A fault detection method for heavy medium coal separation process based on dynamic and static mixing graph

Abstract

The present application belongs to the technical field of industrial process monitoring and fault diagnosis, and proposes a heavy medium coal preparation process fault detection method based on dynamic and static mixed graph, which can be called as HG-STAN based on dynamic and static mixed graph. The method comprises: collecting key variable data of heavy medium coal preparation process and preprocessing; constructing a static graph based on process prior knowledge, and learning dynamic association from data by using self-attention mechanism to form a dynamic and static mixed graph; constructing a time-space synchronous graph attention autoencoder (STGAAE) to synchronously aggregate the spatial neighborhood information and historical time information of the nodes; training the model by using a combined loss function containing reconstruction loss and graph structure sparsity regularization; establishing a multi-level monitoring system of data level statistics and graph level statistics; determining the control limit of each statistic by kernel density estimation to realize online fault detection. The present application significantly improves the accuracy and sensitivity of heavy medium coal preparation process fault detection.

Description

Technical Field

[0001] This invention relates to the field of industrial process monitoring and fault diagnosis technology, specifically a fault detection method for heavy medium coal preparation process based on dynamic and static hybrid diagrams. Background Technology

[0002] Heavy media coal preparation technology is widely used due to its high separation accuracy. Its system consists of multiple key pieces of equipment physically connected together, such as a mixing tank, heavy media cyclone separator, and arc screen. The process is highly nonlinear and dynamically coupled, posing a severe challenge to the accuracy and timeliness of fault detection. Existing fault detection methods have the following limitations: 1. Limitations of linear models and isolated modeling: Traditional linear methods, such as Principal Component Analysis (PCA) and Partial Least Squares (PLS), are difficult to characterize the nonlinear features of the heavy media coal preparation process. While deep learning methods such as autoencoders (AE) and Long Short-Term Memory (LSTM) networks improve the ability to extract nonlinear features, they typically treat sensor variables as isolated time-series inputs, ignoring the topological relationships between variables determined by the process flow. This results in a lack of physical interpretability in the model and makes it difficult to locate fault propagation paths.

[0003] 2. The limited scope of graph structure construction: While graph neural networks (GNNs) can explicitly model relationships between variables, in heavy media coal preparation scenarios, static graphs built solely based on prior knowledge struggle to adapt to dynamic changes such as operating condition drift. Dynamic graphs built solely based on data correlations are prone to generating spurious associations and losing physical meaning in noisy environments. Existing methods cannot simultaneously guarantee both the physical interpretability and adaptability of graph structures to operating conditions.

[0004] 3. The fragmentation of spatiotemporal feature extraction: Existing spatiotemporal graph neural networks mostly use a serial (space first, then time or time first, then space) architecture to process spatiotemporal data, which artificially severs the inherent synchronicity between spatial interaction and temporal evolution in the heavy medium coal preparation process, making it difficult to effectively capture complex spatiotemporal coupling dependencies.

[0005] 4. Incomplete design of monitoring indicators: Existing methods mainly rely on data-level statistics such as T² and SPE, failing to fully utilize the characteristic that process failures are often accompanied by changes in system topology, lacking monitoring of graph-level structural anomalies, resulting in insufficient sensitivity to detect structural failures.

[0006] Therefore, there is an urgent need for a fault detection method that can integrate physical knowledge, dynamically adapt to working conditions, synchronously model spatiotemporal characteristics, and provide multi-level monitoring indicators to improve the reliability and safety of the heavy medium coal preparation process. Summary of the Invention

[0007] The purpose of this invention is to provide a fault detection method for heavy medium coal preparation process based on dynamic and static hybrid diagrams, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a fault detection method for heavy medium coal preparation process based on dynamic and static hybrid graphs, hereinafter referred to as HG-STAN (Hybrid Graph based Spatial-Temporal Synchronous Attention Network), whose core model is a spatiotemporal synchronous graph attention autoencoder (STGAAE).

[0009] The method includes the following steps: S1, Data preprocessing and construction of a hybrid dynamic-static graph: For the heavy media coal preparation process, process monitoring variable data and prior knowledge are collected, and the process monitoring variable data is preprocessed to obtain a standardized time series data for each process variable; dynamic time series data is generated through a fixed-length sliding window to form standardized dynamic input time series data; the adjacency matrix of the static graph is defined using prior knowledge, and the preprocessed time series data is input into a spatial self-attention mechanism. The attention weights between variables are obtained by calculating the query matrix and the key matrix, which constitute the adjacency matrix of the dynamic graph. The static graph and the dynamic graph are fused to obtain a hybrid graph, which serves as the topological basis for subsequent feature extraction.

[0010] S2. Spatiotemporal Synchronous Graph Attention Autoencoder Modeling and Training: A spatiotemporal synchronous graph attention autoencoder (STGAAE) is established. The STGAAE includes a spatiotemporal synchronous graph encoder and a spatiotemporal synchronous graph decoder, both of which are composed of stacked multi-layer spatiotemporal synchronous graph attention modules. Using the hybrid graph adjacency matrix and preprocessed time-series data as model input, during node update, a masked lower triangular matrix and a time feature extraction matrix are introduced to achieve synchronous aggregation and feature extraction of spatial neighborhood information and historical time information. The input data of the heavy medium coal preparation process is reconstructed. A combined loss function is constructed based on the data reconstruction error and graph structure sparsity regularization term to train the spatiotemporal synchronous graph attention autoencoder. This forces the model to learn the key features of the process data under normal operating conditions of the heavy medium coal preparation process and constrains the dynamic graph structure to avoid spurious associations, thus obtaining a well-trained monitoring model.

[0011] S3. Construction of Multi-Level Monitoring Statistics and Offline Modeling: Based on the trained model, data-level statistics and graph-level statistics are constructed respectively. The data-level statistics include T² statistics in the feature space and SPE statistics in the residual space. The graph-level statistics include macroscopic topological statistics based on the eigenvalues ​​of the Laplacian matrix. and microstructure statistics based on the number of edges in a dynamic graph The control limits for each statistic are determined by kernel density estimation.

[0012] S4. Online Fault Detection: Real-time acquisition of process data and preprocessing in the same way as the offline stage. The processed data is then input into the trained model to calculate the values ​​of four monitoring statistics in real time. If the value of any statistic exceeds its corresponding control limit, the process is determined to have failed.

[0013] As a further aspect of the present invention: In step S1, the data preprocessing includes outlier removal, missing value imputation, and maximum-minimum standardization. The spatiotemporal process data of all process monitoring variables, after data preprocessing, are in the form of… X represents the training data consisting of N normal samples, and T represents the matrix transpose. Let X represent the t-th normal sample; let the i-th column of X be denoted as . , This represents the time-series data vector of the i-th process variable. This represents the measurement data of the i-th process variable at time t; by reorganizing the data through a sliding window, the time series data matrix input to the model at time t is obtained. , where w represents the length of the sliding window.

[0014] As a further aspect of the present invention: in step S1, the structure of the heavy medium coal preparation process diagram is defined as a directed unweighted graph. ,in It is a set of nodes. It is a directed edge set. Represents the adjacency matrix; nodes It is the i-th variable in the heavy media coal preparation process, and its node attribute is a standardized time-series data vector. Directed edge Representing variables and variables The causal dependency between them.

[0015] As a further aspect of the present invention: in step S1, the construction of the dynamic-static hybrid graph structure specifically includes: defining the adjacency matrix of the static graph based on prior knowledge. The element is 1 when there is a physical connection or causal relationship between nodes, and 0 otherwise; a self-attention mechanism is used to extract data from the time-series data matrix. Inferring the adjacency matrix of a dynamic graph : ; where the query matrix Key-value matrix , and These are two trainable weight matrices for the self-attention mechanism. The sigmoid activation function is used; the static and dynamic graphs are superimposed and merged to obtain the adjacency matrix of the hybrid graph. : ;in, Indicates taking and Performing a bitwise OR operation on the edges of two adjacency matrices yields the final result. This serves as the topological basis for subsequent spatiotemporal synchronization feature extraction.

[0016] As a further aspect of the present invention: In step S2, the spatiotemporal synchronized graph attention autoencoder (STGAAE) is composed of a spatiotemporal synchronized graph attention mechanism embedded in a graph encoder-graph decoder architecture. To achieve efficient matrix operations and avoid future information leakage, spatial attention weights are first calculated for hybrid graphs. any node at time With nodes The input data are respectively and Synchronous attention coefficient Subsequently, by analyzing the nodes All neighbors The attention scores are standardized to obtain normalized attention coefficients. Then, by combining the learnable temporal feature extraction matrix and the masking lower triangular matrix, which are applied to the input data, a spatiotemporal aggregation matrix expression is obtained: ;in The adjacency matrix of the hybrid graph. The identity matrix is ​​used to add self-loops to the adjacency matrix of the hybrid graph. for The input data corresponding to the layer, For extracting time features, To mask the lower triangular matrix and to better extract data features, a matrix is ​​used. The updated nodes undergo dimensional transformation, and then nonlinear operations are introduced through activation functions to enhance the ability to extract complex high-dimensional features.

[0017] As a further aspect of the present invention: In step S2, the spatiotemporal synchronization graph encoder is composed of multiple layers of spatiotemporal synchronization graph attention modules stacked together, and its function is to process the input raw spatiotemporal data. Mapped to a representation rich in spatiotemporal dependencies ,in For dimension; the spatiotemporal synchronization graph decoder, as the inverse process of the spatiotemporal synchronization graph encoder, aims to obtain from the representation The original input data is reconstructed from the data. .

[0018] As a further aspect of the present invention: in step S2, the combined loss function Reconstruction loss Sparse regularization loss of graph structure Weighted composition: Among them, the reconstruction loss is used to measure the output of the spatiotemporal synchronization graph decoder. With the original input The differences between them force the model to learn the key features of normal operating condition data, and the mean squared error is used as the reconstruction loss: To prevent the dynamic graph from becoming overly complex and introducing spurious associations, a graph structure sparsity regularization loss is added. This penalizes excessively large sums of edge weights in the dynamic adjacency matrix to encourage its sparsity, consistent with the prior knowledge that associations in real industrial processes are typically sparse. L1 norm of a matrix: ; It is a hyperparameter used to balance the importance between reconstruction accuracy and graph sparsity. Represents a training batch. This refers to the batch size.

[0019] As a further aspect of the present invention: In step S3, offline modeling includes: training a spatiotemporal synchronous graph attention autoencoder STGAAE based on normal training samples, and constructing monitoring statistics at the data level and graph level. The monitoring statistics at the data level include constructing traditional... and the SPE statistic, where the feature space statistic at time t is... The calculation formula is: ;in and and represent the average of the feature vector output by the final hidden layer of the spatiotemporal synchronization graph autoencoder at time t and the feature vector obtained during the offline modeling stage, respectively. It is the covariance matrix obtained in the final hidden layer offline modeling stage; the residual space statistics at time t. The calculation formula is: ;in for The original input vector at time step 1. This is the reconstruction output of the spatiotemporal synchronized graph decoder; the graph-level monitoring statistics include Laplace spectrum monitoring statistics constructed from both macroscopic topology and microscopic structure levels. and dynamic graph structure monitoring statistics For each moment dynamic graph structure First, calculate its corresponding degree matrix. With Laplace matrix : Subsequently, the Laplace matrix was... Perform eigenvalue decomposition to obtain a set of eigenvalues. The variance of these eigenvalues ​​is calculated as a comprehensive macroscopic statistic. : For each moment dynamic graph adjacency matrix , elements in Quantified nodes With nodes At any moment Dynamic association. All The summation yields the statistical dynamic graph adjacency matrix. Statistics on the number of all edges : Four statistical measures SPE and control limits , , and All of these methods are obtained by estimating the kernel density of the corresponding statistic sequence of normal training data at a given significance level.

[0020] An electronic device, characterized in that it includes a processor and a memory, the memory storing program code, which, when executed by the processor, causes the processor to perform the steps of the fault detection method for heavy medium coal preparation process based on dynamic-static hybrid diagram as described above.

[0021] A computer-readable storage medium, characterized in that it stores a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the fault detection method for heavy media coal preparation process based on dynamic and static mixing diagrams as described above.

[0022] Compared with the prior art, the beneficial effects of the present invention are: 1. It takes into account both physical interpretability and working condition adaptability: By constructing a hybrid dynamic and static graph, it integrates static correlations that reflect physical mechanisms, ensuring the interpretability and physical rationality of the model; and introduces dynamic correlations learned from data, enabling the model to adapt to dynamic changes such as working condition drift and raw material fluctuations in the production process, overcoming the limitations of a single graph construction mode.

[0023] 2. Efficient and Synchronous Spatiotemporal Feature Extraction: The proposed spatiotemporal synchronous graph attention mechanism can synchronously aggregate the spatial neighbor information and historical time information of nodes in the same computational step, avoiding the fragmentation of spatiotemporal coupled information in the traditional serial processing architecture. This synchronous modeling method is more in line with the essential characteristics of the spatiotemporal evolution of the heavy media coal preparation process, improving the efficiency and effectiveness of feature extraction.

[0024] 3. Multi-level and comprehensive monitoring: Innovatively, graph-level monitoring statistics (L-Spectral and G-Edge) are introduced to complement data-level statistics (T² and SPE). This dual-perspective monitoring system can simultaneously detect numerical anomalies and topological anomalies, significantly improving the sensitivity and accuracy of detecting minor faults, slowly changing faults, and structural faults.

[0025] 4. Strong robustness: By introducing L1 norm penalty into the graph structure sparsity regularization loss, the dynamic graph is constrained to maintain sparsity, which effectively avoids learning false associations in a noisy environment and improves the robustness of the model. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the network architecture of a spatiotemporal synchronous attention network based on a hybrid dynamic-static graph.

[0027] Figure 2 This is a schematic diagram illustrating the principle of the attention mechanism in spatiotemporal synchronization graphs.

[0028] Figure 3 This is a process flow diagram of the heavy media coal preparation process.

[0029] Figure 4 The following are T² statistics for different methods in the heavy media coal preparation process under fault 4: (a) T² statistics for the AE method under fault 4, (b) T² statistics for the LSTM-AE method under fault 4, (c) T² statistics for the GDAE method under fault 4, and (d) T² statistics for the HG-STAN method under fault 4. The horizontal axis represents "samples / unit," and the vertical axis represents "statistical score / point."

[0030] Figure 5 The graphs show the SPE statistics for fault 4 in the heavy media coal preparation process using different methods. (a) is the SPE statistics for the AE method under fault 4, (b) is the SPE statistics for the LSTM-AE method under fault 4, (c) is the SPE statistics for the KDGCN method under fault 4, (d) is the SPE statistics for the GDAE method under fault 4, (e) is the SPE statistics for the KG-GCBiGCN method under fault 4, and (f) is the SPE statistics for the HG-STAN method under fault 4. The horizontal axis represents "samples / unit", and the vertical axis represents "statistical score / point".

[0031] Figure 6This is a graphical monitoring statistics chart of fault 4 in the HG-STAN heavy media coal preparation process of the present invention. (a) is a chart showing the change in Laplace spectrum monitoring statistics, with the dashed line representing the control limits. (b) is a chart showing the change in dynamic graphical structure monitoring statistics, with the dashed line representing the control limits. The horizontal axis in the chart is in units of "samples / unit", and the vertical axis is in units of "statistical score / point".

[0032] Figure 7 The diagrams show a comparison of the mixing structure before and after a failure in the heavy medium coal preparation process. (a) and (b) are the mixing structure at two different times before the failure, and (c) and (d) are the mixing structure at two different times after the failure. Detailed Implementation

[0033] The technical solution of this application will be further described in detail below with reference to specific embodiments.

[0034] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0035] Please see Figure 1 In one embodiment of the present invention, a fault detection method for a heavy medium coal preparation process based on a dynamic-static hybrid graph includes the following steps: S1, data preprocessing and construction of a dynamic-static hybrid graph: For the heavy medium coal preparation process, process monitoring variable data and prior knowledge are collected, and the process monitoring variable data is preprocessed to obtain a standardized time series data for each process variable; dynamic time series data is generated through a fixed-length sliding window to form standardized dynamic input time series data; the adjacency matrix of the static graph is defined using prior knowledge, the preprocessed time series data is input into a spatial self-attention mechanism, and the attention weights between variables are obtained by calculating the query matrix and the key matrix to form the adjacency matrix of the dynamic graph; the static graph and the dynamic graph are fused to obtain a hybrid graph, which serves as the topological basis for subsequent feature extraction.

[0036] S2. Spatiotemporal Synchronous Graph Attention Autoencoder Modeling and Training: A spatiotemporal synchronous graph attention autoencoder (STGAAE) is established. The STGAAE includes a spatiotemporal synchronous graph encoder and a spatiotemporal synchronous graph decoder, both of which are composed of stacked multi-layer spatiotemporal synchronous graph attention modules. Using the hybrid graph adjacency matrix and preprocessed time-series data as model input, during node update, a masked lower triangular matrix and a time feature extraction matrix are introduced to achieve synchronous aggregation and feature extraction of spatial neighborhood information and historical time information. The input data of the heavy medium coal preparation process is reconstructed. A combined loss function is constructed based on the data reconstruction error and graph structure sparsity regularization term to train the spatiotemporal synchronous graph attention autoencoder. This forces the model to learn the key features of the process data under normal operating conditions of the heavy medium coal preparation process and constrains the dynamic graph structure to avoid spurious associations, thus obtaining a well-trained monitoring model.

[0037] S3. Construction of Multi-Level Monitoring Statistics and Offline Modeling: Based on the trained model, data-level statistics and graph-level statistics are constructed respectively. The data-level statistics include T² statistics in the feature space and SPE statistics in the residual space. The graph-level statistics include macroscopic topological statistics based on the eigenvalues ​​of the Laplacian matrix. and microstructure statistics based on the number of edges in a dynamic graph The control limits for each statistic are determined by kernel density estimation.

[0038] S4. Online Fault Detection: Real-time acquisition of process data and preprocessing in the same way as the offline stage. The processed data is then input into the trained model to calculate the values ​​of four monitoring statistics in real time. If the value of any statistic exceeds its corresponding control limit, the process is determined to have failed.

[0039] In step S1, the data preprocessing includes outlier removal, missing value imputation, and max-min standardization. The spatiotemporal process data of all process monitoring variables are processed into the following form: X represents the training data consisting of N normal samples, and T represents the matrix transpose. Let X represent the t-th normal sample; let the i-th column of X be denoted as . , This represents the time-series data vector of the i-th process variable. This represents the measurement data of the i-th process variable at time t; by reorganizing the data through a sliding window, the time series data matrix input to the model at time t is obtained. , where w represents the length of the sliding window.

[0040] In step S1, the graph structure of the heavy medium coal preparation process is defined as a directed unweighted graph. ,in It is a set of nodes. It is a directed edge set. Represents the adjacency matrix; nodes It is the i-th variable in the heavy media coal preparation process, and its node attribute is a standardized time-series data vector. Directed edge Representing variables and variables The causal dependency between them.

[0041] In step S1, the construction of the dynamic-static hybrid graph structure specifically includes: defining the adjacency matrix of the static graph based on prior knowledge. The element is 1 when there is a physical connection or causal relationship between nodes, and 0 otherwise; a self-attention mechanism is used to extract data from the time-series data matrix. Inferring the adjacency matrix of a dynamic graph : ; where the query matrix Key-value matrix , and These are two trainable weight matrices for the self-attention mechanism. The sigmoid activation function is used; the static and dynamic graphs are superimposed and merged to obtain the adjacency matrix of the hybrid graph. : ;in, Indicates taking and Performing a bitwise OR operation on the edges of two adjacency matrices yields the final result. This serves as the topological basis for subsequent spatiotemporal synchronization feature extraction.

[0042] Please see Figure 2 In step S2, the spatiotemporal synchronized graph attention autoencoder (STGAAE) is constructed by embedding a spatiotemporal synchronized graph attention mechanism into a graph encoder-graph decoder architecture. When performing feature aggregation and updating on a node, the spatiotemporal synchronized graph attention mechanism considers not only the influence of its neighboring nodes at the same time but also incorporates the state information of its neighboring nodes at historical moments. For the input reconstructed 3D data... The data is in The node corresponding to the time Chinese characteristics The spatiotemporal aggregation output is represented as: ;in, and Indicates in layer The node corresponding to the time data With nodes In features The input data has the following characteristics when the layer number is 0: , . Represents nodes in a hybrid graph The gathering of neighbors, This indicates the time dimension of the data.

[0043] To achieve efficient matrix operations and avoid future information leakage, spatial attention weights are first calculated. For hybrid graphs... any node at time With nodes The input data are respectively and Synchronous attention coefficient Calculated in the following way: Where LeakyReLU represents a nonlinear activation function. This represents the weight vector of the attention mechanism. This represents a vector concatenation operation. and Indicates in Nodes corresponding to the layer With nodes The input features, when the number of layers is 0, have , .

[0044] Subsequently, by analyzing the nodes All neighbors The attention scores are standardized to obtain normalized attention coefficients. : Then, by combining the learnable temporal feature extraction matrix and the masking lower triangular matrix, which are applied to the input data, a spatiotemporal aggregation matrix expression is obtained: ;in The adjacency matrix of the hybrid graph. It is the identity matrix, used to add self-loops to the adjacency matrix of the hybrid graph. for The input data corresponding to the layer, For extracting time features, This is a masking lower triangular matrix. To better extract features from the data, a matrix is ​​used. The updated nodes undergo dimensional transformation, and then nonlinear operations are introduced through activation functions to enhance the ability to extract complex high-dimensional features.

[0045] In step S2, the spatiotemporal synchronized graph attention autoencoder (STGAAE) is composed of a spatiotemporal synchronized graph attention mechanism embedded in a graph encoder-graph decoder architecture: the spatiotemporal synchronized graph encoder is composed of multiple layers of spatiotemporal synchronized graph attention modules stacked together, and its function is to process the input raw spatiotemporal data... Mapped to a representation rich in spatiotemporal dependencies ,in For the dimension. layer The data at time t is obtained through aggregation and updating using the following formula: The spatiotemporal synchronization graph decoder, as the inverse process of the spatiotemporal synchronization graph encoder, aims to extract from the representation... The original input data is reconstructed from the data. Therefore, the forward propagation calculation formula for the spatiotemporal synchronization graph decoder is as follows: .

[0046] In step S2, the combined loss function Reconstruction loss Sparse regularization loss of graph structure Weighted composition: Among them, the reconstruction loss is used to measure the output of the spatiotemporal synchronization graph decoder. With the original input The differences between these data forces the model to learn key features from normal operating condition data. Mean squared error is used as the reconstruction loss. To prevent the dynamic graph from becoming overly complex and introducing spurious associations, a graph structure sparsity regularization loss is added. This penalizes excessively large sums of edge weights in the dynamic adjacency matrix to encourage its sparsity, consistent with the prior knowledge that associations in real industrial processes are typically sparse. L1 norm of a matrix: ; It is a hyperparameter used to balance the importance between reconstruction accuracy and graph sparsity. Represents a training batch. This refers to the batch size.

[0047] In step S3, offline modeling includes training a spatiotemporally synchronized graph attention autoencoder (STGAAE) based on normal training samples and constructing monitoring statistics at the data and graph levels. The monitoring statistics at the data level include constructing traditional... and the SPE statistic, where the feature space statistic at time t is... The calculation formula is: ;in and These represent the average of the feature vector output by the final hidden layer of the spatiotemporal synchronous graph autoencoder (STGAAE) at time t and the feature vector obtained during the offline modeling stage, respectively. It is the covariance matrix obtained in the final hidden layer offline modeling stage; the residual space statistics at time t. The calculation formula is: ;in for The original input vector at time step 1. This is the reconstruction output of the spatiotemporal synchronized graph decoder; the graph-level monitoring statistics include Laplace spectrum monitoring statistics constructed from both macroscopic topology and microscopic structure levels. and dynamic graph structure monitoring statistics For each moment dynamic graph structure First, calculate its corresponding degree matrix. With Laplace matrix : Subsequently, the Laplace matrix was... Perform eigenvalue decomposition to obtain a set of eigenvalues. The variance of these eigenvalues ​​is calculated as a comprehensive macroscopic statistic. : For each moment dynamic graph adjacency matrix , elements in Quantified nodes With nodes At any moment Dynamic association. All The summation yields the statistical dynamic graph adjacency matrix. Statistics on the number of all edges : Four statistical measures SPE and control limits , , and All of these methods are obtained by estimating the kernel density of the corresponding statistic sequence of normal training data at a given significance level.

[0048] In step S4, online monitoring is performed at every moment. The input matrix is ​​obtained by reorganizing the process data through a sliding window. The trained spatiotemporal synchronization graph attention autoencoder (STGAAE) calculates the values ​​of four monitoring statistics. Finally, the current process state is determined by comparing the statistics with the control limits. If any monitoring statistic exceeds the control limit, a fault is determined to have occurred. .

[0049] Please see Figure 3The heavy medium coal preparation process is complex, mainly including core equipment such as a mixing tank, heavy medium hydrocyclone, arc screen, clean coal / gangue desliming screen, dilute medium tank, combined medium tank, and magnetic separator. After the raw coal is deslimed, it is mixed with qualified medium in the mixing tank and pumped into the hydrocyclone for separation. The overflow and underflow enter the subsequent screening and medium recovery stages, respectively. In order to achieve accurate fault detection of this process, this embodiment selects 18 key process variables as monitoring nodes, including the output density / flow rate of the mixing tank, the overflow / underflow density and flow rate of the hydrocyclone, the flow rate after each screen, and the liquid level / diversion flow rate / output density of the dilute medium tank and the combined medium tank (specific variables are shown in Table 1).

[0050] Table 1. Variables in the Heavy Medium Washing Process

[0051]

[0052] The first step was data acquisition and preprocessing. A total of 1750 samples of actual production data from the coal preparation plant were collected. The first 500 samples were normal operating data used for model training, and subsequent samples introduced faults for testing. In the data preprocessing stage, outlier removal and linear interpolation were first performed on the 18 collected variables, followed by min-max standardization to eliminate the influence of dimensions. To capture temporal characteristics, a sliding window technique was used to reconstruct the standardized time series data. The sliding window length w was set to 20, meaning that the model input matrix at each time step contained data from the current time step and the previous 19 time steps, forming a dynamic spatiotemporal input matrix.

[0053] The second step is to construct a hybrid dynamic-static graph. First, a static graph adjacency matrix is ​​defined based on the process flow diagram and physical mechanism of the heavy medium coal preparation process. For example, since the output of the mixing tank flows directly to the heavy medium cyclone separator, a directed connection (edge ​​weight of 1) is defined between these two nodes in the static graph, and so on, to construct a static graph reflecting physical connections. Second, a dynamic graph is inferred from the process data within the sliding window using a self-attention mechanism. By calculating the query matrix and key-value matrix of the input data, a dynamic correlation matrix reflecting the real-time changes between variables under different operating conditions is obtained. Finally, the static and dynamic graphs are weighted and fused to generate a hybrid graph adjacency matrix, which serves as the topological basis for subsequent spatiotemporal feature extraction.

[0054] The third step was to build and train a spatiotemporally synchronized graph attention autoencoder (STGAAE). In model construction, the hidden state dimension was set to 30. Multiple layers of spatiotemporally synchronized graph attention modules were stacked within the encoder. These modules, when updating node features, simultaneously aggregated the spatial neighbor information and historical state information in the temporal dimension of the mixed graph. During training, a combined loss function was constructed, incorporating data reconstruction loss (MSE) and graph structure sparsity regularization loss (L1 norm). To prevent overfitting of the dynamic graph and the generation of spurious associations, the hyperparameter of the graph structure sparsity regularization loss was set to 0.01. The model was trained using the Adam optimizer, with the following training parameters: batch size of 175, learning rate of 0.002, and training epochs of 100. By minimizing the combined loss function, the model learned the spatiotemporal feature distribution and sparse dynamic graph structure under normal operating conditions in the heavy media coal preparation process.

[0055] The fourth step in implementation is fault detection and verification. Based on the trained model, four monitoring statistics are calculated during the online monitoring phase: data-level... The SPE statistic, as well as the Laplace spectrum monitoring statistic (L-Spectral) and dynamic graph structure monitoring statistic (G-Edge, number of dynamic graph edges) at the graph level, were used. Control limits were determined using kernel density estimation (KDE), with a significance level set at 0.01. In the test, for faults such as "magnetic separator running out of medium" (fault 4), the monitoring results showed that after the fault occurred, not only... The rapid exceedance of SPE and the significant increase in edge count shown by the graph-level G-Edge statistics, along with the dramatic fluctuations in the L-Spectral statistics, indicate that the fault caused a sudden change in the correlation between variables. Through comparative experiments, the method in this embodiment achieved an F1 score of 0.960 in fault detection during the heavy media coal preparation process, significantly outperforming traditional methods such as LSTM-AE (0.908) and KDGCN (0.930), which relies solely on static graphs. This verifies the effectiveness of the present invention under the specified parameter settings.

[0056] Table 2 shows the monitoring performance indicators of each comparative method in the heavy media sorting process. From the overall performance perspective, the proposed HG-STAN method performs best among all comparative methods, achieving an F1 score of 0.960, maintaining the highest detection rate while achieving the lowest false alarm rate. The KDGCN method achieved a second-best F1 score, but it only uses a static graph and cannot capture the dynamic relationships that evolve with operating conditions and raw material characteristics during production. This limits further performance improvement. GDAE, because it only uses a dynamic graph constructed from process data for fault monitoring, has a high detection rate, but its false alarm rate is significantly high, limiting its practical application value. The KG-GCBiGCN method, due to its quasi-static graph structure, has unsatisfactory detection performance. LSTM-AE shows a significant performance improvement compared to traditional AE, indicating that temporal modeling is crucial for process monitoring; however, due to its failure to fully model the spatial relationships between variables, its overall performance still has room for improvement. The above results consistently verify the effectiveness and superiority of the proposed hybrid graph structure and spatiotemporal synchronous modeling mechanism in the monitoring of the heavy media sorting process.

[0057] Table 2. Performance Comparison of All Methods in Fault Detection in DMS

[0058]

[0059] Figure 4 and Figure 5 The T² statistic and SPE statistic for all methods are compared for fault 4. Figure 4 In the table, (a)-(d) represent AE, LSTM-AE, GDAE and HG-STAN respectively. KDGCN and KG-GCBiGCN did not use the T² monitoring statistic. Figure 5 In the figure, (a)-(f) represent AE, LSTM-AE, KDGCN, GDAE, KG-GCBiGCN, and HG-STAN, respectively. The red dashed line represents the control limit, and the fault is introduced from the 501st sample. It can be seen that HG-STAN detects the fault initiation point most promptly and makes the most definitive judgment. The statistic curve quickly exceeds the control limit after the fault occurs. In contrast, the responses of other methods have varying degrees of delay or fluctuation, and the false alarm rate is relatively high. The HG-STAN proposed in this invention successfully detected fault 4 in the heavy medium sorting process using T² and SPE statistics, highlighting the robustness and accuracy of this method.

[0060] Figure 6 and Figure 7 A comparison of the graphical monitoring results for the heavy media sorting process. Figure 6The changes in the Laplace spectrum and dynamic graph structure of the two graph structure monitoring statistics of HG-STAN under fault 4 are shown. As shown in the figure, after the fault occurred, both statistics responded synchronously and significantly exceeded the control limits, indicating that anomalies occurred in both the macroscopic topology and microscopic connectivity of the system. Figure 7 Further visual evidence is provided by comparing the hybrid graph structures before (a, b) and after (c, d) the fault. It is clearly observed that the number of connections between nodes significantly increases after the fault, altering the overall topological characteristics. This visual result is consistent with... Figure 6 The statistical trends are highly consistent, jointly verifying the effectiveness and sensitivity of the graph structure monitoring statistics proposed in this invention in capturing structural anomalies.

[0061] This fault detection method for heavy medium coal preparation based on dynamic-static hybrid graphs balances physical interpretability and operating condition adaptability. By constructing a dynamic-static hybrid graph, it integrates static correlations reflecting physical mechanisms, ensuring the interpretability and physical rationality of the model, while introducing dynamic correlations learned from data, enabling the model to adapt to dynamic changes such as operating condition drift and raw material fluctuations in the production process, overcoming the limitations of a single graph construction mode. The proposed spatiotemporal synchronous graph attention mechanism can synchronously aggregate the spatial neighbor information and historical time information of nodes in the same calculation step, avoiding the fragmentation of spatiotemporal coupling information in the traditional serial processing architecture. This synchronous modeling approach better aligns with the inherent spatiotemporal evolution of the heavy media coal preparation process, improving the efficiency and effectiveness of feature extraction. It innovatively introduces graph-level monitoring statistics (L-Spectral and G-Edge) to complement data-level statistics (T² and SPE). This dual-perspective monitoring system can simultaneously detect numerical anomalies and topological anomalies, significantly improving the sensitivity and accuracy of detecting minor, slowly varying, and structural faults. By introducing L1 norm penalty into the graph structure sparsity regularization loss, the dynamic graph is constrained to maintain sparsity, effectively avoiding the learning of spurious associations in noisy environments and improving the model's robustness. This invention has been validated in actual coal preparation plants. Experimental results show that compared to traditional AE, LSTM-AE, and GNN methods using only static or dynamic graphs, the method of this invention significantly improves the detection rate (FDR) and F1 score while maintaining a low false alarm rate (FAR).

Claims

1. A fault detection method for heavy medium coal preparation process based on dynamic-static hybrid diagrams, characterized in that, Includes the following steps: S1. Data preprocessing and construction of dynamic and static hybrid graphs: For the heavy media coal preparation process, process monitoring variable data and prior knowledge are collected, and the process monitoring variable data are preprocessed to obtain a standardized time series data for each process variable; Dynamic time-series data is generated by a fixed-length sliding window, forming standardized dynamic input time-series data. The adjacency matrix of the static graph is defined using prior knowledge. The preprocessed temporal data is input into the spatial self-attention mechanism. The attention weights between variables are obtained by calculating the query matrix and the key matrix, which constitute the adjacency matrix of the dynamic graph. The static graph and the dynamic graph are fused to obtain a hybrid graph, which serves as the topological basis for subsequent feature extraction. S2. Spatiotemporal Synchronous Graph Attention Autoencoder Modeling and Training: A spatiotemporal synchronous graph attention autoencoder (STGAAE) is established. The STGAAE includes a spatiotemporal synchronous graph encoder and a spatiotemporal synchronous graph decoder, both of which are composed of stacked multi-layer spatiotemporal synchronous graph attention modules. Using the hybrid graph adjacency matrix and preprocessed time-series data as model input, during node update, a masked lower triangular matrix and a time feature extraction matrix are introduced to achieve synchronous aggregation and feature extraction of spatial neighborhood information and historical time information. The input data of the heavy medium coal preparation process is reconstructed. Based on the data reconstruction error and graph structure sparsity regularization term, a combined loss function is constructed to train the spatiotemporal synchronous graph attention autoencoder. This forces the model to learn the key features of the process data under normal operating conditions of the heavy medium coal preparation process and constrains the dynamic graph structure to avoid false associations, thus obtaining a well-trained monitoring model. S3. Construction of Multi-Level Monitoring Statistics and Offline Modeling: Based on the trained model, data-level statistics and graph-level statistics are constructed respectively. The data-level statistics include T² statistics in the feature space and SPE statistics in the residual space. The graph-level statistics include macroscopic topological statistics based on the eigenvalues ​​of the Laplacian matrix. and microstructure statistics based on the number of edges in a dynamic graph The control limits for each statistic are determined by kernel density estimation; S4. Online Fault Detection: Real-time acquisition of process data and preprocessing in the same way as the offline stage. The processed data is then input into the trained model to calculate the values ​​of four monitoring statistics in real time. If the value of any statistic exceeds its corresponding control limit, the process is determined to have failed.

2. The method for fault detection in heavy medium coal preparation process based on dynamic and static hybrid diagrams according to claim 1, characterized in that, In step S1, the data preprocessing includes outlier removal, missing value imputation, and max-min standardization. The spatiotemporal process data of all process monitoring variables are processed into the following form: X represents the training data consisting of N normal samples, and T represents the matrix transpose. Let X represent the t-th normal sample; let the i-th column of X be denoted as . , This represents the time-series data vector of the i-th process variable. This represents the measurement data of the i-th process variable at time t; by reorganizing the data through a sliding window, the time series data matrix input to the model at time t is obtained. , where w represents the length of the sliding window.

3. The method for fault detection in heavy medium coal preparation process based on dynamic and static hybrid diagrams according to claim 1, characterized in that, In step S1, the graph structure of the heavy medium coal preparation process is defined as a directed unweighted graph. ,in It is a set of nodes. It is a directed edge set. Represents the adjacency matrix; nodes It is the i-th variable in the heavy media coal preparation process, and its node attribute is a standardized time-series data vector. Directed edge Representing variables and variables The causal dependency between them.

4. The method for fault detection in heavy medium coal preparation process based on dynamic and static hybrid diagrams according to claim 1, characterized in that, In step S1, the construction of the dynamic-static hybrid graph structure specifically includes: defining the adjacency matrix of the static graph based on prior knowledge. The element is 1 when there is a physical connection or causal relationship between nodes, and 0 otherwise; a self-attention mechanism is used to extract data from the time series data matrix. Inferring the adjacency matrix of a dynamic graph : ； The query matrix Key-value matrix , and These are two trainable weight matrices for the self-attention mechanism. The sigmoid activation function is used; the static and dynamic graphs are superimposed and merged to obtain the adjacency matrix of the hybrid graph. : ； in, Indicates taking and Performing a bitwise OR operation on the edges of two adjacency matrices yields the final result. This serves as the topological basis for subsequent spatiotemporal synchronization feature extraction.

5. The method for fault detection in heavy medium coal preparation process based on dynamic and static hybrid diagrams according to claim 1, characterized in that, In step S2, the spatiotemporal synchronized graph attention autoencoder (STGAAE) is composed of a spatiotemporal synchronized graph attention mechanism embedded in a graph encoder-graph decoder architecture. To achieve efficient matrix operations and avoid future information leakage, spatial attention weights are first calculated for hybrid graphs. any node at time With nodes The input data are respectively and Synchronous attention coefficient Subsequently, by analyzing the nodes All neighbors The attention scores are standardized to obtain normalized attention coefficients. Then, by combining the learnable temporal feature extraction matrix and the masking lower triangular matrix, which are applied to the input data, a spatiotemporal aggregation matrix expression is obtained: ； in The adjacency matrix of the hybrid graph. The identity matrix is ​​used to add self-loops to the adjacency matrix of the hybrid graph. for The input data corresponding to the layer, For extracting time features, To mask the lower triangular matrix and to better extract data features, a matrix is ​​used. The updated nodes undergo dimensional transformation, and then nonlinear operations are introduced through activation functions to enhance the ability to extract complex high-dimensional features.

6. The method for fault detection in heavy medium coal preparation process based on dynamic and static hybrid diagrams according to claim 1, characterized in that, In step S2, the spatiotemporal synchronization graph encoder is composed of multiple layers of spatiotemporal synchronization graph attention modules stacked together. Its function is to process the input raw spatiotemporal data. Mapped to a representation rich in spatiotemporal dependencies ,in For dimension; the spatiotemporal synchronization graph decoder, as the inverse process of the spatiotemporal synchronization graph encoder, aims to obtain from the representation The original input data is reconstructed from the data. .

7. The method for fault detection in heavy medium coal preparation process based on dynamic and static hybrid diagrams according to claim 1, characterized in that, In step S2, the combined loss function Reconstruction loss Sparse regularization loss of graph structure Weighted composition: ； The reconstruction loss is used to measure the output of the spatiotemporal synchronization graph decoder. With the original input The differences between them force the model to learn the key features of normal operating condition data, and the mean squared error is used as the reconstruction loss: ； To prevent dynamic graphs from becoming overly complex and introducing spurious associations, a graph structure sparsity regularization loss is added. This penalizes excessively large sums of edge weights in the dynamic adjacency matrix to encourage sparsity, consistent with the prior knowledge that associations in real industrial processes are typically sparse. L1 norm of a matrix: ； It is a hyperparameter used to balance the importance between reconstruction accuracy and graph sparsity. Represents a training batch. This refers to the batch size.

8. The method for fault detection in heavy medium coal preparation process based on dynamic and static hybrid diagrams according to claim 1, characterized in that, In step S3, offline modeling includes: training a spatiotemporal synchronized graph attention autoencoder STGAAE based on normal training samples, and constructing data-level and graph-level monitoring statistics. The data-level monitoring statistics include constructing traditional... and the SPE statistic, where the feature space statistic at time t is... The calculation formula is: ； in and and represent the average of the feature vector output by the final hidden layer of the spatiotemporal synchronization graph autoencoder at time t and the feature vector obtained during the offline modeling stage, respectively. It is the covariance matrix obtained in the final hidden layer offline modeling stage; the residual space statistics at time t. The calculation formula is: ； in for The original input vector at time step 1. The reconstructed output of the spatiotemporal synchronization graph decoder; The monitoring statistics at the graph level include Laplace spectrum monitoring statistics constructed from both macroscopic topological and microscopic structural levels. and dynamic graph structure monitoring statistics For each moment dynamic graph structure First, calculate its corresponding degree matrix. With Laplace matrix : ； Subsequently, the Laplace matrix Perform eigenvalue decomposition to obtain a set of eigenvalues. The variance of these eigenvalues ​​is calculated as a comprehensive macroscopic statistic. : ； For each moment dynamic graph adjacency matrix , elements in Quantified nodes With nodes At any moment Dynamic association, to all The summation yields the statistical dynamic graph adjacency matrix. Statistics on the number of all edges ; ； Four statistics SPE and Control Limits , , and All of these methods are obtained by estimating the kernel density of the corresponding statistic sequence of normal training data at a given significance level.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the fault detection method for heavy medium coal preparation process based on dynamic and static mixing diagrams as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, execute the steps of the method for fault detection in a heavy media coal preparation process based on a dynamic-static mixing diagram as described in any one of claims 1-8.

Images