A method and system for coal mine equipment state data graph structured conversion

Abstract

The present application relates to the technical field of coal mine equipment detection, and more particularly to a method and system for graph structure conversion of coal mine equipment state data, the method comprising obtaining multi-dimensional time series state data of a coal mine equipment sensor and preprocessing, and constructing a standardized multi-dimensional state perception matrix; using an adaptive sliding window method to time sequence cut the state perception matrix to generate a graph node with a low-dimensional semantic feature vector; constructing a dynamic correlation graph according to the graph node using a time sequence evolution and state similarity dual connection strategy, and performing consistency verification on the dynamic correlation graph based on a multi-loss function joint constraint mechanism to generate a standardized graph anomaly data set, which effectively solves the problem of systematic anomaly missed detection and time sequence evolution information loss caused by the mutual fragmentation of time sequence modeling and graph structure modeling in the prior art, and significantly improves the recognition ability of complex abnormal patterns.

Description

Technical Field

[0001] This invention relates to the field of coal mine equipment testing technology, and in particular to a method and system for structured conversion of coal mine equipment status data diagrams. Background Technology

[0002] With the deepening of intelligent construction in coal mines, equipment condition monitoring technology has become a key means to ensure safe production in coal mines. Modern coal mine equipment is generally equipped with a variety of sensors, which can collect multi-dimensional condition parameters in real time and form large-scale time-series monitoring data. Traditional monitoring methods mainly rely on manual inspections and simple threshold alarms, which can no longer meet the needs of accurate diagnosis under complex working conditions. Therefore, time-series anomaly detection methods based on deep learning and anomaly detection methods based on graph neural networks have gradually become the mainstream technical paths.

[0003] Currently, time-series anomaly detection mainly uses long short-term memory networks or gated recurrent units to model multidimensional monitoring data and judge the abnormal state through prediction errors. Graph neural network methods construct a static graph structure by treating devices as nodes and physical connections as edges, and aggregate neighbor information through graph convolution operations to identify anomalies. In addition, some hybrid solutions attempt to connect time-series models and graph models in series or in parallel to combine the advantages of time-series modeling and spatial relationship modeling to achieve multi-dimensional feature extraction.

[0004] However, time-series modeling methods only focus on the temporal evolution of individual equipment states, ignoring the complex spatial relationships and physical coupling between various equipment in a fully mechanized mining face. For example, the coordinated fluctuation relationship between the coal mining machine's cutting load changes and the scraper conveyor's operating resistance and motor current. Single-dimensional time-series analysis cannot identify systemic anomalies caused by cross-equipment associated faults, easily leading to false alarms for individual machines. While graph neural network-based methods can model the spatial connections between equipment components, they mostly use static graph structures and node features only utilize parameters at the current moment, losing historical evolution information and state decay trends of mechanical components from slight wear to fatigue spalling. It is difficult to capture the dynamic response of equipment under different cutting intensities. In addition, existing hybrid modeling schemes lack a deep semantic fusion mechanism between time-series features and topological structures, easily losing key nonlinear anomaly information during feature conversion. Furthermore, when facing random spike signals generated by strong electromagnetic interference underground, there is a lack of consistency verification methods to ensure accurate alignment between the graph prediction results and the original physical time period, easily leading to alarm drift. At present, there is a need for a graph-structured conversion method and system for coal mine equipment state data. Summary of the Invention

[0005] To address the problems of insufficient detection accuracy and reliability caused by the separation of temporal information and spatial relationships and the lack of consistency verification mechanisms in existing coal mine equipment anomaly detection technologies, this invention provides a method and system for the structured conversion of coal mine equipment status data graphs.

[0006] In a first aspect, the present invention provides a method for structured conversion of coal mine equipment status data diagrams, which adopts the following technical solution: A method for structured conversion of coal mine equipment status data diagrams includes: Acquire multidimensional time-series state data from sensors of coal mine equipment and preprocess them to construct a standardized multidimensional state perception matrix; An adaptive sliding window method is used to perform temporal segmentation of the state-aware matrix, calculate the information entropy of each window, and filter high-value time windows based on a preset threshold. Principal component analysis is used to perform dimensionality reduction encoding on the high-value time windows to generate graph nodes with low-dimensional semantic feature vectors. A dynamic association graph is constructed based on the graph nodes using a dual connection strategy of temporal evolution and state similarity. This includes establishing temporal evolution edges based on Gaussian kernel function measurement of adjacent node features, and establishing state similarity edges based on dynamic time warping algorithm to measure feature distance between nodes. A mapping relationship between graph nodes and the original time series positions is established. The consistency of the dynamic association graph is verified based on a joint constraint mechanism of multiple loss functions. A standardized graph anomaly dataset is generated and output to the coal mine monitoring terminal for early identification of potential equipment failures.

[0007] Furthermore, the construction of the standardized multidimensional state-aware matrix includes smoothly filling missing breakpoints in the multidimensional time-series state data using a cubic spline interpolation algorithm, ensuring the continuity of the first and second derivatives of the sequence at the connection points by constructing a piecewise cubic polynomial, and applying a sliding window method to the filled data. The criteria perform dynamic noise suppression, calculate the mean and standard deviation of the data within the window and correct them in real time through local median filtering, and perform Z-Score transformation on the corrected data to normalize the dimensions in order to construct the standardized multidimensional state perception matrix.

[0008] Furthermore, the calculation of the information entropy of each window and the filtering of high-value time windows based on a preset threshold includes setting the sliding window length and sliding step size, slicing the multidimensional state perception matrix along the time dimension, and generating a continuous set of time windows. Calculate the probability density distribution of state parameters within each time window, and calculate the information complexity of each window using the Shannon information entropy formula. The information complexity calculation formula is as follows: , Set a significance threshold , will satisfy The time window for the condition is selected as the high-value time window, and redundant windows with information entropy below the significance threshold are removed; in, For the first k Time window Information entropy Let be the probability density function value of the j-th dimension state parameter within the window. The mean of the information entropy of all windows. This is the preset adjustment coefficient.

[0009] Furthermore, generating graph nodes with low-dimensional semantic feature vectors includes extracting multi-dimensional state features within the high-value time window, constructing a high-dimensional feature matrix X, and calculating the covariance matrix of the high-dimensional feature matrix: ; The covariance matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the k largest eigenvalues ​​are selected to form the projection matrix. The original high-dimensional feature vector Linearly projected onto a low-dimensional semantic space, the resulting low-dimensional semantic feature vector is used as the core attribute of the graph node: ; Where C is the covariance matrix, Let be the feature vector of the i-th sample. The sample mean. The number of filtered windows. Let be the low-dimensional semantic feature vector of the i-th graph node. Let be the transpose of the projection matrix, and k be the target dimension after dimensionality reduction.

[0010] Furthermore, the Gaussian kernel function metric based on the features of adjacent nodes is used to establish temporal evolution edges, which includes traversing the graph nodes in chronological order and performing operations on nodes at adjacent times. and A directed temporal connection is established, and the weights of the temporal evolution edges are calculated using a Gaussian kernel function to capture the smooth evolution trend of device state transitions. The formula for calculating the weights of the temporal evolution edges is as follows: , in, For nodes and The temporal evolution edge weights between them and These are the low-dimensional semantic feature vectors of adjacent time nodes. The bandwidth parameter of the Gaussian kernel function is... Let be the Euclidean distance norm of the vector.

[0011] Furthermore, the step of measuring the feature distance between nodes and establishing state similarity edges based on the dynamic time warping algorithm includes using the dynamic time warping algorithm to calculate the feature distance between any two non-adjacent nodes. and Optimal alignment path distance between feature sequences: ; When dynamic time warping distance At the node and Undirected state similarity edges are established between them, and cross-temporal and spatial associations are made between nodes that are close to each other in the feature space through state similarity edges; in, For nodes and The dynamic time-normalized distance between them This represents the optimal curved alignment path between feature sequences. To align coordinate pairs on the path, It is a local distance metric function. and They are nodes and The p-th and q-th eigenvectors in the feature sequence, This is the preset similarity threshold.

[0012] Furthermore, the construction of the dynamic association graph using a dual connection strategy of temporal evolution and state similarity includes constructing a time-dimensional skeleton of the graph structure based on the temporal evolution edges, constructing cross-temporal and spatiotemporal topological connections of the graph structure based on the state similarity edges, and fusing the two types of edge sets to construct the dynamic association graph. The dual-connection strategy explicitly transforms the time-dependent and state-similar characteristics of device operation into a graph topology. in, For a set of graph nodes, Let be the set of edges. For the temporal evolution edge set, Let be the set of edges with similar states. It is an edge weight matrix that includes the edge weights of temporal evolution and the edge weights of state similarity.

[0013] Furthermore, the consistency verification of the dynamic association graph based on the joint constraint mechanism of multiple loss functions includes using a graph neural network model to perform anomaly detection on the dynamic association graph, introducing a weighted cross-entropy loss as a mapping loss, forcing the graph neural network model to perform high-confidence regression on the abnormal nodes mapped to the alarm time window, introducing a time-series smoothing term as a time-series continuity loss, and constructing a joint constraint mechanism of multiple loss functions based on the mapping loss and the time-series smoothing term. The expression for the mapping loss is: , in, For mapping loss, For a set of graph nodes, Let be the weight coefficient of node i. For node i, the true anomaly label. This represents the anomaly probability predicted by the graph neural network model.

[0014] Furthermore, the generation of the standardized graph anomaly dataset includes calculating the attention coefficients between nodes through the attention layer of the graph neural network model, weighting and aggregating the graph node features based on the attention coefficients, and jointly optimizing and training the parameters of the graph neural network model by combining the mapping loss and the temporal smoothing term, ultimately generating a standardized graph anomaly dataset with spatiotemporal consistency. The formula for calculating the attention coefficients is as follows: , in, Let be the attention coefficient between node i and node j. For normalized exponential functions, For the activation function of the linear unit with leakage correction, For attention vectors, The weight matrix is ​​a learnable matrix. and Let be the low-dimensional semantic feature vectors of node i and node j, respectively. This is a vector concatenation operation.

[0015] Secondly, a system for structured conversion of coal mine equipment status data diagrams includes: The data acquisition module is configured to acquire multi-dimensional time-series state data collected by multi-source sensors of coal mine equipment; The data preprocessing module is configured to preprocess multidimensional time series state data, including missing value imputation, outlier detection and correction, and dimensional normalization, and to construct a standardized multidimensional state perception matrix. The temporal segmentation and node mapping module is configured to: use an adaptive sliding window to perform temporal segmentation on the multidimensional state perception matrix, calculate the information entropy of each window and filter high-value time windows based on a preset threshold, perform dimensionality reduction encoding on high-value time windows through principal component analysis, and generate graph nodes with low-dimensional semantic feature vectors. The dynamic graph construction module is configured to: construct a dynamic association graph based on graph nodes using a dual connection strategy of temporal evolution and state similarity, including establishing temporal evolution edges based on Gaussian kernel function measurement of adjacent node features, and establishing state similarity edges based on dynamic time warping algorithm to measure feature distance between nodes. The consistency verification module is configured to: establish a mapping relationship between graph nodes and the original time series positions, and perform consistency verification on the dynamic association graph based on a joint constraint mechanism of multiple loss functions; The graph anomaly dataset generation module is configured to generate a standardized graph anomaly dataset and output it to the coal mine monitoring terminal for early identification of potential equipment failures.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This invention employs an adaptive sliding window technique for time-series segmentation, filters time windows with high information complexity based on the Shannon information entropy threshold, and performs dimensionality reduction encoding through principal component analysis. This achieves an effective mapping from continuous time-series segments to high-dimensional semantic graph nodes, eliminates redundant information in stable operation phases, retains key state segments containing fault initiation features, reduces the complexity of subsequent graph structure construction, and ensures the discriminativeness and representativeness of graph node features.

[0017] 2. This invention constructs a dynamic association graph by employing a dual connection strategy of temporal evolution and state similarity. It establishes temporal evolution edges based on Gaussian kernel function metric to capture the smooth evolution trend of device state transitions, and establishes state similarity edges based on dynamic time warping algorithm to aggregate similar fault modes across time and space. This achieves deep fusion modeling of temporal information and spatial relationships, effectively solving the problems of missed detection of systemic anomalies and loss of temporal evolution information caused by the separation of temporal modeling and graph structure modeling in existing technologies, and significantly improving the ability to identify complex anomaly patterns.

[0018] 3. This invention establishes a precise mapping relationship between graph nodes and the original time series positions, introduces weighted cross-entropy loss as a mapping loss to force high-confidence regression of abnormal nodes, introduces a time series smoothing term as a time series continuity loss to constrain the evolutionary continuity of abnormal scores, and achieves adaptive weight enhancement of heterogeneous abnormal subgraphs through graph neural network attention mechanism. This realizes strict alignment between graph anomaly detection results and original time series anomaly labels in the spatiotemporal dimension, eliminates alarm drift between physical warning labels and graph prediction, and ensures accurate transmission and traceability of abnormal information during the transformation process. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of a method for structural conversion of coal mine equipment status data diagram according to an embodiment of the present invention.

[0020] Figure 2 This is a flowchart of the time-series data processing according to an embodiment of the present invention.

[0021] Figure 3 This is a diagram illustrating the effect of the coal mining machine's detection distance in an embodiment of the present invention.

[0022] Figure 4 This is a graph showing the operating speed data of the coal mining machine according to an embodiment of the present invention.

[0023] Figure 5This is a diagram showing the normal data effect of the CWRU dataset in an embodiment of the present invention.

[0024] Figure 6 This is a diagram showing the effect of fault data in the CWRU dataset according to an embodiment of the present invention.

[0025] Figure 7 This is a normal data effect diagram of the MFPT dataset in an embodiment of the present invention.

[0026] Figure 8 This is a diagram showing the fault data effect of the MFPT dataset in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings.

[0028] Example 1 Reference Figure 1 This embodiment of a method for structured conversion of coal mine equipment status data diagrams includes: S1. Acquire multi-dimensional time series state data of coal mine equipment sensors and preprocess them to construct a standardized multi-dimensional state perception matrix to eliminate high-frequency harmonic glitches generated by underground frequency converters and repair sensor signal packet loss and breakpoints caused by electromagnetic shielding of communication in fully mechanized mining faces. S2. An adaptive sliding window method is used to perform temporal segmentation of the state perception matrix, calculate the information entropy of each window, and filter high-value time windows based on a preset threshold. Specifically, this is used to lock key operating condition fluctuation segments such as coal cutting-rock cutting switching, heavy load start-up and shutdown, or machine body obstruction and impact from massive steady-state operation sequences, thereby eliminating meaningless redundant data such as equipment idling and maintenance shutdown. Principal component analysis is used to perform dimensionality reduction encoding on the high-value time windows to generate graph nodes with low-dimensional semantic feature vectors. S3. Based on the graph nodes, a dynamic association graph is constructed using a dual connection strategy of temporal evolution and state similarity. This includes establishing temporal evolution edges based on Gaussian kernel function metrics of adjacent node features to characterize the mechanical damage degradation path of the bearings in the cutting or traveling parts of the coal mining machine over time; and measuring the feature distance between nodes and establishing state similarity edges based on the dynamic time warping algorithm to capture the recurrence pattern of working conditions under the same load conditions in different work cycles. S4. Establish the mapping relationship between graph nodes and the original time series positions, and perform consistency verification on the dynamic association graph based on the joint constraint mechanism of multiple loss functions to generate a standardized graph anomaly dataset.

[0029] Taking the monitoring of the rocker arm gearbox of a coal mining machine as a specific application scenario: In actual production, when the coal mining machine drum comes into contact with hard interbedded rock, the S1 module of this system first filters out the electromagnetic noise from the surrounding high-power motors; the S2 module identifies this instant as a "high-value window" through entropy calculation; the S3 module finds in the graph that the current state is highly similar to the "tooth surface wear" pattern in the historical database, but it shows a "non-sudden gradual change" on the temporal evolution edge. Finally, the graph anomaly dataset generated by the S4 module can accurately determine that the anomaly is triggered by geological conditions and has induced early mechanical damage, thus prompting the ground monitoring center to conduct local exploration in the next shift to prevent unplanned downtime caused by broken teeth.

[0030] Specifically, a method for structured conversion of coal mine equipment status data diagrams includes the following: like Figure 1 As shown in Figure S1, firstly, a multi-source sensor network deployed at the coal mine's fully mechanized mining face is used to collect real-time equipment operating status data, forming a multi-dimensional time-series status data stream. Specifically, the data collection covers the temperature, pressure, vibration, current, speed, and spatial attitude parameters of key equipment such as the coal mining machine, hydraulic supports, and scraper conveyors. The sampling frequency is set from 1Hz to 12kHz according to monitoring requirements, constructing an original data matrix. Where T is the time series length and n is the number of sensor dimensions, in this embodiment... . To address missing data points in the original data caused by signal loss, sensor malfunctions, and communication interruptions due to the complex electromagnetic environment downhole, a cubic spline interpolation algorithm is used for smooth filling. In practice, the intervals containing the missing data are... Divide into several sub-intervals, and in each sub-interval Construct a piecewise cubic polynomial above: , in, , , and The polynomial coefficients for the i-th interpolation interval are determined by solving the three-moment equation system. This ensures that the interpolation function satisfies the conditions of continuous function value, continuous first derivative, and continuous second derivative at the connection point, thereby effectively eliminating numerical jumps and pseudo-high-frequency components introduced by linear interpolation and maintaining the physical continuity of the equipment state evolution trend.

[0031] After filling in missing values, apply a sliding window to the data stream. The criteria are used for dynamic noise suppression. The sliding window length is set. (Corresponding to 50 sampling times), the local statistical characteristics of the samples within the current window are calculated by sliding point by point with a step size of 1. Specifically, for time... t Observations Calculate the window in which it is located. mean within with standard deviation If the observed values ​​satisfy: , If the observed value is then identified as an outlier, a local median filter is used for real-time correction. This involves replacing the outlier with the median of the data within the window, thereby effectively suppressing electromagnetic interference and extreme value drift caused by sensor oscillations while preserving the characteristics of signal mutations.

[0032] The data cleaned as described above is further subjected to Z-Score transformation to achieve dimensional normalization. For the first... j 3D features, calculate their global sample mean with standard deviation The original state parameters are expressed by the following formula. Mapping to the standard normal distribution space: , in, These are the normalized eigenvalues. For the first i The sampling time of the first sampling moment j The original state parameters of the 3D sensor are transformed through this process. Data with different physical dimensions (such as temperature in degrees Celsius, current in amperes, and vibration acceleration g) are uniformly converted to a statistical space with zero mean and unit variance. This eliminates the influence of dimensional differences on subsequent graph structure edge weight calculation and distance measurement, ultimately constructing a standardized multidimensional state sensing matrix. This provides a high-quality data benchmark for the transformation of time-series data into graph structures.

[0033] S2. After completing the construction of the standardized multidimensional state perception matrix, an adaptive sliding window technique is used to discretize the continuous time-series data to achieve a mapping from the time-series space to the graph node space. This is used to identify key characteristic signals generated by the equipment due to changes in cutting hardness, load switching, or sudden impacts. Specifically, for the coal mining machine operation monitoring scenario, the sliding window length is set. Each time step (corresponding to approximately 17 minutes of continuous monitoring), sliding step size At each time step (i.e., with a 50% overlap), the multidimensional state-aware matrix Z is sliced ​​along the time dimension. This overlapping sampling strategy ensures sufficient feature transition information between adjacent windows, avoids the abrupt state transition boundary effect caused by hard segmentation, and generates a continuous set of time windows. , where m is the total number of windows.

[0034] For each time window The 12-dimensional state parameter submatrix is ​​extracted, and the probability density distribution of each dimension of the state parameter is calculated. In practice, kernel density estimation or histogram statistics methods are used to obtain the probability mass function of each dimension. Then, the information complexity of the window is calculated using Shannon's information entropy formula: , in, For the first k Time window Information entropy Let be the probability density function value of the j-th dimension state parameter within the window. This information entropy measure reflects the uncertainty and complexity of the device's operating state within the window. The higher the entropy value, the more drastic the fluctuation of the device's state within that period, and the richer the fault characteristic information contained therein.

[0035] To eliminate redundant information windows during the stable operation phase of the coal mining machine, a significance threshold is set. ,in The arithmetic mean of the information entropy of all windows. The preset adjustment coefficient (taken in this embodiment) ). Will satisfy The time windows for the conditions are selected as high-value time windows. Only these windows containing significant state changes or potential fault initiation features are retained to participate in the subsequent graph structure construction, thereby significantly reducing computational complexity while ensuring the sensitivity of anomaly detection.

[0036] For the selected high-value time windows, multi-dimensional state features are further extracted to construct high-dimensional feature vectors. Specifically, for each window, the statistical features (mean, variance, kurtosis, skewness) and frequency domain features (FFT spectral energy) of each dimension are calculated to construct a high-dimensional feature matrix. ,in The number of windows after filtering is denoted by d, and d is the feature dimension (d=29 in this embodiment). To eliminate correlation between features and reduce the curse of dimensionality, principal component analysis is used for dimensionality reduction encoding. First, the covariance matrix of the high-dimensional feature matrix is ​​calculated: , Where C is The covariance matrix is ​​dimensional. Let be the d-dimensional feature vector of the i-th sample. Given the sample mean vector, perform eigenvalue decomposition on the covariance matrix C. Obtain eigenvalues and its corresponding eigenvectors Select the k largest eigenvalues ​​(in this embodiment) k The eigenvectors corresponding to (=8) form the projection matrix. The original high-dimensional feature vector Linear projection to a low-dimensional semantic space: , in, Let be the low-dimensional semantic feature vector of the i-th graph node. As the transpose of the projection matrix, this low-dimensional vector encapsulates the nonlinear coupling information between cutoff drag and spatial attitude, while filtering out noise redundancy in the original high-dimensional space, ultimately generating a set of graph nodes. and its characteristic matrix The entities that form the basis for constructing dynamic graph structures.

[0037] S3. After generating the graph node features, a dynamic association graph is constructed using a dual connection strategy of temporal evolution and state similarity to explicitly model the spatiotemporal association characteristics of the coal mining machine's operating state, characterizing the wear or performance degradation trend of the equipment's mechanical components over operating time. First, temporal evolution edges are constructed to depict the continuous evolution law of the equipment state. The graph node set V is traversed in chronological order, and nodes at adjacent times are compared. and Establish directed temporal connections The directed edges form the temporal skeleton of the graph structure, ensuring that the anomaly detection model can capture the continuous evolution trend of the device state over time. To quantify the smoothness of state transitions between adjacent time steps, a Gaussian radial basis function kernel is used to calculate the weights of the temporal evolution edges. , in, For nodes and The temporal evolution edge weights between adjacent time steps range from (0,1). A larger weight value indicates a smoother transition of device state between adjacent time steps. and These are the low-dimensional semantic feature vectors of adjacent time nodes. The bandwidth parameter of the Gaussian kernel function controls the sensitivity of the similarity metric. In this embodiment, it is set through cross-validation. , The Euclidean distance norm of the vector is used to measure the geometric distance in the feature space. Through this Gaussian kernel function, time nodes that are close in physical time and have similar state characteristics are given higher weights, while abnormal jumps caused by sudden noise are given lower weights, thus effectively distinguishing between normal state evolution and abnormal disturbances.

[0038] Secondly, state similarity edges are constructed to capture the state correlation characteristics of the equipment under the same load conditions in different work cycles. The dynamic time warping algorithm is used to calculate the similarity between any two non-adjacent nodes. and The optimal alignment path distance between feature sequences. Considering that coal mining machines may exhibit similar cutting conditions or fault symptoms in different shifts, but may experience scaling or deformation on the time axis due to differences in production rhythm, the dynamic time warping algorithm eliminates the scaling or deformation differences on the time axis through a nonlinear bending alignment method. Its distance metric formula is: , in, For nodes and The dynamic time-normalized distance between them; This represents the optimal curved alignment path between feature sequences, i.e., the index pair sequence that satisfies boundary conditions, monotonicity, and continuity constraints; To align coordinate pairs on the path, representing nodes The p-th feature vector in the feature sequence and the node The alignment relationship of the q-th eigenvector in the feature sequence; For local distance metrics, Euclidean distance is used in this embodiment. When the calculated dynamic time-warped distance Time (Similarity threshold in this embodiment) ), at node and Establish undirected state similarity edges between them This edge will link nodes that are close in distance but far apart in physical time in the feature space across time and space, and realize topological clustering of historically similar working conditions such as collision hazard reproduction or bottom plate indentation and attitude oscillation.

[0039] Finally, a complete dynamic association graph is constructed based on the two types of edge sets mentioned above. ,in For a set of graph nodes, To integrate the temporal evolution edge set Set of edges similar to the state The total set of edges, Let be the edge weight matrix, and its elements are... For edges related to temporal evolution, the Gaussian kernel function is used; for edges related to state similarity, the inverse of the dynamic time-warped distance or the similarity transformation value is used; and for non-connected edges, zero is used. Through this dual-connection strategy, the dynamic association graph preserves the continuous physical process of the coal mining machine operation (through...). ), and through cross-temporal topological bridging (through By clustering historically recurring failure modes within a specific manifold region, the graph structure's ability to discriminate complex failure modes and its generalization performance are greatly enhanced.

[0040] S4. After completing the dynamic association diagram After construction, this embodiment first establishes a precise mapping relationship between graph nodes and the positions of the original time series. This ensures the traceability of anomaly detection results. Specifically, based on the timestamp records of the sliding window segmentation in step 2, each graph node... A unique mapping is established to the center time t or start time of the corresponding time window in the original multidimensional time series, forming a bidirectional index table. This mapping relationship constitutes the alignment benchmark between physical alarm events and abnormal nodes in the graph, providing spatiotemporal coordinate references for subsequent consistency verification.

[0041] To achieve graph structure anomaly detection, this embodiment employs a graph neural network model (specifically, a graph attention network GAT) for end-to-end learning of the dynamic association graph. The graph neural network model aggregates neighbor node information through a message passing mechanism, calculates anomaly probability scores for each node, and, to eliminate alarm drift between physical warning labels and graph anomaly predictions, introduces a weighted cross-entropy loss as a mapping loss, forcing the model to perform high-confidence regression on anomaly nodes mapped to the alarm time window. , in, For mapping loss, For a set of graph nodes, Let be the weight coefficient for node i. Higher weights are assigned to sample nodes marked as anomalies to address the class imbalance problem. This represents the actual anomaly label for node i, where 1 indicates an anomaly and 0 indicates normal operation. The loss function, which is the anomaly probability predicted by the graph neural network model, increases the regression penalty for abnormal samples to ensure that the graph neural network outputs a high confidence score for abnormal nodes during the actual alarm period, thereby suppressing the prediction label drift phenomenon commonly found in traditional deep learning.

[0042] Meanwhile, to effectively combat abnormal scoring label flickering caused by sensor noise interference, a temporal smoothing term is introduced as a temporal continuity loss, utilizing the established temporal evolution edge set in the graph structure. Constraint anomaly scores evolution continuity across adjacent time steps: , in, For time series smoothing, The total number of nodes. and These are the anomaly probability scores for nodes at adjacent time points. For an indicator function, if and only if the edge Belongs to the temporal evolution edge set When the value is 1, this smoothing term ensures that the anomaly detection results conform to the continuous change law of the physical motion of the equipment by penalizing the score jump between adjacent time nodes.

[0043] Furthermore, to enhance the model's ability to focus on features of heterogeneous and anomalous subgraphs, attention coefficients between nodes are calculated through the attention layer of the graph neural network model: , in, Let be the attention coefficient between node i and node j, representing the importance weight of neighbor node j to the central node i; The function is a normalized exponential function, ensuring that the sum of the attention coefficients of node i to all its neighbors is 1; This is a linear unit activation function with leakage correction (negative slope is usually set to 0.2) used to introduce nonlinearity; For attention vectors, The weight matrix is ​​a learnable matrix. Let be the low-dimensional semantic feature vectors of node i and node j, respectively. This represents a vector concatenation operation. Through this attention mechanism, the model automatically learns and enhances the weight allocation of heterogeneous subgraphs associated with abnormal patterns during training, suppressing interference from normal background nodes.

[0044] Finally, based on the mapping loss With time series smoothing term A multi-loss function joint constraint mechanism is constructed, which can optionally be combined with the main loss function of graph anomaly detection (such as reconstruction error based on graph autoencoder or discriminative loss based on contrastive learning). The parameters of the graph neural network model are jointly optimized and trained through the backpropagation algorithm. After training convergence, the model outputs the final anomaly score for each graph node. Combined with the mapping relationship M, the score results are mapped back to the original time axis to generate a standardized graph anomaly dataset with spatiotemporal consistency. This dataset includes anomaly node identifiers, corresponding original time window locations, anomaly probability scores, and consistency verification flags. The dataset is output to the coal mine monitoring terminal for early identification of potential equipment faults. This successfully realizes a deep semantic mapping from the original time series signal to the topological evolution of equipment state, providing reliable data support for accurate fault diagnosis and predictive maintenance of coal mine equipment.

[0045] Example 2 The difference between this embodiment and Embodiment 1 is that this embodiment provides a verification experiment for a method of structural conversion of coal mine equipment status data diagrams; like Figure 2 As shown, 1. Experimental data foundation and construction of multidimensional feature space: This embodiment uses a coal mining machine in a fully mechanized coal mining face as the verification object, and applies the aforementioned coal mine equipment status data diagram structured conversion method to a real industrial scenario to verify the feasibility and effectiveness of the method.

[0046] The data source in this embodiment covers the comprehensive operation logs of the coal mining machine body, the anti-collision early warning system, the support electro-hydraulic control system, and the working face video monitoring system. The data collection spans from August 1, 2025 to September 30, 2025, a total of 60 days of continuous production operation records, with an original time series length of 5,281,946 sampling points and a sampling frequency of 1Hz. In terms of data composition, the number of abnormal records is 4,333,019, accounting for 82.03%, exhibiting a typical unbalanced distribution characteristic of long-term steady-state operation and short-term abnormal sudden occurrences.

[0047] like Figure 3 , Figure 4 As shown, at the physical dimension level, this embodiment constructs a feature space containing 12 key state parameters, aiming to comprehensively reproduce the actual posture and stress conditions of the coal mining machine underground. The 12 parameters are specifically shown in Table 1: Table 1. Parameter information for different dimensions; 2. Construction of the data preprocessing pipeline; To address data packet loss caused by complex electromagnetic interference downhole, extreme value drift due to sensor vibration, and discontinuities during equipment maintenance in the raw data, this embodiment designed and implemented a rigorous preprocessing pipeline. First, a linear interpolation algorithm is used to fill in short-term missing data, ensuring the continuity of the time axis. Second, the 3σ criterion is used to remove outliers that significantly deviate from physical laws; that is, the mean and standard deviation of the data are calculated, and data points deviating from the mean by more than three times the standard deviation are identified as outliers and corrected. Finally, Z-score standardization is applied to data of different dimensions, uniformly mapping them to a standard normal distribution space. This eliminates the impact of dimensional differences on subsequent graph structure weight calculations, laying a high-quality data foundation for the conversion of time-series data to graph structures.

[0048] 3. Transformation of temporal features into graph nodes based on sliding window; This embodiment employs an adaptive sliding window technique to slice the cleaned data points. The sliding window size is set to 1024 time steps, and the sliding step size is set to 512 time steps, ensuring 50% information overlap between adjacent windows and effectively preserving the transitional information of state evolution. Through this transformation mechanism, the originally linear time series is reconstructed into a node set consisting of 10,316 independent units.

[0049] Each graph node is no longer a single scalar value, but a high-dimensional feature matrix encapsulating the 12-dimensional state parameters of the coal mining machine. A time-aware subsequence extraction algorithm is used to extract node features. This algorithm not only calculates the statistical characteristics of each parameter within the window, such as mean, variance, kurtosis, and skewness, but also uses Shapelet transform to extract local morphological features of the current waveform, including the start-up impact waveform and the jamming waveform. Therefore, each of the graph nodes becomes an entity rich in semantic information, enabling subsequent anomaly detection tasks to be transformed into node classification or link prediction tasks in graph theory.

[0050] 4. Topology construction and edge generation of dynamic spatiotemporal relational graph structures; Building upon node generation, this embodiment further implements an edge construction process, aiming to explicitly model the time dependency and working condition similarity of the coal mining machine's operating state through the graph's topological structure. The edge construction follows the design principles of dynamic graphs, primarily involving the generation of two types of connection relationships: The first type is time-dependent edges. Based on the chronological order of physical time, the system automatically establishes directed connections between adjacent time window nodes. These connections form the backbone of the graph structure, ensuring that the anomaly detection model can capture the continuous evolution trend of the device state over time.

[0051] The second type is state similarity edges. The system uses the dynamic time warping algorithm and Pearson correlation coefficient to calculate the distance between any two nodes in the feature space. When the current node state is highly similar to the node state at a certain historical moment, and the similarity exceeds a set threshold of 0.75, a time-spanning connection edge is established between the two. This design enables the graph structure to capture the periodic patterns of the coal mining machine and the recurrence of specific fault modes; 5. Implementation of graph mapping and consistency labeling for abnormal data; This embodiment aims to accurately map collision anomalies in the physical world into topological anomaly representations in graph space, thereby constructing a high-confidence anomaly diagnostic benchmark. For sudden collision hazards recorded in on-site collision avoidance systems, this embodiment constructs a standardized dataset containing anomaly subgraphs by spatiotemporally aligning alarm event time periods with graph node sequences. During the construction process, a topological heterogeneity feature enhancement strategy is used to make anomaly nodes exhibit a break in similarity to normal baseline nodes in the graph structure, thus highlighting fault characteristics at the topological level.

[0052] To further eliminate the alarm drift phenomenon commonly found in traditional deep learning, this embodiment introduces a joint constraint using a mapping loss function and a temporal continuity loss function, forcing the abnormal score distribution output by the graph neural network to maintain strict consistency with the physical alarm time window. Through this mechanism, the implicit nonlinear coupling relationship of device operation is made explicit into a structured graph topology.

[0053] By combining the two types of edge-forming rules described above, this embodiment ultimately constructs a dynamic directed graph. Specific statistical information of the dynamic directed graph is shown in Table 2: Table 2. Data information after the coal mining machine is constructed; Example 3 This embodiment provides a modified design of a structured conversion method for coal mine equipment status data diagrams, aiming to verify the generalization ability and universality of the method across equipment, fields, and changing working conditions.

[0054] This embodiment fully inherits the underlying core logic of data preprocessing, dynamic graph construction, and consistency verification. However, it adaptively reconstructs the input data modes, feature extraction mechanisms, and graph semantic mapping rules to address the unique high-frequency vibration physical properties of rotating machinery, thereby overcoming the challenge of accurately identifying subtle faults in complex rotating machinery under varying loads and strong background noise environments.

[0055] This embodiment aims to achieve time-series graph-based modeling of monitoring data for rolling bearings, a core component of mechanical equipment. It strives to overcome the limitations of traditional one-dimensional signal analysis and significantly improve the accuracy, robustness, and domain adaptability of cross-operating-condition fault diagnosis. To achieve these goals, the key technical problems addressed in this embodiment are consistent with the main solution and specifically mapped to the following three aspects: First, adaptive periodic segmentation and graph node feature vectorization of high-frequency vibration time-series data; second, dynamic graph topology construction based on cross-domain signal feature similarity and local temporal dependence; and third, the construction of a consistency verification model between graph-based fault classification decisions and original time-series physical labels.

[0056] To eliminate interference from multiple classifications and accurately verify whether the proposed time-series graph transformation method can extract a consistent underlying topological manifold across completely different physical test benches, this embodiment introduces two authoritative benchmark datasets recognized by industry and academia: the Case Western Reserve University Bearing Dataset and the Society for Mechanical Failure Prevention Challenge Dataset. The initial construction phase of this embodiment focuses only on one of the most representative basic fault types: bearing outer ring failure and its corresponding normal baseline state. The outer ring is typically fixed to the bearing housing, bearing static loads, and the impact energy transfer path caused by its failure is direct, making it the best benchmark for verifying the topological consistency of the graph structure across datasets.

[0057] To visually demonstrate the data heterogeneity and scale processed in the initial binary classification cross-domain validation stage of this embodiment, and to closely reflect the real monitoring characteristics of long-term steady-state operation and short-term sudden failures of industrial equipment, this embodiment breaks with the traditional presupposition of balanced samples in the laboratory and constructs an imbalanced dataset with an anomaly rate of only 10%. This dataset consists of a large-scale normal baseline node sequence and a small number of discrete failure mode nodes. This setting significantly increases the difficulty of topological extraction of anomaly patterns and can effectively verify the weak anomaly capture capability of this method when facing severe class imbalance. Table 3 details the composition information of the original time-series datasets of CWRU and MFPT: Table 3 Bearing timing data; like Figure 5 , Figure 6 , Figure 7 , Figure 8 As shown, the first 1000 time periods of the original time series data are visualized. For the two types of high-frequency non-stationary time series with different sources and varying hardware acquisition systems, this embodiment first establishes an adaptive cross-domain data preprocessing system. First, the acquired raw vibration acceleration data is cleaned and standardized. Since the vibration amplitude varies greatly under different load conditions, directly inputting it into the model would lead to gradient vanishing or exploding. Therefore, the Z-Score standardization method is used to calculate the mean and standard deviation of the data for each load condition, mapping the raw signal to a standard normal distribution and eliminating the influence of dimensions. For any outliers or acquisition discontinuities that may exist in the data, linear interpolation is used to fill in small fragments, ensuring the continuity of the time axis.

[0058] To accommodate the physical differences in sampling frequencies across different datasets, this embodiment abandons the traditional fixed window length approach and instead implements a frequency-adaptive overlapping sliding window segmentation strategy. Considering that the core of bearing fault signals lies in periodic impacts, the window length setting must ensure that each window effectively covers the complete cycle and characteristic multiples of bearing rotation under different sampling environments. Therefore, this embodiment dynamically adjusts the sliding window and step size parameters based on the original sampling frequency of the datasets. By aligning the physical duration, it ensures that different datasets have similar physical meanings after segmentation. Specific parameter settings are shown in Table 4. Table 4. Dataset processing parameter settings; Next, we perform in-depth feature extraction and vectorization of nodes. The original waveform of the vibration signal is difficult to directly use for graph node representation. This embodiment introduces a hybrid domain feature extraction algorithm. For each segmented time window, i.e., the future graph node, we extract its multi-dimensional feature vector: In the time domain, we calculate the root mean square value to reflect the vibration energy, calculate the kurtosis and margin factor to capture the pulse impact characteristics of early faults, and calculate the skewness to describe the asymmetry of the waveform; In the frequency domain, we extract the centroid frequency, mean square frequency, and frequency variance through fast Fourier transform to identify the energy distribution of fault characteristic frequencies; Through the above processing, each original vibration window is transformed into a vector containing 29-dimensional features. This vector serves as the attribute feature of the node in the dynamic graph, completing the materialization process from a one-dimensional time segment to a high-dimensional graph node.

[0059] After generating a series of graph nodes rich in physical meaning, the core of this embodiment lies in how to define the connection relationships between the nodes, that is, to construct the equipment state association graph. In the bearing dataset, samples are usually considered independent and identically distributed, lacking explicit spatial topology. This embodiment implicitly constructs a dynamic graph structure by mining the inherent temporal evolution patterns and state similarities of the data, thereby connecting the scattered samples into an organic whole.

[0060] First, temporal evolution edges are constructed. The performance degradation or failure of a bearing is a continuous physical process, and the signal characteristics of adjacent time windows exhibit strong Markov properties. Therefore, this embodiment establishes directed connections between adjacent graph nodes based on the chronological order of physical time, i.e., a preceding node points to a succeeding node. The weight of the temporal evolution edge can be determined by calculating the cosine similarity of the feature vectors of two nodes, reflecting the smoothness of the state transition. This construction of temporal edges enables the graph neural network to integrate state information from historical moments when performing feature aggregation, thereby effectively suppressing the interference of instantaneous noise and capturing the gradual degradation trend of faults from their initial stages to severe deterioration.

[0061] Secondly, feature similarity edges, also known as k-nearest neighbor edges, are constructed. This is the key to solving the problem of fault diagnosis under varying operating conditions in this embodiment. Although the vibration responses of faults in the dataset under different loads have different amplitudes, their frequency domain structures or wavelet energy distributions have inherent similarities. This embodiment uses Euclidean distance to measure the mutual distance of all nodes in the feature space. For any node in the graph, the k neighbor nodes with the most similar features are found globally, where k is 5, and undirected connecting edges are established between them. This design breaks the linear constraint of time, clustering similar fault patterns scattered at different times and under different operating conditions together, realizing cross-time and cross-operating condition connections, and making the graph structure successfully explicit in the feature space, greatly enhancing the model's ability to generalize and identify similar faults.

[0062] After constructing the temporal evolution edges and feature similarity edges, this embodiment successfully transforms discrete vibration time-series signals into a dynamic graph anomaly detection dataset with a well-defined topological structure. This dataset not only preserves the physical evolution trajectory of the original vibration signals but also enhances the generalization representation across operating conditions through manifold aggregation in high-dimensional space. Table 5 details the statistical information of the graph structure constructed in this embodiment on two benchmark datasets: Table 6 shows the data for constructing the dynamic graph. The above results demonstrate that this embodiment successfully converted rotating machinery bearing vibration data from time-series signals to a graph structure, and the constructed graph structure exhibits excellent topological properties. This verification fully proves the applicability and robustness of the proposed method in the field of rotating machinery fault diagnosis, as well as its strong generalization ability across different physical test benches.

[0063] Example 4 This embodiment provides a system for the structured conversion of coal mine equipment status data diagrams, including: The data acquisition module is configured to acquire multi-dimensional time-series state data collected by multi-source sensors of coal mine equipment; The data preprocessing module is configured to preprocess multidimensional time series state data, including missing value imputation, outlier detection and correction, and dimensional normalization, and to construct a standardized multidimensional state perception matrix. The temporal segmentation and node mapping module is configured to: use an adaptive sliding window to perform temporal segmentation on the multidimensional state perception matrix, calculate the information entropy of each window and filter high-value time windows based on a preset threshold, perform dimensionality reduction encoding on high-value time windows through principal component analysis, and generate graph nodes with low-dimensional semantic feature vectors. The dynamic graph construction module is configured to: construct a dynamic association graph based on graph nodes using a dual connection strategy of temporal evolution and state similarity, including establishing temporal evolution edges based on Gaussian kernel function measurement of adjacent node features, and establishing state similarity edges based on dynamic time warping algorithm to measure feature distance between nodes. The consistency verification module is configured to: establish a mapping relationship between graph nodes and the original time series positions, and perform consistency verification on the dynamic association graph based on a joint constraint mechanism of multiple loss functions; The graph anomaly dataset generation module is configured to generate standardized graph anomaly datasets.

[0064] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for structured conversion of coal mine equipment status data diagrams, characterized in that, include: Acquire multidimensional time-series state data from sensors of coal mine equipment and preprocess them to construct a standardized multidimensional state perception matrix; An adaptive sliding window method is used to perform temporal segmentation of the state-aware matrix, calculate the information entropy of each window, and filter high-value time windows based on a preset threshold. Principal component analysis is used to perform dimensionality reduction encoding on the high-value time windows to generate graph nodes with low-dimensional semantic feature vectors. A dynamic association graph is constructed based on the graph nodes using a dual connection strategy of temporal evolution and state similarity. This includes establishing temporal evolution edges based on Gaussian kernel function measurement of adjacent node features, and establishing state similarity edges based on dynamic time warping algorithm to measure feature distance between nodes. A mapping relationship between graph nodes and the original time series positions is established. The consistency of the dynamic association graph is verified based on a joint constraint mechanism of multiple loss functions. A standardized graph anomaly dataset is generated and output to the coal mine monitoring terminal for early identification of potential equipment failures.

2. The method for structured conversion of coal mine equipment status data diagrams according to claim 1, characterized in that, The construction of a standardized multidimensional state-aware matrix includes smoothing missing breakpoints in multidimensional time-series state data using cubic spline interpolation, ensuring the continuity of the first and second derivatives of the sequence at connection points by constructing piecewise cubic polynomials, and applying a sliding window method to the imputed data. The criteria perform dynamic noise suppression, calculate the mean and standard deviation of the data within the window and correct them in real time through local median filtering, and perform Z-Score transformation on the corrected data to normalize the dimensions in order to construct the standardized multidimensional state perception matrix.

3. The method for structured conversion of coal mine equipment status data diagrams according to claim 1, characterized in that, The calculation of the information entropy of each window and the filtering of high-value time windows based on a preset threshold include setting the sliding window length and sliding step size, slicing the multidimensional state perception matrix along the time dimension, and generating a continuous set of time windows. Calculate the probability density distribution of state parameters within each time window, and calculate the information complexity of each window using the Shannon information entropy formula. The information complexity calculation formula is as follows: ， Set a significance threshold , will satisfy The time window for the condition is selected as the high-value time window, and redundant windows with information entropy below the significance threshold are removed; in, For the first k Time window Information entropy Let be the probability density function value of the j-th dimension state parameter within the window. The mean of the information entropy of all windows. This is the preset adjustment coefficient.

4. The method for structured conversion of coal mine equipment status data diagrams according to claim 3, characterized in that, The process of generating graph nodes with low-dimensional semantic feature vectors includes extracting multi-dimensional state features within the high-value time window, constructing a high-dimensional feature matrix X, and calculating the covariance matrix of the high-dimensional feature matrix. ; The covariance matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the k largest eigenvalues ​​are selected to form the projection matrix. The original high-dimensional feature vector Linearly projected onto a low-dimensional semantic space, the resulting low-dimensional semantic feature vector is used as the core attribute of the graph node: ; Where C is the covariance matrix, Let be the feature vector of the i-th sample. The sample mean. The number of filtered windows. Let be the low-dimensional semantic feature vector of the i-th graph node. Let be the transpose of the projection matrix, and k be the target dimension after dimensionality reduction.

5. A method for structured conversion of coal mine equipment status data diagrams according to claim 1, characterized in that, The Gaussian kernel function metric based on the characteristics of adjacent nodes is used to establish temporal evolution edges, including traversing the graph nodes in chronological order and performing operations on nodes at adjacent times. and A directed temporal connection is established, and the weights of the temporal evolution edges are calculated using a Gaussian kernel function to capture the smooth evolution trend of device state transitions. The formula for calculating the weights of the temporal evolution edges is as follows: ， in, For nodes and The temporal evolution edge weights between them and These are the low-dimensional semantic feature vectors of adjacent time nodes. The bandwidth parameter of the Gaussian kernel function. Let be the Euclidean distance norm of the vector.

6. The method for structured conversion of coal mine equipment status data diagrams according to claim 5, characterized in that, The method of measuring feature distances between nodes and establishing state similarity edges based on the dynamic time warping algorithm includes using the dynamic time warping algorithm to calculate the feature distance between any two non-adjacent nodes. and Optimal alignment path distance between feature sequences: ; When dynamic time warping distance At the node and Undirected state similarity edges are established between them, and cross-temporal and spatial associations are made between nodes that are close to each other in the feature space through state similarity edges; in, For nodes and The dynamic time-warped distance between them This represents the optimal curved alignment path between feature sequences. To align coordinate pairs on the path, It is a local distance metric function. and They are nodes and The p-th and q-th eigenvectors in the feature sequence, This is the preset similarity threshold.

7. A method for structured conversion of coal mine equipment status data diagrams according to claim 6, characterized in that, The method employs a dual-connection strategy of temporal evolution and state similarity to construct a dynamic association graph. This includes building a temporal-dimensional skeleton of the graph structure based on the temporal evolution edges, constructing cross-temporal and spatiotemporal topological connections based on the state similarity edges, and fusing the two types of edge sets to construct the dynamic association graph. The dual-connection strategy explicitly transforms the time-dependent and state-similar characteristics of device operation into a graph topology. in, For a set of graph nodes, Let be the set of edges. For the temporal evolution edge set, Let be the set of edges with similar states. It is an edge weight matrix that includes the edge weights of temporal evolution and the edge weights of state similarity.

8. A method for structured conversion of coal mine equipment status data diagrams according to claim 1, characterized in that, The consistency verification of the dynamic association graph based on the joint constraint mechanism of multiple loss functions includes using a graph neural network model to detect anomalies in the dynamic association graph, introducing weighted cross-entropy loss as a mapping loss, forcing the graph neural network model to perform high-confidence regression on the abnormal nodes mapped to the alarm time window, introducing a time-series smoothing term as a time-series continuity loss, and constructing a joint constraint mechanism of multiple loss functions based on the mapping loss and the time-series smoothing term. The expression of the mapping loss is as follows: ， in, For mapping loss, For a set of graph nodes, Let be the weight coefficient of node i. For the true anomaly label of node i, This represents the anomaly probability predicted by the graph neural network model.

9. A method for structured conversion of coal mine equipment status data diagrams according to claim 8, characterized in that, The process of generating a standardized graph anomaly dataset includes calculating attention coefficients between nodes through the attention layer of a graph neural network model, weighting and aggregating graph node features based on these attention coefficients, and jointly optimizing and training the parameters of the graph neural network model by combining mapping loss and temporal smoothing terms. The final result is a standardized graph anomaly dataset with spatiotemporal consistency. The formula for calculating the attention coefficients is as follows: ， in, Let be the attention coefficient between node i and node j. For normalized exponential functions, For the activation function of the linear unit with leakage correction, For attention vectors, The weight matrix is ​​a learnable matrix. and Let be the low-dimensional semantic feature vectors of node i and node j, respectively. This is a vector concatenation operation.

10. A system for converting coal mine equipment status data into a structured graph, comprising the method described in claim 1, characterized in that, include: The data acquisition module is configured to acquire multi-dimensional time-series state data collected by multi-source sensors of coal mine equipment; The data preprocessing module is configured to preprocess multidimensional time series state data, including missing value imputation, outlier detection and correction, and dimensional normalization, and to construct a standardized multidimensional state perception matrix. The temporal segmentation and node mapping module is configured to: use an adaptive sliding window to perform temporal segmentation on the multidimensional state perception matrix, calculate the information entropy of each window and filter high-value time windows based on a preset threshold, perform dimensionality reduction encoding on high-value time windows through principal component analysis, and generate graph nodes with low-dimensional semantic feature vectors. The dynamic graph construction module is configured to: construct a dynamic association graph based on graph nodes using a dual connection strategy of temporal evolution and state similarity, including establishing temporal evolution edges based on Gaussian kernel function measurement of adjacent node features, and establishing state similarity edges based on dynamic time warping algorithm to measure feature distance between nodes. The consistency verification module is configured to: establish a mapping relationship between graph nodes and the original time series positions, and perform consistency verification on the dynamic association graph based on a joint constraint mechanism of multiple loss functions; The graph anomaly dataset generation module is configured to generate a standardized graph anomaly dataset and output it to the coal mine monitoring terminal for early identification of potential equipment failures.

Images