A Microseismic Monitoring Method for Mines Based on Large Language Model and LoRA
By combining a large language model with LoRA, efficient and accurate microseismic monitoring of mines was achieved, solving the problems of low efficiency and poor accuracy of traditional methods and improving the robustness and generalization ability of the model in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CCTEG COAL MINING RES INST
- Filing Date
- 2026-03-27
- Publication Date
- 2026-06-30
Smart Images
Figure CN122307676A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of mine microseismic monitoring technology, and in particular to a mine microseismic monitoring method, device, electronic device and computer-readable storage medium based on large language model and LoRA. Background Technology
[0002] In mining environments such as coal mining faces, massive amounts of microseismic data are continuously generated by activities such as coal cutting, tunneling, and roof fracturing. Traditional microseismic monitoring methods have the following shortcomings: manual processing is extremely inefficient, while existing deep learning methods face a series of technical bottlenecks when processing these mine microseismic data characterized by small-scale ruptures.
[0003] Traditional methods, such as short-time / long-time averaging and grid search localization, heavily rely on parameters like feature functions, time window lengths, and thresholds for accuracy. These parameters require repeated adjustments based on the noise characteristics and geological conditions of different mining areas, and need to be recalibrated when the working face environment changes, making it difficult to meet practical needs. In recent years, deep learning methods have been widely used in microseismic monitoring, but they typically require designing dedicated networks for each monitoring task, preventing the use of inter-task correlations to improve overall performance. Furthermore, existing deep learning models lack generalization ability with limited training samples; their performance often degrades significantly when mining areas or geological conditions change. In complex scenarios with strong noise interference from mining machinery, the robustness of these models is also difficult to guarantee, and their ability to identify weak first-arrival signals submerged in strong background noise is limited. Summary of the Invention
[0004] This application aims to at least partially address one of the technical problems in the related art.
[0005] Therefore, the first objective of this application is to propose a mine microseismic monitoring method based on large language models and LoRA, in order to solve the problems of low efficiency, poor accuracy and high cost of existing technologies.
[0006] The second objective of this application is to provide an apparatus.
[0007] The third objective of this application is to propose an electronic device.
[0008] The fourth objective of this application is to provide a computer-readable storage medium.
[0009] To achieve the above objectives, the first aspect of this application proposes a method for monitoring microseismic events in mines based on large language models and LoRA, comprising:
[0010] Acquire raw three-component microseismic waveform data collected from the mine; The original three-component microseismic waveform data is input into a multi-scale convolutional embedder. Multi-scale local features are extracted using convolutional kernels of different sizes, and the sequence length is compressed to obtain a compressed feature sequence. The compressed feature sequence is input into the potential patch module, and a short token sequence is generated by aggregating features from multiple consecutive time steps into a patch. The short token sequence is input into a pre-trained large language model block. The pre-trained large language model block adopts a low-rank adaptive fine-tuning strategy. The original parameters of the pre-trained large language model block, except for the position embedding and layer normalization layer, are frozen, and a trainable low-rank decomposition matrix is inserted into the self-attention layer and the feedforward network for processing to extract deep temporal features. The deep temporal features are input into a specific task output head to generate microseismic monitoring results, which include at least one of the following: P-wave first arrival probability sequence, S-wave first arrival probability sequence, epicentral distance estimate, and reverse azimuth estimate.
[0011] Preferably, the step of extracting multi-scale local features using convolutional kernels of different sizes and compressing the sequence length to obtain a compressed feature sequence includes: Perform a linear projection transformation on the input raw three-component microseismic waveform data; The projected data is input in parallel to multiple convolution branches with different kernel sizes. Each branch performs convolution operations, batch normalization, and activation function processing in sequence. The outputs of each branch are concatenated dimensionally, and the concatenated features are integrated by linear projection to obtain a compressed feature sequence.
[0012] Preferably, the step of generating a short token sequence by aggregating features from multiple consecutive time steps into patches includes: The compressed feature sequence is reshaped into multiple patches, each patch containing features from multiple consecutive time steps; The features within each patch are aggregated to generate short token sequences.
[0013] Preferably, the pre-trained large language model block is a backbone network based on generative pre-trained Transformer, and the low-rank adaptive fine-tuning strategy includes: Freeze the original parameters of the pre-trained large language model block, except for the position embedding and layer normalization layers; Make the position embedding and layer normalization layer trainable; A trainable low-rank decomposition matrix is inserted into the self-attention layer and the feedforward network, and the original weights are updated using the low-rank decomposition matrix.
[0014] Preferably, the specific task output head includes a phase pickup output head and a regression task output head; The phase pickup output head restores the deep temporal features to the original waveform length through upsampling and convolution blocks, and uses the Sigmoid function to generate the first arrival probability sequences of P-waves and S-waves; The regression task output head aggregates information through stacked convolutional layers, and outputs the epicentral distance estimate, inverse azimuth estimate, or magnitude estimate after global average pooling and linear projection.
[0015] Preferably, it also includes model pre-training, which is performed before acquiring the raw three-component microseismic waveform data collected from the mine, including: Obtain tagged mine microseismic waveform samples, the tags including P-wave arrival time, S-wave arrival time, epicentral distance, and reverse azimuth. Data augmentation processing is performed on the microseismic waveform samples of the mine, including at least one of random Gaussian noise addition, time drift, and channel dropping. The data-augmented samples are input into the multi-scale convolutional embedder, the latent patch module, the pre-trained large language model block, and the task-specific output head to obtain the prediction results. Based on the prediction results and the labels, the multi-task joint loss is calculated, and the model parameters are optimized according to the multi-task joint loss.
[0016] Preferably, the multi-task joint loss includes phase picking loss and regression task loss; The phase pickup loss is the sum of the binary cross-entropy losses of the P-wave and S-wave. The regression task loss is the Huber loss.
[0017] To achieve the above objectives, a second aspect of this application proposes a mine microseismic monitoring device based on a large language model and LoRA, comprising: The acquisition module is used to acquire the raw three-component microseismic waveform data collected from the mine. The multi-scale convolutional embedding module is used to input the original three-component microseismic waveform data into the multi-scale convolutional embedder, extract multi-scale local features through convolutional kernels of different sizes, and compress the sequence length to obtain a compressed feature sequence. A latent patch module is used to input the compressed feature sequence into the latent patch module, and generate a short token sequence by aggregating the features of multiple consecutive time steps into patches. The large language model processing module is used to input the short token sequence into the pre-trained large language model block. The pre-trained large language model block adopts a low-rank adaptive fine-tuning strategy. It freezes the original parameters of the pre-trained large language model block except for the position embedding and layer normalization layer, and inserts a trainable low-rank decomposition matrix into the self-attention layer and feedforward network for processing to extract deep temporal features. The output module is used to input the deep temporal features into a specific task output head to generate microseismic monitoring results. The monitoring results include at least one of the following: P-wave first arrival probability sequence, S-wave first arrival probability sequence, epicentral distance estimate, and reverse azimuth estimate.
[0018] To achieve the above objectives, a third aspect of this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method described in any of the preceding descriptions.
[0019] To achieve the above objectives, a fourth aspect of this application provides a computer-readable storage medium, comprising computer-executable instructions stored therein, which, when executed by a processor, are used to implement the method described in any of the above embodiments.
[0020] This application presents a mine microseismic monitoring method based on a large language model and LoRA. By introducing a cross-modal transfer learning strategy, the powerful sequence modeling capabilities learned by a pre-trained large language model on text corpora are transferred to the mine microseismic monitoring task, achieving excellent feature extraction capabilities without relying on large-scale seismic datasets for pre-training. A unified multi-task learning framework is constructed, taking the original three-component waveform as input and directly outputting multiple monitoring results such as P-wave and S-wave first-arrival probability sequences, epicentral distance, inverse azimuth, and magnitude, realizing end-to-end learning from first-arrival acquisition to event localization. Employing low-rank adaptive technology, cross-modal transfer is achieved by fine-tuning only a few parameters in the large language model, significantly reducing training costs and memory usage. Local continuous features are extracted through a multi-scale convolutional embedder, combined with the sequence modeling capabilities of the pre-trained large language model, improving the model's robustness in environments with strong background noise.
[0021] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0022] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a first specific embodiment of a mine microseismic monitoring method based on a large language model and LoRA provided by the present invention; Figure 2 A flowchart illustrating a second specific embodiment of a mine microseismic monitoring method based on a large language model and LoRA provided by the present invention; Figure 3 A schematic diagram of the core architecture of the intelligent microseismic monitoring model for mines; Figure 4 A schematic diagram of microseismic data acquisition results for coal seam roof fracturing based on SeisMoLLM; Figure 5 A schematic diagram illustrating the picking performance of microseismic data without clear S-waves; Figure 6 A two-dimensional image of the positioning results; Figure 7 A three-dimensional schematic diagram of the positioning results; Figure 8 This is a map showing the distribution of positioning prediction errors. Figure 9 This is a structural block diagram of a mine microseismic monitoring device based on a large language model and LoRA, provided as an embodiment of the present invention. Detailed Implementation
[0023] The core of this invention is to provide a method, device, electronic device, and computer-readable storage medium for mine microseismic monitoring based on a large language model and LoRA. By transferring a pre-trained large language model across modes to mine microseismic monitoring, and combining LoRA for efficient fine-tuning and multi-scale feature extraction, multi-task integrated processing and accurate positioning are achieved.
[0024] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] Please refer to Figure 1 , Figure 1 The flowchart illustrates a first specific embodiment of a mine microseismic monitoring method based on a large language model and LoRA provided by the present invention; the specific operation steps are as follows: Step S101: Obtain the raw three-component microseismic waveform data collected from the mine; Step S102: Input the original three-component microseismic waveform data into a multi-scale convolutional embedder, extract multi-scale local features through convolutional kernels of different sizes, and compress the sequence length to obtain a compressed feature sequence; Step S103: Input the compressed feature sequence into the latent patch module, and generate a short token sequence by aggregating the features of multiple consecutive time steps into patches; Step S104: Input the short token sequence into the pre-trained large language model block. The pre-trained large language model block adopts a low-rank adaptive fine-tuning strategy. By freezing the original parameters of the pre-trained large language model block except for the position embedding and layer normalization layer, and inserting trainable low-rank decomposition matrices into the self-attention layer and feedforward network for processing, deep temporal features are extracted. Step S105: Input the deep temporal features into the specific task output head to generate microseismic monitoring results. The monitoring results include at least one of the following: P-wave first arrival probability sequence, S-wave first arrival probability sequence, epicentral distance estimate, and reverse azimuth estimate.
[0026] Based on the above embodiments, this embodiment will provide a detailed description of step S101: Specifically, the raw three-component microseismic waveform data serves as the fundamental input for mine microseismic monitoring, containing waveform records in the east-west, north-south, and vertical directions. This data originates from microseismic records during actual production processes at the coal mine face, encompassing micro-fracture events induced by various production activities such as coal cutting, tunneling operations, and roof fracturing. Each sample is recorded in three-component waveform form, effectively reflecting the true characteristics of microseismic signals in the mining environment. After acquiring this data, further unified data augmentation processing can be performed, including randomly adding Gaussian noise, time drift, and channel discarding, to improve the model's robustness in complex noise environments.
[0027] Based on the above embodiments, this embodiment will provide a detailed description of step S102: In one embodiment, a linear projection transformation is performed on the input raw three-component microseismic waveform data; The projected data is input in parallel to multiple convolution branches with different kernel sizes. Each branch performs convolution operations, batch normalization, and activation function processing in sequence. The outputs of each branch are concatenated dimensionally, and the concatenated features are integrated by linear projection to obtain a compressed feature sequence.
[0028] Specifically, the multi-scale convolutional embedder employs a multi-scale convolutional design with parallel convolutional kernels of varying sizes. These kernels can extract local details in parallel, effectively capturing multi-scale features. Through stacked convolutional layers, the sequence length can be progressively compressed, reducing the input waveform length to 1 / 64 of its original size, significantly reducing subsequent computational costs. The compressed feature sequence output by this module retains key information from the original waveform while removing redundant details, providing an efficient feature representation for subsequent processing.
[0029] Based on the above embodiments, this embodiment will provide a detailed description of step S103: In one embodiment, the compressed feature sequence is reshaped into multiple patches, each patch containing features from multiple consecutive time steps; The features within each patch are aggregated to generate short token sequences.
[0030] Specifically, the latent patch module further aggregates the feature sequence output by the embedder. This module reshapes the feature sequence into patch form, with each patch containing features from multiple consecutive time steps, generating a shorter token sequence through aggregation. This processing method avoids information loss caused by discrete partitioning and further compresses the sequence length, significantly reducing the computational load of subsequent large language model modules. In this embodiment, both the patch size and step size are preset values, reducing the number of tokens to 1 / 8 of the original.
[0031] Based on the above embodiments, this embodiment will provide a detailed description of step S104: In one embodiment, the pre-trained large language model block is a backbone network based on a generative pre-trained Transformer, and the low-rank adaptive fine-tuning strategy includes: Freeze the original parameters of the pre-trained large language model block, except for the position embedding and layer normalization layers; Make the position embedding and layer normalization layer trainable; A trainable low-rank decomposition matrix is inserted into the self-attention layer and the feedforward network, and the original weights are updated using the low-rank decomposition matrix.
[0032] Specifically, the pre-trained large language model block is the core feature extractor in this embodiment, employing a generative pre-trained Transformer as the backbone network. This model, pre-trained on a large-scale text corpus, possesses powerful sequence modeling capabilities. To achieve cross-modal transfer, most of the original parameters in the backbone network are frozen, with only the position embedding and layer normalization layers set as trainable to better adapt to the temporal structure of seismic waveforms. For the self-attention layer and feedforward network, a low-rank adaptive technique is introduced, allowing for efficient fine-tuning by inserting trainable low-rank decomposition matrices. Only a small number of parameters need to be set as trainable parameters to achieve efficient cross-modal fine-tuning.
[0033] Based on the above embodiments, this embodiment will provide a detailed description of step S105: In one embodiment, a specific task output head includes a phase pickup output head and a regression task output head; The phase pickup output head restores the deep temporal features to the original waveform length through upsampling and convolution blocks, and uses the Sigmoid function to generate the first arrival probability sequences of P-waves and S-waves; The regression task output head aggregates information through stacked convolutional layers, and outputs the epicentral distance estimate, inverse azimuth estimate, or magnitude estimate after global average pooling and linear projection.
[0034] Specifically, the output head for specific tasks is designed with a simple decoding structure based on different task requirements. For phase picking tasks, upsampling and convolutional blocks are used to gradually restore the latent features to the original waveform length, and then the first arrival probability sequences of P-waves and S-waves are generated through the Sigmoid function. For regression tasks such as epicentral distance estimation and inverse azimuth estimation, stacked convolutional layers are used to further aggregate information, and the predicted values are output after global average pooling and linear projection. Through this embodiment, a single model can simultaneously complete multiple monitoring tasks such as first arrival picking, epicentral distance estimation, and inverse azimuth estimation, realizing end-to-end learning from first arrival picking to event localization.
[0035] In some possible implementations, model pre-training is also included, which is performed before acquiring the raw three-component microseismic waveform data collected from the mine, including: Obtain tagged microseismic waveform samples from the mine, with tags including P-wave arrival time, S-wave arrival time, epicentral distance, and reverse azimuth. Data augmentation processing is performed on the microseismic waveform samples from the mine, including at least one of random Gaussian noise addition, time drift, and channel dropping. The data-augmented samples are input into a multi-scale convolutional embedder, a latent patch module, a pre-trained large language model block, and a task-specific output head to obtain the prediction results. Based on the prediction results and the labels, the multi-task joint loss is calculated, and the model parameters are optimized according to the multi-task joint loss.
[0036] The multi-task joint loss includes phase picking loss and regression task loss; The phase picking loss is the sum of the binary cross-entropy losses of the P-wave and S-wave; the regression task loss is the Huber loss.
[0037] Specifically, tagged microseismic waveform samples from the mine were acquired. Microseismic monitoring data from a coal mining area was used as the data basis for model training and testing. This dataset originated from microseismic records during actual production processes at the coal mining face, covering micro-fracture events induced by various production activities such as coal cutting, tunneling operations, and roof fracturing. The total data volume was approximately 109 GB, with a sampling frequency of 500 Hz. Each sample was recorded in a three-component waveform format, containing waveform data in the east-west, north-south, and vertical directions. Data augmentation processing was performed on the mine microseismic waveform samples, using methods such as randomly adding Gaussian noise, time drift, or channel dropping. The augmented samples were then input into the model to obtain prediction results, and the multi-task joint loss was calculated to optimize the model parameters.
[0038] This embodiment provides a mine microseismic monitoring method based on a large language model and LoRA. By introducing a cross-modal transfer learning strategy, the powerful sequence modeling capabilities learned by the pre-trained large language model on text corpora are transferred to the mine microseismic monitoring task, achieving excellent feature extraction capabilities without relying on large-scale seismic datasets for pre-training. A unified multi-task learning framework is constructed, taking the original three-component waveform as input and directly outputting multiple monitoring results such as P-wave and S-wave first-arrival probability sequences, epicentral distance, inverse azimuth, and magnitude, realizing end-to-end learning from first-arrival acquisition to event localization. Low-rank adaptive technology is employed, achieving cross-modal transfer by fine-tuning only a few parameters in the large language model, significantly reducing training costs and memory usage. Local continuous features are extracted through a multi-scale convolutional embedder, combined with the sequence modeling capabilities of the pre-trained large language model, improving the model's robustness in environments with strong background noise.
[0039] Based on the above embodiments, this embodiment describes a mine microseismic monitoring method based on large language models and LoRA, such as... Figure 2 As shown, the details are as follows: S201, acquire the raw three-component microseismic waveform data collected from the mine.
[0040] Specifically, microseismic monitoring data from a coal mining area is used as the data basis for model training and testing. Each sample is recorded in the form of a three-component waveform, including waveform data in the east-west, north-south, and vertical directions. Each seismic sample contains 35 attribute labels to describe the basic characteristics of the event, phase information, station information, and waveform quality. This invention selects relevant labels for supervised learning according to task requirements. The phase picking task uses P-wave and S-wave first-arrival time labels, constructing them into a Gaussian-shaped probability sequence. That is, the probability value at the first-arrival time is 1, gradually decaying to 0 within a 0.5-second range according to a Gaussian distribution, enabling the model to learn the continuous probability output at the first-arrival time. The epicentral distance estimation task directly uses the spatial distance from the source to the station as the supervision signal. This distance can be calculated based on the source coordinates (Ns, Es, Ds) and station coordinates (Nr, Er, Dr). The inverse azimuth estimation task uses inverse azimuth angle labels and employs sine and cosine values as learning targets to handle the periodicity of the angle. In addition, the dataset also contains information such as the time of the event and the event label, which can be used for sample screening and result tracing. It should be noted that the focal depth and occurrence time of different events vary in actual mining environments, resulting in significant variations in the waveform signal-to-noise ratio. This embodiment performs uniform data augmentation on all waveform samples during training, including randomly adding Gaussian noise, time drift, and channel dropping, to improve the model's robustness in complex noise environments. Other unused attribute labels in the dataset, such as ray parameters and particle motion linearity, can be used for subsequent model validation and physical interpretability analysis. Combining these labels with a multi-task learning framework enables a single model to simultaneously complete multiple monitoring tasks, including first arrival picking, epicentral distance estimation, inverse azimuth estimation, and magnitude estimation.
[0041] S202 performs a linear projection transformation on the input raw three-component microseismic waveform data.
[0042] In one embodiment, the multi-scale convolutional embedder first performs a linear projection transformation on the input original three-component microseismic waveform data, mapping the original waveform data to a high-dimensional feature space, in preparation for subsequent multi-scale feature extraction.
[0043] Specifically, the multi-scale convolutional embedder module employs a multi-scale convolutional design, featuring parallel convolutional kernels of varying sizes to extract local details. This effectively extracts multi-scale features and compresses sequence length. The calculation formula is as follows:
[0044]
[0045] in, Indicates a linear projection layer. Indicates the kernel size as Convolutional layers, Indicates the batch normalization layer. For activation function, This indicates a dimension concatenation operation. Input First, the projection is transformed by a linear projection layer, and the projection results are input in parallel to... There are several convolutional branches with different kernel sizes. Each branch sequentially undergoes convolution, normalization, and... Activation; then the outputs of each branch are concatenated; the concatenated features are integrated in a linear projection layer, and finally normalized to obtain the final output. .
[0046] S203 inputs the projected data in parallel to multiple convolution branches with different kernel sizes. Each branch performs convolution operations, batch normalization, and activation function processing in sequence.
[0047] Specifically, the multi-scale convolutional embedder employs a multi-scale convolutional design, where the projection results are input in parallel to n convolutional branches with different kernel sizes. Each branch sequentially undergoes convolution, batch normalization, and a GELU activation function. For example, three convolutional branches can be set with kernel sizes of 3, 5, and 7, respectively. Each branch contains a convolutional layer, a batch normalization layer, and an activation function to extract local features at different scales. This parallel multi-scale design can effectively capture features of different frequency components in the waveform.
[0048] S204 concatenates the outputs of each branch by dimension, and then performs linear projection integration on the concatenated features to obtain a compressed feature sequence.
[0049] Specifically, the outputs of each branch are concatenated together, the concatenated features are integrated in a linear projection layer, and finally normalized to obtain a compressed feature sequence.
[0050] S205 reshapes the compressed feature sequence into multiple patches, aggregates the features within each patch, and generates a short token sequence.
[0051] Specifically, the latent patching module further aggregates the feature sequence output by the embedder. This module reshapes the feature sequence into patch form, with each patch containing features from multiple consecutive time steps, generating a shorter token sequence through aggregation. For the input feature sequence... First divide into blocks ,in, Let P be the number of patches, P be the patch size, and the final token sequence dimension be... In this invention, both the patch size and step size are set to 8, reducing the number of tokens to 1 / 8 of the original. Compared with the traditional method of patching directly on the original waveform, this method can avoid information loss caused by discrete partitioning, and further compress the sequence length, thus significantly reducing the computational load of the subsequent LLM module.
[0052] S206, input the short token sequence into the pre-trained large language model block, which adopts a low-rank adaptive fine-tuning strategy to extract deep temporal features.
[0053] Specifically, the pre-trained large language model block uses a minimal version of the generative pre-trained Transformer as its backbone, containing 12 decoder layers and a hidden layer dimension of 768. In the low-rank adaptive fine-tuning strategy, most of the original parameters in the backbone network are frozen, with only the position embedding and layer normalization layers set as trainable to better adapt to the temporal structure of seismic waveforms. For the self-attention layer and the feedforward network, a low-rank adaptive technique is introduced, using the insertion of trainable low-rank decomposition matrices for efficient fine-tuning. Specifically, the update of the original weights W is expressed as... ,in, It is decomposed into the product of two low-rank matrices. , , , where rank much smaller and In this embodiment, the rank Set it to 16. In this way, efficient fine-tuning across modalities can be achieved by setting only about 10% of the parameters as trainable parameters.
[0054] S207 inputs deep temporal characteristics into a specific task output header to generate microseismic monitoring results.
[0055] Specifically, the output heads for specific tasks include a phase picking output head, a regression task output head, and a classification task output head. The phase picking output head recovers deep temporal features to the original waveform length through upsampling and convolutional blocks, and generates P-wave and S-wave first-arrival probability sequences using the Sigmoid function. The regression task output head aggregates information by stacking convolutional layers, and outputs an epicentral distance estimate, an inverse azimuth estimate, or a magnitude estimate after global average pooling and linear projection. For the inverse azimuth estimation task, sine and cosine are used as learning targets, and the Tanh function is used to constrain the output range to handle the periodicity of the angle.
[0056] S208 optimizes model parameters based on training samples.
[0057] Specifically, this embodiment also includes a model training step. First, tagged microseismic waveform samples from the mine are acquired. The tags include the P-wave first arrival time, S-wave first arrival time, epicentral distance, and inverse azimuth. The phase picking task's tags are constructed as a Gaussian-shaped probability sequence, meaning the probability of the location at the first arrival time is 1, gradually decreasing to 0 within a 0.5-second interval following a Gaussian distribution. The epicentral distance tag is calculated based on the source coordinates and station coordinates. The inverse azimuth tag uses sine and cosine values as learning targets.
[0058] Secondly, the microseismic waveform samples from the mine are subjected to data augmentation processing, including randomly adding at least one of Gaussian noise, time drift, and channel dropping, in order to simulate waveform changes under the complex environment of the mine.
[0059] Next, the data-enhanced samples are input into the above model to obtain the prediction results.
[0060] Finally, based on the prediction results and labels, the multi-task joint loss is calculated, and the model parameters are optimized according to the multi-task joint loss. The multi-task joint loss includes phase picking loss and regression task loss. The phase picking loss is the sum of the binary cross-entropy losses of P-waves and S-waves. The regression task loss is the Huber loss; for the three regression tasks of epicentral distance estimation, inverse azimuth estimation, and magnitude estimation, the Huber loss is used for optimization. Set to 1. During training, the Adam optimizer is used, with an initial learning rate of 1e-4. A cyclic learning rate scheduling strategy is adopted to make the learning rate oscillate periodically between 1e-4 and 5e-4, and an early stopping strategy is introduced to prevent overfitting.
[0061] Through the above steps, this embodiment enables integrated intelligent monitoring from initial acquisition to event localization. Experimental results show that, in coal mine microseismic monitoring data, this method achieves a detection accuracy of 98.2% with a mean absolute error of 0.498 seconds in P-wave acquisition; a mean absolute error of only 11.342° with a coefficient of determination of 0.948 in reverse azimuth estimation; a mean absolute error of 2.313 km with a coefficient of determination of 0.990 in epicentral distance estimation; and a mean absolute error of 0.158 in magnitude estimation. These performance characteristics demonstrate the application potential of this method in mine microseismic monitoring.
[0062] To make the mine microseismic monitoring method based on large language model and LoRA provided in this disclosure clearer, the following is combined with... Figure 3 Please provide an explanation. For example... Figure 3 As shown, the specific steps are as follows: 1. Dataset and Labels: The microseismic monitoring data of a coal mining area was used as the data basis for model training and testing. The above embodiments have been described in detail and will not be explained further here. 2. Model Architecture: The intelligent microseismic monitoring model for mines proposed in this embodiment consists of four core components: a multi-scale convolutional embedder, a latent patch module, a pre-trained large language model block, and a task-specific output head. The original three-component waveform is first input into the multi-scale convolutional embedder, where convolutional kernels of different sizes extract local features and compress the sequence length. Subsequently, the latent patch module further aggregates the waveforms into shorter token sequences. The pre-trained LLM (Large Language Model) block is based on a generative pre-trained Transformer 2 (GPT-2) backbone network and uses a LoRA parameter efficient fine-tuning strategy to process the token sequences to extract deep features. Finally, the task-specific output head outputs prediction results such as first arrival probability, epicentral distance, and inverse azimuth.
[0063] First is the multi-scale convolutional embedder. This module adopts a multi-scale convolutional design with parallel convolutional kernels of different sizes to extract local details, which can effectively extract multi-scale features and compress sequence length. Indicates a linear projection layer. Indicates the kernel size as The convolutional layers are defined by Batch Normalization (BN), GELU as the activation function, and Concat as the dimension concatenation operation. The input X is first transformed by a linear projection layer, and the projection results are fed in parallel into n convolutional branches with different kernel sizes. Each branch sequentially undergoes convolution, normalization, and GELU activation. The outputs of each branch are then concatenated. The concatenated features are integrated in the linear projection layer and finally normalized to obtain the final output Y.
[0064]
[0065]
[0066] Compared to existing techniques that directly slice waveforms to a fixed size and embed them linearly, this convolutional embedder can more effectively capture local continuous features and adapt to different tasks through multi-scale design. Furthermore, the stacked convolutional layers can progressively compress the sequence length, reducing the input waveform length to 1 / 64 of its original length, significantly lowering computational costs.
[0067] The latent patching module further aggregates the feature sequence output by the embedder. This module reshapes the feature sequence into patch form, with each patch containing features from multiple consecutive time steps, generating a shorter token sequence through aggregation. For the input feature sequence... First divide into blocks Where N = L / P is the number of patches, P is the patch size, and the final token sequence dimension is... In this embodiment, a patch size and step size of 8 are both used, reducing the number of tokens to 1 / 8 of the original. Compared with the traditional method of patching directly on the original waveform, this method can avoid information loss caused by discrete partitioning, and further compress the sequence length, thus significantly reducing the computational load of the subsequent LLM module.
[0068] The pre-trained large language module is the core feature extractor in this embodiment. It uses a minimal version of GPT-2 as its backbone, containing 12 decoder layers and 768 hidden layers. This model, pre-trained on a large-scale text corpus, possesses powerful sequence modeling capabilities. To achieve cross-modal transfer, most parameters in GPT-2 are frozen, with only the position embedding and layer normalization layers made trainable to better adapt to the temporal structure of seismic waveforms. For the self-attention layer and feedforward network, LoRA technology is introduced, enabling efficient fine-tuning by inserting trainable decomposition matrices. Only 10% of the parameters need to be set as trainable to achieve efficient cross-modal fine-tuning, as shown in the following formula:
[0069] in, For the original weights, , It is a trainable low-rank matrix with rank r=16.
[0070] The task-specific output head features a simple decoding structure designed to meet the specific needs of each task. For phase picking, upsampling and convolutional blocks are used to progressively restore the latent features to the original waveform length, and then the sigmoid function is used to generate the first arrival probability sequences of P-waves and S-waves. For regression tasks such as epicentral distance estimation, inverse azimuth estimation, and magnitude estimation, stacked convolutional layers are used to further aggregate information, and the predicted values are output after global average pooling and linear projection. The inverse azimuth uses sine and cosine as learning targets, and the Tanh function is used to constrain the output range. For the initial motion polarity classification task, a two-dimensional probability is output through Softmax. The task-specific output head has a simple and efficient structure, avoiding the drawbacks of designing complex dedicated networks for each task, and achieving end-to-end mapping of multiple tasks within a unified framework.
[0071] 3. Model Training and Evaluation: In the phase picking task, the sum of the binary cross-entropy losses of P-wave and S-wave is used as the basis for calculation.
[0072]
[0073]
[0074] in, and The binary cross-entropy losses for P-waves and S-waves are respectively. Labels for Gaussian shapes. These are the probability values output by the model.
[0075] For the three regression tasks of epicentral distance estimation, inverse azimuth estimation, and magnitude estimation, Huber loss is used for optimization.
[0076]
[0077] in, For the true value, For predicted values, Set it to 1.
[0078] The Adam optimizer was used during training, with an initial learning rate set to... Simultaneously, a cyclic learning rate scheduling strategy is adopted to ensure that the learning rate... to The model oscillates periodically to help it escape local optima. An early stopping strategy is introduced during training: training stops when the validation set loss does not decrease for 30 consecutive epochs to prevent overfitting. Data augmentation techniques, including randomly adding Gaussian noise and channel dropping, are also used to better simulate waveform changes in the complex environment of a mining site and improve the robustness of the model.
[0079] Precision, recall, and F1 score are used as evaluation metrics for phase picking and initial polarity classification tasks, and the calculation formulas are as follows:
[0080]
[0081]
[0082] Where TP represents the number of correctly identified positive samples, FP represents the number of falsely detected positive samples, and FN represents the number of falsely detected positive samples. Precision reflects the accuracy of the model's identification, recall reflects the model's coverage of positive samples, and the F1 score is the harmonic mean of the two, which can comprehensively evaluate the model's classification performance. The higher the values of these metrics, the better the model performs in phase picking and initial polarity discrimination.
[0083] For the three regression tasks of epicentral distance estimation, inverse azimuth estimation, and magnitude estimation, the coefficient of determination was used. The mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) are used as evaluation metrics, as shown below:
[0084]
[0085]
[0086]
[0087] in, For the true value, For predicted values, The mean of the true values is given. The coefficient of determination reflects the extent to which the model explains the target variable; the closer the value is to 1, the better the model fits. The mean absolute error measures the average deviation between the predicted and true values; the smaller the value, the higher the estimation accuracy. The mean absolute percentage error reflects the magnitude of the prediction error as a percentage; the smaller the value, the higher the estimation accuracy of the model and the better the fit to the actual data. The root mean square error is the core indicator for measuring the deviation between the predicted and true values of the regression model; the smaller the value, the higher the estimation accuracy of the model and the better the fit to the target variable.
[0088] 4. Simulation performance analysis: To better evaluate the effectiveness of this method in mine microseismic monitoring, quantitative tests were conducted using microseismic monitoring data from a coal mine area. The diversity of this dataset effectively reflects the complex and variable noise background in the mining environment, especially in scenarios where mining machinery noise and microseismic signals are highly mixed. A comparative analysis was also performed with existing mainstream methods, including PhaseNet, EQ Transformer, SeisT-L, MagNet, and BAZ network.
[0089] Phase picking is a fundamental task in microseismic monitoring. Accurate P-wave and S-wave arrival times are prerequisites for estimating epicentral distance and reverse azimuth. Especially in the high-noise environment of mines, slight changes in arrival times can lead to positioning results with deviations of tens of meters. Figure 4 This study demonstrates the picking performance of SeisMoLLM on coal seam roof fracturing data. To verify the model's generalization ability under different geological conditions, 11,205 coal seam roof fracturing microseismic records were included in the training dataset. Figure 4 (a) is a microseismic waveform image of the site, which contains strong background noise and low signal-to-noise ratio first arrival phase. The background noise is strong and continuous, and the first arrival of P-waves and S-waves is completely submerged in the noise, reflecting the complexity of the mine monitoring environment and the low signal-to-noise ratio. Figure 4(c) The first arrival of the P-wave and S-wave superimposed on the original waveform is picked up. The yellow and red circles mark the first arrival positions of the P-wave and S-wave, respectively. Under strong noise background, the model can still accurately locate the first arrival time, and the picking position is very close to the waveform start point, which verifies the effectiveness and robustness of the model under low signal-to-noise ratio conditions.
[0090] To address the issue of ambiguous S-wave first arrivals in microseismic monitoring, the performance of the model in microseismic records with unclear S-waves is evaluated. Figure 5 (a) The original waveform contains a typical microseismic event, with obvious P-wave first arrival characteristics but no detectable S-wave phase. Figure 5 (b) demonstrates the picking performance of the SeisMoLLM model for the first arrival of the P wave. In the absence of a significant S wave, the picking position of the P wave is almost consistent with the waveform start point, indicating that the model's advanced feature extraction algorithm can still maintain stable and reliable performance, which is significantly better than traditional methods.
[0091] The detailed comparative results of this embodiment on various tasks of microseismic monitoring data in coal mining areas are as follows: In the P-wave and S-wave picking tasks, this embodiment achieves a detection accuracy of 98.2% and a mean absolute error of 0.498 seconds, superior to PhaseNet's 0.556 seconds, EQ Transformer's 0.590 seconds, and SeisT-L's 0.555 seconds. In the S-wave picking task, this embodiment achieves a detection accuracy of 82.26%, superior to PhaseNet's 80.04%, EQ Transformer's 80.57%, and SeisT-L's 80.91%, with a recall of 81.08% and a mean absolute error of 2.258 seconds. All four indicators are superior to all compared methods. S-wave picking has always been a challenge in phase picking because S-waves are often superimposed on the coda of P-waves and are more susceptible to noise interference. The leading advantage of this embodiment in this task verifies the ability of pre-trained large language models to extract weak seismic phase features from complex waveforms.
[0092] Epicentral distance and inverse azimuth are two key parameters for single-station localization. Experimental results show that this embodiment achieves a significant improvement in inverse azimuth estimation, with a mean absolute error of only 11.342° and a coefficient of determination of 0.948. In comparison, the mean absolute error of the BAZ network is 44.489°, and that of SeisT-L is 23.890°, indicating that the pre-trained large language model has a stronger ability to extract the directional information implicit in the waveform. Looking at the mean and standard deviation, this method has the lowest mean error of -0.076° and a standard deviation of 23.390°, far lower than the 63.491° of the BAZ network and the 42.888° of SeisT-L, indicating that the model not only has high accuracy but also good stability, without significant deviations.
[0093] Regarding epicentral distance estimation, the mean absolute error in this embodiment is 2.313 km, with a coefficient of determination of 0.990, slightly better than SeisT-L's 2.609 km and 0.988; the mean error is 0.169 km, and the standard deviation is 4.843 km, both improvements over SeisT-L's -0.175 km and 5.413 km. This improved accuracy in epicentral distance estimation is particularly important for mine-scale monitoring, as the epicentral distance of microseismic events in mines typically ranges from several hundred meters to several kilometers. Even after correction using a network density and velocity model, the accuracy remains valuable.
[0094] Magnitude estimation is valuable for determining whether an event is likely to trigger a mine disaster. In this embodiment, the mean absolute error for magnitude estimation is 0.158, which is better than MagNet's 0.225 and SeisT-L's 0.172. Simultaneously, the standard deviation is 0.221, the lowest among the compared methods, indicating that the model's prediction results are more stable and do not fluctuate significantly due to noise interference or waveform differences. Improved magnitude estimation accuracy is also significant for mine microseismic monitoring, as micro-rupture events of different energies correspond to different disaster risk levels. Accurate magnitude estimation helps distinguish between natural ruptures and minor events induced by mining activities.
[0095] Figure 6 The image shows a two-dimensional diagram of the location results. The SeisMoLLM model performs better in point location, with a higher degree of overlap and smaller offset between the predicted and actual hypocenters in latitude, longitude, and depth.
[0096] Figure 7 A 3D diagram is used to visually represent the effect of 3D spatial point localization, showing a comparison of the spatial distribution of the actual earthquake source location and the model's predicted location. Different colors or shapes of markers in the diagram represent the actual and predicted locations, respectively. The distribution trend shows that the predicted points largely overlap with the actual points, indicating that the model has good 3D localization capabilities.
[0097] Figure 8 The location prediction error distribution map shows the error probability density distribution of the SeisMoLLM model in latitude, longitude, depth and magnitude. It can be seen from the figure that the errors in each direction are mainly concentrated near zero and are symmetrically distributed. The magnitude error is also concentrated in the small error range, indicating that the model has high positioning accuracy and no systematic bias.
[0098] Based on the experimental results of the five tasks described above, this method achieves leading performance in all tasks related to microseismic monitoring data in coal mining areas. The most significant improvements are observed in the two most challenging tasks: inverse azimuth estimation and S-wave picking. Furthermore, the method maintains near-optimal performance in epicentral distance estimation and magnitude estimation. These performance characteristics demonstrate the application potential of this method in mine microseismic monitoring and provide experimental evidence for subsequent practical deployment.
[0099] Please refer to Figure 9 , Figure 9 A structural block diagram of a mine microseismic monitoring device based on a large language model and LoRA is provided for embodiments of the present invention; the specific device may include: The acquisition module 100 is used to acquire the raw three-component microseismic waveform data collected from the mine. The multi-scale convolutional embedding module 200 is used to input the original three-component microseismic waveform data into the multi-scale convolutional embedder, extract multi-scale local features through convolutional kernels of different sizes, and compress the sequence length to obtain a compressed feature sequence. The latent patch module 300 is used to input the compressed feature sequence into the latent patch module and generate a short token sequence by aggregating the features of multiple consecutive time steps into patches. The large language model processing module 400 is used to input the short token sequence into the pre-trained large language model block. The pre-trained large language model block adopts a low-rank adaptive fine-tuning strategy. It freezes the original parameters of the pre-trained large language model block except for the position embedding and layer normalization layer, and inserts a trainable low-rank decomposition matrix into the self-attention layer and feedforward network for processing to extract deep temporal features. The output module 500 is used to input the deep temporal features to a specific task output head to generate microseismic monitoring results. The monitoring results include at least one of the following: P-wave first arrival probability sequence, S-wave first arrival probability sequence, epicentral distance estimate, and reverse azimuth estimate.
[0100] This embodiment provides a mine microseismic monitoring device based on a large language model and LoRA to implement the aforementioned mine microseismic monitoring method based on a large language model and LoRA. Therefore, the specific implementation of the mine microseismic monitoring device based on a large language model and LoRA can be found in the embodiment section of the mine microseismic monitoring method based on a large language model and LoRA described above. For example, the acquisition module 100, the multi-scale convolutional embedding module 200, the latent patch module 300, the large language model processing module 400, and the output module 500 are respectively used to implement steps S101, S102, S103, S104, and S105 in the aforementioned mine microseismic monitoring method based on a large language model and LoRA. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.
[0101] To implement the above embodiments, this application also proposes an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.
[0102] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.
[0103] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.
[0104] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.
[0105] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.
[0106] This application is intended to provide an implementation scheme for users to selectively prevent the use or access to their personal information data. Specifically, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information is de-identified to protect user privacy.
[0107] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0108] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0109] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.
[0110] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.
[0111] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0112] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
[0113] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0114] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.
Claims
1. A mine microseismic monitoring method based on a large language model and LoRA, characterized in that, include: Acquire raw three-component microseismic waveform data collected from the mine; The original three-component microseismic waveform data is input into a multi-scale convolutional embedder. Multi-scale local features are extracted using convolutional kernels of different sizes, and the sequence length is compressed to obtain a compressed feature sequence. The compressed feature sequence is input into the potential patch module, and a short token sequence is generated by aggregating features from multiple consecutive time steps into a patch. The short token sequence is input into a pre-trained large language model block. The pre-trained large language model block adopts a low-rank adaptive fine-tuning strategy. The original parameters of the pre-trained large language model block, except for the position embedding and layer normalization layer, are frozen, and a trainable low-rank decomposition matrix is inserted into the self-attention layer and the feedforward network for processing to extract deep temporal features. The deep temporal features are input into a specific task output head to generate microseismic monitoring results, which include at least one of the following: P-wave first arrival probability sequence, S-wave first arrival probability sequence, epicentral distance estimate, and reverse azimuth estimate.
2. The method according to claim 1, wherein, The process of extracting multi-scale local features using convolutional kernels of different sizes and compressing the sequence length to obtain a compressed feature sequence includes: Perform a linear projection transformation on the input raw three-component microseismic waveform data; The projected data is input in parallel to multiple convolution branches with different kernel sizes. Each branch performs convolution operations, batch normalization, and activation function processing in sequence. The outputs of each branch are concatenated dimensionally, and the concatenated features are integrated by linear projection to obtain a compressed feature sequence.
3. The method according to claim 1, wherein, The process of aggregating features from multiple consecutive time steps into patches to generate short token sequences includes: The compressed feature sequence is reshaped into multiple patches, each patch containing features from multiple consecutive time steps; The features within each patch are aggregated to generate short token sequences.
4. The mine microseismic monitoring method based on large language model and LoRA according to claim 1, characterized in that, The pre-trained large language model block is a backbone network based on generative pre-trained Transformer, and the low-rank adaptive fine-tuning strategy includes: Freeze the original parameters of the pre-trained large language model block, except for the position embedding and layer normalization layers; Make the position embedding and layer normalization layer trainable; A trainable low-rank decomposition matrix is inserted into the self-attention layer and the feedforward network, and the original weights are updated using the low-rank decomposition matrix.
5. The mine microseismic monitoring method based on large language model and LoRA according to claim 1, characterized in that, The specific task output head includes a phase pickup output head and a regression task output head; The phase pickup output head restores the deep temporal features to the original waveform length through upsampling and convolution blocks, and uses the Sigmoid function to generate the first arrival probability sequences of P-waves and S-waves; The regression task output head aggregates information through stacked convolutional layers, and outputs the epicentral distance estimate, inverse azimuth estimate, or magnitude estimate after global average pooling and linear projection.
6. The mine microseismic monitoring method based on large language model and LoRA according to claim 1, characterized in that, It also includes model pre-training, which is performed before acquiring the raw three-component microseismic waveform data collected from the mine, including: Obtain tagged mine microseismic waveform samples, the tags including P-wave arrival time, S-wave arrival time, epicentral distance, and reverse azimuth. Data augmentation processing is performed on the microseismic waveform samples of the mine, including at least one of random Gaussian noise addition, time drift, and channel dropping. The data-augmented samples are input into the multi-scale convolutional embedder, the latent patch module, the pre-trained large language model block, and the task-specific output head to obtain the prediction results. Based on the prediction results and the labels, the multi-task joint loss is calculated, and the model parameters are optimized according to the multi-task joint loss.
7. The mine microseismic monitoring method based on large language model and LoRA according to claim 7, characterized in that, The multi-task joint loss includes phase picking loss and regression task loss; The phase pickup loss is the sum of the binary cross-entropy losses of the P-wave and S-wave. The regression task loss is the Huber loss.
8. A mine microseismic monitoring device based on large language models and LoRA, characterized in that, include: The acquisition module is used to acquire the raw three-component microseismic waveform data collected from the mine. The multi-scale convolutional embedding module is used to input the original three-component microseismic waveform data into the multi-scale convolutional embedder, extract multi-scale local features through convolutional kernels of different sizes, and compress the sequence length to obtain a compressed feature sequence. A latent patch module is used to input the compressed feature sequence into the latent patch module, and generate a short token sequence by aggregating the features of multiple consecutive time steps into patches. The large language model processing module is used to input the short token sequence into the pre-trained large language model block. The pre-trained large language model block adopts a low-rank adaptive fine-tuning strategy. It freezes the original parameters of the pre-trained large language model block except for the position embedding and layer normalization layer, and inserts a trainable low-rank decomposition matrix into the self-attention layer and feedforward network for processing to extract deep temporal features. The output module is used to input the deep temporal features into a specific task output head to generate microseismic monitoring results. The monitoring results include at least one of the following: P-wave first arrival probability sequence, S-wave first arrival probability sequence, epicentral distance estimate, and reverse azimuth estimate.
9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.