A microseismic monitoring method for ground motion targets in sensitive areas

Abstract

This invention belongs to the field of intelligent sensing technology and relates to a method for microseismic monitoring of ground moving targets in sensitive areas. The method includes collecting seismic signals from five types of moving targets and constructing a dataset; preprocessing the data; constructing an interpretable convolution kernel based on the scalar wavefield equation, embedding partial differential operators to achieve adaptive extraction of key physical parameters and nonlinear dimensionality reduction in high-dimensional feature spaces, forming a compact feature representation with physical traceability; using forward diffusion to perform progressive Gaussian noise enhancement on noisy features; designing a Transformer architecture with a multi-head self-attention mechanism and position encoding, capturing the spatiotemporal correlation characteristics of seismic signals through context modeling, and combining a multi-layer perceptron to achieve end-to-end mapping from physical features to classification decisions; training the model, making event decisions, and outputting decision classification results. This method can achieve an accuracy rate of 95.2% in identifying seismic signals from moving targets, significantly improving robustness.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent perception technology, and specifically relates to a seismic anomaly recognition method dominated by a physically driven diffusion Transformer model network, and especially to a microseismic monitoring method for ground motion targets in sensitive areas. Background Art

[0002] Currently, the use of unattended ground sensor systems to detect moving targets in sensitive areas is an important research area in the security field. Detected targets typically include humans, vehicles, and low-altitude drones. A vibration sensing system consists of a large number of fixed and autonomous ground sensors randomly distributed throughout an area. These sensors are self-organizing and communicate via a multi-hop wireless network to detect abnormal events. This network is designed specifically for detecting moving targets within an alert zone. The nodes in the network are homogeneous, low-cost, compact, and capable of long-term operation.

[0003] Although moving target recognition technology based on seismic sensors has advantages such as wide detection range, strong visibility, and strong anti-interference capabilities, two key challenges still exist in the field of ground moving target detection. First, existing methods generally have difficulty effectively capturing the underlying physical information of seismic signal propagation in heavily disturbed environments. Traditional feature extraction techniques and standard deep learning models cannot fully utilize physical principles, resulting in representations that are susceptible to noise. Second, many methods face difficulties in dealing with unbalanced sample distributions, and some target classes are underrepresented, resulting in low classification and detection accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a microseismic monitoring method for ground moving targets in sensitive areas, so as to solve the problem of anomaly detection failure caused by noise interference and data loss in complex terrain.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A microseismic monitoring method for ground motion targets in sensitive areas comprises the following steps:

[0007] A. Collect seismic signals induced by five types of moving targets and construct three datasets. Dataset 1 contains raw seismic signal data with a high signal-to-noise ratio under uniform geological conditions. Dataset 2 contains data from non-stationary geological noise recorded in the field and artificially injected Gaussian white noise to simulate a strong interference environment. Dataset 3 contains unbalanced sample class data reflecting the scarcity of field data.

[0008] B. Data preprocessing:

[0009] The original vibration signal data is normalized and the DC component is removed to eliminate the DC bias generated during data acquisition. The seismic and acoustic signals with the DC component removed are obtained. Data enhancement is performed on the unbalanced category data. Feature extraction, label extraction, normalization and sliding window processing are completed for each data set, and the time series data is converted into a set of input features and corresponding labels.

[0010] C. Physical feature extraction:

[0011] The features are compressed, and physical prior knowledge is embedded in the feature encoding process of the deep convolutional neural network. The Laplace kernel is extracted based on the scalar wave field equation, and an interpretable convolution kernel is constructed to define the initial convolution layer. By embedding partial differential operators, adaptive extraction of key physical parameters and nonlinear dimensionality reduction of high-dimensional feature space are achieved, forming a deep convolutional neural network model with compact feature representation that is physically traceable.

[0012] D. Diffusion-driven data enhancement:

[0013] Deep physical features are extracted from the deep convolutional neural network model. Forward diffusion is used to perform progressive Gaussian noise enhancement on the noisy features in the deep physical features to simulate the situation where the signal is overwhelmed by interference. Progressive Gaussian perturbation injection is used to achieve explicit modeling of environmental interference and improve the model's tolerance to non-stationary noise.

[0014] E. Global modeling decision:

[0015] A Transformer model architecture with a multi-head self-attention mechanism and positional encoding was constructed. After mapping the raw input into a dense vector through an embedding layer, the positional encoding information was then injected into the sequence. Long-range context modeling was used to capture the spatiotemporal correlation characteristics of the vibration signal. This enabled the model to dynamically perceive and integrate information from all positions in the sequence when processing the input sequence. Combined with a multi-layer perceptron, this model achieved end-to-end mapping from physical features to classification decisions.

[0016] F. Event Decision-making:

[0017] The model is trained, using the Transformer model to capture the global environment, applying a forward feedback network to enhance the nonlinear representation capability of the Transformer model, adding residual connection layers and normalization layers around the multi-head attention and feedforward sublayers to ensure stable training and promote gradient flow, and outputting event decision classification results, using different numbers between 0 and 4 to represent five categories of moving targets.

[0018] Furthermore, in step A, the five types of motion seismic signals include seismic signals induced by human activities, wheeled vehicle driving, tracked vehicle driving, aircraft flying, and natural noise.

[0019] Furthermore, in step A, the maximum single sample size ratio of dataset three reaches 1:400.

[0020] Furthermore, in step B, the signal is subjected to a fast Fourier transform to remove the zero-frequency signal, and then subjected to a fast inverse Fourier transform to obtain the seismic and acoustic signals with the DC component removed.

[0021] Furthermore, in step B, the subsequence window size is 100.

[0022] Furthermore, in step C, the specific steps of extracting the Laplace kernel based on the scalar wave field equation and constructing the interpretable convolution kernel to define the initial convolution layer are as follows:

[0023] The propagation of seismic waves in a homogeneous medium is described by the scalar wave field equation:

[0024] ;

[0025] in, is the displacement field, is the wave velocity, is the Laplace operator;

[0026] The corresponding convolution kernel is:

[0027] ;

[0028] in, is a spatially discrete interval, is the convolution kernel;

[0029] This convolution kernel is defined as the initial convolution layer of a convolutional neural network, which can naturally integrate physical laws and data-driven learning and is guided by the propagation characteristics of seismic wavefields.

[0030] Furthermore, in step D, the specific steps of the forward diffusion process are as follows:

[0031] ;

[0032] in, represents the physical characteristics of the previous time step, represents the physical characteristics of the current time step, represents the controlled noise level, is the identity matrix.

[0033] Furthermore, in step E, the multi-head self-attention mechanism is able to calculate the attention score by comparing the query, key, and value derived from the input feature sequence;

[0034] The formula of the multi-head self-attention mechanism is as follows:

[0035] ;

[0036] in, and are respectively the query matrix, key matrix and value matrix obtained by learning the linear projection of the input, is the dimension of the key vector, used as a scaling factor to prevent extremely large dot products;

[0037] The multi-head self-attention mechanism projects the input into multiple subspaces and applies the attention mechanism in parallel:

[0038] ;

[0039] The calculation formula for each attention head is:

[0040] ;

[0041] in, , , are the learned projection matrices of the query matrix Q, key matrix K, and value matrix V, respectively. is the output projection matrix.

[0042] Furthermore, in step E, the multilayer perceptron approximates complex functions through multi-layer nonlinear transformations. It consists of an input layer, a hidden layer, and an output layer. Through multi-layer nonlinear transformations, the original physical features are mapped layer by layer into high-order abstract representations, and classification decisions are output.

[0043] Furthermore, in step F, a forward feedback network is applied to each position of the sequence, specifically expressed by the following formula:

[0044] ;

[0045] in, and are weight matrix and bias respectively; this forward feedback network enhances the nonlinear representation capability of the Transformer model;

[0046] Add residual connection layers and normalization layers around the multi-head attention and feedforward sublayers. Specifically, given an input and a sublayer function, the output is calculated as follows:

[0047] .

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This method proposes a paradigm that integrates physical constraints with deep learning, and pioneers a denoising and modeling mechanism that collaborates with the diffusion model and Transformer. It guides feature space compression and noise robustness enhancement through physical priors. Verified on three datasets in high-interference environments and sample imbalance scenarios, the monitoring method achieved an average accuracy of 95.2% in identifying seismic signals of moving targets, with significantly improved robustness. This method effectively solves the problem of anomaly detection failure caused by noise interference and data missing in complex terrain, providing a reliable technical solution for mobile target monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 A schematic diagram of the unattended ground vibration signal system of the present invention;

[0052] Figure 2 Schematic diagram of the microseismic monitoring method for ground motion targets in sensitive areas according to the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with embodiment:

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0055] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are used only to distinguish the description and should not be understood as indicating or implying relative importance.

[0056] The present invention provides a microseismic monitoring method for ground moving targets in sensitive areas, comprising the following steps:

[0057] A. Collect seismic signals induced by five types of moving targets and construct three datasets. Dataset 1 contains raw seismic signal data with a high signal-to-noise ratio under uniform geological conditions. Dataset 2 contains non-stationary geological noise from field records and data from a strong interference environment simulated by artificially injecting Gaussian white noise. Dataset 3 contains unbalanced sample category data reflecting the scarcity of field data, with a sample ratio spanning 1:400.

[0058] The five types of motion seismic signals include seismic signals induced by human activities, wheeled vehicle driving, tracked vehicle driving, aircraft flying and natural noise.

[0059] B. Data preprocessing:

[0060] In order to ensure that the initial energy of the seismic and acoustic signal segments is at the same level, the present invention preprocesses the original signal. Signal preprocessing includes normalization of signal units and removal of DC components. DC component removal involves eliminating the DC bias generated by the data acquisition equipment. The signal undergoes a fast Fourier transform to remove the zero-frequency signal, and then a fast inverse Fourier transform is performed to obtain the seismic and acoustic signals with the DC component removed. For unbalanced category data, a data enhancement process is added, and each data set completes feature extraction, label extraction, normalization and sliding window operations separately. The time series data is converted into a set of input features and corresponding labels, where the subsequence window size is 100.

[0061] C. Physical feature extraction:

[0062] The features are compressed, and physical prior knowledge is embedded in the feature encoding process of the deep convolutional neural network. The Laplace kernel is extracted based on the scalar wave field equation, and an interpretable convolution kernel is constructed to define the initial convolution layer. By embedding partial differential operators, adaptive extraction of key physical parameters and nonlinear dimensionality reduction of high-dimensional feature space are achieved, forming a deep convolutional neural network model with compact feature representation and physical traceability.

[0063] Specifically, the propagation of seismic waves in homogeneous media is described by the scalar wave field equation:

[0064] ;

[0065] in, is the displacement field, is the wave velocity, is the Laplace operator;

[0066] The corresponding convolution kernel is:

[0067] ;

[0068] in, Represents the spatial discrete interval, which is the distance interval between adjacent discrete points in space when the continuous medium is discretized; Represents the convolution kernel, which is used to perform convolution operations on the values ​​at discrete grid points when numerically solving the wave equation to approximately simulate the physical process described by the wave equation;

[0069] This convolution kernel is defined as the initial convolution layer of a convolutional neural network, which can naturally integrate physical laws and data-driven learning and is guided by the propagation characteristics of seismic wavefields.

[0070] D. Diffusion-driven data enhancement:

[0071] Deep physical features are extracted from a deep convolutional neural network model. Forward diffusion is used to progressively enhance noisy features within these deep physical features with Gaussian noise, simulating the situation where the signal is overwhelmed by interference. Progressive Gaussian perturbation injection is used to explicitly model environmental interference, improving the model's tolerance to non-stationary noise. Unlike generative diffusion models that require a reverse denoising process, this paper focuses solely on forward diffusion as a means of enhancing classification data, applying the model directly to noisy samples during training.

[0072] Specifically, the specific steps of the forward diffusion process are as follows:

[0073] ;

[0074] in, represents the physical characteristics of the previous time step, represents the physical characteristics of the current time step, represents the controlled noise level, is the identity matrix.

[0075] E. Global modeling decision:

[0076] A Transformer model architecture with a multi-head self-attention mechanism and position encoding is constructed. After mapping the original input into a dense vector through the embedding layer, the position encoding information is injected into the sequence. The spatiotemporal correlation characteristics of the vibration signal are captured through long-range context modeling. When processing the input sequence, the model can dynamically perceive and integrate information at all positions in the sequence. Combined with a multi-layer perceptron, the model realizes end-to-end mapping from deep physical features to classification decisions.

[0077] Among them, the multi-layer perceptron approximates complex functions through multi-layer nonlinear transformations. It consists of an input layer, a hidden layer, and an output layer. Through multi-layer nonlinear transformations, the original physical features are mapped layer by layer into high-order abstract representations, and finally, the classification decision is output.

[0078] The multi-head self-attention mechanism is able to compute attention scores by comparing queries, keys, and values ​​derived from the input feature sequence.

[0079] The formula of the self-attention mechanism is as follows:

[0080] ;

[0081] in, and are respectively the query matrix, key matrix and value matrix obtained by learning the linear projection of the input, is the dimension of the key vector, used as a scaling factor to prevent extremely large dot products.

[0082] The multi-head self-attention mechanism projects the input into multiple subspaces and applies the attention mechanism in parallel:

[0083] ;

[0084] The calculation formula for each attention head is:

[0085] ;

[0086] in, , , are the learned projection matrices of the query matrix Q, key matrix K, and value matrix V, respectively. is the output projection matrix.

[0087] F. Event decision-making.

[0088] The model is trained, using the Transformer model to capture the global environment, applying a forward feedback network to enhance the nonlinear representation ability of the Transformer model, adding residual connection layers and normalization layers around the multi-head attention and feedforward sublayers to ensure stable training and promote gradient flow, and outputting event decision classification results, using different numbers between 0 and 4 to represent five categories of moving targets.

[0089] After the multi-head attention mechanism, a forward feedback network is applied to each position in the sequence. The forward feedback network is applied to enhance the nonlinear representation ability of the Transformer model, which is specifically expressed by the following formula:

[0090] ;

[0091] in, and are the weight matrix and bias, respectively. This forward feedback network enhances the nonlinear representation capability of the Transformer model. Residual connection layers and normalization layers are added around the multi-head attention and feedforward sublayers to ensure stable training and promote gradient flow. Specifically, given an input and a sublayer function, the output is calculated as follows:

[0092] .

[0093] In the present invention, the output decision is processed by a decoding layer that aggregates learned representations to produce powerful decisions. The global environment captured by the Transformer model is used to improve detection accuracy and reliability, especially in complex environments that are susceptible to noise. The global environment means that when the model processes the input sequence, it can dynamically perceive and integrate information at all positions in the sequence and establish long-range dependencies between elements. The classification decision result is output, and microseismic target recognition in complex noisy environments is achieved through feature learning guided by physical models, diffusion-driven environmental adaptation enhancement, and end-to-end training of Transformer global reasoning. Different numbers from 0 to 4 represent humans, wheeled vehicles, tracked vehicles, aircraft, and natural noise. The monitoring method of the present invention achieved an average accuracy of 95.2% in the recognition of seismic signals of moving targets, and the robustness was significantly improved.

[0094] The monitoring device of the microseismic monitoring method for ground moving targets in sensitive areas includes a moving target seismic signal acquisition module, a physical feature extraction module, a diffusion-driven data enhancement module, and a global modeling decision module.

[0095] Among them, the moving target seismic signal acquisition module includes a node seismometer and a control center. The node seismometer collects seismic signals generated by different moving targets and transmits them to the control center for data processing. The physical feature extraction module incorporates the wave field propagation formula into the convolution operation to extract and reduce the dimension of key physical quantities, thereby forming a compact and information-rich physical representation. The diffusion-driven data enhancement module embeds these physical characteristics into the diffusion process, making it applicable to different environments and enhancing robustness. The global modeling decision module adopts a transformer architecture to perform global modeling, in which the final decoding layer outputs a robust classification decision. This integrated approach aims to overcome the limitations of current methods and significantly improve detection accuracy and reliability in complex, noise-susceptible environments.

[0096] Example 1:

[0097] A microseismic monitoring method for ground motion targets in sensitive areas comprises the following steps:

[0098] A. Collect seismic signals from five types of moving targets, including human activity, wheeled vehicle movement, tracked vehicle movement, aircraft flight, and natural noise, and transmit the collected signals to the control center for data processing.

[0099] First, operators walked back and forth at a uniform speed along the central axis of the node deployment area (ranging from -12 m to 12 m) to collect artificially induced seismic signals. Subsequently, wheeled vehicles drove in an S-shaped trajectory between the nodes at speeds of 10-40 km / h, while tracked vehicles moved counterclockwise around the nodes at a constant speed of 5 km / h to collect seismic signals generated by the vehicle's motion. Next, a low-altitude drone flew slowly at a constant speed along a straight, back-and-forth path above the node's central axis, maintaining an altitude between 0 m and 20 m, to capture the seismic signals generated by its flight. Finally, a 40-minute ambient noise recording was conducted, during which irregular events such as fallen branches, rolling rocks, and passing animals were observed to obtain seismic signals induced by background noise. Throughout the data acquisition process, only one moving target was active near the sensor at any given time.

[0100] To address the practical monitoring challenges in noisy and unbalanced environments, this example constructs three specialized datasets: Dataset 1: Raw seismic signals with a high signal-to-noise ratio under uniform geological conditions. Dataset 2: A high-noise environment simulated by injecting Gaussian white noise with a σ value of 1.25, non-stationary geological noise from field records, and artificially added noise. Dataset 3: An unbalanced sample class reflecting the scarcity of field data, with a sample ratio of 1:400.

[0101] B. Data preprocessing:

[0102] The original vibration signal data is processed by normalizing the signal units and removing the DC component to eliminate the DC bias generated during data acquisition, obtaining seismic and acoustic signals with the DC component removed. Data enhancement is performed on the unbalanced category data, and feature extraction, label extraction, normalization and sliding window processing are completed for each data set to convert the time series data into a set of input features and corresponding labels.

[0103] C. Physical feature extraction: In the feature compression stage, based on the wavefield propagation physics equation, i.e., the scalar wavefield equation, an interpretable convolution kernel is constructed. By embedding partial differential operators, adaptive extraction of key physical parameters and nonlinear dimensionality reduction of high-dimensional feature space are achieved, forming a deep convolutional neural network model with compact feature representation that is physically traceable. In this embodiment, physical prior knowledge is embedded into the feature encoding process of the deep neural network for the first time, and an interpretable feature space under the constraints of the physical equation is constructed. The wavelength propagation formula is integrated into the convolutional neural network, and the initial convolution layer is defined by the Laplace kernel. The physical feature extraction module achieves dimensionality reduction representation of key physical quantities by embedding the wavefield propagation formula into the convolution operation, thereby increasing the interpretability of the system.

[0104] The scalar wavefield equation describes the propagation of seismic waves in a homogeneous medium:

[0105] ;

[0106] in, is the displacement field, is the wave velocity, is the Laplacian operator, and the corresponding convolution kernel is:

[0107] ;

[0108] in, It represents the spatial discrete interval, which is the distance between adjacent discrete points in space when the continuous medium is discretized. It stands for convolution kernel. When solving the wave equation numerically, it is used to perform convolution operations on the values ​​at discrete grid points to approximately simulate the physical process described by the wave equation.

[0109] This convolution kernel is defined as the initial convolution layer of a convolutional neural network, which can naturally integrate physical laws and data-driven learning and is guided by the propagation characteristics of seismic wavefields.

[0110] D. Diffusion-driven data augmentation: Deep physical features are extracted from the deep convolutional neural network model. Forward diffusion is used to perform progressive Gaussian noise augmentation on the noisy features in the deep physical features. This data augmentation is achieved by directly applying the deep convolutional neural network model to noisy samples during training. Specifically, the deep physical features extracted from the deep convolutional neural network model are fed into the forward diffusion process to simulate the situation where the signal is overwhelmed by interference. Progressive Gaussian perturbation injection is used to explicitly model environmental interference, improving the model's tolerance to non-stationary noise.

[0111] The forward diffusion process is applied to the deep physical features obtained from the convolutional neural network. This process gradually adds Gaussian noise to the features according to a predefined noise schedule. Mathematically, the forward diffusion process can be described as follows:

[0112] ;

[0113] in, represents the physical characteristics of the previous time step, represents the physical characteristics of the current time step, represents the controlled noise level, is the identity matrix.

[0114] Gaussian noise is gradually added to the above features to simulate signal noise. Unlike generative diffusion models, which include a backward denoising process, this example focuses solely on the forward diffusion process as a means of classification data augmentation. This strategy exposes the network to a diverse set of noisy samples during training, improving its ability to generalize and maintain performance in high-noise settings.

[0115] E. Global Modeling Decision: Long-range dependency modeling is achieved using a Transformer model architecture based on a multi-head self-attention mechanism and position encoding. Specifically, in the semantic decision stage, a Transformer model architecture with a multi-head self-attention mechanism and position encoding is designed. Long-range context modeling is used to capture the spatiotemporal correlation characteristics of vibration signals. This allows the model to dynamically perceive and integrate information from all positions in the input sequence when processing it. Combined with a multi-layer perceptron, this achieves an end-to-end mapping from physical features to classification decisions. This embodiment addresses the limitations of current methods by integrating a multi-head attention mechanism, a feedforward network, and normalization techniques to capture both local and long-range dependencies.

[0116] The core of the Transformer is a multi-head attention mechanism that computes attention scores by comparing queries, keys, and values ​​derived from the input feature sequence.

[0117] The model processes input data by first acquiring the feature sequence, mapping the original input into a dense vector through the embedding layer, and then injecting the sequence information through position encoding. Finally, it forms the input required by the multi-head self-attention mechanism.

[0118] Among them, the self-attention mechanism formula is:

[0119] ;

[0120] in, and are respectively the query matrix, key matrix and value matrix obtained by learning the linear projection of the input, is the dimension of the key vector, used as a scaling factor to prevent extremely large dot products. To further enrich the representation, the transformer employs multi-head attention. Instead of computing a single attention function, the input is projected into multiple subspaces and the attention mechanism is applied in parallel:

[0121] The multi-head attention mechanism projects the input into multiple subspaces and applies the attention mechanism in parallel:

[0122] ;

[0123] The calculation formula for each attention head is:

[0124] ;

[0125] in, , , are the learned projection matrices of the query matrix Q, key matrix K, and value matrix V, respectively. is the output projection matrix.

[0126] The multilayer perceptron approximates complex functions through multi-layer nonlinear transformations. It consists of an input layer, a hidden layer, and an output layer. Through multi-layer nonlinear transformations, it maps the original physical features layer by layer into high-order abstract representations and outputs classification decisions.

[0127] F. Event Decision Making. The model is trained using the Transformer model to capture the global environment. A forward feedback network is applied to enhance the nonlinear representation capabilities of the Transformer model. Residual connection layers and normalization layers are added around the multi-head attention and feedforward sublayers to ensure stable training and promote gradient flow. After the multi-head attention, the forward feedback network is applied independently to each position in the sequence. It is usually defined as:

[0128] ;

[0129] in, and are the weight matrix and bias, respectively. This network enhances the nonlinear representation capability of the Transformer model. To ensure stable training and promote gradient flow, residual connection layers and normalization layers are added around the multi-head attention and feedforward sublayers. Specifically, given an input and a sublayer function, the output is calculated as follows:

[0130] ;

[0131] Ultimately, the output classification result is processed by a decoding layer, which aggregates the learned representations to produce a robust decision. This final step effectively leverages the global context captured by the Transformer model to improve detection accuracy and reliability, especially in complex and noise-prone environments. Specifically, when processing the input sequence, the model dynamically perceives and integrates information from all positions in the sequence, establishing long-range dependencies between elements. The output event decision is a number between 0 and 4 representing humans, wheeled vehicles, tracked vehicles, aircraft, and natural noise.

[0132] To evaluate model performance, this example comprehensively evaluates the performance of this method against baseline methods based on three metrics: accuracy, computational efficiency, and robustness. Ten independent replicates were conducted on three datasets, and the averaged results were analyzed. PDTNet, integrating physical information feature extraction, diffusion model-driven optimization, and global context modeling through a Transformer architecture, achieved consistent accuracy exceeding 95.9% across all datasets. PDTNet's ability to outperform baseline models in accuracy and computational efficiency under challenging conditions establishes its superiority for real-world deployment.

[0133] The present invention's microseismic monitoring method for ground moving targets in sensitive areas achieved an average accuracy of 95.2% in identifying moving target seismic signals, significantly improving robustness. The classification accuracy for dataset one was 95.99%, for dataset two was 95.92%, and for dataset three was 96%. In summary, the model proposed in this embodiment has optimal generalization performance, excellent classification accuracy, and computational efficiency, making it the most suitable algorithm for identifying moving targets in fuzzy environments.

[0134] The monitoring device of the microseismic monitoring method for ground moving targets in sensitive areas includes a moving target seismic signal acquisition module, a physical feature extraction module, a diffusion-driven data enhancement module, and a global modeling decision module.

[0135] The moving target seismic signal acquisition module includes node seismometers and a control terminal to transmit data information. The node seismometers collect seismic signals generated by different moving targets and transmit them to the control center for data processing. The physical feature extraction module incorporates wavefield propagation formulas into convolution operations to extract and reduce the dimensionality of key physical quantities, thereby forming a compact and information-rich physical representation. The diffusion-driven data enhancement module embeds these physical characteristics into the diffusion process, making it applicable to different environments and enhancing robustness. The global modeling decision module uses a transformer architecture to perform global modeling, with the final decoding layer outputting a robust classification decision. This integrated approach aims to overcome the limitations of current methods and significantly improve detection accuracy and reliability in complex, noise-prone environments.

[0136] Note that the above are only preferred embodiments of the present invention and the technical principles employed. Those skilled in the art will appreciate that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions are possible for those skilled in the art without departing from the scope of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments and may include many other equivalent embodiments without departing from the scope of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A microseismic monitoring method for ground motion targets in sensitive areas, characterized in that: The following steps are involved: A. Collect seismic signals induced by five types of moving targets and construct three datasets. Dataset 1 contains raw seismic signal data with a high signal-to-noise ratio under uniform geological conditions. Dataset 2 contains data from non-stationary geological noise recorded in the field and artificially injected Gaussian white noise to simulate a strong interference environment. Dataset 3 contains unbalanced sample class data reflecting the scarcity of field data. B. Data preprocessing: The original vibration signal data is normalized and the DC component is removed to eliminate the DC bias generated during data acquisition. The seismic and acoustic signals with the DC component removed are obtained. Data enhancement is performed on the unbalanced category data. Feature extraction, label extraction, normalization and sliding window processing are completed for each data set, and the time series data is converted into a set of input features and corresponding labels. C. Physical feature extraction: The features are compressed, and physical prior knowledge is embedded in the feature encoding process of the deep convolutional neural network. The Laplace kernel is extracted based on the scalar wave field equation, and an interpretable convolution kernel is constructed to define the initial convolution layer. By embedding partial differential operators, adaptive extraction of key physical parameters and nonlinear dimensionality reduction of high-dimensional feature space are achieved, forming a deep convolutional neural network model with compact feature representation that is physically traceable. D. Diffusion-driven data enhancement: Deep physical features are extracted from the deep convolutional neural network model. Forward diffusion is used to perform progressive Gaussian noise enhancement on the noisy features in the deep physical features to simulate the situation where the signal is overwhelmed by interference. Progressive Gaussian perturbation injection is used to achieve explicit modeling of environmental interference and improve the model's tolerance to non-stationary noise. E. Global modeling decision: A Transformer model architecture with a multi-head self-attention mechanism and positional encoding was constructed. After mapping the raw input into a dense vector through an embedding layer, the positional encoding information was then injected into the sequence. Long-range context modeling was used to capture the spatiotemporal correlation characteristics of the vibration signal. This enabled the model to dynamically perceive and integrate information from all positions in the sequence when processing the input sequence. Combined with a multi-layer perceptron, this model achieved end-to-end mapping from physical features to classification decisions. F. Event Decision-making: The model is trained, using the Transformer model to capture the global environment, applying a forward feedback network to enhance the nonlinear representation capability of the Transformer model, adding residual connection layers and normalization layers around the multi-head attention and feedforward sublayers to ensure stable training and promote gradient flow, and outputting event decision classification results, using different numbers between 0 and 4 to represent five categories of moving targets.

2. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: In step A, the five types of seismic signals include seismic signals induced by human activities, wheeled vehicle driving, tracked vehicle driving, aircraft flying, and natural noise.

3. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: In step A, the maximum single sample size ratio of dataset three reaches 1:

400.

4. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: In step B, the signal is subjected to fast Fourier transform to remove the zero-frequency signal, and then subjected to inverse fast Fourier transform to obtain the seismic and acoustic signals with the DC component removed.

5. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: In step B, the subsequence window size is 100.

6. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: Step C: Extract the Laplace kernel based on the scalar wave field equation and construct an interpretable convolution kernel. The specific steps for defining the initial convolution layer are as follows: The propagation of seismic waves in a homogeneous medium is described by the scalar wave field equation: ； in, is the displacement field, is the wave velocity, is the Laplace operator; The corresponding convolution kernel is: ； in, represents a discrete interval in space, represents the convolution kernel; This convolution kernel is defined as the initial convolution layer of a convolutional neural network, which can naturally integrate physical laws and data-driven learning and is guided by the propagation characteristics of seismic wavefields.

7. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: Step D, the specific steps of the forward diffusion process are as follows: ； in, represents the physical characteristics of the previous time step, represents the physical characteristics of the current time step, represents the controlled noise level, is the identity matrix.

8. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: In step E, the multi-head self-attention mechanism can calculate the attention score by comparing the query, key and value derived from the input feature sequence; The formula of the multi-head self-attention mechanism is as follows: ； in, and are respectively the query matrix, key matrix and value matrix obtained by learning the linear projection of the input, is the dimension of the key vector; The multi-head self-attention mechanism projects the input into multiple subspaces and applies the attention mechanism in parallel: ； The calculation formula for each attention head is: ； in, , , are the learned projection matrices of the query matrix Q, key matrix K, and value matrix V, respectively. is the output projection matrix.

9. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: In step E, the multilayer perceptron approximates complex functions through multi-layer nonlinear transformations. It consists of an input layer, a hidden layer, and an output layer. Through multi-layer nonlinear transformations, the original physical features are mapped layer by layer into high-order abstract representations and the classification decisions are output.

10. The microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: In step F, the forward feedback network is applied to each position of the sequence, which is specifically expressed by the following formula: ； in, and are weight matrix and bias respectively; this forward feedback network enhances the nonlinear representation capability of the Transformer model; Add residual connection layers and normalization layers around the multi-head attention and feedforward sublayers. Specifically, given an input and a sublayer function, the output is calculated as follows: 。

Images