A potential time decomposition and convolution cross-variable interaction mine microseismic prediction method

Abstract

A mine microseismic prediction method of potential time decomposition and convolution cross-variable interaction belongs to the technical field of coal mine safety monitoring. It comprises: reversible instance normalization and inverted embedding preprocessing of multivariate microseismic data; trend, seasonality and noise three components are decoupled in high-dimensional latent space through a latent time decomposition module (LTD); local and global dependencies between variables are captured with linear complexity through a convolution cross-variable interaction module (CVI); a prediction sequence is output through inverse normalization of the fused multi-component features through a prediction head; a composite loss function is used to train the network; and the prediction result is used for mine safety state judgment and early warning. The present application significantly improves the accuracy of mine microseismic prediction and provides reliable and safe prediction basis for mining operations.

Description

Technical Field

[0001] This invention belongs to the field of coal mine safety monitoring and intelligent prediction technology, specifically involving a method for applying deep learning time series prediction to the prediction of multivariate signals of mine microseismic events, and particularly involving an end-to-end deep neural network prediction method based on latent time decomposition and convolutional cross-variable interaction. Background Technology

[0002] Microseismic monitoring is a crucial means of ensuring safe coal mining. During coal mining, disturbances cause the coal and rock mass to continuously break up and release elastic energy, resulting in microseismic events. High-energy microseismic events can trigger dynamic coal and rock disasters such as rockbursts and roof collapses, posing a serious threat to mine safety. Therefore, accurate multivariate time series prediction of microseismic monitoring data to detect disaster risks in advance has significant engineering practical implications.

[0003] Mine microseismic monitoring systems typically deploy multiple sensors to collect multi-dimensional time-series signals such as vibration energy, frequency of occurrence, and spatial coordinates in real time, constituting a typical multivariate time series prediction problem. Compared with univariate prediction, multivariate prediction requires both modeling the temporal evolution of each variable and capturing the complex dependencies between variables from different sensors, significantly increasing the difficulty.

[0004] Existing methods for microseismic time series prediction mainly suffer from the following shortcomings: (1) Traditional statistical methods (such as ARIMA) are difficult to model the nonlinear coupling relationship between variables; (2) Transformer-based self-attention mechanisms have limitations in the variable dimension. The computational complexity is high in multi-sensor scenarios; (3) Existing deep learning methods often perform temporal decomposition directly in the original input space, which cannot make full use of the richer feature expression in the high-dimensional potential space; (4) Mine microseismic signals have obvious non-stationarity and distribution shift characteristics, which are difficult to be effectively addressed by ordinary normalization strategies.

[0005] To address the aforementioned issues, there is an urgent need for a multivariate time series prediction method for mine microseismic events that can effectively decouple temporal components in a high-dimensional latent space, efficiently model intervariate dependencies with linear complexity, and suppress the effects of non-stationary distribution migrations. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a mine microseismic prediction method based on potential time decomposition and convolutional intervariate interaction.

[0007] The technical solution of this invention is: a mine microseismic prediction method based on latent time decomposition and convolutional intervariate interaction, the steps of which are as follows:

[0008] Step 1) Preprocess the multivariate time series data of mine microseismic events;

[0009] The specific method is as follows:

[0010] Multiple sensor channels in the mine were selected to collect data including multi-dimensional indicators such as the vibration energy (J), vibration frequency (times / hour), and source location coordinates of microseismic events, with sampling time intervals set. For missing data caused by signal transmission failures, linear interpolation was used to fill in the gaps.

[0011]

[0012] The collected data is recorded as follows: ,in For batch size, For the number of variables, For time step.

[0013] Apply Reversible Instance Normalization (RevIN) to the input data to calculate the mean of each variable along the time dimension. with standard deviation The data is normalized to a zero-mean, unit-variance distribution, and a learnable affine transformation parameter (scaling factor) is applied. With offset coefficient The formula is shown in (2):

[0014]

[0015] in and These are the mean and standard deviation for each variable. and This data is retained during the inference phase and used in the subsequent denormalization operation in step 4. This process eliminates dimensional differences between data features, alleviates the non-stationary distribution shift of microseismic signals, and improves model training stability and cross-time-period generalization ability.

[0016] Subsequently, each normalized sensor variable is treated as an independent token, and the variables are then embedded from the time dimension through a linear embedding layer. Projecting onto a high-dimensional latent feature space, the formula is shown in (3):

[0017]

[0018] in As a hidden layer dimension, the inverted embedding strategy preserves the independence of each variable and avoids interference from subsequent cross-variable operations on the temporal features within the variable.

[0019] Step 2) Decouple the embedded features into three temporal components using the Latent Temporal Decomposition Module (LTD);

[0020] The specific method is as follows:

[0021] This step decouples the time-series features into three components: trend, seasonality, and noise, in a high-dimensional latent feature space, employing a two-level decomposition strategy that combines coarse and fine elements.

[0022] Step 2.1) Coarse trend extraction based on one-dimensional convolution:

[0023] For input high-dimensional features Apply small kernel one-dimensional convolutional layers initialized with uniform distribution. Simulate a learnable moving average operation to extract preliminary trend components, as shown in formula (4):

[0024]

[0025] The preliminary residuals containing seasonal components and noise are obtained by subtraction, as shown in formula (5):

[0026]

[0027] Step 2.2) Adaptive seasonality extraction based on the frequency domain:

[0028] For residuals Perform a Fast Fourier Transform (FFT) on each variable to map the time-domain residual to the frequency domain, as shown in formula (6):

[0029]

[0030] in Indicates the FFT operation. This indicates the length of the transformed frequency domain sequence. Indicates the number of variables. For the th... Statistical mean of spectral amplitude of each variable with standard deviation Set adaptive threshold The formula is shown in (7):

[0031]

[0032] Construct a binary mask to set the frequency components with amplitudes below the threshold to zero, while retaining the high-frequency components, as shown in formula (8):

[0033]

[0034] in This represents element-wise multiplication. It is a binary mask. This adaptive thresholding mechanism can dynamically adjust the filtering boundary according to the spectral characteristics of different microseismic signals, maximizing the separation of low-frequency trend components from the potential space, while avoiding the inclusion of high-frequency noise and periodic components that vary with the signal.

[0035] Perform an inverse FFT (iFFT) on the filtered spectrum to restore the fine-grained trend components in the time domain. And thus separate the seasonal components. The formula is shown in (9):

[0036]

[0037] Step 2.3) Component Summary:

[0038] Integrating the results of the two-level decomposition, the overall trend component Noise components The final output trend of the LTD module. Seasonal ,noise The three orthogonal components correspond to the low-frequency background trend, periodic vibration pattern, and random noise component of the mine microseismic signal, respectively.

[0039] Step 3) Capture the spatiotemporal relationships between variables using the Convolutional Intervariate Module (CVI);

[0040] The specific method is as follows:

[0041] This step utilizes two-dimensional convolution and a long-term memory matrix to... It efficiently models the local spatiotemporal correlations and global dependencies between multi-sensor variables with linear complexity.

[0042] Step 3.1) Two-dimensional convolutional local feature extraction:

[0043] Given an embedding feature tensor This is then reshaped into a view suitable for two-dimensional convolution processing, as shown in formula (10):

[0044]

[0045] Two layers of two-dimensional convolutions with ReLU activation are applied sequentially. The convolution kernels jointly capture the local spatiotemporal correlation of multi-sensor time series in both the time and variable directions, as shown in formula (11):

[0046]

[0047] in The ReLU activation function is used, and the kernel size is determined by the hyperparameter. and Together, we determine and control the receptive field size in the time direction and the variable direction respectively, in order to adapt to the multi-scale features of different mine datasets.

[0048] Step 3.2) Long-term memory matrix and random pooling:

[0049] right By applying a linear projection along the time dimension, local interaction information is compressed into a compact memory space of a fixed size, resulting in a long-term memory matrix. The formula is shown in (12):

[0050]

[0051] in The dimension of the memory matrix is ​​a hyperparameter. To prevent the model from overfitting due to excessive reliance on fixed channels, a stochastic pooling regularization mechanism is introduced: [The text abruptly ends here, so the translation stops as well.] Softmax weights are calculated along the variable dimension and multinomial sampling is performed, randomly selecting variable indices for each memory location and broadcasting them to the original number of variables. , to obtain context-enhanced features The formula is shown in (13):

[0052]

[0053] This stochastic mechanism allows the model to focus on different channels of the memory matrix from a stochastic perspective during each forward propagation, effectively enhancing the model's robustness to the strong noise and micro-vibration environment of mines.

[0054] Step 3.3) Feature Fusion:

[0055] Context-enhanced features With the original embedded input The features are concatenated along the time dimension and then nonlinearly fused through two layers of MLP to output the final cross-variable interaction features, as shown in formulas (14) and (15):

[0056]

[0057]

[0058] The GELU activation function is used for the nonlinear transformation of the first layer of the MLP. Compress the dimension back to the original time dimension , Each variable's features simultaneously contain its own temporal information and a global cross-variable context from the convergence of long-memory matrices.

[0059] Step 4) Fuse the multi-component features, and output the final predicted sequence after passing through the prediction head and inverse normalization;

[0060] The specific method is as follows:

[0061] In each LTDBlock, cross-variable interaction features will be included. Trend components Seasonal component By splicing along the time dimension, a comprehensive feature is formed, as shown in formula (16):

[0062]

[0063] The dimension is changed from through a fully connected layer. Compress to The current block output is obtained, as shown in formula (17):

[0064]

[0065] through After stacking layers of LTDBlock, the encoder's final output Then, the time dimension is shifted from... through a linear prediction head. Mapping to prediction length The formula is shown in (18):

[0066]

[0067] Finally, the mean saved by RevIN in step 1 is used. with standard deviation The prediction results are denormalized to restore the original microseismic data scale, and the final prediction sequence is obtained, as shown in formula (19):

[0068]

[0069] in The final multivariate time series prediction results for mine microseismic events include future... Predicted values ​​of sensor variables at each time step.

[0070] Step 5) Perform end-to-end training of the network using a composite loss function;

[0071] The specific method is as follows:

[0072] To give the decomposition results of the LTD module physical meaning, this invention designs the total loss function as a weighted combination of the prediction loss and the noise regularization term, as shown in formula (20):

[0073]

[0074] Among them, predicted loss The mean square error between the predicted sequence and the actual microseismic sequence is measured by the formula shown in (21):

[0075]

[0076] Noise regularization term Noise components decomposed from LTD Apply Norm penalty forces the model to attribute effective time series signals to trend components. Seasonal component The formula is shown in (22):

[0077]

[0078] in This is a regularization weight hyperparameter used to balance prediction accuracy and decomposition quality; As an auxiliary supervisory signal, the LTD decomposition process is constrained to make the noise term approach zero-mean Gaussian white noise, ensuring that the trend and seasonal components obtain sufficient and clean time-series information, thereby improving the physical interpretability and predictive robustness of the decomposition. The network is trained end-to-end using the Adam optimizer, and the learning rate is scheduled using a cosine annealing strategy.

[0079] Step 6) Input the real-time acquired multivariate microseismic time-series data into the LTDNet network trained in Step 5 to obtain the predicted sequence of each sensor variable for the next H time steps. Based on the prediction results, assess the mine's safety status: if the predicted microseismic energy exceeds a preset safety threshold, it is determined that there is a risk of coal and rock dynamic disasters such as mine rock bursts or roof collapses, triggering a safety warning signal and notifying on-site personnel to take protective measures; otherwise, it is determined that the current mine microseismic activity is at a normal level, and rolling predictions and real-time monitoring continue for subsequent time periods.

[0080] The beneficial effects of this invention are as follows: Reversible instance normalization effectively alleviates the non-stationary distribution migration problem of mine microseismic signals and improves the model's cross-time-period generalization ability; implementing a coarse-to-fine two-level time decomposition in a high-dimensional latent space can accurately separate the trend component, periodic component, and noise component in the microseismic signal, solving the problem of insufficient decomposition in the original space by traditional methods; the convolutional cross-variable interaction module achieves linear complexity. Modeling the correlation between variables from multiple sensors significantly reduces computational overhead compared to the Transformer-based self-attention mechanism, making it suitable for real-time monitoring needs in mines. The noise regularization loss function ensures the physical interpretability of the decomposition results, improves the model's robustness to strong noise micro-seismic environments, and provides more timely and accurate early warnings for high-energy micro-seismic events. This provides a reliable basis for safety prediction in mining operations, effectively reduces the risk of geological disasters, and ensures the safety of personnel and equipment. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0082] Figure 2 This is a schematic diagram illustrating the overall structure and feature fusion method of LTDBlock;

[0083] Figure 3 This is a schematic diagram of the internal structure of the Convolutional Intervariate Module (CVI).

[0084] Figure 4 This is a schematic diagram of the internal structure of the Latent Time Decomposition Module (LTD). Detailed Implementation

[0085] A mine microseismic prediction method based on latent temporal decomposition and convolutional intervariate interaction, comprising the following steps:

[0086] Step 1) Preprocess the multivariate time series data of mine microseismic events;

[0087] Step 2) Decouple the embedded features into three temporal components using the Latent Temporal Decomposition Module (LTD);

[0088] Step 3) Capture the spatiotemporal relationships between variables using the Convolutional Intervariate Module (CVI);

[0089] Step 4) Fuse the multi-component features, and output the final predicted sequence after passing through the prediction head and inverse normalization;

[0090] Step 5) Perform end-to-end training of the network using a composite loss function;

[0091] Step 6) Apply the trained network model to mine microseismic prediction to judge and warn of mine safety status.

[0092] In the Mine Microseismic Multivariate Time Series Prediction (MTSF) task, given an input multivariate microseismic time series... ,in For batch size, For the number of sensor variables, To review the window length, the goal is to predict future sequences. , To predict the horizon, the LTDNet model proposed in this invention is an end-to-end deep neural network, mainly composed of three core parts connected in series: data normalization and embedding layer, and... The encoder consists of stacked LTDBlock layers, a linear prediction head, and an inverse normalized output layer. Within each LTDBlock, a Latent Temporal Decomposition (LTD) module and a Convolutional Intervariate Module (CVI) run in parallel, extracting complementary features from the temporal and variable dimensions, respectively, and finally fusing them to output the prediction result.

[0093] In step 1), a mine is selected. Several key sensor channels collect data including the vibration energy (J) of microseismic events, vibration frequency (times / hour), and P-wave velocity. S-wave velocity Multidimensional microseismic indicators, such as the location coordinates of the seismic source, were used, and the sampling time interval was set to 1 hour. Since signal transmission failures may lead to data loss, linear interpolation was used to fill in the missing data, as shown in formula (1):

[0094]

[0095] The collected data is recorded as follows: For multidimensional microseismic indices with significant differences in dimensions, directly inputting them into the model can lead to training instability. Therefore, this invention employs Reversible Instance Normalization (RevIN) to independently normalize each variable. Specifically, the mean of each variable is calculated along the time dimension. with standard deviation The data is normalized to a zero-mean, unit-variance distribution, and learnable affine transformation parameters (scaling factors) are applied. With offset coefficient The formula is shown in (2):

[0096]

[0097]

[0098] in, and It is retained during the inference phase and used for denormalization in step 4; , The learnable affine parameters enable the model to retain the relative distribution characteristics of the data while eliminating dimensional differences, effectively alleviating the problem of non-stationary distribution shift of mine microseismic signals at different acquisition stages.

[0099] Subsequently, each normalized sensor variable is treated as an independent token, and the variables are then embedded from the time dimension through a linear embedding layer. Projecting onto a high-dimensional latent feature space, the formula is shown in (4):

[0100]

[0101] in This is the hidden layer dimension. This inverted embedding strategy uses variables as tokens instead of time steps. This preserves the independence of each sensor variable, avoiding interference from subsequent cross-variable operations on the intra-variable temporal features, while compressing the long lookback window into a fixed high-dimensional space. This facilitates the extraction of more expressive temporal features from the latent space. At this point, the raw mine microseismic data has completed preprocessing and enters the subsequent encoding stage.

[0102] In step 2), we focus on the temporal variation patterns of each variable, and implement temporal three-component decoupling in the high-dimensional latent feature space to embed the features of each variable. Separate into trend components Seasonal components With noise components Three orthogonal components.

[0103] The first level is coarse trend extraction based on one-dimensional convolution. This is applied to the input high-dimensional features. Apply small kernel one-dimensional convolutional layers initialized with uniform distribution The weights of the convolutional layer are initialized to a uniform distribution to simulate the effect of a moving average, but can be fine-tuned during training to adapt to the low-frequency background change patterns of different mining stages, thereby extracting the initial trend components, as shown in formula (5):

[0104]

[0105] Then, the preliminary residuals containing seasonal components and noise are obtained by subtraction, as shown in formula (6):

[0106]

[0107] The second level is frequency-domain-based adaptive seasonal extraction. Due to residuals... The residual still contains multi-scale periodic components and random noise, so we introduce a frequency-domain adaptive filtering mechanism for fine separation. First, we process the residual... Perform a Fast Fourier Transform (FFT) on each variable to map the time-domain residual to the frequency domain, as shown in formula (7):

[0108]

[0109] in Indicates the FFT operation. This indicates the length of the transformed frequency domain sequence. This indicates the number of variables. Through frequency domain analysis, different frequency components correspond to different physical signal components: low-frequency components correspond to the gradual trend of microseismic activity, mid-frequency components correspond to periodic vibration patterns, and high-frequency components correspond to random noise interference.

[0110] Next, regarding the first Statistical mean of spectral amplitude of each variable with standard deviation Design adaptive threshold The formula is shown in (8):

[0111]

[0112] Based on this adaptive threshold, a binary mask is constructed to retain frequency components with amplitudes below the threshold (i.e., trending low-frequency components) and set high-frequency components with amplitudes above the threshold to zero, as shown in formula (9):

[0113]

[0114] in This represents element-wise multiplication. A binary mask is used. This adaptive thresholding mechanism can dynamically adjust the filtering boundary according to the spectral characteristics of different mine microseismic signals, eliminating the need for manually setting a fixed threshold. This maximizes the separation of low-frequency trend components in the potential space while avoiding the inclusion of high-frequency noise that varies with the signal. An inverse FFT (iFFT) is then performed on the filtered spectrum to restore the fine-grained trend components in the time domain. The formula is shown in (10):

[0115]

[0116] Subtracting the fine-grained trend component from the residuals yields a purer periodic expression, namely the seasonal component. The formula is shown in (11):

[0117]

[0118] Finally, integrate the results of the two-level decomposition: the overall trend component. Low-frequency background changes reflecting mine microseismic activity; seasonal component Capture regular vibrations and noise components caused by blasting and periodic stress release in the formation during mining operations. This represents random interference signals. The three components are mutually orthogonal and together comprehensively describe the temporal characteristics of mine microseismic signals in a high-dimensional potential space.

[0119] In step 3), we focus on the interdependencies between different sensor variables, using a convolutional intervariate interaction module (CVI) with linear complexity. Efficiently modeling local spatiotemporal correlations and global cross-sample dependencies among multi-sensor variables, overcoming existing Transformer-based self-attention mechanisms. The bottleneck is the excessive computational complexity.

[0120] First, embed the feature tensor Reshape it into a four-dimensional view suitable for two-dimensional convolution processing, as shown in formula (12):

[0121]

[0122] Then, two layers of two-dimensional convolutions with ReLU activation are applied sequentially. The convolution kernels cover a certain receptive field in the temporal extension and variable extension directions, respectively, and establish local correlations in the temporal and sensor dimensions, as shown in formula (13):

[0123]

[0124] in The ReLU activation function is used, and the kernel size of the two-dimensional convolution layers is determined by the hyperparameter. and Together, they determine and control the receptive field sizes in the time and variable directions respectively, to adapt to the multi-scale features of different mining datasets, and output... It contains local interaction information that combines variable and time dimensions.

[0125] However, local convolution can only model short-range dependencies and cannot directly capture the correlation of global variables across acquisition batches. Therefore, we introduce an external long-term memory matrix to... Applying a linear projection along the time dimension compresses the local interaction information into a compact memory space of a fixed size, as shown in formula (14):

[0126]

[0127] in is the dimension hyperparameter of the memory matrix. The memory matrix is ​​independent of the positional mapping of the current input sequence. As an independent global context component, it enables the model to learn long-term variable association patterns across samples from multi-sensor data in the mine, achieving continuous memorization and reuse of regular variable dependencies throughout the training set.

[0128] To prevent the model from overfitting due to excessive reliance on fixed channels in the memory matrix, a stochastic pooling regularization mechanism is further introduced: This mechanism applies a regularization method to the memory matrix. Softmax weights are calculated along the variable dimension. Multinomial sampling is performed using this probability distribution. A variable index is randomly selected for each memory location, and then the data is processed according to that index. Collect corresponding context features and broadcast them along the variable dimension to the original number of variables. The context-enhanced features are obtained, as shown in formula (15):

[0129]

[0130] This stochastic mechanism allows the model to focus on different channels of the memory matrix from a random perspective in each forward propagation, effectively enhancing the model's generalization robustness in the high-noise, micro-seismic environment of a mine. Finally, context-enhanced features are incorporated. With the original embedded input The features are concatenated along the time dimension and then nonlinearly fused through two layers of MLP to output the final cross-variable interaction features, as shown in formulas (16) and (17):

[0131]

[0132]

[0133] The GELU activation function is used for the nonlinear transformation of the first layer of the MLP. The spliced ​​dimensions Compress back to the original time dimension Output Each variable's features simultaneously contain its own temporal information and a global cross-variable context from the convergence of long-memory matrices, resulting in an overall computational complexity of only [missing information]. .

[0134] In step 4), we fuse the three-component features from the time dimension in step 2 with the interaction features from the variable dimension in step 3 to obtain the final microseismic prediction sequence. Within each LTDBlock, the LTD module outputs the trend component. Seasonal component The CVI module outputs cross-variable interaction features. The three carry complementary temporal information from different perspectives: Describe the interdependencies between multi-sensor variables. Characterizing the low-frequency background trend of microseismic activity in mines. Capture the periodic vibration pattern. We concatenate the three elements along the time dimension to form a comprehensive feature representation of the current block, as shown in formula (18):

[0135]

[0136] Subsequently, the dimensions of the synthesized features are increased from [previous dimensions] by a fully connected layer. Compress to After completing feature compression and information fusion, the output of the current LTDBlock is obtained, as shown in formula (19):

[0137]

[0138] Thus The encoder uses a stacked layer LTDBlock architecture, with each layer performing a time decomposition and cross-variable interaction on the features. The final output of the encoder is a highly abstract multivariate feature representation of microseismic events. Subsequently, the time dimension was shifted from... using a linear prediction head. Mapping to prediction length And transpose it to the standard prediction format, as shown in formula (20):

[0139]

[0140] Finally, the mean saved by RevIN in step 1 is used. with standard deviation The prediction results are denormalized to restore the normalized prediction of the potential space to the original microseismic physical dimensions, and the final prediction sequence is obtained, as shown in formula (21):

[0141]

[0142] in The final multivariate time series prediction results for mine microseismic events include future... The predicted values ​​of each sensor variable at each time step can be directly used in the mine safety early warning system.

[0143] In step 5), in order to give the decomposition result of the LTD module physical meaning—that is, the noise component— It should approximate zero-mean Gaussian white noise rather than carry an effective prediction signal—we design the total loss function as a weighted combination of prediction loss and noise regularization term, as shown in formula (22):

[0144]

[0145] Among them, predicted loss The mean square error between the predicted sequence and the actual microseismic sequence is measured to drive the model to learn an accurate prediction mapping, as shown in formula (23):

[0146]

[0147] Noise regularization term Noise components decomposed from LTD Apply Norm penalty forces the model to attribute as many effective time series signals as possible to the trend component. Seasonal component The formula is shown in (24):

[0148]

[0149] in This is a regularization weight hyperparameter used to balance the trade-off between prediction accuracy and decomposition quality. The decomposition process of LTD, which serves as an auxiliary supervisory signal constraint, ensures that the trend and seasonal components obtain sufficient and clean time-series information, thereby simultaneously improving the physical interpretability of the decomposition and the model's predictive robustness in the complex noise environment of mines. The network uses the Adam optimizer for end-to-end training, employing a cosine annealing scheduling strategy for the learning rate, with the initial learning rate set at [value missing]. The weight decay coefficient is set to This is to prevent overfitting.

[0150] Step 6) Apply the trained network model to mine microseismic prediction to judge and warn of mine safety status.

[0151] The specific method is as follows:

[0152] The real-time acquired multivariate microseismic time-series data is input into the LTDNet network trained in step 5 to obtain the predicted sequence of each sensor variable for the next H time steps. The mine safety status is assessed based on the prediction results: if the predicted microseismic energy exceeds a preset safety threshold, it is determined that there is a risk of coal and rock dynamic disasters such as mine rock bursts or roof collapses, triggering a safety warning signal and notifying on-site personnel to take protective measures; otherwise, the current mine microseismic activity is considered to be at a normal level, and rolling predictions and real-time monitoring continue for subsequent time periods.

Claims

1. A mine microseismic prediction method based on latent temporal decomposition and convolutional intervariate interaction, characterized in that, The steps are as follows: Step 1) Perform reversible instance normalization preprocessing on the multivariate time series data of mine microseismic events, and map the normalized variables to a high-dimensional latent feature space through an inverted embedding layer; Step 2) The temporal features are decoupled into three components in the high-dimensional feature space by using the Latent Time Decomposition (LTD) module, and the trend component, seasonality component and noise component are extracted. Step 3) The Convolutional Intervariate Variable Interaction (CVI) module, based on two-dimensional convolution and long-term memory matrix, captures the local spatiotemporal correlations and global dependencies between variables in linear complexity. Step 4) The trend component, seasonal component and cross-variable interaction features are spliced ​​and fused along the time dimension, mapped by the linear prediction head and denormalized to obtain the final microseismic prediction sequence; Step 5) Use the prediction mean square error loss and noise component The network is trained end-to-end using a composite loss function composed of regularization. Step 6) Apply the trained network model to mine microseismic prediction: Input the real-time collected multivariate microseismic time series data into the trained network to obtain the prediction sequence of each sensor variable for the next H time steps; if the predicted microseismic energy exceeds the preset safety threshold, it is determined that there is a risk of dynamic disasters such as mine rock bursts or roof collapse, and a safety warning is triggered. Otherwise, the current microseismic activity in the mine is considered to be at a normal level, and monitoring should continue.

2. The mine microseismic prediction method based on latent time decomposition and convolutional intervariate interaction as described in claim 1, characterized in that, In step 1), the specific method is as follows: Input multivariable microseismic monitoring data Applying Reversible Instance Normalization (RevIN): Calculates the mean of each variable along the time dimension. and standard deviation The normalization formula is shown in formula (1): in , To learnable affine transformation parameters, and Reserved for subsequent denormalization; Subsequently, each normalized variable is treated as an independent token, and the variables are then processed through a linear embedding layer from the time dimension. Projecting to a higher-dimensional latent space, as shown in Equation (2): in For batch size, For the number of sensor variables, To check the window length, For hidden layer dimensions.

3. The mine microseismic prediction method based on latent time decomposition and convolutional intervariate interaction as described in claim 1, characterized in that, In step 2), the specific method of the Latent Time Decomposition (LTD) module is as follows: Step 2.1) One-dimensional convolution coarse trend extraction: for input features Apply a small kernel one-dimensional convolutional layer initialized with uniform distribution to extract the initial trend components and residuals, as shown in formulas (3) and (4): Step 2.2) Frequency Domain Adaptive Seasonal Extraction: For the residuals The frequency domain representation is obtained by performing a Fast Fourier Transform (FFT) on each variable. ; for the first Calculate the mean of the spectral amplitude of each variable. with standard deviation Set adaptive threshold Specifically, as shown in formula (5): A binary mask is constructed and low-amplitude frequency components are filtered out. The time-domain fine trend is restored by inverse FFT, and the seasonal components are separated, as shown in formulas (6) and (7): Where ⊙ represents element-wise multiplication. This is the spectrum representation after mask filtering. It is a binary mask; This represents the length of the frequency domain sequence, i.e., the total number of frequency components. Step 2.3) The overall trend component is The noise component is ; Output trend Seasonal ,noise Three orthogonal components.

4. The mine microseismic prediction method based on latent time decomposition and convolutional intervariate interaction as described in claim 1, characterized in that, In step 3), the specific method of the Convolutional Intervariate Variable Interaction (CVI) module is as follows: Step 3.1) Two-dimensional convolutional local feature extraction: extracting local features from the input features. Remodeling Two layers of two-dimensional convolutions with ReLU activation are applied sequentially to capture local spatiotemporal correlations, as shown in Equation (8): Where B is the batch size, C is the number of sensor variables, and L is the length of the feature time dimension. It is a ReLU activation function; Step 3.2) Long-term memory matrix and random pooling: for Applying a linear projection yields the memory matrix. Then, context-enhanced features are obtained through random pooling, as shown in formulas (9) and (10): in The dimension of the memory matrix is ​​a hyperparameter. The number of intermediate channels in the output of a 2D convolution; Step 3.3) Feature Fusion: [The text abruptly ends here, likely due to an incomplete sentence or a format With the original input The features are concatenated along the time dimension and then transformed using two layers of MLP nonlinear transformation to output cross-variable interaction features, as shown in formulas (11) and (12): The overall computational complexity is .

5. The mine microseismic prediction method based on latent time decomposition and convolutional intervariate interaction as described in claim 1, characterized in that, In step 4), the specific method is as follows: In each LTDBlock, cross-variable interaction features will be included. Trend components Seasonal component The blocks are concatenated along the time dimension and compressed by a fully connected layer to obtain the current block output, as shown in formulas (13) and (14): through After stacking layers of LTDBlock, the encoder output is denoted as Henc (dimension 1). The time dimension is mapped to the prediction length through a linear prediction head. Then, after inverse normalization, the final predicted sequence is obtained, as shown in formulas (15) and (16): in This represents the final multivariate time series prediction result for mine microseismic events; where ⊙ denotes element-wise multiplication. and The mean and standard deviation saved by RevIN in step 1) This is the encoder output after N layers of LTDBlock stacking, with dimensions B×C×dmodel. The number of stacking layers for LTDBlock. For batch size, To predict the length, For the number of sensor variables, For hidden layer dimensions.

6. The mine microseismic prediction method based on latent time decomposition and convolutional intervariate interaction as described in claim 1, characterized in that, In step 5), the composite loss function is shown in formula (17): Among them, predicted loss The mean square error between the predicted sequence and the actual microseismic sequence is shown in formula (18): Noise regularization term For the decomposed noise components Apply Norm penalty, as shown in formula (19): This is a regularization weight hyperparameter used to balance prediction accuracy and decomposition quality; As an auxiliary monitoring signal, it constrains the LTD decomposition process, making the noise term approach zero-mean Gaussian white noise, thereby improving the physical interpretability and prediction robustness of the decomposition; where B is the batch size, H is the prediction length, C is the number of sensor variables, L is the length of the feature time dimension, Y is the real microseismic sequence, and N is the noise component decomposed by the LTD module.

Images