Micro-seismic monitoring method for ground moving target in sensitive area

Abstract

The invention belongs to the technical field of intelligent sensing, and relates to a micro-seismic monitoring method for a ground moving target in a sensitive area, which comprises the following steps: collecting five types of moving target seismic signals, and constructing a data set; preprocessing the data; constructing an interpretable convolution kernel based on a scalar wave field equation, and realizing adaptive extraction of key physical parameters and nonlinear dimensionality reduction of a high-dimensional feature space by embedding a partial differential operator to form compact feature representation with physical traceability; forward diffusion is adopted to carry out progressive Gaussian noise enhancement on the noisy features; designing a Transform architecture with a multi-head self-attention mechanism and position coding, capturing time-space correlation characteristics of vibration signals through context modeling, and realizing end-to-end mapping from physical characteristics to classification decision in combination with a multi-layer perceptron; and carrying out training and event decision making on the model, and outputting a decision classification result. According to the method, the identification accuracy of the seismic oscillation signal of the moving target can reach 95.2%, and the robustness is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent perception, and particularly relates to a method for identifying seismic motion anomalies dominated by a diffusion Transformer model network based on physical driving, and more particularly to a microseismic monitoring method for ground motion targets in sensitive areas. Background Art

[0002] Currently, using unattended ground sensor systems to detect moving targets in sensitive areas is an important research direction in the security field. Generally, the detected targets include humans, vehicles, and low-altitude drones. The vibration sensing system consists of a large number of fixed and autonomous ground sensors randomly distributed in the area. The above sensors are self-organized and communicate through multi-hop wireless networks to detect abnormal events. The network is designed specifically for detecting moving targets within the alert area, and the nodes in the network are homogeneous, low-cost, compact, and capable of operating for a long time.

[0003] Although the motion target recognition technology based on seismic sensors has advantages such as a wide detection range, strong visibility, and strong anti-interference ability. However, there are still two key challenges in the field of ground moving target detection. First, existing methods usually have difficulty effectively capturing the potential physical information of seismic signal propagation in a severely disturbed environment. Traditional feature extraction techniques and standard deep learning models cannot fully utilize physical principles, resulting in being restricted by noise-prone representations. Second, many methods are prone to difficulties when dealing with unbalanced sample distributions, and the lack of representation of certain target categories leads to 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 motion targets in sensitive areas to solve the problem of abnormal detection failure caused by noise interference and data loss in complex landforms.

[0005] The purpose of the present invention is achieved by the following technical solutions: A microseismic monitoring method for ground motion targets in sensitive areas includes the following steps: A. Collect seismic motion signals induced by five types of moving targets and construct three groups of data sets. Among them, data set one is the original vibration signal data with high signal-to-noise ratio under uniform geological conditions, data set two is the non-stationary geological noise from on-site records and the data simulating a strong interference environment by artificially injecting Gaussian white noise; data set three is the unbalanced sample category data reflecting the scarcity of on-site data; B. Data preprocessing: Normalize the signal units and remove the DC component from the original vibration signal data to eliminate the DC bias generated during data acquisition, and obtain seismic and acoustic signals with the DC component removed. For unbalanced category data, perform data augmentation. For each dataset, complete feature extraction, label extraction, normalization, and sliding window processing respectively, and convert the time series data into a set of input features and corresponding labels; C. Physical feature extraction: Compress the features, embed the physical prior knowledge into the feature encoding process of the deep convolutional neural network, extract the Laplacian kernel based on the scalar wave field equation, construct an interpretable convolutional kernel to define the initial convolutional layer, and achieve the adaptive extraction of key physical parameters and the non-linear dimensionality reduction of the high-dimensional feature space by embedding partial differential operators, forming a deep convolutional neural network model with physically traceable compact feature representation; D. Diffusion-driven data augmentation: Extract the deep physical features from the deep convolutional neural network model, use forward diffusion to progressively enhance the noisy features in the deep physical features with Gaussian noise, simulate the situation where the signal is submerged by interference, and achieve the explicit modeling of environmental interference through progressive Gaussian perturbation injection, improving the model's tolerance to non-stationary noise; E. Global modeling decision: Construct a Transformer model architecture with a multi-head self-attention mechanism and positional encoding. After mapping the original input to a dense vector through the embedding layer, inject the positional encoding information into the sequence, and capture the spatio-temporal correlation characteristics of the vibration signal through long-range context modeling, enabling the model to dynamically perceive and integrate the information of all positions in the input sequence when processing the input sequence. Combine a multi-layer perceptron to achieve an end-to-end mapping from physical features to classification decisions; F. Event decision: Train the model, use the Transformer model to capture the global environment, apply a forward feedback network to enhance the non-linear representation ability of the Transformer model, add residual connection layers and normalization layers around the multi-head attention and feed-forward sub-layers to ensure stable training and promote gradient flow, and output the event decision classification results, representing five types of moving targets with different numbers between 0 and 4.

[0006] Furthermore, in step A, the five types of moving seismic signals include seismic signals induced by human activities, wheeled vehicle driving, tracked vehicle driving, aircraft flight, and natural noise.

[0007] Furthermore, in step A, the single-sample ratio of dataset three reaches a maximum of 1:400.

[0008] Further, in step B, the signal is subjected to 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.

[0009] Further, in step B, the subsequence window size is 100.

[0010] Further, in step C, the Laplacian kernel is extracted based on the scalar wavefield equation, and the specific steps for constructing the interpretable convolutional kernel to define the initial convolutional layer are as follows: The propagation of seismic waves in a homogeneous medium is described by the scalar wavefield equation: ; where, is the displacement field, is the wave velocity, is the Laplacian operator; The corresponding convolutional kernel is: ; where, is the spatial discretization interval, is the convolutional kernel; This convolutional kernel is defined as the initial convolutional layer of the convolutional neural network, and this convolution can naturally integrate physical laws and data-driven learning and be guided by the propagation characteristics of the seismic wavefield.

[0011] Further, in step D, the specific steps of the forward diffusion process are as follows: ; where, 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.

[0012] Further, in step E, the multi-head self-attention mechanism can calculate the attention scores by comparing the queries, keys, and values derived from the input feature sequence; The formula for the multi-head self-attention mechanism is as follows: ; where, and are the query matrix, key matrix, and value matrix obtained through the linear projection of the input by learning respectively, is the dimension of the key vector, used as a scaling factor to prevent extremely large dot products; 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, the key matrix K, and the value matrix V, respectively. is the output projection matrix.

[0013] Furthermore, in step E, the multilayer perceptron approximates complex functions through multi-layer nonlinear transformations, and is composed 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.

[0014] Furthermore, in step F, a forward feedback network is applied to each position of the sequence, which is specifically expressed by the following formula: ; in, and They 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: .

[0015] Compared with the prior art, the present invention has the following beneficial effects: This method proposes a fusion paradigm of physical constraints and deep learning, and pioneers a denoising-modeling mechanism that collaborates with the diffusion model and Transformer, guiding feature space compression and noise robustness enhancement through physical priors. After verification by three sets of data sets in high-interference environments and sample imbalance scenarios, the monitoring method achieved an average accuracy of 95.2% in the identification of seismic signals of moving targets, and its robustness was significantly improved. This method effectively solves the problem of anomaly detection failure caused by noise interference and data missing in complex terrain, and provides a reliable technical solution for mobile target monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0017] Figure 1Schematic diagram of the unattended ground vibration signal system of the present invention; Figure 2 Schematic diagram of the microseismic monitoring method for ground moving targets in sensitive areas of the present invention. Specific implementation manner

[0018] The present invention will be further described below in conjunction with embodiments: The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. In addition, it should be noted that, for the sake of description, only parts related to the present invention rather than all structures are shown in the drawings.

[0019] It should be noted that: similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used for distinguishing descriptions, and cannot be understood as indicating or implying relative importance.

[0020] The microseismic monitoring method for ground moving targets in sensitive areas of the present invention includes the following steps: A. Collect seismic ground motion signals induced by five types of moving targets, and construct three groups of data sets. Among them, data set one is the original vibration signal data with high signal-to-noise ratio under uniform geological conditions, data set two is the non-stationary geological noise from on-site records and the data simulating a strong interference environment by artificially injecting Gaussian white noise; data set three is the unbalanced sample category data reflecting the scarcity of on-site data, and the sample ratio ranges from 1:400.

[0021] The five types of moving seismic ground motion signals include seismic signals induced by human activities, wheeled vehicle driving, tracked vehicle driving, aircraft flight, and natural noise.

[0022] B. Data preprocessing: In order to ensure that the initial energies of the seismic and acoustic signal segments are at the same level, the present invention preprocesses the original signals. The signal preprocessing includes the normalization of signal units and the removal of DC components. The removal of DC components involves eliminating the DC bias generated by data acquisition devices. The signal is subjected to 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 DC components removed. For unbalanced category data, a data augmentation process is added, and each data set separately completes feature extraction, label extraction, normalization, and sliding window operations. The time series data is converted into a set of input features and corresponding labels, where the subsequence window size is 100.

[0023] C. Physical feature extraction: Compress the features, embed the physical prior knowledge into the feature encoding process of the deep convolutional neural network, extract the Laplacian kernel based on the scalar wave field equation, construct an interpretable convolutional kernel to define the initial convolutional layer, and achieve the adaptive extraction of key physical parameters and the non-linear dimensionality reduction of the high-dimensional feature space by embedding partial differential operators, forming a deep convolutional neural network model with physically traceable compact feature representations.

[0024] Specifically, the propagation of seismic waves in a homogeneous medium is described by the scalar wave field equation: ; where is the displacement field, is the wave velocity, is the Laplacian operator; The corresponding convolutional kernel is: ; where represents the spatial discretization interval, which is the distance interval between adjacent discrete points in space when discretizing the continuous medium; represents the convolutional kernel, which is used to perform convolutional operations on the numerical values at discrete grid points to approximately simulate the physical process described by the wave equation during the numerical calculation and solution of the wave equation; This convolutional kernel is defined as the initial convolutional layer of the convolutional neural network, and this convolution can naturally integrate physical laws and data-driven learning and be guided by the propagation characteristics of the seismic wave field.

[0025] D. Diffusion-driven data augmentation: Extract deep physical features from the deep convolutional neural network model, use forward diffusion to gradually enhance the Gaussian noise in the noisy features of the deep physical features, simulate the situation where the signal is submerged by interference, and explicitly model the environmental interference through progressive Gaussian perturbation injection to improve the model's tolerance to non-stationary noise. Different from the generative diffusion model that requires a reverse denoising process, the present invention only focuses on forward diffusion as a means of classification data augmentation and directly applies the model to noise samples during training.

[0026] Specifically, the specific steps of the forward diffusion process are as follows: ; where represents the physical features of the previous time step, represents the physical features of the current time step, represents the controlled noise level, is the identity matrix.

[0027] E. Global modeling decision: Construct a Transformer model architecture with multi - head self - attention mechanism and positional encoding. After mapping the original input into a dense vector through the embedding layer, inject the positional encoding information into the sequence, and capture the spatio - temporal correlation characteristics of the vibration signal through long - range context modeling, enabling the model to dynamically perceive and integrate the information of all positions in the input sequence when processing the input sequence. Combine a multi - layer perceptron to achieve an end - to - end mapping of the model from deep physical features to classification decisions.

[0028] Among them, the multi - layer perceptron approximates complex functions through multi - layer non - linear transformations, consisting of an input layer, a hidden layer, and an output layer. Through multi - layer non - linear transformations, the original physical features are gradually mapped into high - order abstract representations, and finally, classification decisions are output.

[0029] The multi - head self - attention mechanism can calculate attention scores by comparing queries, keys, and values derived from the input feature sequence.

[0030] The formula for the self - attention mechanism is as follows: ; Among them, and are the query matrix, key matrix, and value matrix obtained through linear projections of the input, respectively. is the dimension of the key vector, used as a scaling factor to prevent extremely large dot products.

[0031] The multi - head self - attention mechanism projects the input into multiple sub - spaces and applies the attention mechanism in parallel: ; The calculation formula for each attention head is: ; Among them, , , are the learned projection matrices of the query matrix Q, key matrix K, and value matrix V, respectively. is the output projection matrix.

[0032] F, event decision.

[0033] Train the model, use the Transformer model to capture the global environment, apply a forward feedback network to enhance the non - linear representation ability of the Transformer model, add residual connection layers and normalization layers around the multi - head attention and feed - forward sub - layers to ensure stable training and promote gradient flow, and output the event decision classification results, which are represented by different numbers between 0 and 4 for five types of motion targets.

[0034] After the multi-head attention mechanism, a forward feedback network is applied to each position of the sequence. The forward feedback network is applied to enhance the non-linear representation ability of the Transformer model, which is specifically represented by the following formula: ; Wherein, and are the weight matrix and bias respectively. This forward feedback network enhances the non-linear representation ability of the Transformer model. Residual connection layers and normalization layers are added around the multi-head attention and feed-forward sub-layers to ensure stable training and promote gradient flow. Specifically, given an input and a sub-layer function, the output is calculated as follows: .

[0035] In the present invention, the output decision is processed by a decoding layer, which aggregates the learned representations to generate powerful decisions. The global environment captured by the Transformer model is utilized to improve the detection accuracy and reliability, especially in complex and noise-prone environments. The global environment refers to the ability of the model to dynamically perceive and integrate the information at all positions in the input sequence when processing it, and establish long-range dependency relationships between elements. The output classification decision result realizes the micro-seismic target recognition in complex noise environments through feature learning guided by a physical model, enhancement of diffusion-driven environmental adaptation, 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 achieves an average accuracy rate of 95.2% in the recognition of seismic signals of moving targets, and the robustness is significantly improved.

[0036] The monitoring device for the micro-seismic monitoring method of the above-mentioned 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.

[0037] Wherein, the moving target seismic signal acquisition module includes a node seismograph and a control center. The node seismograph acquires 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 dimensions 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 to make them applicable to different environments and enhance the robustness. The global modeling decision module adopts a transformer architecture to perform global modeling, where the final decoding layer outputs a robust classification decision. This integrated method aims to overcome the limitations of current methods and significantly improve the detection accuracy and reliability in complex and noise-prone environments.

[0038] Example 1: A microseismic monitoring method for ground moving targets in sensitive areas, comprising the following steps: A. Collect seismic motion signals of five types of moving targets, including human activities, wheeled vehicle driving, tracked vehicle driving, aircraft flight, and natural noise, and transmit the collected signals to a control center for data processing.

[0039] First, an operator walks 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 human-induced seismic signals. Subsequently, a wheeled vehicle drives in an S-shaped trajectory between nodes at a speed of 10 - 40 km / h, while a tracked vehicle moves counterclockwise around the nodes at a constant speed of 5 km / h to collect seismic signals generated by vehicle movement. Next, a low-altitude unmanned aerial vehicle slowly flies back and forth along a straight-line path above the central axis of the nodes at a constant speed, maintaining a height between 0 m and 20 m to capture the seismic signals generated by its flight. Finally, environmental noise was recorded for 40 minutes, during which irregular events such as falling branches, rolling rocks, and passing animals were observed to obtain seismic signals caused by background noise. During the entire data collection process, only one moving target was active near the sensor at any given time.

[0040] To address the practical monitoring challenges in a noisy and imbalanced environment, in this embodiment, three specialized datasets are constructed: Dataset 1: Raw vibration signals with a high signal-to-noise ratio under uniform geological conditions. Dataset 2: Simulate a strong interference environment by injecting Gaussian white noise with σ = 1.25, non-stationary geological noise from on-site records, and artificially added noise. Dataset 3: Imbalanced sample classes reflecting the scarcity of on-site data, with a sample ratio spanning 1:400.

[0041] B. Data preprocessing: Perform normalization of signal units and removal of DC components on the raw vibration signal data, eliminate the DC bias generated during data collection, and obtain seismic and acoustic signals with DC components removed. For imbalanced class data, perform data augmentation, and for each dataset, complete feature extraction, label extraction, normalization, and sliding window processing to convert time series data into a set of input features and corresponding labels.

[0042] C. Physical Feature Extraction: In the feature compression stage, based on the physical equation of wave field propagation, i.e., the scalar wave field equation, an interpretable convolutional kernel is constructed. By embedding partial differential operators, the adaptive extraction of key physical parameters and the non-linear dimensionality reduction of the high-dimensional feature space are realized, forming a deep convolutional neural network model with physically traceable compact feature representation. In this embodiment, physical prior knowledge is first embedded into the feature encoding process of the deep neural network, constructing an interpretable feature space under the constraint of physical equations. The wavelength propagation formula is integrated into the convolutional neural network, and the initial convolutional layer is defined by the Laplace kernel. The physical feature extraction module realizes the dimensionality reduction representation of key physical quantities by embedding the wave field propagation formula into the convolutional operation, increasing the interpretability of the system.

[0043] The scalar wave field equation describes the propagation of seismic waves in a homogeneous medium: ; where is the displacement field, is the wave velocity, is the Laplace operator, and the corresponding convolutional kernel is: ; where represents the spatial discretization interval, which is the distance interval between adjacent discrete points in space when discretizing the continuous medium. represents the convolutional kernel, which is used to perform convolutional 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.

[0044] This convolutional kernel is defined as the initial convolutional layer of the convolutional neural network, and this convolution can naturally integrate physical laws and data-driven learning and be guided by the propagation characteristics of the seismic wave field.

[0045] D. Diffusion-driven Data Augmentation: Deep physical features are extracted from the deep convolutional neural network model, and forward diffusion is used to progressively enhance the noisy features in the deep physical features with Gaussian noise to achieve data augmentation by forward diffusion, and the deep convolutional neural network model is directly applied to the noise samples during training. Specifically, the deep physical features extracted from the deep convolutional neural network model are input into the forward diffusion process to simulate the situation where the signal is overwhelmed by interference, and explicit modeling of environmental interference is realized through progressive Gaussian perturbation injection, improving the model's tolerance to non-stationary noise.

[0046] 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: ; Among them, represents the physical characteristics of the previous time step, represents the physical characteristics of the current time step, represents the noise level of the control, is the identity matrix.

[0047] Gradually add Gaussian noise to the above features to simulate the situation of signal perturbation and submergence. Different from the generative diffusion model that includes a reverse denoising process, this embodiment only focuses on the forward diffusion process as a means of classification data augmentation. This strategy exposes the network to a set of different noise samples during training, thereby improving its ability to generalize and maintain performance in a high-interference setting.

[0048] E. Global modeling decision: The Transformer model architecture based on the multi-head self-attention mechanism and positional encoding is used to achieve long-range dependence modeling. Specifically, in the semantic decision-making stage, a Transformer model architecture with a multi-head self-attention mechanism and positional encoding is designed to capture the spatio-temporal correlation characteristics of the vibration signal through long-range context modeling, enabling the model to dynamically perceive and integrate the information of all positions in the input sequence during the processing of the input sequence, and combining with a multi-layer perceptron to achieve an end-to-end mapping from physical features to classification decisions. In this embodiment, by integrating the multi-head attention mechanism, the forward feedback network, and the normalization technology, the limitations of the current method are solved by capturing local and remote dependencies.

[0049] The core of the Transformer is the multi-head attention mechanism, which calculates the attention scores by comparing the queries, keys, and values derived from the input feature sequence.

[0050] The model processes the input data. First, it obtains the feature sequence, maps the original input to a dense vector through the embedding layer, and then injects sequence information through positional encoding. Finally, it forms the input required for the multi-head self-attention mechanism.

[0051] Among them, the self-attention mechanism formula is: ; Among them, and are the query matrix, key matrix, and value matrix obtained by learning the linear projection of the input respectively. 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 adopts multi-head attention. Instead of calculating a single attention function, the input is projected into multiple subspaces, and the attention mechanism is applied in parallel: The multi-head attention mechanism projects the input into multiple subspaces and applies the attention mechanism in parallel: ; The calculation formula for each attention head is as follows: ; Among them, , , are the learned projection matrices of the query matrix Q, key matrix K, and value matrix V respectively, is the output projection matrix.

[0052] A multi-layer perceptron approximates complex functions through multi-layer non-linear transformations. It consists of an input layer, hidden layers, and an output layer. Through multi-layer non-linear transformations, the original physical features are mapped layer by layer into high-order abstract representations and the classification decisions are output.

[0053] F, event decision. The model is trained to capture the global environment using the Transformer model, and a forward feedback network is applied to enhance the non-linear representation ability of the Transformer model. Residual connection layers and normalization layers are added around the multi-head attention and feed-forward sub-layers to ensure stable training and promote gradient flow. After the multi-head attention, the forward feedback network is independently applied to each position of the sequence. It is usually defined as: ; Among them, and are the weight matrix and bias respectively. This network enhances the non-linear representation ability 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 feed-forward sub-layers. Specifically, given an input and a sub-layer function, the output is calculated as follows: ; Finally, the output classification result is processed by the decoding layer, which aggregates the learned representations to produce powerful decisions. The last step effectively uses the global environment captured by the Transformer model to improve the detection accuracy and reliability, especially in complex and noise-prone environments. That is, when the model processes the input sequence, it can dynamically perceive and integrate the information at all positions in the sequence, establishing long-range dependencies between elements. The output event decision result is represented by different numbers from 0 to 4 for human, wheeled vehicle, tracked vehicle, aircraft, and natural noise.

[0054] To evaluate the model performance, in this embodiment, the performance of this method and the benchmark method are comprehensively evaluated according to three metrics: accuracy, computational efficiency, and robustness. Ten independent repeated experiments were conducted on three groups of datasets, and the average values were analyzed. PDTNet integrates physical information feature extraction, diffusion model-driven optimization, and global context modeling through the Transformer architecture, achieving a consistent accuracy of over 95.9% in all datasets. PDTNet can outperform the benchmark model in terms of accuracy and computational efficiency under challenging conditions, thus establishing its advantage in practical deployment.

[0055] The microseismic monitoring method of the present invention for ground moving targets in sensitive areas achieves an average accuracy of 95.2% in the identification of seismic signals of moving targets, and the robustness is significantly improved. The classification accuracy on Dataset 1 is 95.99%, the classification accuracy on Dataset 2 is 95.92%, and the classification accuracy on Dataset 3 is 96%. In summary, the model proposed in this embodiment has the best generalization performance, excellent classification accuracy, and computational efficiency, and is the most suitable algorithm for mobile target recognition in a fuzzy environment.

[0056] The monitoring device for the above-mentioned microseismic monitoring method of ground moving targets in sensitive areas includes a moving target seismic signal acquisition module, a physical feature extraction module, a diffusion-driven data augmentation module, and a global modeling decision module.

[0057] The moving target seismic signal acquisition module includes a nodal seismograph and a control terminal for transmitting data information. The nodal seismograph 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 dimensions of key physical quantities, thereby forming a compact and information-rich physical representation; the diffusion-driven data augmentation module embeds these physical characteristics into the diffusion process to make it applicable to different environments and enhance robustness; the global modeling decision module uses the transformer architecture to perform global modeling, where the final decoding layer outputs a robust classification decision. This integrated method aims to overcome the limitations of current methods and significantly improve the detection accuracy and reliability in complex and noise-prone environments.

[0058] Note that the above is only a preferred embodiment of the present invention and the technical principles applied. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, re-adjustments and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments. Without departing from the concept of the present invention, more other equivalent embodiments can be included, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A microseismic monitoring method for ground moving targets in sensitive areas, characterized in that, It includes the following steps: A. Collect the ground motion signals induced by five types of moving targets, and construct three groups of datasets. Among them, Dataset 1 is the original vibration signal data with high signal-to-noise ratio under uniform geological conditions, Dataset 2 is the non-stationary geological noise from on-site records and the data simulating a strong interference environment by injecting Gaussian white noise artificially; Dataset 3 is the unbalanced sample category data reflecting the scarcity of on-site data; B. Data preprocessing: Normalize the signal units and remove the DC component of the original vibration signal data to eliminate the DC bias generated during data acquisition, and obtain the seismic and acoustic signals with the DC component removed. For the unbalanced category data, perform data augmentation, and respectively complete feature extraction, label extraction, normalization and sliding window processing for each dataset, and convert the time series data into a set of input features and corresponding labels; C. Physical feature extraction: Compress the features, embed the physical prior knowledge into the feature encoding process of the deep convolutional neural network, extract the Laplacian kernel based on the scalar wave field equation, construct an interpretable convolutional kernel to define the initial convolutional layer, and realize the adaptive extraction of key physical parameters and the non-linear dimensionality reduction of the high-dimensional feature space by embedding partial differential operators, and form a deep convolutional neural network model with physically traceable and compact feature representations; D. Diffusion-driven data augmentation: Extract the deep physical features from the deep convolutional neural network model, and use forward diffusion to gradually enhance the noisy features in the deep physical features with Gaussian noise, simulate the situation where the signal is submerged by interference, and realize the explicit modeling of environmental interference through the injection of progressive Gaussian perturbations, and improve the model's tolerance to non-stationary noise; E. Global modeling decision: Construct a Transformer model architecture with a multi-head self-attention mechanism and positional encoding. After mapping the original input into a dense vector through the embedding layer, inject the positional encoding information into the sequence, and capture the spatio-temporal correlation characteristics of the vibration signal through long-range context modeling, so that the model can dynamically perceive and integrate the information of all positions in the input sequence when processing the input sequence, and combine a multi-layer perceptron to realize the end-to-end mapping from physical features to classification decisions; F. Event decision: Train the model, use the Transformer model to capture the global environment, apply a forward feedback network to enhance the non-linear representation ability of the Transformer model, add residual connection layers and normalization layers around the multi-head attention and feed-forward sub-layers to ensure stable training and promote gradient flow, and output the event decision classification result, which is represented by different numbers between 0 and 4 for the five types 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 moving ground motion signals include seismic signals induced by human activities, wheeled vehicle driving, tracked vehicle driving, aircraft flight, 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 single-sample ratio of Dataset 3 reaches a maximum of 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 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.

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, wherein Step C, based on the scalar wave field equation, extract the Laplacian kernel and construct an interpretable convolutional kernel to define the initial convolutional layer. The specific steps are as follows: Describe the propagation of seismic waves in a homogeneous medium using the scalar wave field equation: ； wherein, is the displacement field, is the wave velocity, is the Laplace operator; The corresponding convolutional kernel is: ； Among them, represents the spatial discretization interval, represents the convolution kernel; This convolutional kernel is defined as the initial convolutional layer of the convolutional neural network. This convolution can naturally integrate physical laws and data-driven learning and is guided by the propagation characteristics of the seismic wave field.

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: ； wherein, represents the physical characteristics of the previous time step, represents the physical characteristics of the current time step, represents the noise level of the control, is the identity matrix.

8. A microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: The multi-head self-attention mechanism can calculate attention scores by comparing queries, keys, and values derived from the input feature sequence; The formula for the multi-head self-attention mechanism is as follows: ； wherein, and are the query matrix, key matrix, and value matrix obtained by learning linear projections 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: ； Among them, , , are the learned projection matrices of the query matrix Q, the key matrix K, and the value matrix V respectively, is the output projection matrix.

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

10. A microseismic monitoring method for ground moving targets in sensitive areas according to claim 1, characterized in that: Step F, the forward feedback network is applied to each position of the sequence, specifically represented by the following formula: ； Among them, and are the weight matrix and bias respectively; the positive feedback network enhances the non-linear representation ability of the Transformer model; Add residual connection layers and normalization layers around the multi-head attention and feed-forward sub-layers. Specifically, given an input and a sub-layer function, the output is calculated as follows: 。

Images