Dynamic graph fault detection method and device based on mechanism and data fusion

Abstract

The invention provides a mechanism and data fused dynamic graph fault detection method and device, and relates to the technical field of fault detection. The method comprises the following steps: acquiring historical data of a key variable, constructing a mechanism graph, a static data graph and a dynamic data graph, and fusing to obtain a fused graph; performing spatial feature extraction on the fused image through an image convolutional network, and performing time feature extraction on the spatial features through a gating loop unit; and integrating the spatial features and the time features, and obtaining a predicted value of each node through a full connection layer to train a fault detection model. According to the method, the dynamic challenge of industrial production data is effectively solved, and the time sequence dependence and association mode of key production variables are accurately captured; the advantages of mechanism knowledge and a data driving method are combined, so that key production variables can be effectively aggregated, and the accuracy and robustness of fault detection are improved. Faults can be efficiently recognized, the adaptive capacity to a dynamic system is enhanced, and therefore the safety and stability of the industrial production process are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a dynamic graph fault detection method and device based on mechanism and data fusion. Background Art

[0002] With the development of big data technology, the steel industry has accumulated a large amount of real-time data by deploying a large number of sensors at the work site. These data record various parameter changes and equipment states in the industrial process, laying a foundation for realizing intelligent fault detection. By combining these data with technologies such as machine learning and deep learning, hidden fault patterns can be effectively discovered, and intelligent monitoring and prediction of complex industrial processes can be achieved.

[0003] Currently, fault detection methods can be mainly divided into two categories: mechanism-based methods and data-based methods. Mechanism-based methods achieve an accurate description of the normal state of the system by constructing a physical model of the system, mainly including residual analysis, state estimation method, analytical redundancy method, adaptive model method, etc. However, when the system involves multivariable control and nonlinear characteristics, traditional physical models are difficult to comprehensively describe all dynamic behaviors of the system and face difficulties in processing large data sets. In addition, industrial equipment may be affected by environmental disturbances, aging, and noise during actual operation, resulting in deviations. Data-based methods do not require constructing an accurate physical model but rely on historical data and, through deep learning, mine potential rules and patterns from them, with flexibility and adaptability. However, such methods lack an in-depth understanding of the internal structure and operating mechanism of the system and do not have enough mechanism knowledge to effectively diagnose the root cause of faults. In addition, in a complex production environment, equipment and process flows change over time, and industrial data poses challenges with high dynamics. This means that fault detection not only needs to focus on current performance but also needs to be sensitive to dynamic changes.

[0004] In an industrial process, sensors do not exist independently in an isolated manner but are interconnected and influenced by complex networks. Therefore, the industrial production data collected not only has temporal characteristics but also has spatial relationships. These spatio-temporal data contain the synergy and potential correlation information between various parts of the system, which has extremely important research significance and practical value for the fault diagnosis of the system. However, traditional fault diagnosis methods mainly focus on feature extraction in Euclidean space and have certain limitations in processing data in non-Euclidean space, which restricts the practical application of fault detection technology in the industrial production process. Summary of the Invention

[0005] To solve the technical problems that the existing fault detection method lacks certain mechanism knowledge in data-driven graph structure learning and the spatial coupling relationship between variables is time-varying, an embodiment of the present invention provides a dynamic graph fault detection method and device that combines mechanism and data. The technical solution is as follows:

[0006] On the one hand, a dynamic graph fault detection method that combines mechanism and data is provided. This method is implemented by a dynamic graph fault detection device, and the method includes:

[0007] S1. Obtain historical data of key variables of the finishing mill system in the steel rolling process through sensors.

[0008] S2. Construct a mechanism graph, a static data graph, and a dynamic data graph based on the historical data.

[0009] S3. Fuse the mechanism graph, the static data graph, and the dynamic data graph to obtain a fused graph.

[0010] S4. Extract spatial features from the fused graph through a graph convolutional network to obtain spatial features, and extract temporal features from the spatial features through a gated recurrent unit to obtain temporal features.

[0011] S5. Integrate the spatial features and the temporal features to obtain integrated features, input the integrated features into a fully connected layer to obtain the predicted value of each node, and train a fault detection model according to the predicted value of each node and the historical data to obtain a trained fault detection model.

[0012] S6. Obtain the steel rolling process data to be detected, input it into the trained fault detection model, and obtain the fault detection result.

[0013] Optionally, constructing a mechanism graph, a static data graph, and a dynamic data graph based on the historical data in S2 includes:

[0014] S21. Determine the causal relationship between key variables, and construct a mechanism graph according to the causal relationship; wherein, the mechanism graph includes nodes and directed edges, the nodes represent key variables, and the directed edges represent the causal relationship between key variables.

[0015] S22. Perform normalization processing on the historical data, extract the feature embedding of each node from the normalized data using an autoencoder, calculate the cosine similarity between nodes according to the feature embedding, and construct a static data graph according to the cosine similarity.

[0016] S23. Construct a dynamic data graph for adaptively learning the variable spatial coupling relationship between key variables.

[0017] Optionally, the graph structure of the mechanism graph is represented by the adjacency matrix shown in the following formula (1):

[0018] (1)

[0019] Wherein, represents the adjacency matrix of the graph structure for representing the mechanism diagram, , represents a node.

[0020] Optionally, the static data graph is represented by the adjacency matrix shown in the following formula (2):

[0021] (2)

[0022] Wherein, represents a node and the similarity score of, and represents the and node embedding representations obtained from the autoencoder, represents the adjacency matrix for representing the static data graph, represents selecting the edge with the highest similarity from all edges edge operation, The value of is determined by the total number of edges in the mechanism diagram, represents a node and the similarity score of, and represents the index of the node.

[0023] Optionally, the mechanism diagram, static data graph, and dynamic data graph in S3 are fused to obtain a fused graph, including:

[0024] By solving the optimization problem shown in the following formula (3), the weights of the mechanism diagram, static data graph, and dynamic data graph are obtained, and the mechanism diagram, static data graph, and dynamic data graph are fused according to the weights to obtain a fused graph:

[0025] (3)

[0026] Wherein, represents the mechanism diagram, represents the static data graph, represents the dynamic data graph, represents weight, represents the Frobenius norm, represents matrix transpose, represents the fused graph adjacency matrix the th column vector, represents the real number space, represents the total number of nodes in the graph, represents the th element of.

[0027] Optionally, obtaining the rolling process data to be detected in S6, inputting it into the trained fault detection model, and obtaining the fault detection result, including:

[0028] Obtaining the rolling process data to be detected, inputting it into the trained fault detection model, and obtaining the predicted adjacency matrix and the predicted value.

[0029] Obtaining the graph structure deviation according to the preset adjacency matrix and the predicted adjacency matrix.

[0030] Obtaining the prediction deviation according to the preset threshold and the predicted value.

[0031] Obtaining the total deviation of each node according to the graph structure deviation and the prediction deviation, and obtaining the fault detection result according to the total deviation.

[0032] Optionally, obtaining the fault detection result according to the total deviation, including:

[0033] According to the total deviation, using the maximum value function to obtain the anomaly score at each time point, as shown in the following formula (4):

[0034] (4)

[0035] In the formula, represents the anomaly score at each time point and represents the total deviation of each node.

[0036] According to the anomaly score at each time point and the preset anomaly score threshold, obtaining the fault time, as shown in the following formula (5):

[0037] (5)

[0038] In the formula, represents the fault time, represents the anomaly score threshold.

[0039] On the other hand, a dynamic graph fault detection device for mechanism and data fusion is provided. This device is applied to the dynamic graph fault detection method for mechanism and data fusion. The device includes:

[0040] A data acquisition module, configured to acquire historical data of key variables of the finishing mill system in the rolling process through sensors.

[0041] The graph structure construction module is used to construct a mechanism graph, a static data graph, and a dynamic data graph based on historical data.

[0042] The graph structure fusion module is used to fuse the mechanism graph, the static data graph, and the dynamic data graph to obtain a fused graph.

[0043] The feature extraction module is used to extract spatial features from the fused graph through a graph convolutional network to obtain spatial features, and extract temporal features from the spatial features through a gated recurrent unit to obtain temporal features.

[0044] The training module is used to integrate the spatial features and the temporal features to obtain integrated features, input the integrated features into a fully connected layer to obtain the predicted value of each node, and train a fault detection model according to the predicted value of each node and historical data to obtain a trained fault detection model.

[0045] The fault detection module is used to obtain the rolling process data to be detected, input it into the trained fault detection model, and obtain a fault detection result.

[0046] Optionally, the graph structure construction module is further used for:

[0047] S21. Determine the causal relationship between key variables, and construct a mechanism graph according to the causal relationship; wherein, the mechanism graph includes nodes and directed edges, the nodes represent key variables, and the directed edges represent the causal relationship between key variables.

[0048] S22. Normalize the historical data, extract the feature embedding of each node from the normalized data by using an autoencoder, calculate the cosine similarity between nodes according to the feature embedding, and construct a static data graph according to the cosine similarity.

[0049] S23. Construct a dynamic data graph for adaptively learning the variable spatial coupling relationship between key variables.

[0050] Optionally, the graph structure of the mechanism graph is represented by the adjacency matrix shown in the following formula (1):

[0051] (1)

[0052] In the formula, represents the adjacency matrix used to represent the graph structure of the mechanism graph, , represents a node.

[0053] Optionally, the static data graph is represented by the adjacency matrix shown in the following formula (2):

[0054] (2)

[0055] In the formula, Represents a node and the similarity score of and represents the and node embedding representation obtained from the autoencoder represents the adjacency matrix for representing the static data graph represents the operation of selecting the edge with the highest similarity from all edges number of edges The value of represents a node and the similarity score of and represents the index of the node

[0056] Optionally, the graph structure fusion module is further configured to:

[0057] By solving the optimization problem shown in the following equation (3), obtain the weights of the mechanism graph, the static data graph, and the dynamic data graph, and fuse the mechanism graph, the static data graph, and the dynamic data graph according to the weights to obtain a fused graph:

[0058] (3)

[0059] In the formula, represents the mechanism graph represents the static data graph represents the dynamic data graph represents weight represents the Frobenius norm represents the matrix transpose represents the adjacency matrix of the fused graph the th column vector of represents the real number space represents the total number of nodes in the graph represents the th element of

[0060] Optionally, the fault detection module is further configured to:

[0061] Obtain the rolling process data to be detected, input it into the trained fault detection model, and obtain the predicted adjacency matrix and the predicted value

[0062] Obtain the graph structure deviation according to the preset adjacency matrix and the predicted adjacency matrix

[0063] Obtain a prediction deviation based on a preset threshold and a predicted value.

[0064] Obtain the total deviation of each node based on the graph structure deviation and the prediction deviation, and obtain a fault detection result based on the total deviation.

[0065] Optionally, obtaining a fault detection result based on the total deviation includes:

[0066] Based on the total deviation, use the maximum value function to obtain the anomaly score at each time point, as shown in the following formula (4):

[0067] (4)

[0068] In the formula, represents the anomaly score at each time point and represents the total deviation of each node.

[0069] Obtain the fault time based on the anomaly score at each time point and a preset anomaly score threshold, as shown in the following formula (5):

[0070] (5)

[0071] In the formula, represents the fault time, and

[0072] represents the anomaly score threshold.

[0072] On the other hand, a dynamic graph fault detection device is provided, and the dynamic graph fault detection device includes: a processor; a memory, and computer-readable instructions are stored on the memory. When the computer-readable instructions are executed by the processor, any one of the methods in the above-mentioned dynamic graph fault detection method of mechanism and data fusion is implemented.

[0073] On the other hand, a computer-readable storage medium is provided, and at least one instruction is stored in the storage medium. The at least one instruction is loaded and executed by a processor to implement any one of the methods in the above-mentioned dynamic graph fault detection method of mechanism and data fusion.

[0074] The beneficial effects brought by the technical solutions provided in the embodiments of the present invention at least include:

[0075] In the present invention, a dynamic graph fault detection method of mechanism and data fusion is proposed. The present invention effectively solves the challenge of the dynamic nature of industrial production data, accurately captures the time-series dependence and correlation patterns of key production variables; combines the advantages of mechanism knowledge and data-driven methods, enables key production variables to be effectively aggregated, and improves the accuracy and robustness of fault detection. This method can efficiently identify faults, enhance the adaptability to dynamic systems, and thus improve the safety and stability of the industrial production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0077] Figure 1 is a flowchart of a dynamic graph fault detection method for mechanism and data fusion provided by an embodiment of the present invention;

[0078] Figure 2 is a schematic diagram of the network model of the dynamic graph fault detection method for mechanism and data fusion provided by an embodiment of the present invention;

[0079] Figure 3 is the dynamic graph fault detection result of mechanism and data fusion provided by an embodiment of the present invention Figure 1 ;

[0080] Figure 4 is the dynamic graph fault detection result of mechanism and data fusion provided by an embodiment of the present invention Figure 2 ;

[0081] Figure 5 is the dynamic graph fault detection result of mechanism and data fusion provided by an embodiment of the present invention Figure 3 ;

[0082] Figure 6 is the dynamic graph fault detection result of mechanism and data fusion provided by an embodiment of the present invention Figure 4 ;

[0083] Figure 7 is the visualization of the true value and the predicted value provided by an embodiment of the present invention Figure 1 ;

[0084] Figure 8 is the visualization of the true value and the predicted value provided by an embodiment of the present invention Figure 2 ;

[0085] Figure 9 is the visualization of the true value and the predicted value provided by an embodiment of the present invention Figure 3 ;

[0086] Figure 10 is the visualization of the true value and the predicted value provided by an embodiment of the present invention Figure 4 ;

[0087] Figure 11 is a block diagram of a dynamic graph fault detection device for mechanism and data fusion provided by an embodiment of the present invention;

[0088] Figure 12 It is a schematic structural diagram of a dynamic graph fault detection device provided by an embodiment of the present invention. Specific embodiments

[0089] The technical solutions in the present invention will be described below with reference to the accompanying drawings.

[0090] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0091] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same. "Of", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same.

[0092] In the embodiments of the present invention, sometimes subscripts such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meanings they express are the same.

[0093] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0094] The embodiments of the present invention provide a dynamic graph fault detection method for mechanism and data fusion. This method can be implemented by a dynamic graph fault detection device, and the dynamic graph fault detection device can be a terminal or a server. As Figure 1 shown in the flowchart of the dynamic graph fault detection method for mechanism and data fusion, the processing flow of this method can include the following steps:

[0095] S1. Obtain historical data of key variables of the finishing system in the steel rolling process through sensors.

[0096] In a feasible implementation, actual steel rolling process data is obtained as training data.

[0097] S2. Construct a mechanism graph, a static data graph, and a dynamic data graph based on the historical data.

[0098] Optionally, the above step S2 can include the following steps S21 - S23:

[0099] S21. Determine the causal relationships between key variables and construct a mechanism diagram according to the causal relationships. Among them, the mechanism diagram includes nodes and directed edges, where the nodes represent key variables and the directed edges represent the causal relationships between key variables.

[0100] In a feasible implementation manner, a mechanism-driven method is adopted. Based on the mining of the association mode between mechanism analysis and sensors in the actual industrial process, the key variables and their interactions in the thickness subsystem of the finishing rolling system are analyzed, the dynamic operation of the system is clarified, and a block diagram of the closed-loop control system is constructed.

[0101] Based on the one-to-one correspondence between the deployed sensors and the monitored variables, qualitative and quantitative analysis of the variables is carried out. Determine the causal relationships between the variables and construct a mechanism diagram, providing a model basis for analyzing the operation mechanism of the process flow and revealing the variable mapping association. Compared with the data-driven black box model, the causal relationship network clarifies the signal transmission path and influence range, which helps to accurately locate faults in the system. By integrating all deployed sensors, the clear variable relationship can avoid invalid variable aggregation and prevent masking of effective information. In the mechanism diagram, each node represents a variable, and the influence relationship between variables is represented by a directed edge. Specifically, the adjacency matrix is defined as follows:

[0102] (1)

[0103] In the formula, represents the adjacency matrix used to represent the graph structure of the mechanism diagram, , represents the nodes of the graph, is the total number of nodes.

[0104] S22. Perform normalization processing on the historical data, extract the feature embeddings of each node from the normalized data using an autoencoder, calculate the cosine similarity between the nodes according to the feature embeddings, and construct a static data graph according to the cosine similarity.

[0105] In a feasible implementation manner, in addition to the mechanism diagram, the present invention also adopts the idea of data-driven for graph structure learning, builds a graph structure to capture the implicit relationships between data. Perform normalization processing on the data, extract the feature embeddings of each node using an autoencoder, calculate the cosine similarity between the nodes, construct an adjacency matrix, and construct a static data graph, thereby effectively capturing the similarity in the data.

[0106] Optionally, the static data graph is represented by the adjacency matrix shown in the following formula (2):

[0107] (2)

[0108] In the formula, Represents the node and The similarity score of and Represents the and Node embedding representation obtained from the autoencoder Represents the adjacency matrix used to represent the static data graph Represents the operation of selecting the edge with the highest similarity from all edges Edges The numerical value of is determined by the total number of edges in the mechanism graph Represents the node and The similarity score of and Represents the index of the node

[0109] This process simplifies the graph structure, emphasizes the edges with the strongest correlations, and is crucial for revealing potential patterns in the data

[0110] S23: Construct a dynamic data graph for adaptively learning the variable spatial coupling relationship between key variables

[0111] In a feasible implementation, due to the influence of complex internal and external factors, the spatial coupling relationship between variables has variable characteristics, that is, it changes over time and cannot be correctly reflected by the fixed representation of the static graph. Therefore, only constructing a static graph is likely to ignore the variable characteristics of the spatial coupling relationship, making it difficult to effectively model the variation law of process variables and possibly resulting in unsatisfactory results. To overcome the limitations of the static graph, a dynamic graph capable of adaptively learning the variable spatial coupling relationship between variables is further introduced, where the adjacency matrix of the graph changes with the historical data in the sliding time window

[0112] S3: Fuse the mechanism graph, static data graph, and dynamic data graph to obtain a fused graph

[0113] In a feasible implementation, based on the graph structure fusion module, the obtained static mechanism graph, data graph, and dynamic data graph are fused in a mutually reinforcing manner, automatically assigning weights to the adjacency matrix of each graph, and jointly learning the fused graph through an alternating iterative optimization algorithm, thereby obtaining the final unified graph adjacency matrix, enabling the final fused graph to optimally integrate and reflect the information from the original graph structures

[0114] Denote the static mechanism graph as the static data graph as the dynamic data graph as where It will change dynamically over time. First, the initial weights of each graph are set to the same value, that is is a uniform distribution. To enable the fused graph to contain as much information as possible about all graph structures, it is obtained by solving the following optimization problem:

[0115] (3)

[0116] In the formula, represents the mechanism graph, represents the static data graph, represents the dynamic data graph, represents weight, represents the Frobenius norm, represents matrix transpose, represents the adjacency matrix of the fused graph of the th column vector, represents the real number space, represents the total number of nodes in the graph, represents of the th element.

[0117] Through the optimization formula, the weights and the fused graph are updated alternately and determined automatically. The adaptive weight update mechanism in the graph structure fusion module allows the algorithm to dynamically adjust the contributions of different graphs, comprehensively characterize the structural information of multiple graphs, and ensure that the final fused graph accurately reflects the prior information in the mechanism graph and the true statistical patterns in the data graph, thus eliminating the subjectivity of manually setting weights.

[0118] S4. The spatial features are obtained by performing spatial feature extraction on the fused graph through a graph convolutional network, and the temporal features are obtained by performing temporal feature extraction on the spatial features through a gated recurrent unit.

[0119] In a feasible implementation, at each time step , the fused graph is processed by a graph convolutional network GCN for spatial feature extraction to capture spatial correlations. GCN updates the attributes of nodes by aggregating information between nodes and their neighbor nodes, effectively capturing the spatial dependencies in the graph structure:

[0120] (4)

[0121] Among them, represents the feature matrix of the nodes in the th layer, represents the non-linear activation function, is the degree matrix, is the adjacency matrix, represents an adjacency matrix containing self-loops, is the identity matrix, represents the trainable weight matrix of the

[0122] After extracting the spatial features, they are processed by a GRU (Gated Recurrent Unit) for time feature extraction. The GRU models the temporal relationships by sequentially processing the spatial features at time step to capture the temporal correlations. The GRU is a type of recurrent neural network architecture that uses a gating mechanism to capture short-term and long-term dependencies in sequential data. The update gate is responsible for capturing the long-term dependencies in the sequence, and the reset gate is responsible for modeling the short-term dependencies in the sequence.

[0123] (5)

[0124] where, represents the hidden state at time step , represents the input feature vector at time step , represents the candidate hidden state at the current time. The operation represents the matrix multiplication between the input and the weight matrix, is the element-wise multiplication, is the concatenation operation, and represent the trainable weight matrix and the bias vector of the network respectively.

[0125] The structure of the GRU helps to effectively model time series data while minimizing the computational complexity. The integration of non-linear activation functions and trainable parameters further enhances the ability to learn complex temporal patterns, enabling it to effectively extract temporal features.

[0126] S5. Integrate the spatial features and the temporal features to obtain the integrated features, input the integrated features into the fully connected layer to obtain the predicted value of each node, and train the fault detection model based on the predicted value of each node and the historical data to obtain the trained fault detection model.

[0127] In a feasible implementation, after integrating the spatial correlation and the temporal correlation, the output is passed to a fully connected layer to generate the predicted value for each node. The objective of this training process is to minimize the error between the predicted value and the actual data. Therefore, the loss function is defined as follows:

[0128] (6)

[0129] Among them, is the true value, is the predicted value, is a hyperparameter, is the regularization term, which can effectively avoid the overfitting problem by penalizing the sum of squares of the weights of the fully connected layer.

[0130] If the system operating in the normal state is represented as , and the system operating in the fault state is represented as , then the basic principle of fault detection based on GNN is:

[0131] (7)

[0132] Among them, and are the adjacency matrix in the normal case (specifically, the static data graph of the test set in this embodiment) and the node feature matrix (specifically, the time series feature in this embodiment) respectively, and are the adjacency matrix and the feature matrix in the fault case respectively. Generally speaking, the feature matrix describes the attribute features of each node. In the present invention, represents time series data to capture the dynamic changes of node features over time.

[0133] When the system is in the normal state, the relationship between variables will remain stable. However, when a fault occurs in the system, it may affect the topological structure, may also affect the node features, and is more likely to affect both at the same time. Therefore, when the time series data contains faults, the relationship between variables will change significantly. At the same time, variables with large prediction deviations also indicate anomalies. Therefore, by combining the prediction deviation and the graph structure deviation to obtain the total deviation of each node, the state of the variables can be better reflected:

[0134] (8)

[0135] Among them, and are the adjacency matrices representing the mechanism graph and the graph structure learned from the test sequence respectively, and represent the column vectors of the true values and predicted values of all variables at time respectively, where is predicted through the fusion graph;

[0136] Thus, it can be seen that It includes both the graph structure deviation and its own behavior deviation, and can effectively reflect the state of variables. When a fault occurs, the deviation of the fault node can be further amplified through the weighting of the graph structure deviation; during the normal operation of the system, this method can smooth out these deviations to ensure the stable state of each node. In addition, the numerical value in at each time point reflects the state of each node of the system or device. Specifically, when a fault occurs, the deviation of the abnormal node is large, while the deviation of the unaffected node is small. Since the fault may only affect a subset of the sensors, the maximum value function is used to obtain the anomaly score at each time point:

[0137] (9)

[0138] In the formula, represents the anomaly score at each time point , and represents the total deviation of each node.

[0139] Taking the maximum anomaly score among all time points of the validation set (normal data) as the anomaly threshold, the threshold and the test set are used to implement fault detection. If the anomaly score at time point exceeds the threshold, then this moment is defined as a fault:

[0140] (10)

[0141] In the formula, represents the fault moment, and represents the anomaly score threshold.

[0142] As Figure 2 shown is the schematic diagram of the network model of the dynamic graph fault detection method for mechanism and data fusion provided by the embodiment of the present invention. By acquiring the actual rolling process data, analyzing the operation mechanism of the actual industrial process and the correlation pattern between sensors, and constructing a causal mechanism diagram between variables, this method can effectively capture the spatial relationship between variables. Then, based on the mutual correlation tendency between sensors, data-driven graph structure learning is carried out, and a dynamic graph in which the adjacency matrix changes with historical data is introduced. Through the graph structure fusion module, the mechanism graph and the data-driven dynamic graph are fused with an adaptive and alternating iterative optimization strategy to maximize the information containing all graph structures. Then, the prediction deviation and the graph structure deviation are combined to effectively reflect the state of variables, calculate the anomaly score, and obtain the fault detection result.

[0143] Two simulation experiments are set up to verify the effectiveness of this fault detection method respectively.

[0144] First, verify whether the present invention can effectively perform fault detection. The data set comes from the finishing thickness control subsystem of the actual 2150 production line, which adopts the long-stroke hydraulic AGC system and the hydraulic loop technology. The data acquisition system therein includes thickness, rolling force, temperature, and roller table speed sensors, with a sampling frequency of 0.1 second, covering production parameters such as strip thickness, rolling force, stand speed, roller table temperature, work roll pressure, and water cooling flow rate. The data is monitored in real time by the automated control system and is applicable to the research on strip thickness control and production process optimization. This experimental data set consists of four different faults. The training set and the validation set only contain normal data, while the test set contains both normal data and fault data. Among them, the fault duration of F2-0116 is 2301-2500, the fault duration of F3-0104 is 2071-2250, and the fault durations of F3-0111 and F3-0113 are both 2415-2625.

[0145] To accurately evaluate the effectiveness of the model, it is evaluated by the false alarm rate and the miss rate. As Figures 3 - 6 shown, in actual fault detection, this method can simultaneously ensure a low false alarm rate and a low miss rate, showing excellent fault detection accuracy. This method has practical significance in ensuring safe production and preventing economic losses and is crucial for industrial production processes.

[0146] Second, verify whether the method of combining mechanism knowledge and data-driven in the present invention can make the prediction performance better. Take the data set of the finishing thickness control subsystem of the actual 2150 production line, and predict the 4th node in the fault F2-0116, and visualize the true value and the predicted value. Thus, the prediction performance is verified.

[0147] As Figures 7 - 10 shown, when not using the graph structure fusion module, but separately adopting the mechanism graph, the static data graph, and the dynamic data graph, the prediction performance deteriorates. The graph structure fusion combines the advantages of mechanism and data, reducing the possibility of fault false alarms and omissions ( Figure 7 ). Mechanism knowledge enables the model to show good prediction performance, but it is still not accurate enough ( Figure 8 ). The data-driven method captures data fluctuations, but it is difficult to accurately predict spikes ( Figure 9 ). The performance of the dynamic data graph is slightly better than that of the static data graph, indicating that sensitivity to dynamic changes can improve performance ( Figure 10). This indicates that when dealing with complex prediction tasks, a single graph structure often has limitations. By integrating graph structure fusion methods, the complementary advantages of making full use of mechanism knowledge to provide a stable theoretical basis for the system and data-driven to make a flexible response to dynamic changes can be realized. The method of the present invention verifies that the prediction performance of combining mechanism knowledge and data-driven methods is significantly improved compared with a single graph structure, indicating that the method of the present invention for prediction through mechanism and data fusion has strong usability and authenticity.

[0148] S6. Obtain the rolling process data to be detected, input it into the trained fault detection model, and obtain the fault detection result.

[0149] Optionally, the above step S6 may include:

[0150] Obtain the rolling process data to be detected, input it into the trained fault detection model, and obtain the predicted adjacency matrix and predicted values.

[0151] Obtain the graph structure deviation according to the preset adjacency matrix and the predicted adjacency matrix.

[0152] Obtain the prediction deviation according to the preset threshold and the predicted values.

[0153] Obtain the total deviation of each node according to the graph structure deviation and the prediction deviation, and obtain the fault detection result according to the total deviation.

[0154] In the embodiment of the present invention, a dynamic graph fault detection method integrating mechanism and data is proposed. The present invention effectively solves the challenge of the dynamic nature of industrial production data, accurately captures the time-series dependencies and correlation patterns of key production variables; combines the advantages of mechanism knowledge and data-driven methods, enables effective aggregation of key production variables, and improves the accuracy and robustness of fault detection. This method can efficiently identify faults, enhance the adaptability to dynamic systems, and thus improve the safety and stability of the industrial production process.

[0155] Figure 11 is a block diagram of a dynamic graph fault detection device integrating mechanism and data shown according to an exemplary embodiment. This device is used for the dynamic graph fault detection method integrating mechanism and data. Refer to Figure 11 , this device includes a data acquisition module 310, a graph structure construction module 320, a graph structure fusion module 330, a feature extraction module 340, a training module 350, and a fault detection module 360. Among them:

[0156] The data acquisition module 310 is used to obtain the historical data of the key variables of the finishing system of the rolling process through sensors.

[0157] The graph structure construction module 320 is used to construct a mechanism graph, a static data graph, and a dynamic data graph according to the historical data.

[0158] The graph structure fusion module 330 is used to fuse the mechanism graph, the static data graph, and the dynamic data graph to obtain a fused graph.

[0159] The feature extraction module 340 is used to extract spatial features from the fused graph through a graph convolutional network to obtain spatial features, and extract temporal features from the spatial features through a gated recurrent unit to obtain temporal features.

[0160] The training module 350 is used to integrate the spatial features and the temporal features to obtain integrated features, input the integrated features into a fully connected layer to obtain the predicted value of each node, and train a fault detection model according to the predicted value of each node and historical data to obtain a trained fault detection model.

[0161] The fault detection module 360 is used to obtain the rolling process data to be detected, input it into the trained fault detection model, and obtain a fault detection result.

[0162] In an embodiment of the present invention, a dynamic graph fault detection method based on mechanism and data fusion is proposed. The present invention effectively solves the challenge of the dynamic nature of industrial production data, accurately captures the temporal dependencies and correlation patterns of key production variables; combines the advantages of mechanism knowledge and data-driven methods, enables effective aggregation of key production variables, and improves the accuracy and robustness of fault detection. This method can efficiently identify faults, enhance the adaptability to dynamic systems, and thus improve the safety and stability of the industrial production process.

[0163] Figure 12 It is a schematic structural diagram of a dynamic graph fault detection device provided by an embodiment of the present invention. As Figure 12 shown, the dynamic graph fault detection device may include the above Figure 11 shown dynamic graph fault detection device for mechanism and data fusion. Optionally, the dynamic graph fault detection device 410 may include a first processor 2001.

[0164] Optionally, the dynamic graph fault detection device 410 may further include a memory 2002 and a transceiver 2003.

[0165] Wherein, the first processor 2001 is connected to the memory 2002 and the transceiver 2003, such as through a communication bus.

[0166] Next, in conjunction with Figure 12 specific introductions will be made to the various components of the dynamic graph fault detection device 410:

[0167] Among them, the first processor 2001 is the control center of the dynamic graph fault detection device 410, which can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), or can be an application specific integrated circuit (ASIC), or can be one or more integrated circuits configured to implement the embodiments of the present invention. For example: one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).

[0168] Optionally, the first processor 2001 can execute various functions of the dynamic graph fault detection device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0169] In a specific implementation, as an embodiment, the first processor 2001 can include one or more CPUs. For example Figure 12 the CPU0 and CPU1 shown in

[0170] In a specific implementation, as an embodiment, the dynamic graph fault detection device 410 can also include multiple processors. For example Figure 12 the first processor 2001 and the second processor 2004 shown in. Each of these processors can be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). Here, the processor can refer to one or more devices, circuits, and / or processing cores for processing data (such as computer program instructions).

[0171] Among them, the memory 2002 is used to store the software program for executing the solution of the present invention and is controlled by the first processor 2001 for execution. The specific implementation method can refer to the above method embodiment and will not be elaborated here.

[0172] Optionally, the memory 2002 can be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or it can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 can be integrated with the first processor 2001 or exist independently, and is coupled to the first processor 2001 through the interface circuit of the dynamic graph fault detection device 410 ( Figure 12 not shown in the figure), and the embodiments of the present invention do not make specific limitations on this.

[0173] The transceiver 2003 is used to communicate with a network device or with a terminal device.

[0174] Optionally, the transceiver 2003 can include a receiver and a transmitter ( Figure 12 not shown separately in the figure). Among them, the receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0175] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and is coupled to the first processor 2001 through the interface circuit of the dynamic graph fault detection device 410 ( Figure 12 not shown in the figure), and the embodiments of the present invention do not make specific limitations on this.

[0176] It should be noted that Figure 12 the structure of the dynamic graph fault detection device 410 shown in the figure does not constitute a limitation on the router. The actual knowledge structure recognition device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0177] In addition, the technical effects of the dynamic graph fault detection device 410 can refer to the technical effects of the dynamic graph fault detection method of mechanism and data fusion described in the above method embodiments, and will not be elaborated here.

[0178] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0179] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM) or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM) and direct rambus RAM (DR RAM).

[0180] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, or a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0181] It should be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. Here, A and B can be singular or plural. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be specifically understood by referring to the context.

[0182] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or its similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or plural.

[0183] It should be understood that in various embodiments of the present invention, the magnitudes of the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0184] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in connection with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.

[0185] Those skilled in the art can clearly understand that, for the sake of convenience and brevity of description, the specific working processes of the devices, apparatuses, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0186] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0187] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0188] In addition, the functional units in the various embodiments of the present invention can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0189] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0190] As described above, the above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A dynamic graph fault detection method integrating mechanism and data, characterized in that, The method includes: S1. Obtain historical data of key variables of the finishing mill system during the steel rolling process through sensors; S2. Construct a mechanism diagram, a static data diagram, and a dynamic data diagram based on the historical data; S3. Fuse the mechanism diagram, the static data diagram, and the dynamic data diagram to obtain a fused diagram; S4. Extract spatial features from the fused diagram through a graph convolutional network to obtain spatial features, and extract temporal features from the spatial features through a gated recurrent unit to obtain temporal features; S5. Integrate the spatial features and the temporal features to obtain integrated features, input the integrated features into a fully connected layer to obtain prediction values for each node, and train a fault detection model based on the prediction values for each node and the historical data to obtain a trained fault detection model; S6. Obtain the steel rolling process data to be detected, input it into the trained fault detection model, and obtain a fault detection result.

2. The dynamic graph fault detection method based on mechanism and data fusion according to claim 1, characterized in that The constructing of the mechanism diagram, the static data diagram, and the dynamic data diagram according to the historical data in S2 includes: S21. Determine the causal relationship between key variables, and construct a mechanism diagram according to the causal relationship; wherein, the mechanism diagram includes nodes and directed edges, the nodes represent key variables, and the directed edges represent the causal relationship between key variables; S22. Perform normalization processing on the historical data, use an autoencoder to extract the feature embeddings of each node from the normalized data, calculate the cosine similarity between nodes according to the feature embeddings, and construct a static data diagram according to the cosine similarity; S23. Construct a dynamic data diagram for adaptively learning the variable spatial coupling relationship between key variables.

3. The dynamic graph fault detection method based on mechanism and data fusion according to claim 1, characterized in that The graph structure of the mechanism diagram is represented by the adjacency matrix shown in the following formula (1): （1） In the formula, represents the adjacency matrix of the graph structure for representing the mechanism diagram, , represents a node.

4. The dynamic graph fault detection method based on mechanism and data fusion according to claim 1, characterized in that, The static data diagram is represented by the adjacency matrix shown in the following formula (2): （2） In the formula, represents the similarity score of nodes and ; and represent the node embedding representations of and obtained from the autoencoder; represents the adjacency matrix for representing the static data graph; represents an operation of selecting the edges with the highest similarity from all edges; The value of is determined by the total number of edges in the mechanism graph; and represent the similarity score of nodes and represent the indices of the nodes.

5. The dynamic graph fault detection method based on mechanism and data fusion according to claim 1, characterized in that The fusing of the mechanism diagram, the static data diagram, and the dynamic data diagram in S3 to obtain a fused diagram includes: By solving the optimization problem shown in the following formula (3), obtain the weights of the mechanism diagram, the static data diagram, and the dynamic data diagram, and fuse the mechanism diagram, the static data diagram, and the dynamic data diagram according to the weights to obtain a fused diagram: （3） In the formula, represents the mechanism diagram, represents the static data diagram, represents the dynamic data diagram, represents the weight, represents the Frobenius norm, represents the matrix transpose, represents the th column vector of the fusion graph adjacency matrix represents the real number space, represents the total number of nodes in the graph, represents the th element of 6. The dynamic graph fault detection method based on mechanism and data fusion according to claim 1, characterized in that The obtaining of the steel rolling process data to be detected, inputting it into the trained fault detection model, and obtaining a fault detection result in S6 includes: Obtain the steel rolling process data to be detected, input it into the trained fault detection model, and obtain a predicted adjacency matrix and prediction values; Obtain the graph structure deviation according to the preset adjacency matrix and the predicted adjacency matrix; Obtain the prediction deviation according to the preset threshold and the prediction values; Obtain the total deviation of each node according to the graph structure deviation and the prediction deviation, and obtain the fault detection result according to the total deviation.

7. The dynamic graph fault detection method based on mechanism and data fusion according to claim 6, characterized in that, The obtaining of the fault detection result according to the total deviation includes: According to the total deviation, use the maximum value function to obtain the anomaly score at each time point, as shown in the following formula (4): （4） In the formula, represents the anomaly score at each time point , and represents the total deviation of each node; According to the anomaly score at each time point and the preset anomaly score threshold, obtain the fault time, as shown in the following formula (5): （5） In the formula, represents the fault moment, represents the abnormal score threshold.

8. A dynamic graph fault detection device with mechanism and data fusion, the dynamic graph fault detection device with mechanism and data fusion is used to implement the dynamic graph fault detection method with mechanism and data fusion as described in any one of claims 1-7, characterized in that, The device includes: A data acquisition module for obtaining historical data of key variables of the finishing mill system during the steel rolling process through sensors; A graph structure construction module for constructing a mechanism diagram, a static data diagram, and a dynamic data diagram according to the historical data; A graph structure fusion module, which is used to fuse a mechanism graph, a static data graph, and a dynamic data graph to obtain a fused graph; A feature extraction module, which is used to extract spatial features from the fused graph through a graph convolutional network to obtain spatial features, and extract temporal features from the spatial features through a gated recurrent unit to obtain temporal features; A training module, which is used to integrate the spatial features and the temporal features to obtain integrated features, input the integrated features into a fully connected layer to obtain the predicted value of each node, and train a fault detection model according to the predicted value of each node and historical data to obtain a trained fault detection model; A fault detection module, which is used to obtain the rolling process data to be detected, input it into the trained fault detection model, and obtain a fault detection result.

9. A dynamic graph fault detection device, characterized in that, The dynamic graph fault detection device includes: A processor; A memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that, Program code is stored in the computer-readable storage medium, and the program code can be called by the processor to execute the method described in any one of claims 1 to 7.

Images