Deep learning anomaly detection method and system for underground coal mine time series data

Abstract

The invention discloses a deep learning anomaly detection method and system for underground coal mine time series data, and the method can effectively fuse the spatial-temporal topological relation and frequency domain characteristics of a multi-source sensor, improves the accuracy and robustness of the anomaly detection of the underground coal mine time series data, remarkably reduces the false alarm rate and the missing report rate, and improves the detection precision. And the method has the capability of dynamically adapting to mine environment change.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of coal mine safety production and artificial intelligence, and in particular to a deep learning anomaly detection method and system for underground coal mine time series data. Background Technology

[0002] With increasing underground mining depth, the mining environment becomes increasingly complex, posing various safety hazards such as excessive gas concentration, abnormal roof pressure, and equipment failure. Underground safety monitoring has become a crucial link in ensuring production safety. Traditional coal mine monitoring systems primarily use threshold alarms or simple statistical methods to detect anomalies in individual time-series data collected by sensor networks (such as gas concentration, temperature, vibration, and wind speed sensors). However, these methods have the following limitations: First, threshold alarms are easily affected by environmental noise and data gaps, leading to a high false alarm rate; second, simple statistical methods cannot capture the nonlinear spatiotemporal dependencies in the data, resulting in a high risk of missed alarms; finally, existing methods lack in-depth analysis of the frequency domain characteristics of the data, making it difficult to identify periodic or sudden anomalies.

[0003] In recent years, artificial intelligence (AI) technology has made tremendous progress, achieving significant advancements in areas such as image detection, autonomous driving, speech recognition, and text-based dialogue. Related AI technologies have also been applied to the coal mining industry, with research and work already underway using machine learning algorithms and deep learning models for gas concentration prediction, gas concentration anomaly detection, gas outburst early warning, and working face anomaly pressure detection. For time-series anomaly detection, intelligent methods can detect anomalies faster and more accurately, reducing manual maintenance costs.

[0004] However, current anomaly detection schemes based on machine learning or deep learning for underground coal mine time-series data have the following problems: First, they are basically all based on detection models built from single time-series data. These models cannot overcome the noise and missing data in single time-series data, nor can they utilize information from other related sequences. Furthermore, current deep learning methods that fuse multiple time-series sequences primarily only integrate information from different sensors at the same location, such as gas, temperature, and wind speed, with relatively little utilization of spatial geographic information and temporal information. These simple processing methods can only yield insufficient anomaly detection results, leading to significant missed or false alarms, affecting detection outcomes and potentially causing coal mine operation and maintenance accidents. Summary of the Invention

[0005] The main objective of this invention is to provide a deep learning-based anomaly detection method for underground coal mine time series data, enabling rapid and accurate detection of abnormal data points and patterns in coal mine time series data, overcoming the shortcomings of existing technologies, and solving the problems of high false negatives and high false negatives.

[0006] Another objective of this invention is to propose a deep learning anomaly detection system for underground coal mine time-series data.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a deep learning-based anomaly detection method for underground coal mine time-series data, comprising: S1, collect time-series data from multiple sources in the coal mine and parse it into a unified time format. The time-series data includes sensor type, spatial location information and dynamic characteristics. S2, construct spatiotemporal topological data including relational graph structure, dynamic feature tensor and time information encoding, and mine the frequency domain features of the spatiotemporal topological data through graph Fourier transform; S3 inputs spatiotemporal topological data and frequency domain feature data into a deep learning model, and performs feature extraction and aggregation through a multi-layer spatiotemporal aggregation module to generate the reconstruction result of time series data; S4 performs anomaly detection based on the difference between the reconstruction results and the original time series data, and periodically detects changes in the topology of the mine sensor network. When the detected topology difference exceeds a preset threshold, it triggers model parameter updates.

[0008] In one embodiment of the present invention, the steps of S2 in constructing spatiotemporal topology data and mining frequency domain feature data further include: S21, The method for constructing the relational graph structure is as follows: construct graph nodes based on the geographic coordinates of the sensors. If there is an underground road connection between two nodes, establish an edge. The weight of the edge is the actual distance between the nodes, forming a matrix with a size of the number of nodes × the number of nodes. S22, The implementation of the graph Fourier transform is as follows: the graph Laplacian matrix L is calculated based on the relation graph structure, the eigenvector basis is obtained through eigenvalue decomposition, and the dynamic feature tensor is projected onto the eigenvector basis to extract frequency domain features.

[0009] In one embodiment of the present invention, the feature extraction and aggregation steps of the multi-layer spatiotemporal aggregation module S3 further include: S31, the multi-layer spatiotemporal aggregation module is composed of a gated temporal convolutional neural network, a Chebyshev graph convolutional neural network, and a gated temporal convolutional neural network stacked sequentially, wherein the Chebyshev graph convolutional neural network is constructed using a recursive formula. Calculate Chebyshev polynomials of all orders; S32, the attention mechanism spatiotemporal encoder enhances the representation ability of key features by weighting the input spatiotemporal topological data and frequency domain feature data through a self-attention network.

[0010] In one embodiment of the present invention, the step of anomaly detection based on the difference between the reconstruction result and the original time-series data further includes: S41, calculate the point-by-point difference between the original time series data and the reconstruction result. If the difference exceeds a preset threshold, it is determined to be abnormal. The threshold is dynamically adjusted according to the statistical distribution of historical data. S42, The method for calculating the graph structure similarity index is as follows: For nodes u and v on two graphs, their structure similarity index is... ,in It is the domain of node u. It is the degree of node u. When the number of dissimilar nodes exceeds the threshold, the model is retrained.

[0011] In one embodiment of the present invention, it further includes: S5 performs sliding window framing on the preprocessed time series data, setting the sliding window size to k and the sliding step size to s, converting the variable-length time series data into fixed-length frame data to adapt to the neural network input requirements.

[0012] To achieve the above objectives, a second aspect of the present invention proposes a deep learning anomaly detection system for underground coal mine time-series data, comprising: a multi-source sensor data acquisition and parsing module, used to acquire time-series data from multiple underground coal mine sensors and parse it into a unified time format, wherein the time-series data includes sensor type, spatial location information and dynamic characteristics; The spatiotemporal topology construction and frequency domain feature mining module is used to construct spatiotemporal topology data containing relational graph structure, dynamic feature tensor and time information encoding, and to mine the frequency domain features of the spatiotemporal topology data through graph Fourier transform; The deep learning model feature extraction and aggregation module is used to input spatiotemporal topological data and frequency domain feature data into the deep learning model, and perform feature extraction and aggregation through a multi-layer spatiotemporal aggregation module to generate the reconstruction results of time series data. The anomaly detection and topology update triggering module is used to detect anomalies based on the differences between the reconstruction results and the original time-series data, and periodically detect changes in the topology of the mine sensor network. When the detected topology difference exceeds a preset threshold, the model parameters are updated.

[0013] The deep learning anomaly detection method and system for underground coal mine time-series data in this invention can effectively integrate the spatiotemporal topological relationships and frequency domain features of multi-source sensors, improve the accuracy and robustness of anomaly detection in underground coal mine time-series data, and significantly reduce the false alarm rate and false negative rate.

[0014] The beneficial effects of this invention are as follows: it mines the direct correlation between sensors based on spatiotemporal topology and spatiotemporal frequency relationships; it can accurately identify hidden anomalies with multiple causes in time-series data based on graph neural networks; it trains a model for the entire mine data, rather than training a model for each sensor, reducing the amount of computation and system load; it can continuously update the model for new data; the method has a wide range of applications, strong generalization ability, and higher recognition accuracy. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention 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 of a deep learning anomaly detection method for underground coal mine time-series data provided in this embodiment of the invention; Figure 2 This is an architecture diagram of a deep learning anomaly detection method for underground coal mine time-series data provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a deep learning anomaly detection system for underground coal mine time-series data provided in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only 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 should fall within the scope of protection of the present invention.

[0018] The following describes a deep learning anomaly detection method and system for underground coal mine time-series data according to an embodiment of the present invention, with reference to the accompanying drawings.

[0019] Example 1 This embodiment provides a deep learning-based anomaly detection method for underground coal mine time-series data. For example... Figure 1 As shown, the method includes the following steps: S1. Collect time-series data from multiple sources in the coal mine and parse it into a unified time format. The time-series data includes sensor type, spatial location information, and dynamic characteristics.

[0020] Specifically, in the time-series data acquisition and unified time format parsing steps of multi-source sensors in underground coal mines, this invention constructs a standardized data acquisition and parsing mechanism to achieve unified processing of heterogeneous time-series data, providing a structured and time-aligned input foundation for subsequent deep learning models. This step is the starting point of the entire anomaly detection system's data flow, and its technical implementation quality directly affects the model's training effect and detection accuracy.

[0021] Sensor networks deployed underground in coal mines include, but are not limited to, gas concentration sensors, temperature sensors, wind speed sensors, vibration sensors, and coal mining equipment status monitoring devices. Their data acquisition frequency is typically 10 or 60 seconds, configurable according to monitoring requirements. The acquired data formats are divided into two categories: fixed-interval data and event-based data. Fixed-interval data is recorded at fixed time intervals, exhibiting good temporal consistency; while event-based data is recorded only when the sensor status changes, including the change amount and corresponding timestamp. To achieve unified processing, this invention employs a timestamp alignment strategy, converting event-based data into a fixed-interval format through interpolation or zero-padding, ensuring the comparability and continuity of all sensor data in the time dimension.

[0022] During data parsing, a timestamp reference (such as UTC time or local time zone) needs to be set, and the time format should be standardized to ISO 8601 (YYYY-MM-DD HH:MM:SS). Simultaneously, sensor type, spatial location information (such as latitude and longitude, working face number, tunnel face number, return airway identifier, etc.), and dynamic features (such as sensor values, equipment status, and operating mode) all need to be structured and stored to form a multi-dimensional time series tensor with dimensions of (time steps, number of nodes, and feature dimension), facilitating subsequent graph neural network processing.

[0023] This step is widely used in underground coal mine safety monitoring systems, especially in scenarios such as multi-source heterogeneous data fusion, distributed sensor network management, and intelligent early warning platform construction. By using a unified time format, the system can achieve synchronous analysis of time-series data across regions and devices, providing high-quality input for subsequent spatiotemporal topology modeling and frequency domain feature extraction.

[0024] This step effectively solves the data alignment difficulties caused by inconsistent data formats and timestamps in traditional monitoring systems, improving the efficiency and accuracy of multi-source data fusion. Through standardized processing, the system has stronger compatibility and scalability, adapting to different mine structures and sensor deployment methods, providing stable and reliable input data for subsequent deep learning models, thereby significantly improving the robustness and generalization ability of anomaly detection.

[0025] S2, construct spatiotemporal topological data including relational graph structure, dynamic feature tensor and time information encoding, and mine the frequency domain features of the spatiotemporal topological data through graph Fourier transform.

[0026] Specifically, this step aims to fuse and model the multi-source heterogeneous time-series data and spatial topology of underground coal mines, thereby extracting feature representations with spatiotemporal semantics and further exploring their potential patterns in the frequency domain to enhance the model's ability to perceive abnormal behavior.

[0027] The relational graph structure is constructed by building a graph of size 1000. adjacency matrix To describe the spatial connection relationship between sensor nodes, where This represents the total number of nodes. Edge weights between nodes are typically set based on geographical distance or roadway connections. For example, if two sensor nodes are physically connected by a roadway, the edge weight is the straight-line distance or path length between them. Dynamic Feature Tensor This is used to represent the multidimensional features of each node in the time series, where For time steps, Features include gas concentration, temperature, wind speed, and equipment status. Time information is encoded using One-Hot or Sin-Cos encoding methods to capture periodic characteristics of time, such as diurnal variations or work shift patterns.

[0028] Furthermore, the graphical Fourier transform is achieved through the graphical Laplace matrix. The eigenvalue decomposition is used to construct the frequency domain representation of the graph signal. Specifically, the normalized Laplacian matrix is ​​first calculated. ,in It is the identity matrix. This is the degree matrix. Subsequently, for... Perform eigenvalue decomposition to obtain the eigenvector matrix. and eigenvalue diagonal matrix Dynamic feature tensor The features are projected onto the graph Fourier basis to obtain the frequency domain feature representation. This reveals the global frequency distribution characteristics of sensor data on a graph structure.

[0029] At the parameter level, the construction of graph structures needs to consider the degree of nodes. Neighborhood Isograph theory metrics are used to ensure the accuracy of topological relationships. The order of the graph Fourier transform. The setting is usually based on the size and feature dimensions of the sensor network, and is generally taken as... This is to balance computational complexity with feature representation capability. Furthermore, the period length of the temporal information encoding... It is typically set to a 24-hour or 8-hour work cycle to match the typical time pattern of coal mine operations.

[0030] This step is applicable to scenarios involving multi-sensor collaborative monitoring in underground coal mines, such as joint modeling of multi-dimensional time-series data on gas concentration, temperature, and wind speed in areas like the working face, tunneling face, and return airway. By constructing spatiotemporal topology data, the system can identify spatial dependencies between sensors and extract physically meaningful frequency domain features using graphical Fourier transform, thereby enhancing the model's ability to detect periodic anomalies (such as periodic failures of the ventilation system) and sudden anomalies (such as sudden increases in gas).

[0031] By unifying spatial topology and time series modeling and introducing graph Fourier transform, this invention effectively captures the nonlinear spatiotemporal dependencies and frequency domain features in underground coal mine data, providing richer input features for subsequent deep learning models. Compared to traditional methods based on single time series or simple spatial aggregation, this invention significantly improves the accuracy and robustness of anomaly detection through frequency domain feature enhancement, maintaining high detection performance even in the presence of noise and missing data.

[0032] Furthermore, S2 includes: S21, The method for constructing the relational graph structure is as follows: construct graph nodes based on the geographic coordinates of the sensors. If there is an underground road connection between two nodes, establish an edge. The weight of the edge is the actual distance between the nodes, forming a matrix with a size of number of nodes × number of nodes.

[0033] Specifically, in some implementations, the method for constructing the relational graph structure is based on the geographic coordinate information of underground sensors in coal mines. A connection structure between graph nodes and edges is established through spatial topological relationships, thereby forming an adjacency matrix of size (number of nodes × number of nodes). The core of this method lies in abstracting the sensor network in physical space into a graph structure to support subsequent modeling and analysis of spatiotemporal data by graph neural networks.

[0034] Specifically, each graph node corresponds to a sensor unit, and its spatial location is represented by three-dimensional geographic coordinates (x, y, z) or two-dimensional planar coordinates (x, y), depending on the actual deployment environment. If there is a physical connection path between two sensor nodes (such as underground tunnels, transportation channels, etc.), an edge is established in the graph, and the weight of the edge is defined as the actual Euclidean distance between the two nodes. ,in and This represents two nodes in the diagram. This distance can be obtained through a GIS system or an underground positioning system, with an accuracy typically between 0.1 meters and 1 meter, meeting the modeling requirements of a coal mine environment.

[0035] Furthermore, this graph structure uses an adjacency matrix. Stored in the form of , where The total number of sensor nodes, and the elements in the matrix Represents a node With nodes The connection weights between them. If there are no connections, then... This matrix serves as input to a graph neural network (GNN) to model spatial dependencies between sensors, thereby enhancing the model's ability to detect anomalous events.

[0036] In practical applications, this step is typically completed during the data preprocessing stage of a coal mine safety monitoring system, and is suitable for underground environments with dense sensor deployments and complex roadway structures. By constructing an accurate graph structure, the spatial proximity and topological relationships between sensors can be effectively captured, providing a foundation for subsequent frequency domain feature extraction operations such as graph Fourier transform and graph convolution.

[0037] The technical value of this step lies in introducing explicit spatial constraints to the deep learning model by transforming physical spatial information into a graph structure, thereby improving the model's ability to identify multi-source anomalies. Simultaneously, the method for constructing this graph structure has good scalability, adapting to dynamic changes in the mine structure and providing a structural basis for continuous updates to the model.

[0038] S22, The implementation of the graph Fourier transform is as follows: the graph Laplacian matrix L is calculated based on the relation graph structure, the eigenvector basis is obtained through eigenvalue decomposition, and the dynamic feature tensor is projected onto the eigenvector basis to extract frequency domain features.

[0039] Specifically, in some implementations, the graph Fourier transform is implemented by calculating the graph Laplace matrix based on the graph structure. The process involves obtaining an eigenvector basis through eigenvalue decomposition, and then projecting the dynamic feature tensor onto this eigenvector basis to extract frequency domain features. This step is the core of spatiotemporal frequency domain feature mining in this invention. Its technical principle is based on graph signal processing theory, aiming to map multi-node time-series data from the time domain to the graph frequency domain, thereby revealing the global frequency characteristics of the data in the spatial topology and enhancing the model's ability to perceive abnormal patterns.

[0040] At the technical implementation level, a relational graph structure is first constructed based on the geographical location and connection relationships of sensors underground in the coal mine. This graph structure uses an adjacency matrix. Sum-degree matrix Based on this, calculate the normalized graph Laplacian matrix of an undirected graph. ,in This is the identity matrix. This matrix reflects the topological relationships and spatial distances between nodes in the graph and is a fundamental operator of the graph Fourier transform. Furthermore, for... Eigenvalue decomposition yields a set of orthogonal eigenvectors. and corresponding eigenvalues ,in , representing the frequency components of the signal in the graph.

[0041] In practical operation, dynamic feature tensor (in For time steps, For the number of nodes, (The feature dimension) is projected onto the graph Fourier basis. The frequency domain representation is obtained from the above. This frequency domain feature can effectively capture the global trend of sensor data changes in the graph structure, and is especially suitable for identifying abnormal events with spatial propagation characteristics, such as abnormal gas concentration diffusion and equipment cascading failures.

[0042] The calculation of the graph Laplacian matrix must ensure the symmetry and connectivity of the graph structure. The accuracy of eigenvalue decomposition is typically achieved using high-precision algorithms from numerical libraries (such as NumPy or PyTorch Geometric). The eigenvector basis can be selected based on the distribution of eigenvalues. Principal components are used to reduce computational complexity while preserving key frequency information.

[0043] This step is widely applicable in practical applications for time-series data processing in multi-node sensor networks in underground coal mines, especially in scenarios with spatial dependence and nonlinear variations, where it can significantly improve the robustness and accuracy of anomaly detection. By mapping time-series data to the graph frequency domain, the model can more effectively identify periodic anomalies, sudden anomalies, and anomaly patterns with spatial propagation characteristics, thereby providing a more reliable intelligent monitoring method for coal mine safety production.

[0044] S3 inputs spatiotemporal topological data and frequency domain feature data into a deep learning model, and performs feature extraction and aggregation through a multi-layer spatiotemporal aggregation module to generate the reconstruction result of time series data.

[0045] Specifically, this deep learning model consists of several key modules, including a spatiotemporal encoder, a multi-layer spatiotemporal aggregation module, an aggregation layer, a spatiotemporal decoder, and a linear output layer. The multi-layer spatiotemporal aggregation module employs a combined structure of a gated temporal convolutional network (Gated Temporal Convolutional Network, GateTCN) and a Chebyshev graph convolutional network (ChebGCN) to jointly model temporal features and graph structure features. Specifically, GateTCN extracts local temporal dependencies in time series through a stacked structure of causal convolutions and dilated convolutions, and enhances the model's expressive power through residual connections; ChebGCN is based on the graph Laplacian matrix. Calculate its Chebyshev polynomial Through recursive formula This enables the extraction of frequency domain features from graph structures, thereby enhancing the model's ability to perceive spatial topological relationships.

[0046] The model's input is a three-dimensional dynamic feature tensor with dimensions (sequence length, number of nodes, feature dimension), where the feature dimension includes the raw values ​​acquired by the sensor, spatial coordinates, device status, etc. The order of the Chebyshev graph convolution is also considered. The value is typically set to 3-5 to balance computational complexity and frequency domain modeling accuracy. The expansion factor for GateTCN can be set to 2 or 4 to expand the model's temporal receptive field and enhance its ability to model long-term dependencies. The model's output is reconstructed time-series data with the same dimensions as the original input, used for subsequent anomaly detection.

[0047] This step is applicable to multi-sensor collaborative monitoring systems in underground coal mines, such as joint modeling of multi-dimensional time-series data including working face gas concentration, temperature, and wind speed. By simultaneously inputting spatial topology and frequency domain features into the model, it is possible to effectively identify multi-source anomalies caused by equipment failure, ventilation abnormalities, or geological changes, thereby improving the robustness and accuracy of detection.

[0048] Furthermore, S3 includes: S31, the multi-layer spatiotemporal aggregation module is composed of a gated temporal convolutional neural network, a Chebyshev graph convolutional neural network, and a gated temporal convolutional neural network stacked sequentially, wherein the Chebyshev graph convolutional neural network is constructed using a recursive formula. Calculate Chebyshev polynomials of all orders.

[0049] Specifically, in some implementations, the multi-layer spatiotemporal aggregation module is composed of a gated temporal convolutional network (GatedTemporal Convolutional Network, GateTCN), a Chebyshev graph convolutional network (ChebGCN), and another gated temporal convolutional network stacked sequentially. Its core function is to fuse the local temporal dependencies and global spatial topology of time-series data, thereby extracting high-order feature representations with spatiotemporal semantics. This module achieves progressive aggregation of multi-scale features through a recursive structure, enhancing the model's ability to model multi-source heterogeneous time-series data from underground coal mines.

[0050] GateTCN employs a structure combining causal convolution and dilated convolution to ensure the causality of time series and expand the receptive field. Its output undergoes non-linear filtering via a gating mechanism to preserve key temporal information. ChebGCN, on the other hand, models spatial features based on graph structures. Its core lies in applying Chebyshev polynomials to the graph Laplacian matrix. Perform recursive calculations; the formula is: ,in It is the identity matrix. This is the normalized graph Laplacian matrix. This polynomial is used to construct the graph convolution kernel, thereby performing low-pass filtering on node features in the frequency domain space while preserving frequency domain features related to the global topology.

[0051] Furthermore, in the multi-layer spatiotemporal aggregation module, the stacking order of GateTCN and ChebGCN is: GateTCN → ChebGCN → GateTCN. The number of neurons in each layer can be set according to the actual data dimension. For example, the number of GateTCN layers is 2, and the number of neurons in each layer is 64. The order of ChebGCN is... The value is typically set to 3 or 5 to balance computational complexity with frequency domain modeling accuracy. The feature vector output by the module will serve as input to subsequent aggregation layers (such as self-attention networks and average pooling layers) to further fuse multi-scale spatiotemporal information.

[0052] This step plays a crucial role in underground coal mine temporal anomaly detection systems. By introducing graph structures and frequency domain features, it effectively captures the topological dependencies between sensor nodes and their dynamic changes over time, thereby enhancing the model's ability to identify anomalies with multiple causes. In practical applications, this module is suitable for mine environments with complex spatial layouts, exhibiting stronger robustness and generalization capabilities, especially in scenarios where the sensor network exhibits topological changes or significant frequency domain features.

[0053] S32, the attention mechanism spatiotemporal encoder enhances the representation ability of key features by weighting the input spatiotemporal topological data and frequency domain feature data through a self-attention network.

[0054] Specifically, in some implementations, the spatiotemporal encoder of the attention mechanism uses a self-attention network to weight the input spatiotemporal topological data and frequency domain feature data. Its core purpose is to enhance the representational ability of key features, thereby improving the accuracy and robustness of anomaly detection in underground coal mine time-series data. This step is technically implemented based on the self-attention mechanism in the Transformer architecture, combined with a graph neural network (GNN) and time-series modeling methods to achieve joint encoding of multi-dimensional spatiotemporal features.

[0055] Self-attention networks dynamically allocate attention weights by calculating the correlations between input features, enabling the model to focus on features with higher discriminative power for the current task. Specifically, the input spatiotemporal topological data is represented as a dynamic feature tensor with a size of [missing information]. ,in Indicates the number of time steps. This indicates the number of nodes (i.e., the number of sensors). This represents the feature dimension of each node. Frequency domain feature data is extracted from the graph structure through Graph Fourier Transform, and its form is the projection of the feature vector into the graph space. The self-attention network takes these features as input and calculates the correlation between nodes at different time steps through a multi-head attention mechanism, thereby generating a weighted feature representation.

[0056] The formula for calculating the attention weights in a self-attention network is:

[0057] in , , These are matrices representing queries, keys, and values. is the dimension of the key vector, used to scale the dot product to avoid excessively large values. In this invention, Typically, the number of attention heads is set to 64 or 128, and the number of attention heads is set to 8 to enhance the model's ability to perceive multidimensional features.

[0058] This step is primarily used for the joint modeling of multi-source heterogeneous sensor data in underground coal mines. For example, in the working face area, multiple sensors (such as gas concentration, temperature, and wind speed) have topological relationships in space, and their data exhibit complex dependencies over time. By weighting these data using a self-attention network, the model can automatically identify nodes and features that contribute more to anomaly detection within a specific time window, thereby improving overall detection performance.

[0059] By introducing an attention mechanism, the model can dynamically adjust the importance of different nodes and features, effectively suppressing the influence of noise and redundant information, and enhancing the expressive power of key spatiotemporal features. Furthermore, by combining frequency domain feature processing, the model can capture periodic changes or sudden bursts.

[0060] S4 performs anomaly detection based on the difference between the reconstruction results and the original time series data, and periodically detects changes in the topology of the mine sensor network. When the detected topology difference exceeds a preset threshold, it triggers model parameter updates.

[0061] Specifically, anomaly detection based on the difference between the reconstructed results and the original time-series data is the core component of this invention for achieving high-precision underground safety monitoring in coal mines. The technical principle behind this step is based on the deep learning model's ability to learn normal data patterns. By comparing the differences between the reconstructed data output by the model and the original input data, anomalies deviating from the normal pattern are identified. The specific operation is as follows: First, the preprocessed and spatiotemporally feature-extracted data is input into the trained deep learning model. The model, through a spatiotemporal encoder, a multi-layer spatiotemporal aggregation module, and a decoder structure, performs feature extraction and reconstruction on the input data, ultimately outputting reconstructed time-series data with the same dimensionality as the original data. Subsequently, the difference between the original data and the reconstructed data is calculated, typically using mean squared error (MSE) or mean squared error (MAE) as evaluation metrics.

[0062] in, This indicates the original time series data at time step. The observed values, This represents the reconstructed value of the model at that time step. This represents the length of the time series. If... Exceeding the preset anomaly detection threshold If so, then an anomaly is determined to exist at that time step. Threshold Error distribution can be statistically set using historical data, such as taking the 95th percentile of the error distribution or a dynamic adjustment mechanism, to adapt to data fluctuations under different operating conditions.

[0063] In practical applications, this step is deployed in real-time monitoring systems of underground coal mine sensor networks, suitable for joint anomaly detection of multi-dimensional time-series data such as gas concentration, temperature, and wind speed. By introducing graph neural networks and frequency domain features, the model can capture the spatial dependencies and temporal periodicities between sensors, thereby effectively identifying multi-source anomalies caused by equipment failure, sudden environmental changes, or human interference. Its technical value lies in significantly reducing false alarm and false negative rates, improving the accuracy and response speed of coal mine safety early warnings, and providing a reliable basis for subsequent emergency response and decision support.

[0064] The coal mine underground time series anomaly detection method based on the fusion of spatiotemporal topology and frequency domain features in this invention effectively integrates spatiotemporal topological relationships and frequency domain features, improving the accuracy and robustness of anomaly detection of multi-source time series data in coal mines, and reducing false alarms and false negatives.

[0065] Furthermore, S4 includes: S41, calculate the point-by-point difference between the original time series data and the reconstruction result. If the difference exceeds the preset threshold, it is judged as abnormal. The threshold is dynamically adjusted according to the statistical distribution of historical data.

[0066] In the specific implementation, the original time-series data Reconstruction results compared with model output Perform point-by-point difference calculation between them, that is The difference sequence This reflects the model's reconstruction error for the current data point; the larger the error, the more likely it is that the point deviates from the normal pattern. To avoid the insufficient adaptability of static thresholds to dynamic environments, this invention employs a dynamic threshold adjustment mechanism based on the statistical distribution of historical data. Specifically, the mean value can be calculated based on the reconstruction error distribution within the sliding window. and standard deviation And set the threshold as ,in This is an adjustable parameter used to control the sensitivity of anomaly detection. If at a certain moment... If so, it is considered abnormal.

[0067] In application scenarios, this step is widely used in real-time monitoring systems of underground coal mine sensor networks, such as for anomaly identification of key parameters like gas concentration, temperature, and wind speed. Due to the complex environment of coal mines, sensor data is easily affected by noise, equipment malfunctions, or geological changes. The dynamic threshold mechanism can effectively adapt to data fluctuations and reduce false alarm and false negative rates.

[0068] By combining the reconstruction error of the deep learning model with a dynamic threshold, high-precision identification of anomalies is achieved. Compared with the traditional threshold method, it has stronger adaptability and can cope with various complex situations such as data drift, sudden anomalies, and periodic fluctuations, thereby improving the intelligence level and response efficiency of the coal mine safety monitoring system.

[0069] S42, The method for calculating the graph structure similarity index is as follows: For nodes u and v on two graphs, their structure similarity index is... ,in It is the domain of node u. It is the degree of node u. When the number of dissimilar nodes exceeds the threshold, the model is retrained.

[0070] Specifically, in some implementations, the graph structure similarity index is used to assess the similarity between two graph structures in terms of node neighborhood relationships, thereby determining whether their topological structures have changed significantly. The mathematical expression of this index is:

[0071] in, Represents a node The neighborhood set of the node, that is, the set of the nodes All directly connected nodes; Represents a node The degree of a, i.e., the number of its connected edges; Represents a node and The inner product between eigenvectors is used to measure their similarity in the feature space. This formula combines the structural adjacency information of the graph with the vector similarity of node features, thus comprehensively evaluating the similarity of node pairs across multiple dimensions.

[0072] From a technical implementation perspective, this step first constructs a graph structure for the underground sensor network in the coal mine, where each node represents a sensor and edges represent their spatial adjacency or communication relationships. During model operation, the system periodically collects new sensor data and constructs a new graph structure, which is then compared with the original graph. By traversing all node pairs, its structural similarity index is calculated, and the number of dissimilar nodes is counted. When the number of dissimilar nodes exceeds a preset threshold... When this happens, the system will trigger the model's retraining mechanism to adapt to the dynamic changes in the topology.

[0073] At the parameter level, neighborhood intersection size Reflects the nodes and In graph structures, the higher the value of the direct connection similarity, the more similar the structures. Vector similarity is also relevant. The similarity of node feature vectors is measured using a normalized inner product, with values ​​ranging from -1 to 1. Values ​​closer to 1 indicate greater feature similarity. (Threshold) The setting is usually based on the statistical distribution of historical data, such as setting it to 5% or 10% of the total number of dissimilar nodes, to ensure that the model is updated in a timely manner when the topology changes significantly.

[0074] This step is primarily used for dynamic topology detection in underground sensor networks in coal mines. Due to the complex underground environment, sensor connections may change due to equipment movement, damage, or additions, causing the input structure of the original model to become invalid. By calculating the graph structure similarity index, the system can automatically identify topology changes, thereby triggering model retraining without manual intervention and improving the model's robustness and adaptability.

[0075] By quantizing the similarity of graph structures, this method enables automatic perception of topology changes in underground coal mine sensor networks. This ensures that the deep learning model can be updated promptly when significant structural changes occur, thereby maintaining the accuracy and stability of anomaly detection. This approach has significant practical value in the dynamic environment of underground coal mines and can effectively address the topology drift problem in sensor networks.

[0076] Also includes: S5 performs sliding window framing on the preprocessed time series data, setting the sliding window size to k and the sliding step size to s, converting the variable-length time series data into fixed-length frame data to adapt to the neural network input requirements.

[0077] Specifically, in some implementations, performing sliding window framing on the preprocessed time-series data is a key step in transforming variable-length underground coal mine sensor time-series data into fixed-length frame data suitable for input to deep learning models. This step is based on a sliding window mechanism, where the window size is set... and sliding step size The original time series data is divided into multiple consecutive subsequences, thereby constructing a standardized input format that can be used for model training and inference.

[0078] Sliding window frame splitting typically employs causal convolution or sequence slicing to divide frames into segments of length 10 ... The time series data is divided into Each frame contains [number] frames. Data from consecutive time points. For example, if the original data length is 1000 time steps, the sliding window size... sliding step size Then it can generate There are 10 frames. The input dimension of each frame is 1000. ,in This represents the feature dimensions of each time step, such as sensor type, spatial coordinates, and device status.

[0079] Sliding window size The value is usually set based on the model's receptive field and task requirements, and typically ranges from [value range missing]. This ensures the model can capture sufficient temporal context information. Sliding step size. Normal settings of ,like or This aims to strike a balance between data coverage density and computational efficiency. Furthermore, to avoid losing boundary data, zero-padding or mirror padding can be optionally applied at the beginning and end of the data to ensure that all time points are covered.

[0080] This step is widely used in the modeling and input processing of multi-source heterogeneous sensor data in underground coal mines, such as the joint modeling of multi-dimensional time-series signals like gas concentration, temperature, wind speed, and equipment vibration. By using sliding window framing, the raw data can be adapted to the input format of deep learning models (such as Transformer, LSTM, GateTCN, GCN, etc.), thereby achieving end-to-end learning and detection of abnormal patterns.

[0081] Sliding window frame splitting not only improves the structuring of data but also enhances the model's ability to perceive local temporal patterns. Through proper settings... and This approach can maintain the consistency of model inputs while preserving sufficient temporal context information, providing a high-quality data foundation for subsequent spatiotemporal coding, aggregation, and reconstruction, thereby improving the accuracy and robustness of anomaly detection.

[0082] The coal mine underground time-series anomaly detection method based on spatiotemporal topology and frequency domain feature fusion in this invention introduces a sliding window framing mechanism to transform variable-length multi-source time-series data from underground coal mines into fixed-length frame data that adapts to neural network input. By combining spatiotemporal topology relationships and frequency domain feature analysis, the method further enhances the model's ability to capture local time-series patterns and improves the real-time performance and adaptability of anomaly detection in dynamically changing environments.

[0083] Example 2 like Figure 2 As shown, the deep learning anomaly detection method for underground coal mine time-series data of the present invention may further include the following steps: Step 1: Mine safety monitoring data collection and analysis.

[0084] In one specific embodiment, the acquisition of mine safety monitoring data involves the safety monitoring system collecting data in real time based on actual measurements taken underground, at a certain acquisition frequency (10 seconds or 60 seconds). The acquisition formats include timed data and variable-value / abnormal data. Timed data is the most common format, where data is collected and recorded at fixed time intervals. Variable-value / abnormal data is a data format that records the corresponding change and time point only when the data or sensor state changes.

[0085] In one specific embodiment, the parsing of security monitoring data is as follows: if it is a timed format, it is used directly; if it is a variable or abnormal format, it needs to be parsed into a timed sequence format based on the change records.

[0086] Step 2: Perform data preprocessing on the time series data.

[0087] In one specific embodiment, the preprocessing method includes: temporal noise reduction, alignment, normalization, missing value generation, and outlier removal.

[0088] In one specific embodiment, the noise reduction method addresses the issue that underground sensors are easily affected by noise, vibration, and other factors during coal mining operations, resulting in noisy datasets. Traditional methods such as moving average, exponential average, and wavelet are ineffective. Therefore, a deep learning noise reduction model is used, such as the Conv-TasNet model specifically trained for coal mine safety monitoring environments. This model can reduce human intervention and perform noise reduction in complex coal mine scenarios.

[0089] In a specific embodiment, data alignment involves directly mapping the data from different sensors to the same time point based on the time of each sensor's data.

[0090] In one specific embodiment, outlier removal: Each value in the time series is checked, and if an unreasonable value is found, such as exceeding the sensor's range, or the sensor continuously alarming when the coal mine is normal, the abnormal data is removed and replaced with a NaN value.

[0091] In a specific embodiment, normalization is performed: Normalization can be achieved using the minimum-maximum method or the z-score method, but the z-score method is generally used. The specific process is as follows: For a time series t, the normalized value for each point in the series is: ,in Let be the nth value of sequence t. Let be the mean of sequence t. Let be the standard deviation of sequence t.

[0092] In one specific embodiment, missing value generation: for missing data, traditional linear interpolation or Lagrange interpolation.

[0093] In one specific embodiment, data framing is performed: considering that the time series obtained from target trajectory data processing has an indefinite length, while some neural network models require a fixed input length, a sliding window process is applied to the input sequence data. The sliding window size is set to k, and the sliding step size is s, representing taking k consecutive sampling points as one frame of data. These frames of time series samples are the input to the deep learning model.

[0094] Step 3: Construction of spatiotemporal topology data and spatiotemporal frequency domain data.

[0095] In one specific embodiment, the spatiotemporal topological relationship model in step 3 includes three parts: relationship graph structure, dynamic feature tensor, and temporal information encoding.

[0096] In one specific embodiment, the relationship graph structure is a matrix used to describe spatial relationships. We construct the nodes of the graph based on geographical locations. If there is an underground road connecting two nodes, an edge is constructed. The weight of the edge is the distance between them. Therefore, the relationship graph structure is a matrix of size (number of nodes . number of nodes), and the relationship and weight of the edges are reflected in the values ​​of the matrix. In one specific embodiment, the dynamic feature tensor is a tensor of size (sequence length, number of nodes, feature dimension) that describes the features of each node at different time steps. The features may include time-series values ​​collected by sensors (gas concentration, temperature, etc.), spatial geographic information, node type, nearby equipment operation status, holiday markers, etc.

[0097] In one specific embodiment, time information is encoded using One-Hot or Sin-Cos encoding to describe time periodicity characteristics.

[0098] In one specific embodiment, the spatiotemporal frequency domain feature mining model in step 3 adopts a graph Fourier decomposition model. The specific process is to first construct a graph Laplacian matrix based on the relation graph structure, then perform eigenvalue decomposition on the matrix to obtain the eigenvector basis, and finally project the dynamic feature tensor onto the eigenvector basis to obtain the graph Fourier decomposition result.

[0099] Step 4: Input the spatiotemporal data into the deep learning temporal anomaly detection model to calculate the reconstruction results.

[0100] In one specific embodiment, the deep temporal anomaly detection model in step 4 is constructed as follows: It receives spatiotemporal topological data and spatiotemporal frequency data as input. First, it uses an encoder neural network to extract data features. Then, it inputs the encoded features into a spatiotemporal aggregation module to obtain spatiotemporal aggregated features. Next, it inputs these spatiotemporal aggregated features into the next spatiotemporal aggregation module to obtain further spatiotemporal aggregated features. After multiple stacking operations, it obtains the final spatiotemporal aggregated features. Finally, it inputs the encoded features, the first-layer spatiotemporal aggregated features, and the final spatiotemporal aggregated features together into an aggregation layer neural network for aggregation. The aggregation result is then input into a decoder neural network to obtain the decoding result, and finally input into a linear output layer to obtain the final reconstruction result.

[0101] In one specific embodiment, the spatiotemporal aggregation module is structured as follows: a gated temporal convolutional neural network is connected to a Chebyshev graph convolutional neural network, which is then connected to another gated temporal convolutional neural network.

[0102] In one specific embodiment, the encoder neural network in step 4 is a spatiotemporal encoder neural network employing an attention mechanism.

[0103] In one specific embodiment, the gated temporal convolutional neural network structure in step 4 is a causal convolutional connection with gated mechanism, dilated convolution, filtered, and then residual connection is performed.

[0104] In one specific embodiment, the aggregation layer in step 4 is a self-attention network connection average pooling layer.

[0105] In one specific embodiment, the decoder in step 4 is an inverse graph convolutional network connected to an inverse temporal convolutional network.

[0106] In one specific embodiment, the Chebyshev graphical convolutional neural network in step 4 first calculates the Laplacian matrix L, and then calculates the Chebyshev polynomials of each order. Finally, it is aggregated through a linear neural network. The Chebyshev polynomial is recursively defined. .

[0107] Step 5: Perform temporal anomaly detection based on the reconstruction results.

[0108] In one specific embodiment, the data is first preprocessed, then spatiotemporal topological relationship data and spatiotemporal frequency data are constructed. Next, the constructed spatiotemporal data is input into a trained deep learning temporal anomaly detection model to obtain the data reconstruction result. Finally, the difference between the original data and the reconstructed data is calculated. If it exceeds a threshold, it is judged as an anomaly. The threshold needs to be set.

[0109] Step 6: Periodically detect changes in spatiotemporal topology.

[0110] In one specific embodiment, step 6 detects changes in spatiotemporal topological relationships using a graph structural similarity index. For two nodes u and v on two graphs, their structural similarity index is:

[0111] in It is the domain of node u. This is the degree of node u. If the structural similarity index between two nodes in two graphs exceeds a threshold, the nodes are considered dissimilar. If the number of dissimilar nodes exceeds the threshold, the anomaly detection model is retrained. The threshold needs to be specified.

[0112] According to the method of the present invention, artificial intelligence techniques such as deep learning, spatiotemporal topology learning, and spatiotemporal frequency mining are used on the basis of existing coal mine safety monitoring time series data to solve the problem of data anomaly detection in coal mine safety monitoring. It has a wide range of applications, strong generalization ability, and higher recognition accuracy.

[0113] Example 3 This invention also provides a deep learning anomaly detection system 10 for underground coal mine time-series data, such as... Figure 3 As shown, the system 10 includes: The multi-source sensor data acquisition and parsing module 100 is used to acquire time-series data from multi-source sensors in underground coal mines and parse it into a unified time format. The time-series data includes sensor type, spatial location information and dynamic characteristics. The spatiotemporal topology construction and frequency domain feature mining module 200 is used to construct spatiotemporal topology data containing relational graph structure, dynamic feature tensor and time information encoding, and to mine the frequency domain features of the spatiotemporal topology data through graph Fourier transform; The deep learning model feature extraction and aggregation module 300 is used to input the spatiotemporal topological data and frequency domain feature data into the deep learning model, and perform feature extraction and aggregation through a multi-layer spatiotemporal aggregation module to generate the reconstruction result of time series data. The anomaly detection and topology update triggering module 400 is used to perform anomaly detection based on the difference between the reconstruction result and the original time series data, and periodically detect changes in the topology of the mine sensor network. When the detected topology difference exceeds a preset threshold, the module triggers a model parameter update.

[0114] Furthermore, the spatiotemporal topology construction and frequency domain feature mining module is also used for: Map nodes are constructed based on the geographic coordinates of the sensors. If there is an underground road connecting two nodes, an edge is established. The weight of the edge is the actual distance between the nodes, and a matrix with a size of number of nodes × number of nodes is generated. The graph Laplacian matrix L is calculated based on the relational graph structure. The eigenvector basis is obtained through eigenvalue decomposition. The dynamic feature tensor is then projected onto the eigenvector basis to extract frequency domain features.

[0115] Furthermore, the feature extraction and aggregation module of the deep learning model is also used for: Feature extraction is performed using a structure that stacks gated temporal convolutional neural networks, Chebyshev graphical convolutional neural networks, and gated temporal convolutional neural networks sequentially. The Chebyshev graphical convolutional neural network utilizes a recursive formula... Calculate Chebyshev polynomials of all orders; By using a self-attention network to weight the input spatiotemporal topological data and frequency domain feature data, the representation ability of key features is enhanced.

[0116] Furthermore, the anomaly detection and topology update triggering module is also used for: Calculate the point-by-point difference between the original time-series data and the reconstruction result. If the difference exceeds the preset threshold, it is judged as an anomaly. The threshold is dynamically adjusted according to the statistical distribution of historical data. Calculate the graph structural similarity index. For nodes u and v in two graphs, the structural similarity index is: ,in It is the domain of node u. It is the degree of node u. When the number of dissimilar nodes exceeds the threshold, the model is retrained.

[0117] Furthermore, it also includes: The sliding window framing module is used to perform sliding window framing on the preprocessed time series data. The sliding window size is set to k and the sliding step size is s. The variable-length time series data is converted into fixed-length frame data to adapt to the input requirements of the neural network.

[0118] This invention discloses a deep learning anomaly detection system for underground coal mine time-series data. By employing artificial intelligence techniques such as deep learning, spatiotemporal topology learning, and spatiotemporal frequency mining based on existing coal mine safety monitoring time-series data, it solves the problem of data anomaly detection in coal mine safety monitoring. The system has a wide range of applications, strong generalization ability, and higher recognition accuracy.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0120] In the description of this specification, the references to terms such as "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 the present invention. 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.

[0121] 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 invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A deep learning-based anomaly detection method for underground coal mine time-series data, characterized in that, include: S1, collect time-series data from multiple sources in the coal mine and parse it into a unified time format. The time-series data includes sensor type, spatial location information and dynamic characteristics. S2, construct spatiotemporal topological data including relational graph structure, dynamic feature tensor and time information encoding, and mine the frequency domain features of the spatiotemporal topological data through graph Fourier transform; S3 inputs spatiotemporal topological data and frequency domain feature data into a deep learning model, and performs feature extraction and aggregation through a multi-layer spatiotemporal aggregation module to generate the reconstruction result of time series data; S4 performs anomaly detection based on the difference between the reconstruction results and the original time series data, and periodically detects changes in the topology of the mine sensor network. When the detected topology difference exceeds a preset threshold, it triggers model parameter updates.

2. The method as described in claim 1, characterized in that, The steps for constructing spatiotemporal topology data and mining frequency domain feature data in S2 also include: S21, The method for constructing the relational graph structure is as follows: construct graph nodes based on the geographic coordinates of the sensors. If there is an underground road connection between two nodes, establish an edge. The weight of the edge is the actual distance between the nodes, forming a matrix with a size of the number of nodes × the number of nodes. S22, The implementation of the graph Fourier transform is as follows: the graph Laplacian matrix L is calculated based on the relation graph structure, the eigenvector basis is obtained through eigenvalue decomposition, and the dynamic feature tensor is projected onto the eigenvector basis to extract frequency domain features.

3. The method as described in claim 1, characterized in that, The S3 multi-layer spatiotemporal aggregation module also includes the following steps for feature extraction and aggregation: S31, the multi-layer spatiotemporal aggregation module is composed of a gated temporal convolutional neural network, a Chebyshev graph convolutional neural network, and a gated temporal convolutional neural network stacked sequentially, wherein the Chebyshev graph convolutional neural network is constructed using a recursive formula. Calculate Chebyshev polynomials of all orders; S32, the attention mechanism spatiotemporal encoder enhances the representation ability of key features by weighting the input spatiotemporal topological data and frequency domain feature data through a self-attention network.

4. The method as described in claim 1, characterized in that, The step of anomaly detection based on the difference between the reconstruction result and the original time-series data also includes: S41, calculate the point-by-point difference between the original time series data and the reconstruction result. If the difference exceeds the preset threshold, it is judged as abnormal. The threshold is dynamically adjusted according to the statistical distribution of historical data. S42, The method for calculating the graph structure similarity index is as follows: For nodes u and v on two graphs, their structure similarity index is... ,in It is the domain of node u. It is the degree of node u. When the number of dissimilar nodes exceeds the threshold, the model is retrained.

5. The method as described in claim 1, characterized in that, Also includes: S5 performs sliding window framing on the preprocessed time series data, setting the sliding window size to k and the sliding step size to s, converting the variable-length time series data into fixed-length frame data to adapt to the neural network input requirements.

6. A deep learning anomaly detection system for underground coal mine time-series data, characterized in that, include: The multi-source sensor data acquisition and parsing module is used to acquire time-series data from multi-source sensors in underground coal mines and parse it into a unified time format. The time-series data includes sensor type, spatial location information, and dynamic characteristics. The spatiotemporal topology construction and frequency domain feature mining module is used to construct spatiotemporal topology data containing relational graph structure, dynamic feature tensor and time information encoding, and to mine the frequency domain features of the spatiotemporal topology data through graph Fourier transform; The deep learning model feature extraction and aggregation module is used to input spatiotemporal topological data and frequency domain feature data into the deep learning model, and perform feature extraction and aggregation through a multi-layer spatiotemporal aggregation module to generate the reconstruction results of time series data. The anomaly detection and topology update triggering module is used to detect anomalies based on the differences between the reconstruction results and the original time-series data, and periodically detect changes in the topology of the mine sensor network. When the detected topology difference exceeds a preset threshold, the model parameters are updated.

7. The system as described in claim 6, characterized in that, The spatiotemporal topology construction and frequency domain feature mining module is also used for: Map nodes are constructed based on the geographic coordinates of the sensors. If there is an underground road connecting two nodes, an edge is established. The weight of the edge is the actual distance between the nodes, and a matrix with a size of number of nodes × number of nodes is generated. The graph Laplacian matrix L is calculated based on the relational graph structure. The eigenvector basis is obtained through eigenvalue decomposition. The dynamic feature tensor is then projected onto the eigenvector basis to extract frequency domain features.

8. The system as described in claim 6, characterized in that, The deep learning model feature extraction and aggregation module is also used for: Feature extraction is performed using a structure that stacks gated temporal convolutional neural networks, Chebyshev graphical convolutional neural networks, and gated temporal convolutional neural networks sequentially. The Chebyshev graphical convolutional neural network utilizes a recursive formula... Calculate Chebyshev polynomials of all orders; By using a self-attention network to weight the input spatiotemporal topological data and frequency domain feature data, the representation ability of key features is enhanced.

9. The system as described in claim 6, characterized in that, The anomaly detection and topology update triggering module is also used for: Calculate the point-by-point difference between the original time-series data and the reconstruction result. If the difference exceeds the preset threshold, it is judged as an anomaly. The threshold is dynamically adjusted according to the statistical distribution of historical data. Calculate the graph structural similarity index. For nodes u and v in two graphs, the structural similarity index is: ,in It is the domain of node u. It is the degree of node u. When the number of dissimilar nodes exceeds the threshold, the model is retrained.

10. The system as described in claim 6, characterized in that, Also includes: The sliding window framing module is used to perform sliding window framing on the preprocessed time series data. The sliding window size is set to k and the sliding step size is s. The variable-length time series data is converted into fixed-length frame data to adapt to the input requirements of the neural network.

Images