Coal mine rock burst data enhancement and deep fusion early warning method under small sample condition

Abstract

The application discloses a coal mine rock burst data enhancement and deep fusion early warning method under a small sample condition, and utilizes a generative adversarial network to expand small sample coal mine rock burst data into large data volume samples, so as to solve the problem that accident samples are far less than normal samples in the rock burst monitoring and early warning field, and a large number of accident samples cannot meet the deep learning process; then, a mixed deep neural network model is constructed according to characteristic quantities, an LSTM is utilized to extract the characteristics of a time sequence, and an attention mechanism is utilized to assign weights to different characteristics, so as to solve the imbalance problem of different characteristics among samples, and further enhance the accuracy of rock burst early warning.

Description

Technical Field

[0001] This invention relates to the field of coal mine rockburst monitoring and early warning technology, and in particular to a method for enhancing and deeply fusing coal mine rockburst data under small sample conditions for early warning. Background Technology

[0002] Rockburst refers to a dynamic phenomenon where the stress in the surrounding rock exceeds the bearing capacity of the coal-rock system, causing sudden deformation and failure of the coal-rock mass, releasing a large amount of energy, resulting in the destruction of the mining space and generating strong vibrations. Due to the increasing depth and intensity of coal mining, the frequency and intensity of rockburst events are constantly increasing. Rockbursts are characterized by their suddenness, unpredictability, and high destructiveness, posing a significant threat to coal mine production safety. Therefore, rockburst monitoring and early warning are crucial for rockburst prevention and control.

[0003] The occurrence of rockbursts is accompanied by changes in geophysical signals such as microseisms, ground sounds, electromagnetic radiation, and seismic wave CT. Early warning of rockbursts can be achieved by monitoring these signal changes. Different monitoring technologies reflect different parameters, have different monitoring ranges and characteristics. Microseismic methods are mainly used for monitoring low-frequency (0–150 Hz) and high-energy (greater than 100 J) mine tremors, but their positioning accuracy is affected by objective factors such as the physical and mechanical properties of coal and rock strata and the structure of the mining area, as well as subjective factors such as the accuracy of data acquisition and wave velocity models. Acoustic emission is mainly used for monitoring high-frequency (greater than 150 Hz) and low-energy (less than 100 J) mine tremors; however, due to the high monitoring frequency, it is easily interfered with by underground mechanical operations, resulting in unclear precursory information for rockbursts. Electromagnetic radiation methods use non-contact monitoring technology to collect electromagnetic waves released by coal and rock deformation and fracturing, providing a more comprehensive reflection of the rockburst evolution process, but they are easily interfered with by local operations or high-voltage equipment, affecting the assessment of rockburst hazard. Therefore, relying solely on changes in a single geophysical signal for early warning of rockbursts is insufficient and cannot fully reflect the precursor information of rockbursts. It is necessary to comprehensively utilize multiple monitoring methods to conduct multi-physical quantity fusion monitoring and early warning of rockbursts.

[0004] The limited number of rockburst samples and imbalanced data restrict the improvement of model early warning accuracy. High-quality rockburst disaster data samples are a prerequisite for building rockburst early warning models based on machine learning. The rockburst sample data is far less than the normal sample data, resulting in a severe imbalance in the sample size. This leads to the model being prone to overfitting to the features of normal data, causing precursory features of rockbursts to be ignored. Furthermore, the sample size of different types of rockburst phenomena also varies greatly, making it easy for the model to ignore certain abnormal data. This results in poor performance in the detection and classification of abnormal data, failing to fully explore the different precursory features of rockbursts. The inability to fully and comprehensively explore features causes deep learning models to concentrate on the feature space regions with high activity, while other regions, with insufficient samples, cannot be effectively learned by the model. These factors severely limit the improvement of rockburst early warning accuracy, and solving the problem of accurate rockburst early warning under small sample conditions is an urgent need for current safety production.

[0005] Patent application publication number CN113723595A discloses an intelligent early warning method for coal mine rockbursts based on quantitative prediction of microseismic events. It trains a database of microseismic events monitored underground in coal mines and uses the established microseismic event quantitative prediction model MSNet to predict the temporal and spatial locations of future microseismic events. The MSNet model combines convolutional neural networks, recurrent neural networks, skip recurrent networks, and autoregressive models to achieve integrated intelligent early warning based on data-driven principles. However, its rockburst early warning based solely on changes in microseismic events cannot comprehensively reflect the precursor information of rockbursts. Furthermore, the sample size is small when using microseismic changes for rockburst early warning, failing to fully reflect the characteristics of precursor information and resulting in low accuracy in rockburst early warning. Summary of the Invention

[0006] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a method for enhancing and deeply fusing early warning of coal mine rockburst data under small sample conditions.

[0007] Technical solution: The present invention provides a method for enhancing and deeply fusing coal mine rockburst data under small sample conditions, comprising the following steps:

[0008] Step 1: Acquire the acoustic and electrical data of rockburst in coal mine and perform noise reduction preprocessing. Perform mode decomposition on the preprocessed rockburst acoustic and electrical data and optimize the modal components after mode decomposition.

[0009] Step 2: Convert the optimized modal components into time series data, filter and denoise the time series data according to the multi-scale permutation entropy, and use the filtered and denoised time series data as the real data.

[0010] Step 3: Construct a generative adversarial network (GAN) and train it using real data. The GAN consists of a generator and a discriminator. The generator receives real data and outputs fake data that is similar to the real data. The discriminator distinguishes between real and fake data.

[0011] Step 4: Input real data into the trained generative adversarial network, and use the fake data that the discriminator judges to be real as the output data of the generative adversarial network. Use the fake data to form a large number of samples.

[0012] Step 5: Construct a hybrid deep neural network model, using a large amount of data samples as input and the early warning results as output. Use the trained hybrid deep neural network model to monitor and issue early warnings for coal mines.

[0013] Furthermore, in step 1, the preprocessed rockburst acoustic-electric data undergoes modal decomposition, and the modal components after modal decomposition are optimized, including:

[0014] The preprocessed rockburst acoustic-electric data is decomposed into multiple modal component IMFs and normalized. The modal and center frequencies of the normalized IMFs are obtained using a modal component variational constraint model, expressed as follows:

[0015]

[0016] Where k is the number of mode decomposition layers, u k (t) represents the k-th modal component, ω k x(t) is the center frequency of the k-th modal component, x(t) is the original signal, δ(t) represents the Hilbert transform, and j represents the phase shift.

[0017] The modal components and center frequency are iteratively updated using the alternating direction multiplier method until the required accuracy is met, at which point the iteration stops. The calculation expression during the iterative update process is as follows:

[0018]

[0019]

[0020]

[0021] in, This represents the modal components obtained in the nth iteration. Represents the original signal. This represents the i-th modal component in the nth iteration. Represents the Lagrange multipliers. This represents the center frequency obtained in the (n-1)th iteration. Let α represent frequency, β represent quadratic penalty factor, and β represent penalty factor.

[0022] Furthermore, step 2 specifically includes the following steps:

[0023] Step 21, denote the optimized modal components IMFs as X = {x} i The sequence X is coarsened to obtain a coarse-grained sequence, where i = 1, 2, ..., N, and N represents the length. The expression is:

[0024]

[0025] Where s is the scaling factor, s = 1, 2, ..., [N / s] represents rounding down N / s. When s = 1, the coarse-grained sequence is the original sequence.

[0026] Step 22, for y j (s) Perform time series reconstruction, denoted as Where l represents the l-th reconstructed component, τ represents the delay time, and m represents the embedding dimension;

[0027] Step 23: Sort the obtained reconstructed time series in ascending order, denoted as S(g)=(j1,j2,…,j m ); where g represents the g-th reordered component, g = 1, 2, ..., k, k ≤ m! ;

[0028] Step 24, calculate the probability of each symbol sequence occurring, using the following expression:

[0029] P g ={g=1,2,…,k},

[0030] Step 25: Calculate and normalize the entropy to obtain the multi-scale permutation entropy. At this point, the time series X = {x} i The entropy of the group i = 1, 2, ..., N is defined as follows: H p (m) is normalized, i.e., H p =H p (m) / ln(m!), H p The value range is [0,1];

[0031] Step 26, H p The IMF component is compared with the preset value and removed if it is greater than the preset value, thus completing the filtering and noise reduction.

[0032] Furthermore, step 3 specifically includes:

[0033] A generator is constructed, including a variational autoencoder (VAE) and a long short-term memory recurrent neural network model (LSTM). The generator's loss function consists of the sum of mean squared error loss, binary cross-entropy loss, and KL divergence error.

[0034] Construct discriminators, including VAE and LSTM.

[0035] Furthermore, step 3, training the generative adversarial network using real data, includes:

[0036] After inputting real data into the VAE in the generator, the output of the VAE is input into the LSTM. The LSTM outputs fake data that is similar to the real data, and the fake data is used as training samples.

[0037] Real and fake data are input into the discriminator, and the discriminator parameters are iteratively updated to generate a cluster of error discrimination results based on cross-entropy. The error discrimination results are fed back to the generator through the backpropagation algorithm to update the generator parameters. During the iteration process, the generator always tries to generate more realistic samples, and the discriminator always tries to identify the difference between fake and real data. The generator and discriminator engage in a minimax game and eventually reach Nash equilibrium, thus completing the training of the adversarial generative network.

[0038] Furthermore, when the generator and discriminator engage in a mini-game to optimize the parameters of the generative adversarial network, the loss function used is:

[0039]

[0040] Where, k(x) i ,x j ) represents (x i ,x j The probability distribution of k(y) i ,y j ) represents y i ,y j The probability distribution is given by n, where n represents the number of samples.

[0041] Furthermore, the hybrid deep neural network model includes a convolutional layer, a long short-term memory recurrent neural network, an attention layer, and an activation function layer arranged sequentially. The convolutional layer is used to extract contour features of large data samples, the long short-term memory recurrent neural network is used to extract temporal features of signals, and the attention layer is used to capture important temporal features.

[0042] Furthermore, step 5, constructing the hybrid deep neural network model, includes:

[0043] Construct an input layer, which specifies the batch size, time steps, and feature dimensions of the input data. The input data is denoted as a real number sequence matrix R. t×n Let x be the number of digits. i For R t×n Vector representation of the data at the i-th time step;

[0044] Convolutional layers are constructed by feeding input data through the input layer. The convolutional kernels perform convolution along a single temporal direction, extracting features once for the sequence vector at every k time steps, resulting in a feature O. i The calculation formula is as follows:

[0045]

[0046] Where, x i:i+k-1 W1 is a real matrix from the i-th time step to the (i+k-1)-th time step in t×n, f is a non-linear activation function, and b1∈R is a bias.

[0047] After passing through r convolutional kernels, r feature maps are obtained. Then, max pooling is performed with a pooling size of 2 and a stride of 2, resulting in r feature maps O of shape [(t-k+1) / 2]×1. The calculation formula is as follows:

[0048] O = max{O i O i+1}, i = 1, 3, 5, ..., tk

[0049] The r feature maps are the features extracted by the convolutional layer. They are reduced to a real vector of length r*(t-k+1) / 2, which is then used as input to the long short-term memory recurrent neural network for processing.

[0050] To construct a Long Short-Term Memory Recurrent Neural Network, four linear transformations are required to concatenate the input with the hidden state of the previous time step.

[0051] An attention mechanism is constructed and calculated. The attention layer is used to assign weights to the features extracted from both long and short time series, and the weighted average of the output vectors from the long and short time series layers is calculated. The specific formula is as follows:

[0052]

[0053] Among them, H i For the output of the hidden layer of the Long Short-Term Memory recurrent neural network, α i S is the weighting coefficient. i The score output for each hidden layer, C i This is the result after weighted summation;

[0054] An activation function layer is constructed and used as the output layer to output the early warning results of rockburst.

[0055] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0056] This invention utilizes a generator to produce training samples that approximate the true values. These generated training samples and the true samples are then input into a discriminator. By processing the two sets of data with different true values, a cluster of error discrimination results based on cross-entropy is generated. The error is then fed back to the generator and discriminator via a backpropagation algorithm. The parameters of the generative adversarial network (GAN) are iteratively modified to correct the model. During the iteration process, the generator and discriminator engage in a game-like interaction, eventually reaching Nash equilibrium, resulting in good model training results. The GAN expands the sample size, addressing the problem in rockburst monitoring and early warning where accident samples are far smaller than normal samples, failing to meet the need for a large number of accident samples in deep learning. A hybrid deep neural network model is constructed based on feature quantities. This model consists of convolutional layers capable of extracting contour features, long short-term memory layers capable of extracting signal temporal features, and attention layers capable of capturing important temporal components. LSTM is used to extract time-series features, and the attention mechanism assigns weights to different features, thus solving the problem of imbalance between different features among samples. Compared with traditional methods, this further enhances the accuracy of rockburst early warning. Attached Figure Description

[0057] Figure 1 This is a flowchart of the coal mine rockburst data enhancement and deep fusion early warning method under small sample conditions according to the present invention;

[0058] Figure 2 This is a schematic diagram of the structure of a hybrid deep neural network model. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0060] The flowchart of the coal mine rockburst data enhancement and deep fusion early warning method under small sample conditions described in this embodiment is as follows: Figure 1 The above includes the following steps:

[0061] Step 1: Acquire the acoustic and electrical data of rockburst in coal mine and perform noise reduction preprocessing. Perform mode decomposition on the preprocessed rockburst acoustic and electrical data and optimize the modal components after mode decomposition.

[0062] Specifically, in step 1, the preprocessed impact pressure acoustic-electric data is subjected to modal decomposition, and the modal components after modal decomposition are optimized, including:

[0063] The preprocessed rockburst acoustic-electric data is decomposed into multiple modal component IMFs and normalized. The modal and center frequencies of the normalized IMFs are obtained using a modal component variational constraint model, expressed as follows:

[0064]

[0065] Where k is the number of mode decomposition layers, u k (t) represents the k-th modal component, ω k x(t) is the center frequency of the k-th modal component, x(t) is the original signal, δ(t) represents the Hilbert transform, and j represents the phase shift.

[0066] The modal components and center frequency are iteratively updated using the alternating direction multiplier method until the required accuracy is met, at which point the iteration stops. The calculation expression during the iterative update process is as follows:

[0067]

[0068]

[0069]

[0070] in, This represents the modal components obtained in the nth iteration. Represents the original signal. This represents the i-th modal component in the nth iteration. Represents the Lagrange multipliers. This represents the center frequency obtained in the (n-1)th iteration. Here, α represents the frequency, β represents the quadratic penalty factor, and β represents the penalty factor.

[0071] The iteration continues until the function satisfies the following condition, at which point the iteration ends:

[0072]

[0073] In the formula, ε represents the judgment precision.

[0074] Step 2: Convert the optimized modal components into time series data, filter and denoise the time series data according to the multi-scale permutation entropy, and use the filtered and denoised time series data as the real data.

[0075] Specifically, step 2 above includes the following steps:

[0076] Step 21, denote the optimized modal components IMFs as X = {x} iThe sequence X is coarsened to obtain a coarse-grained sequence, where i = 1, 2, ..., N, and N represents the length. The expression is:

[0077]

[0078] Where s is the scaling factor, s = 1, 2, ..., [N / s] represents rounding down N / s. When s = 1, the coarse-grained sequence is the original sequence.

[0079] Step 22, for y j (s) Perform time series reconstruction, denoted as Where l represents the l-th reconstructed component, τ represents the delay time, and m represents the embedding dimension;

[0080] Step 23: Sort the obtained reconstructed time series in ascending order, denoted as S(g)=(j1,j2,…,j m ); where g represents the g-th reordered component, g = 1, 2, ..., k, k ≤ m!, and S(g) has m! different arrangements, with the probability of being arranged in ascending order being one of m!.

[0081] Step 24, calculate the probability of each symbol sequence occurring, using the following expression:

[0082] P g ={g=1,2,…,k},

[0083] Step 25, when P g When the probability of each symbol sequence is equal to 1 / m!, the time series has the highest complexity and the largest permutation entropy, which is lnm!. For ease of representation, entropy calculation and normalization are performed to obtain the multi-scale permutation entropy. At this time, the time series X = {x} i The entropy of the group i = 1, 2, ..., N is defined as follows: H p (m) is normalized, i.e., H p =H p (m) / ln(m!), H p The value range is [0,1];

[0084] Step 26, H p The IMF components that are greater than the preset value (e.g., 0.6) are removed to complete the filtering and noise reduction.

[0085] Step 3: Construct a generative adversarial network (GAN) and train it using real data. The GAN consists of a generator and a discriminator. The generator receives real data and outputs fake data that is similar to the real data. The discriminator distinguishes between real and fake data.

[0086] Specifically, step 3 above includes:

[0087] A generator is constructed, including a variational autoencoder (VAE) and a long short-term memory recurrent neural network model (LSTM). The generator's loss function consists of the sum of mean squared error loss, binary cross-entropy loss, and KL divergence error.

[0088] Construct discriminators, including VAE and LSTM.

[0089] Furthermore, a variational autoencoder (VAE) is constructed. Using reparameterization, the latent variable z follows a Gaussian distribution, with the formula: z = μ·N(0,1) + σ, where σ and μ represent the mean and variance of the posterior distribution of the latent variable z.

[0090] Furthermore, a Long Short-Term Memory Recurrent Neural Network (LSTM) model is constructed as the encoder. First, the four linear transformations required by the LSTM are defined to concatenate the input and the hidden state of the previous time step. An attention mechanism is constructed and calculated. Then, the input gate, forget gate, output gate, and memory gate are calculated to calculate the new unit state and hidden state.

[0091] The formula used to calculate the input gate is as follows:

[0092] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0093] In the formula, i t W represents the signal output by the input gate at time t. i and b i h represents the parameter of the i-th sigmoid activation function. t-1 x represents the previous output signal of the LSTM at time t-1. t This indicates that the input signal at time t is received.

[0094] The forget gate consists of a sigmoid neural network layer and a bitwise multiplication operation. The formula used to calculate the forget gate is as follows:

[0095] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0096] In the formula, W f and b f These are the parameters of the sigmoid neural network layer.

[0097] The memory gate will determine the newly input information x t and h t-1 Which information will be retained, including a parameter W f and b f A sigmoid neural network layer and a parameter W c and b c The formula used to calculate the memory gates in the tanh neural network layer is as follows:

[0098] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0099]

[0100] The output f of the forget gate t Compared to the previous cell state C t-1 Multiplication is used to select between forgetting and retaining some information. The output of the memory gate is added to the information selected from the forget gate to obtain the new cell state C. t The formula used is as follows:

[0101]

[0102] This indicates the cell state C at time t. t It already contains the information transmitted at time t-1 that needs to be discarded, and the new information acquired from the input signal at time t that needs to be added. C t It will continue to be passed on to the LSTM at time t+1, and will be passed on as a new cell state.

[0103] The output gate is the cell state C transmitted from time t-1 after passing through the forget gate and memory gate. t-1 The output signal h at time t-1 t-1 and the input signal x at time t t The signals are combined and used as the output signal at the current moment. The formula used to calculate the output gate is as follows:

[0104] o t =σ(W o [h t-1 ,x t ]+b o )

[0105] h t =o t *tanh(C t )

[0106] In the formula, W o and b o These are the parameters of the sigmoid neural network layer, C t The signal is passed through a tanh function to a value between -1 and 1, and then multiplied by this value to obtain the output signal h. t .

[0107] Furthermore, the formula for calculating the mean squared error loss (MSE) is as follows:

[0108]

[0109] In the formula, y and y' are the real data label and the training sample label, respectively.

[0110] Furthermore, the formula for calculating the binary cross-entropy loss (BCE) is as follows:

[0111]

[0112] In the formula, y is the true label, and p(y) is the probability that the model output belongs to the label y.

[0113] Furthermore, the formula for calculating the KL divergence error is as follows:

[0114]

[0115] In the formula, P represents the distribution of the real data, and Q represents the distribution of the training samples.

[0116] Specifically, step 3, training the generative adversarial network using real data, includes:

[0117] After inputting real data into the VAE in the generator, the output of the VAE is input into the LSTM. The LSTM outputs fake data that is similar to the real data, and the fake data is used as training samples.

[0118] Real and fake data are input into the discriminator, and the discriminator parameters are iteratively updated to generate a cluster of error discrimination results based on cross-entropy. The error discrimination results are fed back to the generator through the backpropagation algorithm to update the generator parameters. During the iteration process, the generator always tries to generate more realistic samples, and the discriminator always tries to identify the difference between fake and real data. The generator and discriminator engage in a minimax game and eventually reach Nash equilibrium, thus completing the training of the adversarial generative network.

[0119] Furthermore, when the generator and discriminator engage in a mini-game to optimize the parameters of the generative adversarial network, the loss function used is:

[0120]

[0121] Where, k(x) i ,x j ) represents (x i ,x j The probability distribution of k(y) i ,y j ) represents y i ,y j The probability distribution is given by n, where n represents the number of samples.

[0122] Step 4: Input real data into the trained generative adversarial network, and use the fake data that the discriminator judges to be real as the output data of the generative adversarial network. Use the fake data to form a large number of samples.

[0123] Step 5: Construct a hybrid deep neural network model, using a large amount of data samples as input and the early warning results as output. Use the trained hybrid deep neural network model to monitor and issue early warnings for coal mines.

[0124] like Figure 2 As shown, the hybrid deep neural network model includes a convolutional layer, a long short-term memory recurrent neural network, an attention layer, and an activation function layer arranged sequentially. The convolutional layer is used to extract contour features of large data samples, the long short-term memory recurrent neural network is used to extract temporal features of signals, and the attention layer is used to capture important temporal features.

[0125] Specifically, the construction of the hybrid deep neural network model in step 5 above includes:

[0126] Construct an input layer, which specifies the batch size, time steps, and feature dimensions of the input data. The input data is denoted as a real number sequence matrix R. t×n Let x be the number of digits. i For R t×n Vector representation of the data at the i-th time step;

[0127] Convolutional layers are constructed by feeding input data through the input layer. The convolutional kernels perform convolution along a single temporal direction, extracting features once for the sequence vector at every k time steps, resulting in a feature O. i The calculation formula is as follows:

[0128]

[0129] Where, x i:i+k-1W1 is a real matrix from the i-th time step to the (i+k-1)-th time step in t×n, f is a non-linear activation function, and b1∈R is a bias.

[0130] After passing through r convolutional kernels, r feature maps are obtained. Then, max pooling is performed with a pooling size of 2 and a stride of 2, resulting in r feature maps O of shape [(t-k+1) / 2]×1. The calculation formula is as follows:

[0131] O = max{O i O i+1}, i = 1, 3, 5, ..., tk

[0132] The r feature maps are the features extracted by the convolutional layer. They are reduced to a real vector of length r*(t-k+1) / 2, which is then used as input to the long short-term memory recurrent neural network for processing.

[0133] To construct a Long Short-Term Memory Recurrent Neural Network, four linear transformations are required to concatenate the input with the hidden state of the previous time step.

[0134] An attention mechanism is constructed and calculated. The attention layer is used to assign weights to the features extracted from both long and short time series, and the weighted average of the output vectors from the long and short time series layers is calculated. The specific formula is as follows:

[0135]

[0136] Among them, H i For the output of the hidden layer of the Long Short-Term Memory recurrent neural network, α i C is the weighting coefficient. i This is the result after weighted summation;

[0137] An activation function layer is constructed and used as the output layer to output the early warning result of rockburst. Its calculation expression is as follows:

[0138]

[0139] Where M is the maximum value of the vector, M = max(C), C i and C j It is an element in the vector.

Claims

1. A method for enhancing and deeply fusing coal mine rockburst data for early warning under small sample conditions, characterized in that, Includes the following steps: Step 1: Acquire the acoustic and electrical data of rockburst in coal mine and perform noise reduction preprocessing. Perform mode decomposition on the preprocessed rockburst acoustic and electrical data and optimize the modal components after mode decomposition. Step 2: Convert the optimized modal components into time series data, filter and denoise the time series data according to the multi-scale permutation entropy, and use the filtered and denoised time series data as the real data. Step 3: Construct a generative adversarial network (GAN) and train it using real data. The GAN consists of a generator and a discriminator. The generator receives real data and outputs fake data that is similar to the real data. The discriminator distinguishes between real and fake data. Step 4: Input real data into the trained generative adversarial network, and use the fake data that the discriminator judges to be real as the output data of the generative adversarial network. Use the fake data to form a large number of samples. Step 5: Construct a hybrid deep neural network model, using a large amount of data samples as input and the early warning results as output. Use the trained hybrid deep neural network model to monitor and issue early warnings for coal mines. The hybrid deep neural network model includes a convolutional layer, a long short-term memory recurrent neural network, an attention layer, and an activation function layer arranged sequentially. The convolutional layer is used to extract the contour features of large data samples, the long short-term memory recurrent neural network is used to extract the temporal features of the signal, and the attention layer is used to capture important temporal features. Building a hybrid deep neural network model includes: Construct an input layer that specifies the batch size, time steps, and feature dimensions of the input data, and denotes the input data as a real number sequence matrix. ,remember for The Middle Vector representation of data at each time step; Convolutional layers are constructed by feeding input data through the input layer. The convolutional kernels perform convolutions along a single temporal direction, processing each... Perform feature extraction on the sequence vectors at each time step to obtain a feature. The calculation formula is as follows: ； in, for The Middle The time step to the 1 A real matrix at each time step. yes A real matrix, It is a non-linear activation function. It is a bias; pass After one convolution kernel, we get Each feature map is convolved and then max pooled with a pooling size of 2 and a sliding stride of 2, resulting in... indivual Feature map of shape The calculation formula is as follows: ； The feature maps are features extracted by the convolutional layer, and their dimensionality is reduced to a length of [value missing]. A real vector, which is used as input to the Long Short-Term Memory Recurrent Neural Network for processing; To construct a Long Short-Term Memory Recurrent Neural Network, four linear transformations are required to concatenate the input with the hidden state of the previous time step. An attention mechanism is constructed and calculated. The attention layer is used to assign weights to the features extracted from both long and short time series, and the weighted average of the output vectors from the long and short time series layers is calculated. The specific formula is as follows: ； in, This is the output of the hidden layer of a long short-term memory recurrent neural network. These are the weighting coefficients. The score output for each hidden layer. This is the result after weighted summation; An activation function layer is constructed and used as the output layer to output the early warning results of rockburst.

2. The method for enhancing and deeply fusing coal mine rockburst data under small sample conditions according to claim 1, characterized in that, Step 1 involves performing modal decomposition on the preprocessed rockburst acoustic-electric data, and then optimizing the modal components after modal decomposition, including: The preprocessed rockburst acoustic-electric data is decomposed into multiple modal component IMFs and normalized. The modal and center frequencies of the normalized IMFs are obtained using a modal component variational constraint model, expressed as follows: ； in, The modal decomposition level is 1. For the first One modal component, For the first The center frequency of each modal component The original signal, Represents the Hilbert transform. Indicates phase shift ; The modal components and center frequency are iteratively updated using the alternating direction multiplier method until the required accuracy is met, at which point the iteration stops. The calculation expression during the iterative update process is as follows: ； ； ； in, This represents the modal components obtained in the nth iteration. Represents the original signal. This represents the i-th modal component in the nth iteration. Represents the Lagrange multipliers. This represents the center frequency obtained in the (n-1)th iteration. Indicates frequency, This represents the secondary penalty factor. Indicates the penalty factor.

3. The method for enhancing and deeply fusing coal mine rockburst data under small sample conditions as described in claim 2, characterized in that, Step 2 specifically includes the following steps: Step 21, denote the optimized modal components IMFs as N represents the length. Coarsening X yields a coarse-grained sequence, expressed as: ； Where s is the scale factor, , This indicates that the integer part of N / s is taken as the integer part of N / s. At that time, the coarse-grained sequence is the original sequence; Step 22, for Perform time series reconstruction, denoted as ;in, Indicates the first A reconstructed component, The delay time is represented by m, and the embedding dimension is represented by m. Step 23: Sort the obtained reconstructed time series in ascending order, denoted as... ;in, Indicates the first A reordered component, ; Step 24, calculate the probability of each symbol sequence occurring, using the following expression: ， ； Step 25: Calculate and normalize the entropy to obtain the multi-scale permutation entropy. At this point, the time series... Defined in the form of entropy ,Will Perform normalization processing, that is , The range of values ​​is ; Step 26, The IMF component is compared with the preset value and removed if it is greater than the preset value, thus completing the filtering and noise reduction.

4. The method for enhancing and deeply fusing coal mine rockburst data under small sample conditions according to claim 3, characterized in that, Step 3 specifically includes: A generator is constructed, including a variational autoencoder (VAE) and a long short-term memory recurrent neural network model (LSTM). The generator's loss function consists of the sum of mean squared error loss, binary cross-entropy loss, and KL divergence error. Construct discriminators, including VAE and LSTM.

5. The method for enhancing and deeply fusing coal mine rockburst data under small sample conditions as described in claim 4, characterized in that, Step 3, training the generative adversarial network using real data, includes: After inputting real data into the VAE in the generator, the output of the VAE is input into the LSTM. The LSTM outputs fake data that is similar to the real data, and the fake data is used as training samples. Real and fake data are input into the discriminator, and the discriminator parameters are iteratively updated to generate a cluster of error discrimination results based on cross-entropy. The error discrimination results are fed back to the generator through the backpropagation algorithm to update the generator parameters. During the iteration process, the generator always tries to generate more realistic samples, and the discriminator always tries to identify the difference between fake and real data. The generator and discriminator engage in a minimax game and eventually reach Nash equilibrium, thus completing the training of the adversarial generative network.

6. The method for enhancing and deeply fusing coal mine rockburst data under small sample conditions as described in claim 5, characterized in that, When the generator and discriminator engage in a mini-game to optimize the parameters of the generative adversarial network, the loss function used is: ； in, express The probability distribution, express The probability distribution, Indicates the number of samples.

Images