A microseismic signal recognition method based on transfer learning and BiLSTM-DCNN

Abstract

The application discloses a kind of based on transfer learning and BiLSTM-DCNN's microseismic signal identification method, belongs to mining engineering microseismic monitoring and signal processing technical field, solve the problem of low identification precision under the condition of small sample with data scarcity in the initial stage of mine monitoring. First, the microseismic signal is extracted for Mel spectrum feature;BiLSTM-DCNN model containing bidirectional long short-term memory network and deep convolutional neural network is constructed, and large-scale source mine data is used for pre-training;Then, through transfer learning, the model is adapted to the target mine small sample data, and the model parameters are fine-tuned to achieve high-precision classification of signals. The application significantly improves the recognition accuracy of microseismic signals and the model generalization ability under the condition of small sample, and the test accuracy reaches 0.9444, which is improved by 80.85% compared with the non-transfer model, and is suitable for mine dynamic disaster intelligent early warning and safety monitoring system.

Description

Technical Field

[0001] This invention relates to the field of microseismic monitoring and signal processing technology in mining engineering, specifically to an intelligent microseismic signal recognition method based on transfer learning and bidirectional long short-term memory network-deep convolutional neural network (BiLSTM-DCNN). Background Technology

[0002] Rock fracture signals are typically stress waves released by minute vibrations caused by rock mass fracturing. These waves contain rich information about the mechanics and deformation of rock fractures, and their spatiotemporal evolution is often used as precursory evidence for rockburst warnings. Microseismic monitoring technology, as an efficient and real-time geological hazard early warning method, is widely used in mine safety, earthquake prediction, and engineering structure health monitoring. However, in complex excavation environments, microseismic monitoring systems also record a large number of interference signals caused by blasting or mechanical vibrations during construction. Accurately identifying rock fracture signals from microseismic signals allows for rockburst assessment and location, reducing the risk of accidents and providing a scientific basis for mining activities. This is crucial for predicting geological hazards, assessing mine stability, and ensuring personnel safety.

[0003] Traditional identification methods mainly rely on manually set thresholds, time-frequency analysis, or matched filtering techniques. These methods are not only cumbersome and highly dependent on expert experience, but also suffer from low identification accuracy and poor generalization ability in low signal-to-noise ratio environments. For newly established mines or mining areas with a short monitoring history, labeled microseismic signal samples are scarce, making data-driven deep learning methods prone to overfitting on small samples, severely limiting model performance.

[0004] Existing research attempts to address small-sample problems using transfer learning, but most methods use earthquake datasets as the source domain. Although earthquake signals and mine microseismic signals both belong to rupture events, they differ fundamentally in energy magnitude, propagation medium, inducing mechanisms, and background noise. Direct transfer learning can easily lead to "negative transfer," meaning that model performance deteriorates instead of improving. Furthermore, existing deep learning models either focus only on temporal dependencies (such as LSTM) or only on spatial features (such as CNN), failing to fully leverage the temporal evolution and local frequency domain features of microseismic signals, thus limiting the model's recognition potential.

[0005] Therefore, there is an urgent need to propose a high-precision and robust microseismic signal identification method that can overcome domain differences, effectively integrate temporal and spatial characteristics, and is suitable for small sample environments in mines. Summary of the Invention

[0006] This invention aims to overcome the shortcomings of existing technologies and provide a microseismic signal recognition method based on transfer learning and BiLSTM-DCNN. This method constructs a temporal-spatial collaborative hybrid deep learning model and innovatively adopts a "mine-to-mine" transfer learning paradigm. It effectively utilizes large-scale source domain data from similar geological environments to provide prior knowledge for small-sample identification tasks of target mines, thereby significantly improving the model's recognition accuracy and generalization ability under data-scarce conditions.

[0007] To achieve the above objectives, this application provides the following solution:

[0008] A method for identifying microseismic signals based on transfer learning and BiLSTM-DCNN includes the following steps:

[0009] S1. Collect various types of microseismic signals from the source mine and construct the original signal dataset;

[0010] S2. After performing preprocessing and short-time Fourier transform on the original signal, Mel spectral features are extracted to obtain the Mel time spectrum and construct a waveform feature dataset.

[0011] S3. Construct a BiLSTM-DCNN hybrid neural network model, including a temporal feature extraction module, a spatial feature extraction module, and a classification output module, using the Mel-time spectrum as the input signal, and finally outputting the predicted labels for each signal segment;

[0012] S4. Use the source mine waveform feature dataset to pre-train the BiLSTM-DCNN hybrid neural network model and optimize the model parameters;

[0013] S5. Using a transfer learning strategy, the pre-trained model is transferred to the target mine, and the parameters of the pre-trained model are fine-tuned using a small sample dataset to obtain the Tr-BiLSTM-DCNN model.

[0014] S6. Input the microseismic signal to be identified into the Tr-BiLSTM-DCNN model and output the signal category identification result.

[0015] Furthermore, S2 specifically refers to:

[0016] The preprocessing process includes: using a fixed time window to truncate each signal to a uniform length; resampling the signal segments; pre-emphasizing the signal segments using a first-order high-pass filter; dividing the continuous signal stream of each signal segment into multiple short-time analysis frames, and setting appropriate overlap between adjacent frames to maintain signal continuity; and applying a Hanning window function to smooth the data in each frame.

[0017] After preprocessing, a short-time Fourier transform is performed on each frame of signal to map it from the time domain to the time-frequency domain, obtaining a complex spectrum that characterizes the time-frequency distribution of the signal. The short-time amplitude spectrum is obtained by calculating the modulus of the complex spectrum.

[0018] The short-time amplitude spectrum is filtered using a filter bank to transform it to the Mel-scale, yielding the Mel-time spectrum. The transformation formula is as follows:

[0019]

[0020] In the formula, k is the Mel-time spectrum frequency in Hz; f is the short-time amplitude spectrum frequency in Hz, and 0 ≤ f ≤ 22050.

[0021] Furthermore, in S3, the BiLSTM-DCNN hybrid neural network model includes:

[0022] The temporal feature extraction module first uses a two-layer bidirectional long short-term memory network (BiLSTM) to learn the input signal, so as to capture the forward and backward dependencies simultaneously. After each BiLSTM, a batch normalization operation is introduced to stabilize the training process, and a dropout layer is set at key nodes to reduce the risk of overfitting. Finally, the temporal embedding vector is output.

[0023] The spatial feature extraction module, after obtaining the temporal embedding vector, first uses a one-dimensional convolutional layer to encode the local receptive field, and then combines batch normalization and pooling operations to reduce the feature dimension; then, it stacks convolutional layers and uses the same normalization and pooling operations to extract higher-level discriminative features.

[0024] In the classification output module, after feature extraction, the time dimension is compressed into a fixed-length feature vector through global average pooling, and then input into two fully connected layers in sequence. After each fully connected layer, a batch normalization and dropout layer is added to further improve the generalization performance. Finally, the classification task is completed through the fully connected layers, and the probability distribution of the three types of signals is output. The category with the highest probability is taken as the final predicted label.

[0025] Furthermore, S5 specifically refers to:

[0026] Multiple target signals from the target mine are collected, and waveform feature datasets of the target mine are obtained through S1 and S2 processing. These datasets are then input into the BiLSTM-DCNN hybrid neural network model for transfer learning, and the model parameters are trained and fine-tuned to obtain the Tr-BiLSTM-DCNN model. The model parameters that are fine-tuned include the parameters of all layers except those after the global average pooling layer in the classification output module.

[0027] Furthermore, S6 specifically refers to:

[0028] The microseismic signals of the target mine to be identified are processed by S1 and S2 to obtain Mel spectrum features, which are then input into the Tr-BiLSTM-DCNN model. The model outputs the probability distribution of various types of signals, and the category corresponding to the highest probability is taken as the final identification result.

[0029] Compared with the prior art, the present invention has at least the following beneficial effects:

[0030] This invention proposes a microseismic signal recognition method based on transfer learning and BiLSTM-DCNN, addressing the challenge of microseismic signal recognition in small-sample environments in mines. It innovates in feature extraction, model structure, and training strategies. This application constructs a temporal-spatial collaborative hybrid deep learning model, capturing bidirectional dependencies in signals through BiLSTM, extracting local frequency domain features through DCNN, and combining Mel spectrum analysis to achieve noise-robust feature representation. Furthermore, a "mine-to-mine" transfer learning mechanism is introduced, utilizing pre-trained models with large amounts of source mine data to significantly improve recognition performance and generalization ability under small-sample conditions in target mines. Experiments show that this method improves recognition accuracy by 80.85% under 100-sample conditions, achieving a test accuracy of 0.9444, which is superior to traditional CNN, LSTM, and non-transfer learning models, providing reliable technical support for intelligent microseismic monitoring in mines. Attached Figure Description

[0031] Figure 1 This is a diagram of the BiLSTM-DCNN model structure.

[0032] Figure 2 Here is the small sample confusion matrix before transfer learning in the BP neural network; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type.

[0033] Figure 3 Here is the small sample confusion matrix before the CNN model transfer; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type.

[0034] Figure 4 Here is the small sample confusion matrix before the BiLSTM model transfer; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type.

[0035] Figure 5 Here is the small sample confusion matrix before the BiLSTM-DCNN model transfer; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type.

[0036] Figure 6 Here is the small-sample confusion matrix after transfer learning in a BP neural network; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type.

[0037] Figure 7 Here is the small sample confusion matrix after the CNN model transfer; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type.

[0038] Figure 8 Here is the small-sample confusion matrix after the BiLSTM model transfer; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type.

[0039] Figure 9 Here is the small sample confusion matrix after the BiLSTM-DCNN model transfer; where (a) represents 100 samples for each signal type; (b) represents 200 samples for each signal type; and (c) represents 300 samples for each signal type. Detailed Implementation

[0040] To further understand the implementation method of the present invention, the technical solutions in the embodiments of the present invention are clearly and completely described. Furthermore, the examples described below are merely a part of the present invention, and not all of it. The following description is only for further illustrating the advantages and features of the present invention, and not for claiming limitations on the present invention. Other embodiments obtained by those skilled in the art without inventive effort are all within the scope of protection of the present invention.

[0041] This embodiment uses the Xiadian gold mine, which has a long monitoring history and a large number of labeled microseismic signal samples, as the source mine and the Hongtoushan copper-zinc mine as the target mine. The invention will be further explained with reference to the accompanying drawings. Compared with the seismic dataset, the source mine sample set has a higher similarity to the target mine in terms of geological conditions, rock mass characteristics and noise background, thus avoiding "negative transfer" in transfer learning.

[0042] This embodiment provides a microseismic signal recognition method based on transfer learning and BiLSTM-DCNN, including the following steps:

[0043] S1. Collect source mine microseismic signals, including rock fracture signals RSsignals, blasting signals BSsignals, and mechanical vibration signals MVSsignals, and construct the original signal dataset; specifically in this embodiment, 3000 signals of each type are collected.

[0044] S2. After performing preprocessing and short-time Fourier transform on the original signal, Mel spectrum features are extracted to obtain waveform feature dataset;

[0045] The preprocessing process includes: uniformly truncating the original signal into multiple signal segments with a length of 9000 sampling points, padding with zeros for any insufficient segments, and labeling each signal segment; resampling the signal segments to balance the number of samples; using a first-order high-pass filter to pre-emphasize the signal segments to effectively compensate for the attenuation of high-frequency components and enhance the salience of rupture details; based on the short-time stationary characteristics of the microseismic signal, dividing the continuous signal stream of each signal segment into multiple short-time analysis frames, and setting appropriate overlap between adjacent frames to maintain signal continuity; and applying a Hanning window function to smooth the data in each frame to suppress spectral leakage caused by framing.

[0046] After preprocessing, a short-time Fourier transform is performed on each frame of signal to map it from the time domain to the time-frequency domain, obtaining a complex spectrum that characterizes the time-frequency distribution of the signal. The short-time amplitude spectrum is obtained by calculating the modulus of the complex spectrum.

[0047] The short-time amplitude spectrum is filtered using a triangular filter bank of equal height to transform it to the Mel-scale, yielding the Mel-time spectrum. The transformation formula is as follows:

[0048]

[0049] In the formula, k is the Mel-time spectrum frequency in Hz; f is the short-time amplitude spectrum frequency in Hz, and 0 ≤ f ≤ 22050. Specifically, in this embodiment, the equal-height triangular filter bank adopts a Mel-time filter bank with 40 filters.

[0050] The role of extracting Mel spectral features is twofold: firstly, by significantly enhancing the perception weight of low-frequency components of the signal through nonlinear mapping, it helps to focus on the main energy distribution frequency band of rock mass fracture and effectively suppress high-frequency interference; secondly, the obtained Mel time spectrum can effectively compress the feature dimension while preserving key monitoring features to the maximum extent, significantly reducing the sample data size required for subsequent analysis, and laying the foundation for rapid training of deep learning models and improved efficiency of real-time on-site identification.

[0051] S3. Construct a BiLSTM-DCNN hybrid neural network model, including a temporal feature extraction module, a spatial feature extraction module, and a classification output module, using the Mel-time spectrum as the input signal, and finally outputting the predicted labels for each signal segment;

[0052] In the temporal feature extraction stage, a Bidirectional Long Short-Term Memory (BiLSTM) network is first used to learn the input signal to simultaneously capture forward and backward dependencies. A BiLSTM layer with 64 hidden units and a BiLSTM layer with 128 hidden units are stacked sequentially to enhance the representation ability of long-term temporal features. After each BiLSTM, a batch normalization operation is introduced to stabilize the training process, and a dropout layer with a dropout rate of 0.3 is set at key nodes to reduce the risk of overfitting. The final output is a temporal embedding vector. Compared with a single-layer structure, the stacked BiLSTM has stronger expressive power and can learn complex dynamic patterns, thus providing a more accurate temporal representation. Each layer learns the input sequence from a different perspective and passes the results to subsequent layers to achieve progressive feature extraction.

[0053] In the local feature extraction stage, after obtaining the temporal embedding vector, the model further extracts local patterns through a convolutional module. First, a one-dimensional convolutional layer with 64 kernels is used to encode the local receptive field. Then, batch normalization and pooling operations (one-dimensional max pooling layer) are combined to reduce the feature dimensionality. Next, a convolutional layer with 128 kernels is stacked and combined with the same normalization and pooling operations to extract higher-level discriminative features. Through multiple combinations of convolution and pooling, the model can gradually form a hierarchical representation from low-order features to high-order features, thereby better capturing local anomalies in surface deformation signals.

[0054] In the classification stage, after feature extraction, the time dimension is compressed into a fixed-length feature vector through global average pooling (a one-dimensional global average pooling layer), and then input into two fully connected layers. After each fully connected layer, batch normalization and Dropout with a dropout rate of 0.15 are added to further improve generalization performance. Finally, the classification task is completed through a fully connected layer with three neurons, outputting the probability distribution of the three types of signals, and taking the category corresponding to the highest probability as the final predicted label.

[0055] In summary, the BiLSTM-DCNN hybrid neural network model organically combines the long-term dependency modeling capability of BiLSTM with the local pattern recognition advantages of convolutional networks. It achieves comprehensive discrimination of high-level features through fully connected layers, thus balancing global temporal dynamics with fine-grained local features. This results in good adaptability and application potential, effectively serving complex time series classification tasks. The initial parameter values ​​of the BiLSTM-DCNN hybrid neural network model are shown in Table 1.

[0056] Table 1 Initial parameter values ​​of the BiLSTM-DCNN hybrid neural network model

[0057]

[0058] S4. Use the source mine waveform feature dataset to pre-train the BiLSTM-DCNN hybrid neural network model and optimize the model parameters;

[0059] During the pre-training phase, 80% of the waveform feature dataset was randomly selected as the training set and 20% as the test set. These were then input into the constructed BiLSTM-DCNN hybrid neural network model for training. The training run consisted of 400 epochs with an initial learning rate of 0.001, a batch size of 1000, and a weight decay of 0.0005. After pre-training, the BiLSTM-DCNN hybrid neural network model parameters were saved for subsequent training and transfer learning.

[0060] S5. Using a transfer learning strategy, the pre-trained model is transferred to the target mine, and the parameters of the pre-trained model are fine-tuned using a small sample dataset to obtain the Tr-BiLSTM-DCNN model.

[0061] Three types of target signals from the target mine were collected. After processing through S1 and S2, waveform feature datasets of the target mine were obtained. These datasets were then input into a BiLSTM-DCNN hybrid neural network model for transfer learning. The model parameters were trained and fine-tuned to obtain the Tr-BiLSTM-DCNN model. The fine-tuned model parameters included all layer parameters except those after the one-dimensional global average pooling layer in the classification output module. The advantage of transfer learning is that it can utilize feature representations trained on large datasets, thereby achieving better recognition results on small datasets while reducing training time and avoiding overfitting. In this process, the parameters of some network layers were adjusted to adapt to the characteristics of small datasets.

[0062] S6. Input the microseismic signal to be identified into the Tr-BiLSTM-DCNN model and output the signal category identification result.

[0063] The microseismic signal to be identified from the target mine is processed by S1 and S2 to obtain Mel spectrum features, which are then input into the Tr-BiLSTM-DCNN model. The model outputs the probability distribution of three types of signals, and the category corresponding to the highest probability is taken as the final identification result.

[0064] Experimental verification

[0065] Comparison models were set up, including: a traditional BP neural network (BP), a single CNN model (CNN), and a single BiLSTM model (BiLSTM). Based on the same source mine dataset (divided into source mine training set, source mine validation set, and source mine test set in a 7:1.5:1.5 ratio), the comparison models and a BiLSTM-DCNN model without transfer learning were pre-trained with the same number of iterations. Training set accuracy, test set accuracy, training set loss, and test set loss were used as the main evaluation metrics to evaluate the performance of each model. In addition, data was collected... Three types of signals from the target mine form a small-sample target mine dataset, which is divided into a target mine training set, a target mine validation set, and a target mine test set in a 7:1.5:1.5 ratio. Based on the same target mine dataset, the same transfer learning strategy as S5 is used to transfer the trained models to the target mine, resulting in Tr-BP, Tr-CNN, Tr-BiLSTM, and Tr-BiLSTM-DCNN. The same evaluation metrics are used to evaluate the performance of each transferred model when the number of samples for each class in the target mine dataset is 100, 200, and 300, respectively. At the same time, the recognition and misjudgment of each type of signal are visualized and analyzed through a confusion matrix.

[0066] Tables 2 and 3 show the comparison of the transfer learning results of different models before and after the transfer learning process under small sample conditions. The results show that without transfer learning, the performance of all models drops significantly under small sample conditions, especially with 100 samples per class. The test accuracy of the BiLSTM-DCNN model is only 0.5222. After introducing the transfer learning strategy, the performance of all models is significantly improved, with the Tr-BiLSTM-DCNN model showing the largest improvement. With 100 samples per class, the test set accuracy jumped from 0.5222 to 0.9444, a relative performance improvement of 80.85%, while the test set loss decreased from 1.2148 to 0.2214. Table 4 shows the percentage change in model performance before and after the transfer learning. It can be seen that when the sample size increases to 300 samples per class, the Tr-BiLSTM-DCNN model still maintains the highest recognition accuracy (0.9074), significantly outperforming the other comparative models.

[0067] Table 2 Comparison of model performance before transfer learning under small sample conditions

[0068]

[0069] Table 3 Comparison of model performance after transfer learning under small sample conditions

[0070]

[0071] Table 4 Comparison of results before and after transfer for the four models

[0072]

[0073] As the number of signal samples per class increased from 100 to 300, the percentage improvement in test set accuracy for BP decreased from 35.29% to 1.54%, indicating that the transfer gain gradually weakens with sufficient data, consistent with the expectation that "the model's own training can cover the transferred knowledge after the data volume increases." CNN maintained significant improvement across all data volumes, with a 74.40% improvement in test accuracy at 100 samples per class and still reaching 12.38% at 300 samples per class, demonstrating its strong dependence on and adaptability to transfer learning, and the pre-trained feature extraction capability continuing to provide value for small-sample training. BiLSTM-DCNN achieved over 69% improvement in test accuracy at 100 and 200 samples per class, and still reached 24.50% at 300 samples, reflecting that the hybrid model structure is more likely to leverage the value of pre-trained knowledge—the joint transfer of temporal and spatial features effectively compensates for the shortcomings of feature learning under small data conditions.

[0074] To further reveal the discriminative features of each model under different sample sizes and the effects of transfer learning, confusion matrices of the four models before and after transfer learning were plotted and compared under conditions of 100 samples / class, 200 samples / class, and 300 samples / class, as shown below. Figures 2 to 9 As shown.

[0075] Under the transfer learning strategy, the confusion matrices of each model showed varying degrees of change for the small sample dataset. Overall, models with stronger feature modeling capabilities exhibited significantly improved diagonal concentration in their confusion matrices and a marked reduction in off-diagonal misclassification regions after incorporating prior information from the source domain. This indicates that the models gradually formed clearer and more stable discrimination boundaries among the three types of typical microseismic signals from mines. This phenomenon suggests that the model structure's ability to collaboratively express temporally dependent features and local pattern features is a key factor determining the class discrimination effect. When the model itself possesses sufficient feature learning capabilities, the prior features provided by the source domain pre-training can be effectively absorbed and transformed into stable discriminative capabilities, thus manifesting as a more concentrated diagonal distribution in the confusion matrix.

[0076] For Tr-BP, the diagonal distribution of its confusion matrix is ​​improved, especially with 100 samples per class of signals, where the recognition accuracy of RFsignals is significantly improved. This indicates that more reasonable parameter initialization in small sample scenarios helps shallow models alleviate the problem of insufficient generalization to some extent. However, there is still a relatively obvious confusion phenomenon between BLsignals and MVsignals, indicating that due to the expressive power of the network structure itself, shallow models are unable to fully characterize the complex differences between different microseismic signals, and their discrimination performance still has a significant upper limit.

[0077] Tr-CNN shows a significant improvement in the confusion matrix. With 100 samples per class, its confusion matrix gradually transforms from an approximately random distribution to a diagonally concentrated feature matrix, and the misclassification ratio between BLsignals and MVsignals is significantly reduced. This indicates that the convolutional structure, after obtaining stable local feature representations, can improve class discrimination ability under small sample conditions to some extent. However, due to the lack of modeling ability for the temporal evolution characteristics of microseismic signals, Tr-CNN is still structurally limited in terms of class discrimination stability.

[0078] The confusion matrix of Tr-BiLSTM shows that its recognition accuracy for RF signals exceeds 93.33%, and its ability to distinguish between BL signals and MV signals is also significantly enhanced. However, from the overall distribution, there are still a small number of cross-class misclassifications, indicating that relying solely on time-series features is insufficient to completely distinguish signal types with similar energy evolution characteristics.

[0079] The Tr-BiLSTM-DCNN model exhibits the best overall performance. Its confusion matrix displays a highly concentrated diagonal distribution even with only 100 samples per class, showing minimal cross-class misclassification. This advantage is further solidified with 200 and 300 samples per class, maintaining balanced accuracy across classes without significant class bias. This demonstrates that by jointly modeling temporal-dependent features and local pattern features, the model can construct a stable and highly discriminative feature space under conditions of minimal mine samples, thereby significantly improving the consistency and reliability of microseismic signal classification results in complex mine environments.

[0080] Experimental results demonstrate that the Tr-BiLSTM-DCNN network structure exhibits significant advantages in microseismic signal recognition tasks within complex mining environments. Especially under conditions of small sample sizes and high noise, its temporal-spatial joint modeling mechanism effectively improves feature discriminativeness and classification stability. These results further illustrate that for small-sample microseismic recognition problems, structural design, rather than the training strategy itself, is a key factor determining the upper limit of model performance, and Tr-BiLSTM-DCNN is a network structure highly adapted to this application scenario.

[0081] The embodiments described above are merely illustrative of implementation methods of the present invention and should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for identifying microseismic signals based on transfer learning and BiLSTM-DCNN, characterized in that, Includes the following steps: S1. Collect various types of microseismic signals from the source mine to construct the original signal dataset; S2. After performing preprocessing and short-time Fourier transform on the original signal, extract the Mel-frequency spectral features to obtain the Mel-time spectrum and construct a waveform feature dataset; S3. Construct a BiLSTM-DCNN hybrid neural network model, including a temporal feature extraction module, a spatial feature extraction module, and a classification output module, using the Mel-time spectrum as the input signal, and finally outputting the predicted labels for each signal segment; The BiLSTM-DCNN hybrid neural network model includes: The temporal feature extraction module first uses a two-layer bidirectional long short-term memory network (BiLSTM) to learn the input signal, so as to capture the forward and backward dependencies simultaneously. After each BiLSTM, a batch normalization operation is introduced to stabilize the training process, and a dropout layer is set at key nodes to reduce the risk of overfitting. Finally, the temporal embedding vector is output. The spatial feature extraction module, after obtaining the temporal embedding vector, first uses a one-dimensional convolutional layer to encode the local receptive field, and then combines batch normalization and pooling operations to reduce the feature dimension; then, it stacks convolutional layers and uses the same normalization and pooling operations to extract higher-level discriminative features. After feature extraction, the classification output module compresses the time dimension into a fixed-length feature vector through global average pooling and then inputs it into two fully connected layers. After each fully connected layer, a batch normalization and dropout layer is added to further improve the generalization performance. Finally, the classification task is completed through the fully connected layers, and the probability distribution of the three types of signals is output. The category with the highest probability is taken as the final predicted label. S4. Use the source mine waveform feature dataset to pre-train the BiLSTM-DCNN hybrid neural network model and optimize the model parameters; S5. Using a transfer learning strategy, the pre-trained model is transferred to the target mine, and the parameters of the pre-trained model are fine-tuned using a small sample dataset to obtain the Tr-BiLSTM-DCNN model. S6. Input the microseismic signal to be identified into the Tr-BiLSTM-DCNN model and output the signal category identification result.

2. The microseismic signal recognition method based on transfer learning and BiLSTM-DCNN according to claim 1, characterized in that, Specifically, S2 is: The preprocessing process includes: truncating each signal to a uniform length using a fixed time window; resampling the signal segments; pre-emphasizing the signal segments using a first-order high-pass filter; dividing the continuous signal stream of each signal segment into multiple short-time analysis frames, and setting appropriate overlap between adjacent frames to maintain signal continuity; and applying a Hanning window function to smooth the data in each frame. After preprocessing, a short-time Fourier transform is performed on each frame of signal to map it from the time domain to the time-frequency domain, obtaining a complex spectrum that characterizes the time-frequency distribution of the signal. The short-time amplitude spectrum is obtained by calculating the modulus of the complex spectrum. The short-time amplitude spectrum is filtered using a filter bank to transform it to the Mel scale, thus obtaining the Mel-time spectrum.

3. The microseismic signal recognition method based on transfer learning and BiLSTM-DCNN according to claim 2, characterized in that, The transformation formula for converting the short-time amplitude spectrum to the Mel scale is as follows: In the formula, k is the Mel-time spectrum frequency; f is the short-time amplitude spectrum frequency.

4. The microseismic signal recognition method based on transfer learning and BiLSTM-DCNN according to claim 1, characterized in that, Specifically, S5 involves: collecting multiple target signals from the target mine, processing them through S1 and S2 to obtain the waveform feature dataset of the target mine, inputting it into the BiLSTM-DCNN hybrid neural network model for transfer learning, and training and fine-tuning the model parameters to obtain the Tr-BiLSTM-DCNN model.

5. A method for identifying microseismic signals based on transfer learning and BiLSTM-DCNN according to claim 1 or 4, characterized in that, The parameters of the fine-tuned BiLSTM-DCNN hybrid neural network model include the parameters of all layers except those after the global average pooling layer in the classification output module.

Images