Rockburst risk prediction method and system in combination with LSTM and convolutional neural network

Abstract

The invention relates to the technical field of rockburst risk prediction, and particularly provides a rockburst risk prediction method and system combining LSTM and a convolutional neural network, and the method comprises the steps: collecting rockburst sample data, and carrying out the Kendall correlation analysis of the rockburst sample data, and obtaining a correlation analysis value; based on the correlation analysis value, obtaining a rockburst risk feature which meets a preset correlation threshold value with the rockburst risk level label column; constructing a data set based on the rockburst risk features and the rockburst risk level label column; constructing a rockburst risk prediction model based on a long short-term memory (LSTM) network and a convolutional neural network; and based on the data set and the rockburst risk prediction model, obtaining a rockburst risk level, and completing rockburst risk prediction. By adopting the technical scheme of the invention, rockburst can be accurately predicted, so that engineering accidents can be reduced and even avoided, and personal and property safety can be guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of rockburst risk prediction, and specifically relates to a rockburst risk prediction method and system based on combining LSTM with convolutional neural network. Background Art

[0002] Rockburst is a dynamic instability process caused by the instantaneous release of energy along the free surface of the energetic rock mass after excavation. It is an uncontrollable rock mass destruction phenomenon, which is accompanied by extremely violent energy release, posing a huge threat to on-site workers and equipment, and seriously affecting the progress of the project. Therefore, accurate prediction of rockburst risk is of great significance for reducing or avoiding accidents. With the development of society and the progress of science and technology, rockburst prediction has gradually evolved from traditional mathematical model method prediction to deep learning method prediction. The existing deep learning model rockburst prediction method has been optimized from different angles, and has achieved certain results, which has a significant role in promoting rockburst risk prediction. However, these methods still have certain shortcomings in prediction accuracy. Summary of the invention

[0003] In order to solve the problems existing in the prior art, the present invention provides a rockburst risk prediction method and system based on the combination of LSTM and convolutional neural network, which analyzes and predicts rockburst data, improves the accuracy of rockburst risk prediction, can reduce or even avoid engineering accidents, and ensure personal and property safety.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] The rockburst risk prediction method combining LSTM and convolutional neural network includes the following steps:

[0006] Collect rockburst sample data, and perform Kendall correlation analysis on the rockburst sample data to obtain correlation analysis values; wherein the rockburst sample data includes maximum tangential stress of surrounding rock of cave wall, uniaxial compressive strength of rock, uniaxial tensile strength of rock, stress index, brittleness index, rock elastic energy index and rockburst risk level label column;

[0007] Based on the correlation analysis value, a rockburst risk feature that satisfies a preset correlation threshold with the rockburst risk level label column is obtained;

[0008] Based on the rockburst risk characteristics and the rockburst risk level label column, construct a data set; divide the data set into a training set and a test set;

[0009] Based on the long short-term memory network LSTM and the convolutional neural network CNN, a rockburst risk prediction model is constructed; and the rockburst risk prediction model is trained using the training set; wherein the long short-term memory network LSTM is a double-layer long short-term memory network Double LSTM that introduces the attention mechanism Attention;

[0010] Based on the trained rockburst risk prediction model, the rockburst risk level of the test set is predicted to complete the rockburst risk prediction.

[0011] Preferably, in the data set,

[0012] Rockburst risk characteristics include stress index, brittleness index and rock elastic energy index;

[0013] The rockburst risk level label columns include no rockburst, mild rockburst, moderate rockburst, and severe rockburst.

[0014] Preferably, the structure of the rockburst risk prediction model includes: a CNN layer, a LSTM-ATT-LSTM layer and an output layer;

[0015] The CNN layer is used to extract the features of the training set based on the one-dimensional convolution layer, and downsample the extracted features using the maximum pooling method to obtain key features; the calculation formula is as follows:

[0016] X conv [i,j,k] = ReLU(X input [i,j:j+k,:]*W conv [k,:,:]+b conv ),

[0017] X pool [i,j]=max(X conv [i,2j,:],X conv [i,2j+1,:]),

[0018] In the formula, X input is the data sequence of the training set of the input model, N is the number of samples, T is the sequence length, and F is the number of features of each sequence; W conv is the convolution kernel, with a shape of k×F×F conv , F conv is the number of convolution kernels, b conv is the bias term; ReLU(z)=max(0,z) is the ReLU activation function, X conv [i,j,k] is the output of the convolutional layer, X pool [i,j] is the key feature output after pooling;

[0019] The LSTM-ATT-LSTM layer includes a first LSTM layer, an attention mechanism layer and a second LSTM layer, which are used to process the key features and obtain feature vectors; wherein the flow of information is controlled by the internal gating structure of the double-layer LSTM, and weights are dynamically assigned to the key features by the attention mechanism Attention; wherein the gating structure includes a forget gate, an input gate and an output gate, and the gating structure controls the flow of information through a sigmoid function and a tanh function;

[0020] The output layer is used to obtain a prediction result based on the feature vector, and use a classification cross entropy loss function to obtain the difference between the prediction result and the true value.

[0021] Preferably, the method for the LSTM-ATT-LSTM layer to process the key features includes:

[0022] Based on the first-layer LSTM processing of the key features, the first-layer LSTM output is obtained, and the calculation formula is as follows:

[0023] X lstm1 =LSTM(X pool ,W lstm1 ,b lstm1 )

[0024] Where W lstm1 is the first layer LSTM weight matrix, b lstm1 is the first layer LSTM bias term, X lstm1 is the output of the first layer of LSTM;

[0025] Based on the attention mechanism layer, the weighted sum of the first layer LSTM output is performed to obtain the output of the attention mechanism layer. The calculation formula is as follows:

[0026] e t =tanh(X lstm1 [t]W att +b att )

[0027]

[0028]

[0029] In the formula, X lstm1 [t] is the output of the first layer LSTM at the tth time step, with a shape of N×U, W att is the attention weight matrix, with shape U×U, b att is the attention bias, with shape U, α t is the attention weight at the t-th time step, X attis the output after weighted summation, with a shape of N×U, which serves as the input of the second layer of LSTM; t is the attention score at time step t;

[0030] Based on the output of the attention mechanism layer processed by the second layer LSTM, the feature vector is obtained, and the calculation formula is as follows:

[0031] X lstm2 =LSTM(X att ,W lstm2 ,b lstm2 )

[0032] Where W lstm2 is the weight matrix of the second layer LSTM, b lstm2 is the bias term of the second layer LSTM, X lstm2 is the feature vector.

[0033] Preferably, it also includes evaluating the prediction performance of the rockburst risk prediction model based on evaluation indicators, wherein the evaluation indicators include accuracy, recall rate and F1 value; wherein the F1 value is the harmonic mean of accuracy and recall rate.

[0034] The present invention also provides a rockburst risk prediction system combining LSTM and convolutional neural network, which is used to implement the method, including:

[0035] A sample collection module is used to collect rock burst sample data and perform Kendall correlation analysis on the rock burst sample data to obtain a correlation analysis value; wherein the rock burst sample data includes the maximum tangential stress of the surrounding rock of the cave wall, the uniaxial compressive strength of the rock, the uniaxial tensile strength of the rock, the stress index, the brittleness index, the rock elastic energy index and the rock burst risk level label column;

[0036] A correlation analysis module, configured to obtain, based on the correlation analysis value, a rockburst risk feature that satisfies a preset correlation threshold with the rockburst risk level label column;

[0037] A data set construction module is used to construct a data set based on the rockburst risk characteristics and the rockburst risk level label column; divide the data set into a training set and a test set;

[0038] A rockburst risk prediction model construction module is used to construct a rockburst risk prediction model based on a long short-term memory network (LSTM) and a convolutional neural network (CNN); and train the rockburst risk prediction model using the training set; wherein the long short-term memory network (LSTM) is a double-layer long short-term memory network (Double LSTM) that introduces an attention mechanism (Attention);

[0039] The rockburst risk prediction module is used to predict the rockburst risk level of the test set based on the trained rockburst risk prediction model to complete the rockburst risk prediction.

[0040] Preferably, the structure of the rockburst risk prediction model includes a CNN layer, an LSTM-ATT-LSTM layer and an output layer; wherein the LSTM-ATT-LSTM layer includes a first LSTM layer, an attention mechanism layer and a second LSTM layer;

[0041] The key feature extraction unit is used to extract the features of the input sequence based on the one-dimensional convolution layer of the CNN layer, and downsample the extracted features using the maximum pooling method to obtain the key features; the calculation formula is as follows:

[0042] X conv [i,j,k] = ReLU(X input [i,j:j+k,:]*W conv [k,:,:]+b conv ),

[0043] X pool [i,j]=max(X conv [i,2j,:],X conv [i,2j+1,:]),

[0044] In the formula, X input is the data sequence of the training set of the input model, N is the number of samples, T is the sequence length, and F is the number of features of each sequence; W conv is the convolution kernel, with a shape of k×F×F conv , F conv is the number of convolution kernels, b conv is the bias term; ReLU(z)=max(0,z) is the ReLU activation function, X conv [i,j,k] is the output of the convolutional layer, X pool [i,j] is the key feature output after pooling;

[0045] A feature vector acquisition unit, used to process the key features and obtain feature vectors; wherein the flow of information is controlled by the internal gating structure of the double-layer LSTM, and weights are dynamically assigned to the key features by the attention mechanism Attention; wherein the gating structure includes a forget gate, an input gate and an output gate, and the gating structure controls the flow of information through a sigmoid function and a tanh function;

[0046] The prediction unit is used to obtain a prediction result based on the feature vector in the output layer, and obtain the difference between the prediction result and the true value by using a classification cross entropy loss function.

[0047] Preferably, the feature vector acquisition unit includes:

[0048] The first-layer LSTM subunit is used to process the key features based on the first-layer LSTM to obtain the first-layer LSTM output. The calculation formula is as follows:

[0049] X lstm1 =LSTM(X pool ,W lstm1 ,b lstm1 )

[0050] Where W lstm1 is the first layer LSTM weight matrix, b lstm1 is the first layer LSTM bias term, X lstm1 is the output of the first layer of LSTM;

[0051] The attention mechanism subunit is used to perform weighted summation on the output of the first layer LSTM based on the attention mechanism layer to obtain the output of the attention mechanism layer. The calculation formula is as follows:

[0052] e t =tanh(X lstm1 [t]W att +b att )

[0053]

[0054]

[0055] In the formula, X lstm1 [t] is the output of the first layer LSTM at the tth time step, with a shape of N×U, W att is the attention weight matrix, with shape U×U, b att is the attention bias, with shape U, α t is the attention weight at the t-th time step, X att is the output after weighted summation, with a shape of N×U, which serves as the input of the second layer of LSTM; t is the attention score at time step t;

[0056] The second layer LSTM subunit is used to process the output of the attention mechanism layer based on the second layer LSTM to obtain the feature vector. The calculation formula is as follows:

[0057] X lstm2 =LSTM(X att ,W lstm2 ,b lstm2 )

[0058] Where W lstm2 is the weight matrix of the second layer LSTM, blstm2 is the bias term of the second layer LSTM, X lstm2 is the feature vector.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention selects features based on correlation analysis of rockburst sample data, divides the data set, uses a method combining long short-term memory network and convolutional neural network model for training, and uses the trained rockburst risk prediction model to make accurate predictions, thereby reducing or even avoiding engineering accidents and ensuring personal and property safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0061] Figure 1 This is a flow chart of a rockburst risk prediction method combining LSTM and convolutional neural network according to an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram of the structure of a rockburst risk prediction system combining LSTM and convolutional neural network according to an embodiment of the present invention;

[0063] Figure 3 It is a schematic diagram of the structure of the rockburst risk prediction model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] Embodiment 1

[0067] like Figure 1 As shown in the figure, the rockburst risk prediction method combining LSTM and convolutional neural network includes the following steps:

[0068] S1: Collect rock burst sample data, and perform Kendall correlation analysis on the rock burst sample data to obtain correlation analysis values; in this embodiment, the collected rock burst sample data is data generated when rock bursts occur in major mines at home and abroad. Among them, the maximum tangential stress of the surrounding rock of the cave wall (σ θ ), uniaxial compressive strength of rock (σ c ), uniaxial tensile strength of rock (σ t ), stress index (σ θ / σ c ), brittleness index (σ c / σ t ), rock elastic energy index (W et ) and a rockburst risk level (grade) label column.

[0069] Kendall correlation analysis includes:

[0070] Kendall correlation analysis determines the degree of correlation between variables by calculating the concordant pairs and inverse pairs between variables (rockburst sample data). Concordant pair: When the order of two variables is the same (that is, the pair with a larger value in variable X also has a larger value in variable Y, or vice versa), that is, the data pair (x i ,y i ) and (x j ,y j ) satisfies: (x i >x j and y i >y j ) or (x i <x j and y i <y j Discordant Pair: When the order of two variables is the same (i.e., the pair with a larger value in variable X also has a smaller value in variable Y), that is, the data pair (x i ,y i ) and (x j ,y j ) satisfies: (x i >x j and y i <y j ) or (x i <x j and y i >y j ), Kendall Tau coefficient τ is obtained by comparing the number of synergistic pairs and inverse pairs. Its formula is as follows:

[0071]

[0072] Where: C is the number of synergistic pairs; D is the number of inverse pairs; n is the total number of sample data pairs. The meaning of the above formula is: if the order of the variables is highly consistent (many synergistic pairs), τ is close to 1; if the order is completely opposite (many inverse pairs), τ is close to -1; if there is no obvious order relationship, τ is close to 0.

[0073] S2: Based on the correlation analysis value, obtain the rockburst risk features that meet the preset correlation threshold with the rockburst risk level label column; specifically, select the three features most relevant to the rockburst risk level label column according to the correlation analysis value, and the three most relevant features are stress index (σ θ / σ c ), brittleness index (σ c / σ t ) and rock elastic energy index (W et ).

[0074] S3: constructing a data set based on the rockburst risk characteristics and the rockburst risk level label column; dividing the data set into a training set and a test set.

[0075] A further implementation method is that in the data set,

[0076] Rockburst risk characteristics include stress index, brittleness index and rock elastic energy index, which are the three characteristics most relevant to the rockburst risk level label column mentioned above.

[0077] The rockburst risk level label columns include no rockburst, mild rockburst, moderate rockburst, and severe rockburst.

[0078] Specifically, the data set is divided into a test set and a training set according to a preset ratio.

[0079] The training set and the test set are normalized to 0-1 data, and the normalization process is shown in the following formula:

[0080]

[0081] Where: x' is the standardized value of the desired eigenvalue, x is the desired eigenvalue, μ is the mean of the original data, and δ is the standard deviation of the original data;

[0082] The training set data is trained using a combined long short-term memory network and convolutional neural network model to optimize the model parameters; the test set is predicted using the trained combined long short-term memory network and convolutional neural network model to obtain accurate prediction results.

[0083] S4: Based on the long short-term memory network LSTM and the convolutional neural network CNN, a rock burst risk prediction model is constructed. The above model is referred to as the CADL model; and the rock burst risk prediction model is trained using the training set; wherein the long short-term memory network LSTM is a double-layer long short-term memory network Double LSTM that introduces the attention mechanism Attention; model principle: first, the local features in the input sequence are learned in the CNN layer, and different features are extracted through multiple convolution kernels. The data is downsampled using the maximum pooling method to retain key features and reduce model parameters. Secondly, two layers of LSTM are used in the LSTM-ATT-LSTM layer to process the sequence memory characteristics, and the attention mechanism is added to dynamically learn and adapt to the features that need special attention in the sequence, enhance the interpretability of the model, and improve the prediction performance. Then, the Softmax activation function and the classification cross entropy loss function are used in the output layer to help the gradient descent algorithm find the optimal solution more stably. Finally, ten-fold cross validation is used to select the final hyperparameters and output the final prediction results. Figure 3 shown.

[0084] A further implementation method is that the structure of the rockburst risk prediction model includes: a CNN layer, a LSTM-ATT-LSTM layer and an output layer;

[0085] Among them, the CNN layer is used to extract the features X of the training set based on the one-dimensional convolution layer input ∈R N×T×F , and use the maximum pooling method to downsample the extracted features to obtain the key features; the calculation formula is as follows:

[0086] X conv [i,j,k] = ReLU(X input [i,j:j+k,:]*W conv [k,:,:]+b conv ),

[0087] Perform a maximum pooling operation on the output data, retaining the maximum value in each window to reduce the dimension of the data while retaining key feature information. Output length information:

[0088] X pool [i,j]=max(X conv [i,2j,:],X conv [i,2j+1,:]),

[0089] In the formula, X input is the data sequence of the training set of the input model, N is the number of samples, T is the sequence length, and F is the number of features of each sequence; W conv is the convolution kernel, with a shape of k×F×F conv , F convis the number of convolution kernels, b conv is the bias term; ReLU(z)=max(0,z) is the ReLU activation function, X conv [i,j,k] is the output of the convolutional layer, X pool [i,j] is the key feature of the output after pooling; combined with the training and tuning of this model, the output shape is X conv ∈R 32×8×150 , X pool ∈R 32×4×150 .

[0090] The LSTM-ATT-LSTM layer includes a first LSTM layer, an attention mechanism layer and a second LSTM layer, which are used to process the key features and obtain feature vectors; wherein the flow of information is controlled by the internal gating structure of the double-layer LSTM, and weights are dynamically assigned to the key features by the attention mechanism Attention; wherein the gating structure includes a forget gate (Forget Gate), an input gate (Input Gate) and an output gate (Output Gate), and the gating structure controls the flow of information through a sigmoid function and a tanh function; specifically, the input gate determines which new information is stored in the cell state, and the input gate includes two parts: one part is the sigmoid layer to determine which values ​​will update the cell state, and the other part is the tanh layer to create a new candidate value vector To add to the state, the calculation formula is:

[0091] i t =σ(W i X t +U i h t-1 +b i )

[0092]

[0093] Where: i t is the output of the input gate, is the new candidate cell state, X t is the input current time step data, U i , U C is the weight matrix of the hidden state of the previous time step, with a shape of U×U, σ is the sigmoid function, tanh is the hyperbolic tangent function, and h t-1 is the output of the previous time step, W i , W C and the corresponding bias b i , b C is the weight matrix and bias vector of the input gate.

[0094] The forget gate determines which information is discarded from the cell. The output range is 0 to 1, where 0 means completely forgotten and 1 means completely retained. The calculation formula is:

[0095] f t =σ(W f X t +U f h t-1 +b f )

[0096] Where: f t is the output of the forget gate, σ is the sigmoid function, W f and b f is the weight matrix and bias vector of the forget gate, U f is the weight matrix of the hidden state at the previous time step, with a shape of U×U,.

[0097] Cell State C t The update is performed by the forget gate f t and input gate i t The output of is updated. The update calculation process is:

[0098]

[0099] The output gate determines the hidden state h at the current moment t , which is based on the current input X t and the updated cell state C t The calculation formula is:

[0100] o t =σ(W o X t +U o h t-1 +b o )

[0101] h t =o t tanh(C t )

[0102] In the formula: o t is the output of the output gate, tanh is the hyperbolic tangent function, W o and b o is the weight matrix and bias vector of the output gate, U o is the weight matrix of the hidden state at the previous time step, with shape U×U.

[0103] A further implementation method is to process the key features based on the first layer LSTM to obtain the first layer LSTM output. Specifically, the output of the first layer LSTM is X lstm1 , the shape is N×T'×U, where is the length after pooling, and the calculation formula is:

[0104] X lstm1 =LSTM(X pool ,W lstm1 ,b lstm1 )

[0105] Where W lstm1 is the first layer LSTM weight matrix, b lstm1 is the first layer LSTM bias term, X lstm1 The output of the first layer of LSTM; combined with the training and tuning of this model, the output shape is X lstm1 ∈R 32×4×45 .

[0106] Adding an attention mechanism layer in the middle of the two-layer LSTM layer can enhance the model's attention to the input data and dynamically assign weights to different features so that more attention can be focused on the problem-solving part, thereby improving the model's expressiveness and performance. Perform attention weight calculation and weighted sum calculation operations in the attention layer. Based on the attention mechanism layer, perform weighted summation on the output of the first layer of LSTM to obtain the output of the attention mechanism layer. The calculation formula is as follows:

[0107] e t =tanh(X lstm1 [t]W att +b att )

[0108]

[0109]

[0110] In the formula, X lstm1 [t] is the output of the first layer LSTM at the tth time step, with a shape of N×U, W att is the attention weight matrix, with shape U×U, b att is the attention bias, with shape U, α t is the attention weight at the t-th time step, X att is the output after weighted summation, with a shape of N×U, which serves as the input of the second layer of LSTM; t represents the attention score at time step t, which reflects the degree to which time step t should be paid attention to in the model; combined with the training and tuning of this model, the output shape is X att ∈R 32 ×4×45 .

[0111] Based on the output of the attention mechanism layer processed by the second layer LSTM, the feature vector is obtained, and the calculation formula is as follows:

[0112] X lstm2 =LSTM(X att ,W lstm2 ,b lstm2 )

[0113] Where W lstm2 is the weight matrix of the second layer LSTM, b lstm2 is the bias term of the second layer LSTM, X lstm2 is the feature vector. Combined with the training and tuning of this model, the output shape is X lstm2 ∈R 32×45 .

[0114] The output layer is used to obtain the prediction result based on the feature vector, and use the classification cross entropy loss function to obtain the difference between the prediction result and the true value. Specifically, the softmax activation function is used on the feature vector output by the upper layer to convert the input of each neuron into a probability value. The calculation formula is:

[0115] X output =softmax(X lstm2 W dense +b dense )

[0116] Where: is the output of the second LSTM layer, with a shape of N×U, W dense is the weight matrix of the fully connected layer, with a shape of U×C,b dense is the bias term, with a shape of C. The softmax activation function converts the output into the probability of each category, and the output shape is N×C. Combined with the training and tuning of this model, the output shape is x output ∈R 32×2 .

[0117] The categorical cross entropy loss function is used to measure the difference between the model prediction results and the true labels.

[0118] S5: Based on the trained rockburst risk prediction model, predict the rockburst risk level of the test set to complete the rockburst risk prediction.

[0119] A further implementation method is to evaluate the prediction performance of the rockburst risk prediction model based on evaluation indicators, wherein the evaluation indicators include accuracy, recall and F1 value; wherein the F1 value is the harmonic mean of accuracy and recall. Specifically, the accuracy evaluates the quality of the classification model, which represents the percentage of the correct prediction results in the total samples; the recall represents the probability of being predicted as a positive sample in the actual positive sample; the F1 value is a statistical quantity, which is numerically equal to the harmonic mean of the precision and recall.

[0120] The trained CADL model is used to predict the test set and evaluated using evaluation indicators to prove the accuracy of the CADL model prediction and predict the accurate rockburst risk level.

[0121] Embodiment 2

[0122] like Figure 2 As shown, the present invention also provides a rockburst risk prediction system combining LSTM and convolutional neural network, which is used to implement the method, including:

[0123] A sample collection module is used to collect rock burst sample data and perform Kendall correlation analysis on the rock burst sample data to obtain a correlation analysis value; wherein the rock burst sample data includes the maximum tangential stress of the surrounding rock of the cave wall, the uniaxial compressive strength of the rock, the uniaxial tensile strength of the rock, the stress index, the brittleness index, the rock elastic energy index and the rock burst risk level label column;

[0124] A correlation analysis module, configured to obtain, based on the correlation analysis value, a rockburst risk feature that satisfies a preset correlation threshold with the rockburst risk level label column;

[0125] A data set construction module is used to construct a data set based on the rockburst risk characteristics and the rockburst risk level label column; divide the data set into a training set and a test set;

[0126] A rockburst risk prediction model construction module is used to construct a rockburst risk prediction model based on a long short-term memory network (LSTM) and a convolutional neural network (CNN); and train the rockburst risk prediction model using the training set; wherein the long short-term memory network (LSTM) is a double-layer long short-term memory network (Double LSTM) that introduces an attention mechanism (Attention);

[0127] The rockburst risk prediction module is used to predict the rockburst risk level of the test set based on the trained rockburst risk prediction model to complete the rockburst risk prediction.

[0128] Preferably, the structure of the rockburst risk prediction model includes a CNN layer, an LSTM-ATT-LSTM layer and an output layer; wherein the LSTM-ATT-LSTM layer includes a first LSTM layer, an attention mechanism layer and a second LSTM layer;

[0129] The key feature extraction unit is used to extract the features of the input sequence based on the one-dimensional convolution layer of the CNN layer, and downsample the extracted features using the maximum pooling method to obtain the key features; the calculation formula is as follows:

[0130] X conv [i,j,k] = ReLU(X input[i,j:j+k,:]*W conv [k,:,:]+b conv ),

[0131] X pool [i,j]=max(X conv [i,2j,:],X conv [i,2j+1,:]),

[0132] In the formula, X input is the data sequence of the training set of the input model, N is the number of samples, T is the sequence length, and F is the number of features of each sequence; W conv is the convolution kernel, with a shape of k×F×F conv , F conv is the number of convolution kernels, b conv is the bias term; ReLU(z)=max(0,z) is the ReLU activation function, X conv [i,j,k] is the output of the convolutional layer, X pool [i,j] is the key feature output after pooling;

[0133] A feature vector acquisition unit, used to process the key features and obtain feature vectors; wherein the flow of information is controlled by the internal gating structure of the double-layer LSTM, and weights are dynamically assigned to the key features by the attention mechanism Attention; wherein the gating structure includes a forget gate, an input gate and an output gate, and the gating structure controls the flow of information through a sigmoid function and a tanh function;

[0134] The prediction unit is used to obtain a prediction result based on the feature vector in the output layer, and obtain the difference between the prediction result and the true value by using a classification cross entropy loss function.

[0135] Preferably, the feature vector acquisition unit includes:

[0136] The first-layer LSTM subunit is used to process the key features based on the first-layer LSTM to obtain the first-layer LSTM output. The calculation formula is as follows:

[0137] X lstm1 =LSTM(X pool ,W lstm1 ,b lstm1 )

[0138] Where W lstm1 is the first layer LSTM weight matrix, b lstm1 is the first layer LSTM bias term, X lstm1 is the output of the first layer of LSTM;

[0139] The attention mechanism subunit is used to perform weighted summation on the output of the first layer LSTM based on the attention mechanism layer to obtain the output of the attention mechanism layer. The calculation formula is as follows:

[0140] e t =tanh(X lstm1 [t]W att +b att )

[0141]

[0142]

[0143] In the formula, X lstm1 [t] is the output of the first layer LSTM at the tth time step, with a shape of N×U, W att is the attention weight matrix, with shape U×U, b att is the attention bias, with shape U, α t is the attention weight at the t-th time step, X att is the output after weighted summation, with a shape of N×U, which serves as the input of the second layer of LSTM; t It represents the attention score at time step t, which reflects the degree to which time step t should be paid attention to in the model;

[0144] The second layer LSTM subunit is used to process the output of the attention mechanism layer based on the second layer LSTM to obtain the feature vector. The calculation formula is as follows:

[0145] X lstm2 =LSTM(X att ,W lstm2 ,b lstm2 )

[0146] Where W lstm2 is the weight matrix of the second layer LSTM, b lstm2 is the bias term of the second layer LSTM, X lstm2 is the feature vector.

[0147] Embodiment 3

[0148] Due to the complexity of rockburst mechanism and numerous influencing factors, the selection of characteristic indicators for rockburst risk prediction is particularly important. However, there is no clear unified standard for rockburst risk in related studies. This paper collects 350 sets of rockburst sample data from home and abroad, conducts Kendall correlation analysis, and selects three appropriate characteristic indicators based on the correlation heat map of the analysis results, namely: stress index (σ θ / σ c ), brittleness index (σ c / σ t), rock elastic energy index (W et ), where σ θ is the maximum tangential stress of the surrounding rock of the cave wall; c is the uniaxial compressive strength of rock; t is the uniaxial tensile strength of rock.

[0149] According to the selected rockburst risk characteristic indicator results, the selected characteristic indicators and rockburst risk label indicators are retained from the original data set as a new data set. Some data are shown in Table 1. The rockburst risk is divided into four levels: no rockburst, mild rockburst, moderate rockburst and severe rockburst, represented by 1, 2, 3 and 4 respectively. The collected data includes 51 cases of no rockburst data, 104 cases of mild rockburst, 138 cases of moderate rockburst and 57 cases of severe rockburst. The three characteristic indicators are standardized and the model training prediction is carried out.

[0150] Table 1

[0151]

[0152] In order to verify the performance of the CADL model proposed in this paper in rockburst risk prediction, the accuracy, precision, recall rate and F1 are used as model evaluation indicators to conduct comparative experiments with KNN, SVM, and DNN models, and ablation experiments are conducted with CNN, LSTM, and CNN-LSTM for comparative analysis. The results are shown in Table 2.

[0153] Table 2

[0154]

[0155] Comparative analysis of results: The CADL model proposed in this paper predicts rockburst risk with an accuracy of 98.6%, a precision of 98.8%, a recall rate of 98.6%, and an F1 of 0.986. In the comparative experiment, all evaluation indicators are far superior to the traditional machine learning algorithms KNN, SVM, and deep learning algorithm DNN, with an accuracy increase of 48.6%, 41.5%, and 31.1%, respectively. In the ablation experiment, all indicators are improved to varying degrees by CNN, LSTM, and CNN-LSTM models, with an accuracy increase of 31.5%, 34.2%, and 28.6%, respectively. The rockburst risk prediction results of different models in the ablation experiment are compared and displayed through confusion matrix diagrams.

[0156] All models can accurately predict no rockburst. The inaccurate predictions of CNN and CNN-LSTM models are mainly concentrated in moderate rockburst and severe rockburst. The LSTM model has prediction errors when rockburst occurs (mild rockburst, moderate rockburst and severe rockburst), and the effect is not good. The CADL model proposed in the present invention basically predicts correctly at all levels of rockburst risk, with only one case of prediction deviation, which greatly improves the accuracy of rockburst risk prediction.

[0157] The CADL model proposed in this paper has made significant progress in the rockburst risk prediction method. Through experimental comparison with multiple models such as KNN, SVM, DNN, CNN, LSTM and CNN-LSTM under the same data set, the results show that the prediction performance of the CADL model is significantly better than other models, successfully solving the problem of low accuracy in rockburst risk prediction. This achievement provides more reliable technical support for rockburst risk management.

[0158] The CADL model can not only accurately predict rockburst risks, but can also be widely used in a variety of underground projects, thereby effectively reducing or even avoiding damage to engineering equipment and effectively protecting the lives of on-site workers. This research provides an important theoretical basis and practical guidance for improving the safety and reliability of underground projects.

[0159] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. The rockburst risk prediction method combining LSTM and convolutional neural network is characterized by: The following steps are involved: Collect rockburst sample data, and perform Kendall correlation analysis on the rockburst sample data to obtain correlation analysis values; wherein the rockburst sample data includes maximum tangential stress of surrounding rock of cave wall, uniaxial compressive strength of rock, uniaxial tensile strength of rock, stress index, brittleness index, rock elastic energy index and rockburst risk level label column; Based on the correlation analysis value, a rockburst risk feature that satisfies a preset correlation threshold with the rockburst risk level label column is obtained; Based on the rockburst risk characteristics and the rockburst risk level label column, construct a data set; divide the data set into a training set and a test set; Based on the long short-term memory network LSTM and the convolutional neural network CNN, a rockburst risk prediction model is constructed; and the rockburst risk prediction model is trained using the training set; wherein the long short-term memory network LSTM is a double-layer long short-term memory network Double LSTM that introduces the attention mechanism Attention; Based on the trained rockburst risk prediction model, the rockburst risk level of the test set is predicted to complete the rockburst risk prediction.

2. The method according to claim 1, characterized in that: In the data set, Rockburst risk characteristics include stress index, brittleness index and rock elastic energy index; The rockburst risk level label columns include no rockburst, mild rockburst, moderate rockburst, and severe rockburst.

3. The method according to claim 1, characterized in that The structure of the rockburst risk prediction model includes: a CNN layer, a LSTM-ATT-LSTM layer and an output layer; The CNN layer is used to extract the features of the training set based on the one-dimensional convolution layer, and downsample the extracted features using the maximum pooling method to obtain key features; the calculation formula is as follows: X conv [i,j,k]=ReLU(X input [i,j:j+k,:]*W conv [k,:,:]+b conv ), X pool [i,j]=max(X conv [i,2j,:],X conv [i,2j+1,:]), Where, X input is the data sequence of the training set of the input model, N is the number of samples, T is the sequence length, and F is the number of features of each sequence; W conv is the convolution kernel, with a shape of k×F×F conv , F conv is the number of convolution kernels, b conv is the bias term; ReLU(z)=max(0,z) is the ReLU activation function, X conv [i,j,k] is the output of the convolutional layer, X pool [i,j] is the key feature output after pooling; The LSTM-ATT-LSTM layer includes a first LSTM layer, an attention mechanism layer and a second LSTM layer, which are used to process the key features and obtain feature vectors; wherein the flow of information is controlled by the internal gating structure of the double-layer LSTM, and weights are dynamically assigned to the key features by the attention mechanism Attention; wherein the gating structure includes a forget gate, an input gate and an output gate, and the gating structure controls the flow of information through a sigmoid function and a tanh function; The output layer is used to obtain a prediction result based on the feature vector, and use a classification cross entropy loss function to obtain the difference between the prediction result and the true value.

4. The method according to claim 3, characterized in that: The method for the LSTM-ATT-LSTM layer to process the key features includes: Based on the first-layer LSTM processing of the key features, the first-layer LSTM output is obtained, and the calculation formula is as follows: X lstm1 =LSTM(X pool ,W lstm1 ,b lstm1 ) Where W lstm1 is the first layer LSTM weight matrix, b lstm1 is the first layer LSTM bias term, X lstm1 is the output of the first layer of LSTM; Based on the attention mechanism layer, the weighted sum of the first layer LSTM output is performed to obtain the output of the attention mechanism layer. The calculation formula is as follows: e t =tanh(X lstm1 [t]W att +b att ) Where, X lstm1 [t] is the output of the first layer LSTM at the tth time step, with a shape of N×U, W att is the attention weight matrix, with shape U×U, b att is the attention bias, with shape U, α t is the attention weight at the t-th time step, X att is the output after weighted summation, with a shape of N×U, which serves as the input of the second layer of LSTM; t represents the attention score at time step t; Based on the output of the attention mechanism layer processed by the second layer LSTM, the feature vector is obtained, and the calculation formula is as follows: X lstm2 =LSTM(X att ,W lstm2 ,b lstm2 ) Where W lstm2 is the weight matrix of the second layer LSTM, b lstm2 is the bias term of the second layer LSTM, X lstm2 is the feature vector.

5. The method according to claim 1, characterized in that It also includes evaluating the prediction performance of the rockburst risk prediction model based on evaluation indicators, wherein the evaluation indicators include accuracy, recall rate and F1 value; wherein the F1 value is the harmonic mean of the accuracy rate and the recall rate.

6. A rockburst risk prediction system combining LSTM and convolutional neural network, used to implement the method according to any one of claims 1 to 5, characterized in that: include: A sample collection module is used to collect rock burst sample data and perform Kendall correlation analysis on the rock burst sample data to obtain a correlation analysis value; wherein the rock burst sample data includes the maximum tangential stress of the surrounding rock of the cave wall, the uniaxial compressive strength of the rock, the uniaxial tensile strength of the rock, the stress index, the brittleness index, the rock elastic energy index and the rock burst risk level label column; A correlation analysis module, configured to obtain, based on the correlation analysis value, a rockburst risk feature that satisfies a preset correlation threshold with the rockburst risk level label column; A data set construction module is used to construct a data set based on the rockburst risk characteristics and the rockburst risk level label column; divide the data set into a training set and a test set; A rockburst risk prediction model construction module is used to construct a rockburst risk prediction model based on a long short-term memory network (LSTM) and a convolutional neural network (CNN); and train the rockburst risk prediction model using the training set; wherein the long short-term memory network (LSTM) is a double-layer long short-term memory network (Double LSTM) that introduces an attention mechanism (Attention); The rockburst risk prediction module is used to predict the rockburst risk level of the test set based on the trained rockburst risk prediction model to complete the rockburst risk prediction.

7. The system according to claim 6, characterized in that The structure of the rockburst risk prediction model includes a CNN layer, an LSTM-ATT-LSTM layer and an output layer; wherein the LSTM-ATT-LSTM layer includes a first LSTM layer, an attention mechanism layer and a second LSTM layer; The key feature extraction unit is used to extract the features of the input sequence based on the one-dimensional convolution layer of the CNN layer, and downsample the extracted features using the maximum pooling method to obtain the key features; the calculation formula is as follows: X conv [i,j,k]=ReLU(X input [i,j:j+k,:]*W conv [k,:,:]+b conv ), X pool [i,j]=max(X conv [i,2j,:],X conv [i,2j+1,:]), Where, X input is the data sequence of the training set of the input model, N is the number of samples, T is the sequence length, and F is the number of features of each sequence; W conv is the convolution kernel, with a shape of k×F×F conv , F conv is the number of convolution kernels, b conv is the bias term; ReLU(z)=max(0,z) is the ReLU activation function, X conv [i,j,k] is the output of the convolutional layer, X pool [i,j] is the key feature output after pooling; A feature vector acquisition unit, used to process the key features and obtain feature vectors; wherein the flow of information is controlled by the internal gating structure of the double-layer LSTM, and weights are dynamically assigned to the key features by the attention mechanism Attention; wherein the gating structure includes a forget gate, an input gate and an output gate, and the gating structure controls the flow of information through a sigmoid function and a tanh function; The prediction unit is used to obtain a prediction result based on the feature vector in the output layer, and obtain the difference between the prediction result and the true value by using a classification cross entropy loss function.

8. The system according to claim 7, characterized in that The feature vector acquisition unit comprises: The first-layer LSTM subunit is used to process the key features based on the first-layer LSTM to obtain the first-layer LSTM output. The calculation formula is as follows: X lstm1 =LSTM(X pool ,W lstm1 ,b lstm1 ) Where W lstm1 is the first layer LSTM weight matrix, b lstm1 is the first layer LSTM bias term, X lstm1 is the output of the first layer of LSTM; The attention mechanism subunit is used to perform weighted summation on the output of the first layer LSTM based on the attention mechanism layer to obtain the output of the attention mechanism layer. The calculation formula is as follows: e t =tanh(X lstm1 [t]W att +b att ) Where, X lstm1 [t] is the output of the first layer LSTM at the tth time step, with a shape of N×U, W att is the attention weight matrix, with shape U×U, b att is the attention bias, with shape U, α t is the attention weight at the t-th time step, X att is the output after weighted summation, with a shape of N×U, which serves as the input of the second layer of LSTM; t represents the attention score at time step t; The second layer LSTM subunit is used to process the output of the attention mechanism layer based on the second layer LSTM to obtain the feature vector. The calculation formula is as follows: X lstm2 =LSTM(X att ,W lstm2 ,b lstm2 ) Where W lstm2 is the weight matrix of the second layer LSTM, b lstm2 is the bias term of the second layer LSTM, X lstm2 is the feature vector.

Images