Abnormality diagnosis model construction method, abnormality diagnosis method and related device

Abstract

The invention discloses a construction method of an anomaly diagnosis model, an anomaly diagnosis method, a computer program product, electronic equipment and a computer readable storage medium, and the construction method of the anomaly diagnosis model comprises the steps: obtaining vibration signal data collected by a vibration sensor array in a historical time period, the real operation state of the vibration sensor array corresponding to the vibration signal data is obtained; based on the vibration signal data, a heterogeneous graph associated with the vibration sensor array is constructed, nodes in the heterogeneous graph are vibration sensors, node types are operation states of the vibration sensors, and a side relation is connection weights among different vibration sensors in the vibration sensor array; and taking the heterogeneous graph as a training data set to train an anomaly diagnosis model, wherein the anomaly diagnosis model is obtained based on error training between the output anomaly diagnosis result and the real operation state. The diagnosis accuracy of the abnormal state of the vibration sensor can be improved.

Description

Technical Field

[0001] The present application relates to the field of anomaly diagnosis, and particularly to a method for constructing an anomaly diagnosis model, an anomaly diagnosis method, a computer program product, an electronic device, and a computer-readable storage medium. Background Art

[0002] On some devices that can generate regular vibrations, such as the flat wire motor which is a key component in industrial applications such as new energy vehicles, multiple vibration sensors are generally configured. The multiple vibration sensors can collect the vibration signals generated when the flat wire motor vibrates, and the operating state of the flat wire motor can be evaluated through the collected vibration signals.

[0003] The vibration sensor itself may have installation anomalies, such as position deviation or looseness, etc., which will cause the vibration signals of the flat wire motor collected by the vibration sensor to be inaccurate, thereby affecting the accurate evaluation of the operating state of the flat wire motor.

[0004] In related technologies, the installation of the vibration sensor is often analyzed by manually analyzing the vibration signals collected by the vibration sensor itself. However, the current diagnostic method for the abnormal state of the vibration sensor is inefficient and has low accuracy. Summary of the Invention

[0005] In view of the above technical problems, the present application provides a method for constructing an anomaly diagnosis model, an anomaly diagnosis method, a computer program product, an electronic device, and a computer-readable storage medium. The technical solutions are as follows: According to a first aspect of the present application, there is provided a method for constructing an anomaly diagnosis model, which is applicable to the diagnosis scenario of a vibration sensor array arranged on a target device. The target device can generate regular vibrations, and the vibration sensor array includes multiple vibration sensors. The method includes: Obtaining the vibration signal data collected by the vibration sensor array during a historical time period, and the true operating state corresponding to the vibration signal data of the vibration sensor array. The true operating state includes a normal operating state or an abnormal operating state; Based on the vibration signal data, constructing a heterogeneous graph associated with the vibration sensor array. The nodes in the heterogeneous graph are vibration sensors, the node type is the operating state of the vibration sensor, and the edge relationship is the connection weight between different vibration sensors in the vibration sensor array; Using the heterogeneous graph as a training data set to train an anomaly diagnosis model. The anomaly diagnosis model is trained based on the error between the output anomaly diagnosis result and the true operating state.

[0006] According to a second aspect of the present application, there is provided an anomaly diagnosis method applicable to a diagnostic scenario of a vibration sensor array arranged on a target device, where the target device is capable of generating vibrations with a fixed pattern, and the vibration sensor array includes a plurality of vibration sensors; the method includes: Obtain vibration signal data collected in real time by the vibration sensor array; Based on the vibration signal data, construct a heterogeneous graph associated with the vibration sensor array, where the nodes in the heterogeneous graph are vibration sensors, the node type is the operating state of the vibration sensor, and the edge relationship is the connection weight between different vibration sensors in the vibration sensor array; Input the heterogeneous graph into the anomaly diagnosis model as described in the first aspect, and output an anomaly diagnosis result for the vibration sensor array through the anomaly diagnosis model.

[0007] According to a third aspect of the present application, there is provided a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the method as described in the first aspect or the second aspect.

[0008] According to a fourth aspect of the present application, there is provided an electronic device, which includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to implement the method as described in the first aspect or the second aspect.

[0009] According to a fifth aspect of the present application, there is provided a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps in the method as described in the first aspect or the second aspect.

[0010] The technical solution provided by this application is directed to the diagnostic scenario of a vibration sensor array arranged on a target device capable of generating vibrations with a fixed pattern. An heterogeneous graph is constructed using the vibration signal data historically collected by the vibration sensor array. In this heterogeneous graph, vibration sensors are used as nodes, the operating states of the vibration sensors are used as node types, and the connection weights between different vibration sensors in the vibration sensor array are used as edge relationships. The constructed heterogeneous graph is used as a training data set, and the true operating state corresponding to the vibration signal data historically collected by the vibration sensor array is used as a label to train an anomaly diagnosis model, obtaining a trained anomaly diagnosis model to diagnose the abnormal state of the vibration sensor array. The construction of the heterogeneous graph fully considers the different operating states of different vibration sensors and the connection weights between different vibration sensors, and can more accurately capture the complex relationship between the abnormal state of the vibration sensor and the vibration signal, thereby improving the diagnostic accuracy of the abnormal state of the vibration sensor. The abnormal diagnosis result of the vibration sensor array is directly output through the constructed abnormal diagnosis model, improving the diagnostic efficiency of the abnormal state of the vibration sensor.

[0011] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the drawings in the following description are only some embodiments recorded in this application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0013] Figure 1 is a schematic diagram of a sensor installation scenario according to an embodiment of this application; Figure 2 is a flowchart of a method for constructing an anomaly diagnosis model according to an embodiment of this application; Figure 3 is a schematic diagram of a node type combination according to an embodiment of this application; Figure 4 is a schematic diagram of the structure of an anomaly diagnosis model according to an embodiment of this application; Figure 5 is a schematic diagram of the structure of an anomaly diagnosis model according to an embodiment of this application; Figure 6 is a flowchart of an anomaly diagnosis method according to an embodiment of this application; Figure 7 is a schematic diagram of the structure of an electronic device according to an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] To enable those skilled in the art to better understand the technical solutions in this application, the following will describe the technical solutions in the embodiments of this application in detail with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application shall fall within the scope of protection of this application.

[0015] On some devices that can generate fixed regular vibrations, such as the flat wire motor which is a key component in industrial applications such as new energy vehicles, the stability of its operating state directly affects the performance and safety of the entire system. During the vibration test of the flat wire motor, the vibration sensor plays a crucial role. It can monitor the vibration of the flat wire motor in real time and provide data support for the state evaluation of the motor.

[0016] As Figure 1 shown, multiple vibration sensors are generally configured on the flat wire motor. The multiple vibration sensors can collect the vibration signals generated when the flat wire motor vibrates, and the operating state of the flat wire motor can be evaluated through the collected vibration signals. The vibration sensor itself may have abnormal installations, such as abnormal position deviation or looseness, which will lead to inaccurate vibration signals of the flat wire motor collected by the vibration sensor, thus affecting the accurate evaluation of the operating state of the flat wire motor. For example: at different positions on the flat wire motor, the vibration response accuracy of the vibration sensor to the flat wire motor varies greatly. If the vibration sensor deviates from the predetermined position, the collected vibration signal will change significantly; and if the vibration sensor becomes loose, it may cause the resonance frequency of the vibration sensor and the structure to shift, affecting the authenticity of the collected vibration signal.

[0017] In related technologies, the installation of the vibration sensor is often analyzed by manually analyzing the vibration signal itself collected by the vibration sensor. For example, a simple analysis based on a single feature or statistical method, such as the time-domain amplitude or frequency-domain peak value (such as peak frequency analysis) of the vibration signal. However, the current diagnostic method for the abnormal state of the vibration sensor is inefficient and has low accuracy.

[0018] To address the above problems, this application provides a method for constructing an abnormal diagnosis model, which is applicable to the diagnostic scenario of a vibration sensor array arranged on a target device. The target device can generate fixed regular vibrations, and the vibration sensor array includes multiple vibration sensors. It can improve the diagnostic accuracy of the abnormal state of the vibration sensor and improve the diagnostic efficiency of the abnormal state of the vibration sensor. As Figure 2 shown, the method includes the following steps: S201. Obtain the vibration signal data collected by the vibration sensor array within a historical time period, and the true operating state of the vibration sensor array corresponding to the vibration signal data.

[0019] The true operating state includes a normal operating state or an abnormal operating state.

[0020] S202. Based on the vibration signal data, construct a heterogeneous graph associated with the vibration sensor array.

[0021] The nodes in the heterogeneous graph are vibration sensors, the node type is the operating state of the vibration sensor, and the edge relationship is the connection weight between different vibration sensors in the vibration sensor array.

[0022] S203. Use the heterogeneous graph as a training data set to train an anomaly diagnosis model.

[0023] The anomaly diagnosis model is trained based on the error between the output anomaly diagnosis result and the true operating state.

[0024] The technical solution provided in the embodiments of the present application is directed to the diagnosis scenario of a vibration sensor array arranged on a target device that can generate vibrations with a fixed pattern. A heterogeneous graph is constructed using the vibration signal data historically collected by the vibration sensor array. In this heterogeneous graph, the vibration sensors are used as nodes, the operating states of the vibration sensors are used as node types, and the connection weights between different vibration sensors in the vibration sensor array are used as edge relationships. The constructed heterogeneous graph is used as a training data set, and the true operating state corresponding to the vibration signal data historically collected by the vibration sensor array is used as a label to train an anomaly diagnosis model, obtaining a trained anomaly diagnosis model to diagnose the abnormal state of the vibration sensor array. The construction of the heterogeneous graph fully considers the different operating states of different vibration sensors and the connection weights between different vibration sensors, and can more accurately capture the complex relationship between the abnormal state of the vibration sensor and the vibration signal, thereby improving the diagnosis accuracy of the abnormal state of the vibration sensor. The anomaly diagnosis result of the vibration sensor array is directly output through the constructed anomaly diagnosis model, improving the diagnosis efficiency of the abnormal state of the vibration sensor.

[0025] It can be understood that the vibration signal data collected by the vibration sensor array within a historical time period may refer to the vibration signal data collected by each vibration sensor in the vibration sensor array respectively; the true operating state of the vibration sensor array corresponding to the vibration signal data may refer to the true operating states of each vibration sensor respectively represented by the vibration signal data collected by each vibration sensor in the vibration sensor array within the above historical time period.

[0026] As an example, the above historical time period can be the past year, the past month, the past week, or the past day, or it can be other past time periods, which are not specifically limited herein. As another example, the above vibration signal data can be time series data, that is, the vibration signals are arranged in chronological order within the historical time period.

[0027] It can be understood that the true operating state corresponding to the above vibration sensor array and the vibration signal data serves as the label of the above training data set. This true operating state can be obtained through various means. As an example, the above true operating state can be obtained from the maintenance records of the above vibration sensor array within the historical time period, or it can be obtained through other means, which are not specifically limited herein.

[0028] There can be various specific implementations of the abnormal operating state included in the above true operating state. As an example, the abnormal operating state can include: the vibration sensors in the vibration sensor array are in a position deviation state, or the vibration sensors in the vibration sensor array are in a loose state. It can be understood that the heterogeneous graph associated with the above vibration sensor array can contain multiple nodes, and each node can respectively represent each vibration sensor in the vibration sensor array, and the attributes of each node can have various specific implementations. As an example, the attributes of each node can at least include the vibration signal characteristics corresponding to the vibration sensor, the position deviation characteristics of the vibration sensor, and the loose characteristics of the vibration sensor.

[0029] As an example, the above vibration signal characteristics can be determined from the vibration signal data collected by the vibration sensor, and the above position deviation characteristics and the above loose characteristics can be determined from the above true operating state.

[0030] It can be understood that the heterogeneous graph associated with the above vibration sensor array can contain various node types. The node type is the operating state of the vibration sensor, and the operating state of the vibration sensor can be preset to three possible states in this heterogeneous graph, including: the vibration sensor is in a normal operating state, the vibration sensor is in a position deviation state, and the vibration sensor is in a loose state. Then, there can be three node types corresponding to the three operating states in this heterogeneous graph.

[0031] It is worth noting that the above introduction to the specific implementation of the node type is only an exemplary display. In actual applications, other specific implementations are not excluded, which are not limited herein.

[0032] The above edge relationship can have multiple specific implementations. It can be understood that the edge relationship is the association relationship (i.e., connection weight) between different vibration sensors in the above vibration sensor array. As an example, this association relationship can be determined in the following way: perform similarity analysis on the vibration signals respectively corresponding to different vibration sensors in the vibration sensor array, that is, the vibration signal data collected respectively, and determine the connection weights between different vibration sensors based on the analysis results of the similarity analysis; determine the above association relationship based on the connection weights between different vibration sensors. For example, the higher the similarity of the vibration signals respectively corresponding to two vibration sensors, the greater the connection weight between these two vibration sensors, and the closer the association between these two vibration sensors; while the lower the similarity of the vibration signals respectively corresponding to two vibration sensors, the smaller the connection weight between these two vibration sensors, and the smaller the association between these two vibration sensors.

[0033] It should be noted that the above introduction to the specific implementation of the edge relationship is only an exemplary display. In actual applications, there may be other specific implementations, and no limitation is made thereto.

[0034] There are multiple ways to construct a heterogeneous graph associated with the above vibration sensor array. As an example, when the multiple node types included in the heterogeneous graph include the vibration sensor in a normal operating state, the vibration sensor in a position deviation state, or the vibration sensor in a loose state, different node types can be combined, and the composite relationship between different vibration sensors in the vibration sensor array can be determined based on the combination result, and a meta-path connecting different vibration sensors can be defined based on this composite relationship; based on the meta-path connecting different vibration sensors and the vibration signal data respectively collected by each vibration sensor in the vibration sensor array during the historical time period, the above heterogeneous graph is constructed. It should be noted that the above introduction to the construction method of the heterogeneous graph associated with the vibration sensor array is only an exemplary display. In actual applications, there may be other construction methods, and no limitation is made thereto.

[0035] As an example, the above meta-path can be defined by the following expression: (1) Where represents a node (i.e., a vibration sensor); , represents different nodes and between, that is, the composite relationship between different vibration sensors, to is the combination result of combining different node types, and ∘ represents the composite operator in this composite relationship.

[0036] Such as Figure 3As shown, taking three node types: the vibration sensor is in a normal operating state (such as Figure 3 "Instance1" in ), the vibration sensor is in a position deviation state (such as Figure 3 "Instance2" in ), and the vibration sensor is in a loose state (such as Figure 3 "Instance3" in ) as examples, by combining different node types, the following combination results can be obtained: ; ; ; ; ; .

[0037] The structure of the above abnormal diagnosis model can have multiple specific implementations. As an example, the abnormal diagnosis model can at least include a feature embedding layer, a self-attention network layer, a feature fusion layer, a feature extraction layer, and an output layer; the feature embedding layer is used to convert the vibration signal data corresponding to each node in the above heterogeneous graph into feature vectors in a unified dimension; the self-attention network layer is used to calculate the attention values between each node in the heterogeneous graph based on the meta-paths connecting different vibration sensors, and determine the association degree between each node based on the attention values; the feature fusion layer is used to fuse the above feature vectors corresponding to each node based on the association degree between each node to obtain a fusion result; the feature extraction layer is used to perform feature extraction based on the fusion result of the feature vectors corresponding to each node to obtain an extraction result; the output layer is used to output an abnormal diagnosis result based on the extraction result, and the abnormal diagnosis result includes that the vibration sensor is in a normal operating state, the vibration sensor is in a position deviation state, or the vibration sensor is in a loose state.

[0038] It should be noted that the above introduction to the specific implementation of the structure of the abnormal diagnosis model is only an exemplary display. In actual applications, there may be other specific implementations, which are not limited herein.

[0039] As an example, the above feature embedding layer can specifically be used to convert the vibration signal data corresponding to each node in the above heterogeneous graph into feature vectors in a unified dimension based on matrix transformation.

[0040] As another example, the formula for the self-attention network layer to calculate the attention values between each node in the heterogeneous graph can be as follows: (2) (3) (4) Among them, represents the attention function of the node pair (i, j) under the meta-path Φ, which can characterize the degree of association between nodes. For a given meta-path Φ, the weights assigned to different adjacent nodes can be shared. represents the sigmoid activation function, || represents the concatenation operation, and represents the node-level attention vector. As another example, after the self-attention network layer calculates the attention values between each pair of nodes in the heterogeneous graph, it can determine the weights between each pair of nodes based on the attention values between each pair of nodes, and determine the degree of association between each pair of nodes based on these weights.

[0041] As another example, the above-mentioned feature fusion layer is specifically used to fuse the feature vectors corresponding to each node by means of average pooling operation and non-linear transformation to obtain the fusion result of the feature vectors corresponding to each node. As another example, the above-mentioned average pooling operation and non-linear transformation can be implemented by the following formula: (5) Among them, in the process of multimodal fusion, different node types belong to different modalities, and multimodality means multiple node types. represents the fused feature representation under modality u. represents the feature set of modality u, and LeakyReLU represents the non-linear activation function. represents the d-dimensional representation of the adjacent information of embedding j in modality m. represents the weight transformation matrix, which can extract useful features after being trained.

[0042] As another example, the above-mentioned feature extraction layer is specifically used to extract features from the fusion result of the feature vectors corresponding to each node by means of graph convolution calculation to obtain the above-mentioned extraction result. As another example, the above-mentioned graph convolution calculation can be implemented by the following formula: (6) (7) Among them, represents the node representation or embedding output of the (k + 1)-th layer (for classification or prediction); represents the input feature matrix of the k-th layer (which can be regarded as the output of the previous layer, or initially the node features), and its dimension is , is the number of nodes, is the feature dimension; represents the trainable weight matrix of the k-th layer, and its dimension is ; Denotes the bias term of the k-th layer, whose dimension is ; Denotes normalizing the output of each node so that its output becomes a probability distribution (commonly used in classification tasks), which can be changed to ReLU or other non-linear functions for regression tasks; Denotes the symmetric normalized graph Laplacian, which is used for information propagation and balancing the adjacency structure; Denotes the adjacency matrix of the graph, with self-loops added (i.e., , where is the identity matrix), representing the connection relationship between nodes; Represents the degree matrix of the nodes, which is the result of diagonal summation of .

[0043] As Figure 4 shown, the left side can represent the fusion result obtained by the above-mentioned feature fusion layer based on the degree of association between each node, which fuses the above-mentioned feature vectors corresponding to each node. It can be represented by a hypernode. In a heterogeneous graph structure, a hypernode can be used to represent a virtual node that aggregates information of multiple modalities, multiple states, or incomplete data. Through the processing of subsequent hidden layers, three abnormal diagnosis results (normal operating state, position deviation state, loosening state) can be obtained.

[0044] The vibration signal data collected by the vibration sensor array can have multiple specific implementations. As an example, as Figure 5 shown, the vibration signal data has temporality, that is, it can be data representing the change of amplitude over time within a period. As another example, the vibration signal data itself can have multi-modal characteristics. For example, the vibration signal data can be data collected when the target device is in different working conditions. For another example, as Figure 5 shown, the vibration signal data can be data representing the change of amplitude over time in the tangential direction, or data representing the change of amplitude over time in the radial direction. Therefore, the specific implementation of the vibration signal data is not limited.

[0045] As Figure 5As shown, through hypermode embedding, the information aggregated in the hypernode can be encoded by the neural network in the anomaly diagnosis model and converted into a feature vector representation form that can be used for the calculation of the anomaly diagnosis model. The role of hypermode embedding in a specific application scenario of the embodiment of the present application can be to represent the multi-modal data of the fused vibration sensor (including the part that replaces the missing mode) as a unified feature vector, facilitating participation in the learning and propagation of the neural network in the anomaly diagnosis model together with other nodes. Through information embedding or feature vector representation, the original vibration time series signal (i.e., vibration signal data, such as acceleration, spectral features) can be encoded into a low-dimensional learnable representation vector by the neural network and used as the input of the graph neural network in the anomaly diagnosis model. Its role in a specific application scenario of the embodiment of the present application can be to convert the time series signal of the vibration sensor into a "node feature" form that can participate in graph learning, and then operations such as neighbor node aggregation and attention weighting can be performed.

[0046] There are various ways to train the above anomaly diagnosis model. As an example, the anomaly diagnosis model can be trained based on the error between the anomaly diagnosis result output by the anomaly diagnosis model and the above real operating state as the loss function.

[0047] As an example, the anomaly diagnosis model can include a classification model. The classification model can be used to output the above anomaly diagnosis results (for example, output three anomaly diagnosis results: the vibration sensor is in a normal operating state, the vibration sensor has a position deviation state, or the vibration sensor has a loosening state), and the error between the anomaly diagnosis result and the above real operating state can be used as the cross-entropy loss function to train the anomaly diagnosis model. As another example, the anomaly diagnosis model can also include a regression model. The regression model can be used to predict the degree of anomaly, and the predicted degree of anomaly can be compared with a set threshold for classification to output different anomaly diagnosis results, and the anomaly diagnosis model can be trained using loss functions such as mean squared error. As another example, the anomaly diagnosis model can also be based on the method of contrastive learning. By learning the embedding distance between different operating states of the vibration sensor, the division of different anomaly diagnosis results can be realized, and the anomaly diagnosis model can be trained using loss functions such as Triplet Loss or Contrastive Loss.

[0048] It should be noted that the above introduction to the method of outputting the anomaly diagnosis result and the selection method of the loss function type is only an exemplary display. In actual applications, there may be other output methods and selection methods, which are not limited herein.

[0049] Taking the anomaly diagnosis model including a classification model as an example, during the process of training the above anomaly diagnosis model, the operating state of the vibration sensor characterized by the vibration signal data collected by the vibration sensor can be regarded as a signal classification task. Guided by the cross-entropy loss function, the anomaly diagnosis model is trained by minimizing the cross-entropy loss between the true operating state of each node (vibration sensor) in the heterogeneous graph during the above historical time period and the anomaly diagnosis result (normal operating state or abnormal operating state) output by the anomaly diagnosis model for each node. The model parameters are optimized using backpropagation, and finally, the model can accurately output the anomaly diagnosis result.

[0050] There can be various specific implementations of the loss function during the training process of the anomaly diagnosis model. As an example, the loss function can be expressed by the following formula: (8) Where, is the classifier parameter, is the set of labeled node indices (the true operating states of each vibration sensor characterized by the vibration signal data collected by the vibration sensor array during the historical time period), and are the label and embedding of the labeled node, respectively.

[0051] The above target device can have various specific implementations. As an example, the target device can be a flat wire motor, and the vibration sensor array can be arranged on the stator surface of the flat wire motor. The target device can also be other types of devices, such as round wire motors, etc., and specific details are not limited here.

[0052] It should be noted that in the vibration sensor anomaly diagnosis scenario of the motor targeted by the embodiments of the present application, there are several unique challenges in this specific scenario: First, the vibration signal data collected by the vibration sensor has strong temporal and multi-modal characteristics (such as tangential / radial, different working conditions), which is different from traditional static node attributes. For the vibration signal characteristics in multiple directions and multiple working conditions, the embodiments of the present application construct a multi-modal graph structure (a heterogeneous graph structure including multiple different node types), which can extract and fuse the vibration mode characteristics in the tangential and radial directions respectively; Second, although the anomaly of the vibration sensor is a local problem of the target device, it will cause complex non-linear effects in the frequency domain, and it is necessary to introduce weights and attention mechanisms in the edge relationship to model the anomaly propagation relationship; Third, the node and edge types in the heterogeneous graph are complex, involving the operating state and spatial position characteristics of the vibration sensor, and it is necessary to construct the above Meta-path analysis to achieve more effective feature learning and anomaly pattern recognition.

[0053] As an example, the vibration signal data collected by the above vibration sensor array when the target device is in different working conditions can be obtained, and the abnormal diagnosis model can be trained under different working conditions by using the vibration signal data corresponding to different working conditions respectively. Taking a 106KW flat wire motor as the target device, a plurality of vibration sensors are arranged around the stator surface of the flat wire motor, and the vibration signal data of the flat wire motor under different working conditions are collected through the vibration sensors. Under the working condition of 24N and 2000rpm, the above heterogeneous graph is constructed by using the collected vibration signal data, and the abnormal diagnosis model is trained based on the heterogeneous graph. After testing, the diagnosis accuracy of the abnormal diagnosis model under this working condition reaches 92%.

[0054] Based on the abnormal diagnosis model construction method described in any of the above embodiments, that is, the preset abnormal diagnosis model, an embodiment of the present application further provides an abnormal diagnosis method, which is applicable to the diagnosis scenario of the vibration sensor array arranged on the target device. The target device can generate vibrations with a fixed pattern, and the vibration sensor array includes a plurality of vibration sensors; as Figure 6 shown, the method includes the following steps: S601. Obtain the vibration signal data collected in real time by the vibration sensor array.

[0055] S602. Based on the vibration signal data, construct a heterogeneous graph associated with the vibration sensor array.

[0056] The nodes in the heterogeneous graph are vibration sensors, the node type is the operating state of the vibration sensor, and the edge relationship is the connection weight between different vibration sensors in the vibration sensor array.

[0057] S603. Input the heterogeneous graph into the preset abnormal diagnosis model, and output the abnormal diagnosis result for the vibration sensor array through the abnormal diagnosis model.

[0058] For the construction method of the above preset abnormal diagnosis model, reference can be made to the abnormal diagnosis model construction method described in any of the above embodiments, and details are not described here again.

[0059] The technical solution provided by the embodiments of the present application is directed to the diagnostic scenario of a vibration sensor array arranged on a target device capable of generating vibrations with a fixed pattern. An heterogeneous graph is constructed using the vibration signal data historically collected by the vibration sensor array. In this heterogeneous graph, vibration sensors are used as nodes, the operating states of the vibration sensors are used as node types, and the connection weights between different vibration sensors in the vibration sensor array are used as edge relationships. The constructed heterogeneous graph is used as a training data set, and the true operating state corresponding to the vibration signal data historically collected by the vibration sensor array is used as a label to train an anomaly diagnosis model, obtaining a trained anomaly diagnosis model to diagnose the abnormal state of the vibration sensor array. The construction of the heterogeneous graph fully considers the different operating states of different vibration sensors and the connection weights between different vibration sensors, and can more accurately capture the complex relationship between the abnormal state of the vibration sensor and the vibration signal, thereby improving the diagnostic accuracy of the abnormal state of the vibration sensor. The abnormal diagnosis result of the vibration sensor array is directly output through the constructed anomaly diagnosis model, improving the diagnostic efficiency of the abnormal state of the vibration sensor.

[0060] As an example, based on the abnormal diagnosis result output by the anomaly diagnosis model, a calibration strategy for the above-mentioned vibration sensor array can be determined; among them, when the abnormal diagnosis result is that the vibration sensor is in a normal operating state, the calibration strategy is not to perform calibration; when the abnormal diagnosis result is that the vibration sensor is in a position deviation state, the calibration strategy is to perform vibration sensor position calibration; when the abnormal diagnosis result is that the vibration sensor is in a loose state, the calibration strategy is to perform vibration sensor tightening calibration.

[0061] Corresponding to the above method embodiments, the present application also provides a computer program product, including a computer program, which when executed by a processor, implements the construction method or the anomaly diagnosis method of the anomaly diagnosis model described in any one of the above embodiments.

[0062] The present application also provides an electronic device, as Figure 7 shown, the electronic device includes: a processor 701; a memory 702 for storing instructions executable by the processor; wherein, the processor 701 is configured to implement the construction method or the anomaly diagnosis method of the anomaly diagnosis model described in any one of the above embodiments.

[0063] The present application also provides a computer-readable storage medium, on which a computer program is stored, which when executed by a processor, implements the construction method or the anomaly diagnosis method of the anomaly diagnosis model described in any one of the above embodiments.

[0064] The above are only specific embodiments of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A method for constructing an anomaly diagnosis model, characterized in that, A diagnostic scenario applicable to a vibration sensor array arranged on a target device, the target device being capable of generating vibrations with a fixed pattern, the vibration sensor array including a plurality of vibration sensors; the method includes: Obtaining vibration signal data collected by the vibration sensor array during a historical time period, and the true operating state of the vibration sensor array corresponding to the vibration signal data, the true operating state including a normal operating state or an abnormal operating state; Based on the vibration signal data, constructing a heterogeneous graph associated with the vibration sensor array, where the nodes in the heterogeneous graph are vibration sensors, the node type is the operating state of the vibration sensor, and the edge relationship is the connection weight between different vibration sensors in the vibration sensor array; Using the heterogeneous graph as a training data set to train an anomaly diagnosis model, the anomaly diagnosis model being trained based on the error between the output anomaly diagnosis result and the true operating state.

2. The method according to claim 1, wherein The connection weights between different vibration sensors in the vibration sensor array are determined by the following method: Performing a similarity analysis on the vibration signals respectively corresponding to different vibration sensors in the vibration sensor array, and determining the connection weights between different vibration sensors based on the analysis results.

3. The method according to claim 1, characterized in that, The abnormal operating state includes: the vibration sensor is in a position deviation state or the vibration sensor is in a loose state; the attributes of the node at least include the vibration signal characteristics corresponding to the vibration sensor, the position deviation characteristics of the vibration sensor, and the loose characteristics of the vibration sensor.

4. The method according to claim 1, wherein The node types include: the vibration sensor is in a normal operating state, the vibration sensor is in a position deviation state, or the vibration sensor is in a loose state; The constructing of the heterogeneous graph associated with the vibration sensor array includes: Combining different node types, determining the composite relationship between different vibration sensors in the vibration sensor array based on the combination result, and defining a meta-path connecting different vibration sensors based on the composite relationship; Based on the meta-path and the vibration signal data, constructing the heterogeneous graph.

5. The method according to claim 4, wherein The anomaly diagnosis model at least includes a feature embedding layer, a self-attention network layer, a feature fusion layer, a feature extraction layer, and an output layer; The feature embedding layer is used to convert the vibration signal data corresponding to each node in the heterogeneous graph into feature vectors in a unified dimension; The self-attention network layer is used to calculate the attention values between each node in the heterogeneous graph based on the meta-path, and determine the association degree between each node based on the attention values; The feature fusion layer is used to fuse the feature vectors corresponding to each node based on the association degree to obtain a fusion result; The feature extraction layer is used to perform feature extraction based on the fusion result to obtain an extraction result; The output layer is used to output an anomaly diagnosis result based on the extraction result, the anomaly diagnosis result including that the vibration sensor is in a normal operating state, the vibration sensor is in a position deviation state, or the vibration sensor is in a loose state.

6. The method according to claim 5, wherein The feature embedding layer is specifically configured to convert the vibration signal data corresponding to each node in the heterogeneous graph into feature vectors in a unified dimension based on matrix transformation; The feature fusion layer is specifically configured to fuse the feature vectors through average pooling operation and non-linear transformation to obtain the fusion result; The feature extraction layer is specifically configured to perform feature extraction on the fusion result through graph convolution calculation to obtain the extraction result.

7. The method according to claim 1, characterized in that The target device is a flat wire motor, and the vibration sensor array is arranged on the stator surface of the flat wire motor.

8. An abnormal diagnosis method, characterized in that, Applicable to the diagnostic scenario of the vibration sensor array arranged on the target device, the target device can generate vibrations with a fixed pattern, and the vibration sensor array includes multiple vibration sensors; the method includes: Obtain the vibration signal data collected in real time by the vibration sensor array; Based on the vibration signal data, construct a heterogeneous graph associated with the vibration sensor array, where the nodes in the heterogeneous graph are vibration sensors, the node type is the operating state of the vibration sensor, and the edge relationship is the connection weight between different vibration sensors in the vibration sensor array; Input the heterogeneous graph into the anomaly diagnosis model in the method according to any one of claims 1 to 7, and output the anomaly diagnosis result for the vibration sensor array through the anomaly diagnosis model.

9. The method according to claim 8, wherein The method further includes: Determine a calibration strategy for the vibration sensor array based on the anomaly diagnosis result; Wherein, when the anomaly diagnosis result is that the vibration sensor is in a normal operating state, the calibration strategy is not to perform calibration; When the anomaly diagnosis result is that the vibration sensor is in a position deviation state, the calibration strategy is to perform vibration sensor position calibration; When the anomaly diagnosis result is that the vibration sensor is in a loose state, the calibration strategy is to perform vibration sensor tightness calibration.

10. A computer program product, including a computer program, where the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

11. An electronic device, characterized in that, Including: A processor; A memory for storing processor-executable instructions; Wherein, the processor is configured to implement the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when executed by the processor, implements the steps in the method according to any one of claims 1 to 9.

Images