An unknown working condition equipment state anomaly detection method based on a graph neural network

Abstract

This invention discloses a method for detecting abnormal conditions in equipment under unknown operating conditions based on graph neural networks, belonging to the field of mechanical equipment fault diagnosis. The method includes: preprocessing normal and abnormal state data of the equipment under unknown operating conditions to obtain low-frequency data at different layers under each state, and determining the analysis layer; segmenting the data corresponding to the analysis layer under each state; extracting features from each segmented data from multiple perspectives; obtaining a multimodal feature node set from the segmented data under each state, and generating an operating condition graph feature set using a feature centroid hierarchical clustering graph construction method; and using the operating condition graph feature set to obtain a trained message-passing graph neural network model. Based on the constructed operating condition graph feature set, this invention effectively achieves abnormal detection of equipment under unknown operating conditions using the constructed message-passing graph neural network model.

Description

TECHNICAL FIELD

[0001] The application relates to a graph neural network-based unknown working condition equipment state anomaly detection method and belongs to the field of mechanical equipment fault diagnosis. BACKGROUND

[0002] In modern industry, equipment anomaly detection is crucial to guarantee production efficiency and reduce maintenance costs. Traditional methods rely on manual experience or rule systems, which are difficult to adapt to complex equipment and dynamic working conditions, and are prone to misjudgment or omission. Although existing machine learning and deep learning methods have certain effects, they still tend to learn statistical features that are strongly related to specific working conditions (such as speed and load) in nature, so when facing working condition shifts, there is a common problem of significantly reduced generalization ability.

[0003] Graph neural networks (GNN) have potential in handling complex data structures, but existing models still have the following problems in the specific task of state detection: first, the computational complexity is high when processing large-scale graph data, making it difficult to meet the real-time requirements of industrial scenarios; second, the graph construction process relies on pre-defined topological structures, which are difficult to dynamically adjust to adapt to changes in equipment operating states, limiting the modeling ability of the model for complex working conditions; in addition, traditional methods rely heavily on accurate working condition labels for model training, making it difficult to cope with complex scenarios in actual industry where working conditions are unknown, variable or continuously changing, resulting in poor model generalization ability; finally, a large amount of abnormal data is required for training in the task of equipment state anomaly detection, but abnormal data is difficult to obtain and the labeling cost is high. Therefore, designing a graph neural network-based unknown working condition equipment state anomaly detection method that adapts to unknown working conditions, efficiently processes large-scale data and does not require a large amount of abnormal data has important theoretical and practical significance for solving the problem of equipment anomaly detection. SUMMARY

[0004] The application provides a graph neural network-based unknown working condition equipment state anomaly detection method, which effectively realizes anomaly detection of unknown working condition equipment based on a constructed working condition graph feature set and a constructed message passing graph neural network model.

[0005] The technical scheme of the application is as follows:

[0006] A graph neural network-based unknown working condition equipment state anomaly detection method comprises the following steps:

[0007] Step 1: Obtain the vibration data of the unknown working condition equipment; wherein the vibration data of the unknown working condition equipment comprises normal state data and abnormal state data;

[0008] Step two: pre-process the normal state data and the abnormal state data of the unknown working condition equipment respectively to obtain low-frequency data of different layers under each state; compare the frequency spectrum of the normal state data and the frequency spectrum of the abnormal state data to determine the difference frequency range; select a layer covering the largest difference frequency range from the low-frequency data of different layers under each state as an analysis layer;

[0009] Step three: segment the data corresponding to the analysis layer under the two states according to the pre-set segmentation point number to obtain a plurality of segmented processing data; extract a first pre-set number of features from multiple angles for each segmented processing data; group the segmented processing data under each state according to a second pre-set number, and construct a feature node for the same feature of each group of segmented processing data, so that each feature constructs a feature node with a second pre-set number of dimensions; the feature nodes constructed by the first pre-set number of features form a multi-modal feature node set of ; wherein, is the first pre-set number;

[0010] Step four: according to the multi-modal feature node set, a graph construction method of feature centroid hierarchical clustering is used to first construct a feature centroid node, and then a graph feature structure is constructed according to the feature centroid node, and finally a working condition graph feature set composed of graph feature structures is generated;

[0011] Step five: build a message passing graph neural network model with a fusion message passing mechanism; use the working condition graph feature set to train and learn the message passing graph neural network model to obtain a trained message passing graph neural network model for detecting the state of the unknown working condition equipment.

[0012] Further, the pre-processing of the normal state data and the abnormal state data of the unknown working condition equipment is specifically: the normal state data and the abnormal state data of the unknown working condition equipment are sequentially subjected to normalization processing and n-layer wavelet decomposition, and the wavelet basis function used for wavelet decomposition is db4, thereby obtaining low-frequency data of different layers under each state.

[0013] Further, step four is specifically: a feature centroid node is constructed by selecting centroids for each feature node in the multi-modal feature node set according to the K-means clustering method, and feature centroid nodes form a feature centroid node set of ; each row of the feature centroid node set has ​​​​Using the centroid nodes of each feature as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each centroid node to construct the graph feature structure. Each graph feature structure forms a working condition graph feature set.

[0014] Furthermore, the set of characteristic centroid nodes for each row Using the feature centroid nodes as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each feature centroid node to construct the graph feature structure. Specifically, the similarity matrix of each row in the feature centroid node set is used as the basis for graph construction. Based on the feature centroid nodes, each feature centroid node is treated as a cluster. Only the two clusters with the maximum similarity matrix are selected and connected by an edge to form a new cluster 1. A new similarity matrix 1 is formed based on this new cluster 1. Similarly, based on the new similarity matrix 1, only the two clusters with the maximum similarity matrix 1 are selected and connected by an edge to form a new cluster 2. A new similarity matrix 2 is formed based on this new cluster 2, and so on. -1 compression of the similarity matrix yields the graph feature structure.

[0015] Furthermore, the two clusters are connected by an edge, specifically: when the number of feature centroid nodes in a cluster is less than or equal to 2, one representative node is selected for edge connection; when the number of feature centroid nodes in a cluster is greater than 2, two representative nodes are selected for edge connection.

[0016] Furthermore, the selection of representative nodes is specifically as follows: assuming the two clusters are cluster 1 and cluster 2; when both cluster 1 and cluster 2 select one representative node, the two feature centroid nodes with the highest similarity are selected from the two clusters as a whole as representative nodes; when both cluster 1 and cluster 2 select two representative nodes, the four feature centroid nodes with the maximum similarity of four edges are selected from the two clusters as representative nodes; when one cluster selects one representative node and the other selects two representative nodes, the three feature centroid nodes with the maximum similarity of two edges are selected from the two clusters as representative nodes.

[0017] Furthermore, the message passing graph neural network model includes: a first branch module, a second branch module, an adaptive feature fusion layer, and two fully connected layers; the working condition graph feature set is used as the input of the first branch module, and the output of the first branch module is used as the input of the second branch module; the output of the first branch module and the output of the second branch module are adaptively fused through the adaptive feature fusion layer; finally, the two fully connected layers are used for dimensionality reduction output to realize the anomaly detection of equipment under unknown working conditions.

[0018] Furthermore, the first branch module and the second branch module have the same structure, both including a graph convolutional layer, a batch normalization layer, a ReLU activation layer, a message passing layer, a pooling layer, and a readout layer connected in sequence. The inputs to the first branch module and the second branch module are processed by the graph convolutional layer to extract spatial graph features. Simultaneously, the graph convolutional layer aggregates neighbor node features, and the extracted spatial graph features and aggregated neighbor node features are combined into the feature centroid node to obtain the comprehensive features of the working condition graph. Next, the output data of the graph convolutional layer is normalized by the batch normalization layer. Then, the ReLU activation layer learns complex nonlinear mapping relationships in the data. Next, the message passing layer transmits messages between feature centroid nodes and updates nodes, capturing the local topological structure between nodes. Then, the pooling layer performs dimensionality reduction while retaining important features. Finally, the readout layer merges the features of all nodes into graph-level features, achieving a global representation of the entire working condition graph.

[0019] The beneficial effects of this invention are:

[0020] This invention enables the detection of equipment anomalies without the need for operating condition labels. Specifically, it achieves autonomous construction of an adaptive graph topology through a feature centroid hierarchical clustering method. This guides the message-passing graph neural network model to break through its dependence on specific operating condition features and learn the essential features that are stable across operating conditions. As a result, it exhibits excellent generalization ability and detection accuracy when facing new and unknown operating conditions. This provides a new and effective technical path to solve the problem of "model failure when operating conditions change". Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 These are time-domain diagrams and spectrum diagrams of two types of state data provided according to embodiments of the present invention;

[0023] Figure 3 This is a flowchart illustrating the construction process of the graphical feature structure of the present invention;

[0024] Figure 4 This is a schematic diagram of the message passing graph neural network model structure of the present invention;

[0025] Figure 5 A visualization of the prediction results of the message passing graph neural network model. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0027] Example 1: As Figures 1-5 As shown, according to a first aspect of the present invention, a method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks is provided, comprising:

[0028] Step 1: Obtain vibration data of equipment under unknown operating conditions; the vibration data of equipment under unknown operating conditions includes normal state data and abnormal state data.

[0029] Step 2: Preprocess the normal state data and abnormal state data of the equipment under unknown operating conditions to obtain low-frequency data of different layers under each state; compare the differences between the spectrum diagrams of the normal state data and the abnormal state data to determine the frequency range of the differences; select the layer with the largest coverage of the frequency range of differences from the low-frequency data of different layers under each state as the analysis layer.

[0030] Furthermore, the preprocessing of the normal state data and abnormal state data of the unknown operating condition equipment is as follows: the normal state data and abnormal state data of the unknown operating condition equipment are respectively normalized and n-level wavelet decomposition is performed. The wavelet basis function used for wavelet decomposition is db4 to obtain low-frequency data of different layers under each state, namely A1-A4 layer data.

[0031] Step 3: Based on the preset number of segmentation points, segment the data corresponding to the analysis layer in both states to obtain multiple segmented data; extract a first preset number of features from each segmented data from multiple perspectives; group the segmented data in each state according to a second preset number, and construct feature nodes for the same feature in each grouped segmented data, so that each feature has a second preset number of dimensions. A number of feature nodes; feature nodes are constructed by extracting a first preset number of features. × A multimodal feature node set; among which, This is the first preset quantity.

[0032] Step 4: Based on the multimodal feature node set, a graph construction method using feature centroid hierarchical clustering is used to first construct feature centroid nodes, and then a graph feature structure is constructed based on the feature centroid nodes, ultimately generating the graph. The feature set of the working condition diagram is composed of the feature structure of the diagram.

[0033] Further, step four specifically involves: selecting the feature vectors of the second preset number of dimensions of each feature node in the multimodal feature node set using the K-means clustering method. Construct feature centroid nodes from individual centroids, through... × A characteristic centroid node is formed × The set of feature centroid nodes; based on each row of the feature centroid node set. Using the centroid nodes of each feature as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each centroid node to construct the graph feature structure. Each graph feature structure forms a working condition graph feature set.

[0034] Furthermore, the set of characteristic centroid nodes for each row Using the feature centroid nodes as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each feature centroid node to construct the graph feature structure. Specifically, the similarity matrix of each row in the feature centroid node set is used as the basis for graph construction. Based on the feature centroid nodes, each feature centroid node is treated as a cluster. Only the two clusters with the maximum similarity matrix are selected and connected by an edge to form a new cluster 1. A new similarity matrix 1 is formed based on this new cluster 1. Similarly, based on the new similarity matrix 1, only the two clusters with the maximum similarity matrix 1 are selected and connected by an edge to form a new cluster 2. A new similarity matrix 2 is formed based on this new cluster 2, and so on. -1 compression of the similarity matrix yields the graph feature structure.

[0035] Furthermore, the two clusters are connected by an edge, specifically: when the number of feature centroid nodes in a cluster is less than or equal to 2, one representative node is selected for edge connection; when the number of feature centroid nodes in a cluster is greater than 2, two representative nodes are selected for edge connection.

[0036] Furthermore, the selection of representative nodes is specifically as follows: assuming the two clusters are cluster 1 and cluster 2; when both cluster 1 and cluster 2 select one representative node, the two feature centroid nodes with the highest similarity are selected from the two clusters as a whole as representative nodes; when both cluster 1 and cluster 2 select two representative nodes, the four feature centroid nodes with the maximum similarity of four edges are selected from the two clusters as representative nodes; when one cluster selects one representative node and the other selects two representative nodes, the three feature centroid nodes with the maximum similarity of two edges are selected from the two clusters as representative nodes.

[0037] Step 5: Build a message passing graph neural network model with integrated message passing mechanism; use the working condition graph feature set to train the message passing graph neural network model to obtain a trained message passing graph neural network model to detect the status of unknown working conditions equipment.

[0038] Furthermore, the message passing graph neural network model includes: a first branch module, a second branch module, an adaptive feature fusion layer, and two fully connected layers; the working condition graph feature set is used as the input of the first branch module, and the output of the first branch module is used as the input of the second branch module; the output of the first branch module and the output of the second branch module are adaptively fused through the adaptive feature fusion layer; finally, the two fully connected layers are used for dimensionality reduction output to realize the anomaly detection of equipment under unknown working conditions.

[0039] Furthermore, the first and second branch modules have the same structure, both including a graph convolutional layer, a batch normalization layer, a ReLU activation layer, a message passing layer, a pooling layer, and a readout layer connected in sequence. The inputs to the first and second branch modules are processed by the graph convolutional layer to extract spatial graph features. Simultaneously, the graph convolutional layer aggregates neighbor node features, and the extracted spatial graph features and aggregated neighbor node features are combined into the feature centroid node to obtain the comprehensive features of the working condition graph. Next, the batch normalization layer normalizes the output data of the graph convolutional layer. Then, the ReLU activation layer learns complex nonlinear mapping relationships in the data. Next, the message passing layer transmits messages between feature centroid nodes and updates nodes, capturing the local topological structure between nodes. Then, the pooling layer performs dimensionality reduction while retaining important features. Finally, the readout layer merges the features of all nodes into graph-level features, achieving a global representation of the entire working condition graph. Specifically, as shown... Figure 4 As shown, the first branch module includes a graph convolutional layer 1, a batch normalization layer 1, a ReLU activation layer 1, a message passing layer 1, a pooling layer 1, and a readout layer 1 connected in sequence; the second branch module includes a graph convolutional layer 2, a batch normalization layer 2, a ReLU activation layer 2, a message passing layer 2, a pooling layer 2, and a readout layer 2 connected in sequence.

[0040] According to a second aspect of the present invention, a system for detecting abnormal equipment status under unknown operating conditions based on graph neural networks is provided, comprising a module of the method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks described in any one of the above embodiments.

[0041] Example 2: As Figures 1-5 As shown in the accompanying drawings and experimental data, optional embodiments of the present invention will be described below. A method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks includes:

[0042] Step 1, Data Acquisition and Generation: This embodiment uses data from the stand end of an industrial roughing mill as an example. The roughing mill includes a stand, gearbox, and bearings. A single-axis accelerometer installed at a measuring point on the stand end is used, combined with a data acquisition device to collect vibration data. The sampling frequency is set to 25.6kHz. The data includes normal and abnormal state data of equipment under unknown operating conditions, categorized into two types, such as... Figure 2 The images shown, from top to bottom, are the time-domain plot of normal state data, the spectrum plot of normal state data, the time-domain plot of abnormal state data, and the spectrum plot of abnormal state data for equipment under unknown operating conditions.

[0043] Step two: Normalize the normal state data and abnormal state data. Then, perform wavelet decomposition on the normalized data at four levels with the wavelet basis function "db4" to obtain the low-frequency data of different levels under each state, namely the A1-A4 level data. Based on the spectrum diagrams of the normal state data and the abnormal state data, compare the differences to determine the frequency range of differences. Select the layer with the largest coverage of the frequency range of differences from the data of different levels under each state as the analysis layer.

[0044] In this embodiment, the difference comparison based on the spectrum diagrams of normal state data and abnormal state data specifically involves: In the spectrum diagrams, if the amplitude difference between the two values ​​falls within a preset range across consecutive frequency intervals, then it is considered a non-difference frequency range; the intervals within the frequency band range of the spectrum diagram that are not non-difference frequency ranges are taken as the difference frequency ranges; Figure 2 As shown in the example, the frequency range of the spectrum is 0Hz-12800Hz. When the frequency is 2531.86Hz, the normal data amplitude is 0.00424851m / s. 2 The abnormal state data amplitude is 0.000293 m / s 2 The amplitude difference is 0.00395551 m / s 2 The non-differential frequency range is (2531.86Hz, 12800Hz); the differential frequency range is [0, 2531.86Hz]. The preset range can be [0, 0.0039m / s]. 2 In the above description, the frequency range of layer A1 is 0-6400Hz, layer A2 is 0-3200Hz, layer A3 is 0-1600Hz, and layer A4 is 0-800Hz. Layer A3 covers the largest range of frequency differences; therefore, layer A3 is selected as the analysis layer.

[0045] Step 3: Based on the preset number of segmentation points, the data corresponding to the analysis layer in both states are segmented into 1024 points with 50% overlap, resulting in multiple segmented data. Features of each segmented data are extracted from multiple perspectives using a first preset number of features. The segmented data in each state are grouped according to a second preset number of features, and feature nodes are constructed for the same feature in each group of segmented data. This ensures that each feature has a second preset number of dimensions. A number of feature nodes; feature nodes are constructed by extracting a first preset number of features. × A multimodal feature node set; among which, This is the first preset quantity;

[0046] In this embodiment, the 3,187,200 data points corresponding to the analysis layer in the two states are segmented into 1,024 points with 50% overlap, resulting in 6,224 segmented data points. Ten features are extracted from the segmented data points in the two states from four perspectives: time domain, frequency domain, time-frequency domain, and morphology (i.e., the first preset number is 10). The segmented data points in each state are grouped according to the second preset number 30. Feature nodes are constructed for the same feature in each group of segmented data points, i.e., each feature node is a 30-dimensional feature vector composed of the same feature at different times (it needs to be explained). It is clear that the 6224 segmented data are processed into feature nodes in groups of 30, and the last group with less than 30 is discarded and does not participate in the construction of feature nodes. By constructing feature nodes for the same feature of each group of segmented data, each feature is constructed into 207 feature nodes with 30 dimensions. Finally, the feature nodes constructed from the first preset number of extracted features form a 207×10 multimodal feature node set. The features are specifically: kurtosis, effective value, interquartile range, frequency band energy distribution, spectral bandwidth, logarithmic energy entropy, wavelet energy, coefficient energy sum, wavelet entropy, and waveform entropy.

[0047] It should be noted that selecting the A3 layer for feature extraction has the following advantages: The A3 layer can provide a multi-scale representation of the signal and capture the low-frequency components of the signal, which usually contain the main trends and periodic features of the signal; at the same time, the A3 layer can effectively smooth the signal and improve the signal quality by removing high-frequency noise components; in addition, the A3 layer has a relatively small number of wavelet coefficients, which not only improves computational efficiency but also enhances the interpretability of features, making these features exhibit good performance and adaptability in anomaly detection tasks under unknown conditions, and improving the accuracy, robustness and efficiency of detection.

[0048] The formulas for the 10 features mentioned above are shown in Table 1:

[0049] Table 1 Calculation formulas for each feature

[0050]

[0051] in, This represents the number of data points processed in each segment, with a value of 1024. For the first One data point; , They are the 75th and 25th percentiles, respectively; The mean; Standard deviation; For the frequency spectrum signal The amplitude at that point; This represents the total number of frequency points, with a value of 512. This represents the number of frequency bands, with a value of 5. For the first One frequency band; For frequency The energy normalized frequency at that point is given by the formula: ; The total number of wavelet coefficients. =128; These are wavelet coefficients; The normalized frequency of the wavelet coefficients is given by the formula: ; The normalized frequency of the signal amplitude is given by the formula: .

[0052] Step 4: Construction of the feature set of the working condition diagram.

[0053] like Figure 3 As shown, for each feature node in the multimodal feature node set, three centroids are selected from the feature vectors of the second preset number of dimensions according to the K-means clustering method to construct feature centroid nodes, forming a 207×10 feature centroid node set.

[0054] Each row of the feature centroid node set Using the centroid nodes of each feature as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each centroid node to construct the graph feature structure. Each graph feature structure forms a working condition graph feature set; that is, when a 10×10 similarity matrix is ​​compressed to 1×1, a graph feature structure that can represent working condition information and feature information is successfully constructed.

[0055] The rows of the feature centroid node set Using the feature centroid nodes as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each feature centroid node to construct the graph feature structure. Specifically, the similarity matrix of each row in the feature centroid node set is used as the basis for graph construction. Based on the feature centroid nodes, each feature centroid node is treated as a cluster. Only the two clusters with the maximum similarity matrix are selected and connected by an edge to form a new cluster 1. A new similarity matrix 1 is formed based on this new cluster 1. Similarly, based on the new similarity matrix 1, only the two clusters with the maximum similarity matrix 1 are selected and connected by an edge to form a new cluster 2. A new similarity matrix 2 is formed based on this new cluster 2, and so on. -1 compression of the similarity matrix constructs a graph feature structure from 10 feature centroid nodes, which can represent working condition information and feature information.

[0056] The similarity matrix is ​​specifically as follows:

[0057] First, calculate the similarity value between each pair of feature centroid nodes in each row of the feature centroid node set, using the following formula:

[0058] ;

[0059] in, For the feature centroid node set of the first In the middle of the line The feature centroid node and the first The similarity value of each feature centroid node. The number of centroids =3; , These are the feature centroid node sets of the th In the middle of the line The feature centroid node and the first The first feature centroid node The center of mass; For the feature centroid node set of the first In the middle of the line The feature centroid node and the first The first feature centroid node The Euclidean distance between the centroids; where... =1,2,..., .

[0060] Secondly, a 10×10 similarity matrix is ​​generated from the 10 feature centroid nodes of each row in the feature centroid node set, using the following formula:

[0061] ;

[0062] in, For the first characteristic centroid node set The 10 feature centroid nodes of the row generate a similarity matrix with a dimension of 10×10; This is the number of feature centroid nodes in each row of the feature centroid node set, i.e., the first preset number, which is 10.

[0063] The two clusters are connected by an edge, specifically: when the number of feature centroid nodes in a cluster is less than or equal to 2, one representative node is selected for edge connection; when the number of feature centroid nodes in a cluster is greater than 2, two representative nodes are selected for edge connection.

[0064] The selection of representative nodes is as follows: Assuming the two clusters are cluster 1 and cluster 2; when both cluster 1 and cluster 2 select one representative node, the two feature centroid nodes with the highest similarity are selected from the two clusters as a whole as representative nodes (i.e., two feature centroid nodes are selected from the set of feature centroid nodes formed by cluster 1 and cluster 2 as representative nodes for edge connection); when both cluster 1 and cluster 2 select two representative nodes, the four feature centroid nodes with the maximum similarity of four edges are selected from the two clusters as representative nodes; when one cluster selects one representative node and the other selects two representative nodes, the three feature centroid nodes with the maximum similarity of two edges are selected from the two clusters as representative nodes.

[0065] The advantages of this invention in constructing a graph feature set representing working condition information and feature information through a feature centroid hierarchical clustering method are as follows: it does not rely on specific data distributions or preset anomaly models, possesses high adaptability, and can effectively handle nonlinear and high-dimensional data; at the same time, it avoids the limitation of KNN graph construction in ordinary graph neural networks ignoring global structure, and by capturing the global distribution pattern of features, it can more robustly uncover the inherent distinguishing rules between steady-state and shock states, thereby enhancing the flexibility and adaptability of message-passing graph neural network models; in addition, by calculating feature centroids, the data dimensionality is effectively reduced, which not only improves clustering efficiency, but also effectively captures and represents the complex relationships between data by constructing graph feature structures; finally, this method also enhances the interpretability of anomaly detection results, thereby helping to quickly identify and respond to anomaly patterns in the data, and improving the accuracy and efficiency of detection.

[0066] For example, based on the data given in steps two and three, the number of data points in each segment of processed data is 1024, and the number of segments is 6224. The above 10 features are extracted from each segment of processed data. Every 30 segments of processed data with the same feature form a 30-dimensional feature node, that is, each feature generates 207 feature nodes. 10 feature nodes construct a graph feature structure. Therefore, 207 graph feature structures can be constructed in the end, which can be used as the working condition graph feature set.

[0067] Step 5: Construction and training of the message passing graph neural network model.

[0068] like Figure 4 As shown, the message-passing graph neural network model includes: a first branch module, a second branch module, an adaptive feature fusion layer, and two fully connected layers; the first branch module and the second branch module have the same structure, both including a graph convolutional layer, a batch normalization layer, a ReLU activation layer, a message-passing layer, a pooling layer, and a readout layer connected in sequence (specifically: the first branch module includes a graph convolutional layer 1, a batch normalization layer 1, a ReLU activation layer 1, a message-passing layer 1, a pooling layer 1, and a readout layer 1 connected in sequence; the second branch module includes a graph convolutional layer 2, a batch normalization layer 2, a ReLU activation layer 2, a message-passing layer 2, a pooling layer 2, and a readout layer 2 connected in sequence). The specific settings are as follows: the input dimension is 320×30; the graph convolutional layer dimension is 128×128; the batch normalization layer dimension is 128×128; the ReLU activation layer dimension is 128×128; the message passing layer dimension is 128×128; the pooling layer dimension is 128×128; the readout layer dimension is 128×128; the feature fusion layer dimension is 128×128; and the two fully connected layers have dimensions of 128×64 and 64×number of data types, respectively. In this example, the number of data types is 2.

[0069] Next, graph data augmentation techniques involving edge perturbation and feature perturbation are applied to the constructed working condition graph feature set. The augmented feature set is then divided into training and testing sets in an 8:2 ratio and labeled, where 0 represents normal state data of equipment under unknown working conditions and 1 represents abnormal state data. The Adam optimizer is used during training, with cross-entropy loss function, batch size set to 32, training epochs set to 100, and learning rate set to 0.005. The message-passing graph neural network model is trained and validated using the training set to obtain the optimal model with the best parameters. The model is then tested using the testing set, thus achieving anomaly detection for unknown working conditions.

[0070] Next, the model was trained using the training set data from the industrial roughing mill stand end in this embodiment of the invention to obtain the optimal model accuracy under the optimal parameters. The confusion matrix was then used to visualize and analyze the model's test set data results, as shown below. Figure 5 As shown, the horizontal axis represents the predicted label, the vertical axis represents the true label, and the value represents the number of correctly predicted samples. Figure 5 As can be seen, all normal and abnormal state data were correctly identified.

[0071] As can be seen from the above technical solution, the present invention eliminates the reliance on complex operating condition labels. It does not require prior knowledge or labeling of specific operating condition information (such as speed, load) of the data. Only the status label of the data (normal, abnormal, etc.) is required to realize the status detection of equipment with unknown operating conditions.

[0072] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks, characterized in that, include: Step 1: Obtain vibration data of equipment under unknown operating conditions; the vibration data of equipment under unknown operating conditions includes normal state data and abnormal state data; Step 2: Preprocess the normal state data and abnormal state data of the equipment under unknown operating conditions to obtain low-frequency data of different layers under each state; compare the differences between the spectrum diagrams of the normal state data and the abnormal state data to determine the frequency range of the differences; select the layer with the largest coverage of the frequency range of differences from the low-frequency data of different layers under each state as the analysis layer. Step 3: Based on the preset number of segmentation points, segment the data corresponding to the analysis layer in both states to obtain multiple segmented data; extract a first preset number of features from each segmented data from multiple perspectives; group the segmented data in each state according to a second preset number, and construct feature nodes for the same feature in each grouped segmented data, so that each feature has a second preset number of dimensions. A number of feature nodes; feature nodes are constructed by extracting a first preset number of features. × A multimodal feature node set; among which, This is the first preset quantity; Step 4: Based on the multimodal feature node set, a graph construction method using feature centroid hierarchical clustering is used to first construct feature centroid nodes, and then a graph feature structure is constructed based on the feature centroid nodes, ultimately generating the graph. A set of working condition diagram features composed of individual diagram feature structures; Step 5: Build a message passing graph neural network model with integrated message passing mechanism; use the working condition graph feature set to train the message passing graph neural network model to obtain a trained message passing graph neural network model to detect the status of unknown working conditions equipment; Step four specifically involves: selecting the feature vectors of the second preset number of dimensions of each feature node in the multimodal feature node set using the K-means clustering method. Construct feature centroid nodes from individual centroids, through... × A characteristic centroid node is formed × The set of characteristic centroid nodes; Each row of the feature centroid node set Using the centroid nodes of each feature as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each centroid node to construct the graph feature structure. Each graph feature structure forms a working condition graph feature set; The rows of the feature centroid node set Using the feature centroid nodes as the basis for graph construction, hierarchical clustering is used to calculate the compressed similarity matrix of each feature centroid node to construct the graph feature structure. Specifically, the similarity matrix of each row in the feature centroid node set is used as the basis for graph construction. Based on the feature centroid nodes, each feature centroid node is treated as a cluster. Only the two clusters with the maximum similarity matrix are selected and connected by an edge to form a new cluster 1. A new similarity matrix 1 is formed based on this new cluster 1. Similarly, based on the new similarity matrix 1, only the two clusters with the maximum similarity matrix 1 are selected and connected by an edge to form a new cluster 2. A new similarity matrix 2 is formed based on this new cluster 2, and so on. -1 compression of the similarity matrix yields the graph feature structure.

2. The method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks according to claim 1, characterized in that, The preprocessing of normal and abnormal state data of unknown operating equipment is as follows: normalization and n-level wavelet decomposition are performed on the normal and abnormal state data of unknown operating equipment respectively. The wavelet basis function used for wavelet decomposition is db4 to obtain low-frequency data of different levels under each state.

3. The method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks according to claim 1, characterized in that, The two clusters are connected by an edge, specifically: when the number of feature centroid nodes in a cluster is less than or equal to 2, one representative node is selected for edge connection; when the number of feature centroid nodes in a cluster is greater than 2, two representative nodes are selected for edge connection.

4. The method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks according to claim 3, characterized in that, The selection of representative nodes is as follows: Assuming the two clusters are cluster 1 and cluster 2; when both cluster 1 and cluster 2 select one representative node, the two feature centroid nodes with the highest similarity are selected from the two clusters as a whole as representative nodes; when both cluster 1 and cluster 2 select two representative nodes, the four feature centroid nodes with the maximum similarity of the four edges are selected from the two clusters as representative nodes; when one cluster selects one representative node and the other selects two representative nodes, the three feature centroid nodes with the maximum similarity of the two edges are selected from the two clusters as representative nodes.

5. The method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks according to claim 1, characterized in that, The message passing graph neural network model includes: a first branch module, a second branch module, an adaptive feature fusion layer, and two fully connected layers; the working condition graph feature set is used as the input of the first branch module, and the output of the first branch module is used as the input of the second branch module; the output of the first branch module and the output of the second branch module are adaptively fused through the adaptive feature fusion layer; finally, the two fully connected layers are used for dimensionality reduction output to realize the anomaly detection of equipment under unknown working conditions.

6. The method for detecting abnormal equipment status under unknown operating conditions based on graph neural networks according to claim 5, characterized in that, The first and second branch modules have identical structures, each including a graph convolutional layer, a batch normalization layer, a ReLU activation layer, a message passing layer, a pooling layer, and a readout layer connected in sequence. The inputs to the first and second branch modules are processed by the graph convolutional layer to extract spatial graph features. Simultaneously, the graph convolutional layer aggregates neighbor node features, and the extracted spatial graph features and aggregated neighbor node features are combined into the feature centroid node to obtain the comprehensive features of the working condition graph. Next, the output data of the graph convolutional layer is normalized by the batch normalization layer. Then, the ReLU activation layer learns complex nonlinear mapping relationships in the data. Next, the message passing layer transmits messages between feature centroid nodes and updates nodes, capturing the local topological structure between nodes. Finally, the pooling layer performs dimensionality reduction while retaining important features. Next, the features of all nodes are merged into graph-level features through the readout layer to achieve a global representation of the entire working condition diagram.

Images