x l (q)+C1×(x l (q)-Lb b )+C2×(x l (q)-Ub b ) (23) Among them, C1 is a random number that follows a normal distribution, and C2 is a random vector belonging to the interval (0,1); The stealing strategy is: The method for updating the position of the stealing dung beetle is: x l (q+1)=X b +S×g×|x l (q)-X * |+|x l (q)-X b | (24) Where g is a random vector of size 1×D, which follows a normal distribution, D is the spatial dimension, and S is a constant value; Step 6.7: The current iteration number is increased by 1; Step 6.8: Calculate the fitness value of each dung beetle individual, and take the dung beetle individual with the smallest fitness value as the current optimal solution. Compare the current optimal solution with the optimal solution of the previous iteration, and take the one with the smallest fitness value as the global optimal solution. Step 6.9: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the current global optimal solution and its fitness value. Otherwise, return to step 6.

5.

7. The open-pit mine slope stress prediction method based on VMD-DBO optimized GRU according to claim 1, characterized in that: Step 8 is specifically as follows: inputting several normalized intrinsic mode function components corresponding to a stress history data into the trained stress prediction model respectively, obtaining the prediction result of each intrinsic mode function component, adding the prediction results of the intrinsic mode function components according to the time point, and obtaining the stress prediction result.

8. An open-pit mine slope stress prediction system based on VMD-DBO optimized GRU, used to implement the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU according to any one of claims 1 to 7, characterized in that: include: A stress history data acquisition module is used to acquire stress history data; A data preprocessing module is used to preprocess the stress history data to obtain the preprocessed stress history data; The variational mode decomposition and normalization module is used to decompose the preprocessed stress history data using the variational mode decomposition method to obtain a number of intrinsic mode function components and normalize them, thereby obtaining a number of normalized intrinsic mode function components corresponding to each stress history data; The Dung Beetle Optimization module is used to determine the hyperparameters of the stress prediction model; The stress prediction model is an improved gated recurrent unit GRU-A, which uses the hyperparameters determined by the dung beetle optimization module, inputs a normalized intrinsic mode function component, and outputs the prediction result of the intrinsic mode function component; A summing module adds the prediction results of all intrinsic mode function components according to the time points to obtain the prediction results of stress; The denormalization module is used to denormalize the stress prediction results to obtain the final stress prediction results.

9. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate via the bus. When the machine-readable instructions are executed by the processor, the steps of the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU as described in any one of claims 1 to 7 are performed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of the open-pit mine slope stress prediction method based on VMD-DBO optimized GRU as described in any one of claims 1 to 7.

Images