Intelligent production line rotating machine health assessment method based on hypergraph convolutional fuzzy network

Abstract

The invention relates to an intelligent production line rotating machine health assessment method based on a hypergraph convolution fuzzy network, and the method comprises the steps: collecting time domain signals generated by various faults during the operation of a rotating machine system, and obtaining corresponding time frequency representation; dividing time-domain signals and time-frequency representations according to fault types, and adding labels to form vibration data; constructing a hypergraph by using the hypergraph model and taking the time domain signal and the time frequency representation as graph nodes, and obtaining two hypergraphs under each fault type; performing data fusion on nodes of the two hypergraphs under each fault type by using a weight fuzzy module; the fused hypergraph features are input into a multi-layer perception mixing module; and the probability of all fault classes is calculated by outputting a full connection layer and a classification function, and training, optimization and updating of the hypergraph convolutional fuzzy network are realized. According to the method, the interpretability of the model is improved, and a more robust health assessment result of the rotating mechanical system can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of rotary machinery system health assessment, especially an intelligent production line rotary machinery health assessment method based on hypergraph convolutional fuzzy network. Background Art

[0002] Catastrophic failures of rotating machinery systems in smart production lines will bring huge economic losses. In order to prevent potential risks that may lead to such failures, health assessment is one of the core tasks of smart mechanical equipment maintenance. The Industrial Internet of Things can provide big data generated by equipment operation for intelligent evaluation. At present, the focus of the Industrial Internet of Things is mainly on the development of hardware and software for tracking the health status of monitored system components. Accurate health assessment can create the premise for promoting this development process. Existing health assessment methods for rotating machinery systems are mainly divided into two categories, including fault identification based on mathematical statistics and fault detection based on machine learning. Most fault identification based on mathematical statistics uses statistical means to extract identifiable feature vectors in the original data and distinguish them using classifiers. Common statistical features include feature entropy, kurtosis, mean, spectrum, etc. This type of method requires advance analysis based on prior knowledge, which may lose many rich features. The artificially extracted feature indicators are closely related to the accuracy of health assessment, and the loss of feature indicators will affect the final evaluation results. The fault detection method based on the deep neural network model aims to develop a deep nonlinear mapping structure to achieve gradient calculation for a given complex problem. That is, an iterative training process is established based on a large data stream to achieve knowledge extraction. Deep network structures are complex, and generally speaking, the interpretability of such deep models is extremely weak.

[0003] In the current fault detection methods based on deep neural network models, feature extraction in deep learning problems lacks rich fault semantics, that is, there is a lack of relevant descriptions of the interpretability of the model, which makes it impossible for researchers to further understand the working mechanism of the deep model and improve the rationality of the model structure. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a health assessment method for rotating machinery in an intelligent production line based on a hypergraph convolutional fuzzy network, aiming to increase the interpretability of the model and thereby obtain more robust health assessment results of the rotating machinery system.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A method for health assessment of rotating machinery in an intelligent production line based on a hypergraph convolutional fuzzy network, wherein the hypergraph convolutional fuzzy network includes a hypergraph model, a weight fuzzy module, a multi-layer perception hybrid module, an output fully connected layer and a classification function;

[0007] The method comprises:

[0008] S1. Collect time domain signals generated by various faults during the operation of the rotating machinery system, divide the time domain signals according to the fault type and add labels, obtain the time-frequency representation of each time domain signal, and mark the corresponding labels to form vibration data;

[0009] S2. Enter the vibration data into the hypergraph model, and use the time domain signal and its time frequency to construct the hypergraph to graph nodes to obtain two hypergraphs under each type of fault type;

[0010] S3. Use the weight fuzzy module to fusion data on the two hypergraph nodes under each type of fault type to realize the fusion of the two hypergraph features and obtain the fused hypergraph;

[0011] S4. The fused hypergraph features are input into the multi-layer perception hybrid module to achieve knowledge divergence suppression in the graph structure;

[0012] S5. Calculate the probability of all fault classes through the output full connection layer and classification function, and calculate the loss function of fault evaluation to realize the training, optimization and update of hypergraph convolutional fuzzy network;

[0013] S6. Use the trained hypergraph convolutional fuzzy network to perform a health assessment of the rotating mechanical system.

[0014] Further technical solutions are:

[0015] In step S3, the weight fuzzy module is used to perform data fusion on the nodes of two hypergraphs under each type of fault type, including:

[0016] Initialize the connection between the nodes of the two hypergraphs;

[0017] The fuzzy entropy of nodes between the two hypergraphs is calculated in turn:

[0018]

[0019] Among them, a is the adjustment factor, v i , The i-th node in the hypergraph constructed by time domain signals and the j-th node in the hypergraph constructed by time domain signals respectively;

[0020] The fuzzy entropy of the node is used as the fuzzy membership degree and the fusion weight is calculated:

[0021]

[0022] Among them, m and n are the hypergraphs constructed by time domain signals and the time frequency represent the total number of nodes in the constructed hypergraph;

[0023] Based on the fusion weight I, the node relationship in the two hypergraphs is determined according to the following fuzzy rules:

[0024]

[0025] Among them, 1 indicates a strong relationship, 0 indicates a slight relationship, and ξ is a predefined threshold;

[0026] Two or more nodes with a strong relationship are connected through a hyperedge to replace the node initialization connection, so as to transfer the most critical semantics of the two hypergraphs to the fused hypergraph.

[0027] The structure of the multi-layer perception hybrid module includes a first graph convolution layer, a self-attention mechanism, a second graph convolution layer, a first multi-layer perceptron and a second multi-layer perceptron, a third graph convolution layer and an embedding vector;

[0028] The first multilayer perceptron and the second multilayer perceptron are composed of multiple fully connected layers of the same size;

[0029] The features output by the first convolutional layer Carry out the dependency construction of the self-attention mechanism:

[0030]

[0031] Among them, ψ1 is the output feature of the self-attention mechanism, conv() represents the convolution operation, and its subscript represents the corresponding graph convolution layer;

[0032] The features of the second layer hypergraph convolution output After linear transformation, it is mapped in the first multi-layer perceptron:

[0033]

[0034] Where ψ2 is the output feature of the first multilayer perceptron, w1 and w2 are the weights of the first and second fully connected layers in the first multilayer perceptron, σ(·) is the unit nonlinear Gaussian error linear unit; LayerNorm() is the layer regularization operation;

[0035] The second multilayer perceptron And ψ2 are mapped:

[0036]

[0037] Where ψ3 is the output feature of the second multilayer perceptron, w3 and w4 are the weights of the first and second fully connected layers in the second multilayer perceptron respectively;

[0038] Align the dimensions of ψ1, ψ2 and ψ3, and complete the embedding vector through feed-forward combination The embedding vector is also used as the input of the third graph convolutional layer to achieve the accumulation of the first layer of knowledge in the last layer.

[0039] In step S2, the information transmission mechanism of the graph convolutional neural network used to construct the hypergraph is as follows:

[0040]

[0041] In the above formula:

[0042] X l and X l-1 represents the input and output of the graph convolutional neural network, σ(·) represents the ReLU activation function, and θ l-1 are network parameters;

[0043] is the smallest non-negative eigenvector, which can be calculated by the identity matrix I, where:

[0044] D v is a diagonal matrix, representing the degree matrix of the node, and the elements on the diagonal are the degree d(v) of each node v in the hypergraph;

[0045] D e is a diagonal matrix, representing the degree matrix of the hyperedge, and the elements on the diagonal are the degree δ(e) of each hyperedge e in the hypergraph;

[0046] H is dimensional point-edge incidence matrix, ε、 They are the node set, hyperedge set, and weight set assigned to the hyperedges of the hypergraph respectively.

[0047] The degree of a node is calculated as follows:

[0048]

[0049] Among them, v j ,e j are the jth node and hyperedge in the hypergraph, respectively, w(e j ) represents the hyperedge e j The average value of the hot core distance between any two vertices on it;

[0050] The degree of a hyperedge is calculated as follows:

[0051]

[0052] In step S2, the optimization of the hypergraph is transformed into solving the eigenvector corresponding to the minimum eigenvalue of the matrix bundle (L, A) to achieve the segmentation of the graph;

[0053] L is the Laplacian matrix of the hypergraph,

[0054] A is the adjacency matrix of the hypergraph, The value a of each element in the adjacency matrix i,j =w(e n ), w(e n ) represents the nth hyperedge e n Any two vertices on the node v i ,v j The average value of the hot core distance is:

[0055]

[0056] x i ,x j and v i ,v j The corresponding time frequency represents or time domain signal, u is the Gaussian core size.

[0057] Before step S4, the method further includes:

[0058] Repeat S1 to S3, use more vibration data to iteratively train the hypergraph model, expand the number of graph nodes and layers of the fused hypergraph, and construct the hypergraph model into a deep neural network with high fault feature extraction capability.

[0059] The vibration data obtained in step S1 is expressed as follows:

[0060]

[0061] in, is the time domain signal set The corresponding time frequency represents, c represents a collection of all fault classes, and the lth fault class It is the jth time domain signal and its time frequency represent vectors under the lth fault class, and N represents the number of data.

[0062] The time domain signals generated by various faults during operation of the rotating mechanical system include:

[0063] The acceleration sensor is installed on the cylinder head of the fuel and valve system of the rotating mechanical system, and the reference standard reflective label of the dead point signal measured by the photoelectric speed sensor is pasted on the main shaft end of the rotating mechanical system. According to the dead point signal, the time domain signal with the same time length is collected.

[0064] The cross entropy function is used as the loss function for calculating fault evaluation to realize the training, optimization and updating of the hypergraph convolutional fuzzy network.

[0065] The beneficial effects of the present invention are as follows:

[0066] The present invention introduces the idea of ​​fuzzy reasoning into the traditional graph convolutional neural network method, and proposes a hypergraph convolutional fuzzy architecture with a multi-layer perception hybrid module and a weighted fuzzy module, which can well solve the problem of lack of rich fault semantics in feature extraction in traditional deep learning problems, and successfully realizes the fusion of different fault features.

[0067] The present invention designs a fuzzy rule-driven graph construction method to induce the cascading behavior of dynamically changing data constructed by two different input feature subsets, and perform data fusion in the graph, thereby achieving the transfer of the most critical semantics of the two graphs to the final constructed hypergraph.

[0068] The present invention utilizes a novel weighted coding to generate heterogeneous embedding, namely, multi-layer perceptual hybrid embedding, to suppress knowledge divergence in the graph structure, so that the model can further learn fault characteristics, better enhance the semantic transfer between graphs, and effectively capture the key features between data, thereby obtaining more robust rotating machinery system health assessment results.

[0069] Other features and advantages of the present invention will be set forth in the following description or may be learned by practicing the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention.

[0071] Figure 2 Schematic diagram of the hypergraph convolutional fuzzy network structure of an embodiment of the present invention.

[0072] Figure 3 The diagram is a schematic diagram of the fuzzy rule-driven graph construction principle of an embodiment of the present invention.

[0073] Figure 4 It is a vibration signal collected during the experimental verification process of the embodiment of the present invention.

[0074] Figure 5 This is the prediction result obtained by experimental verification of the embodiment of the present invention. DETAILED DESCRIPTION

[0075] The specific implementation of the present invention is described below with reference to the accompanying drawings.

[0076] See also Figure 1 and Figure 2 , a method for health assessment of rotating machinery in an intelligent production line based on a hypergraph convolutional fuzzy network in this embodiment, wherein the hypergraph convolutional fuzzy network includes a hypergraph model, a weight fuzzy module, a multi-layer perception hybrid module, an output fully connected layer and a classification function;

[0077] See also Figure 1 , the method comprising:

[0078] S1. Collect time domain signals, i.e., vibration signals, generated by various faults during the operation of the rotating machinery system, divide the time domain signals according to the fault type and add labels, obtain the time-frequency representation of each time domain signal, and mark the corresponding labels to form vibration data;

[0079] As a specific embodiment, the acquisition of vibration signals includes:

[0080] Install the acceleration sensor on the cylinder head of the fuel and gas distribution system of the rotating mechanical system, paste the reference standard reflective label of the dead point signal measured by the photoelectric speed sensor on the main shaft end of the rotating mechanical system, perform key time positioning according to the dead point signal, and obtain a vibration signal with the same time length

[0081]

[0082] The vibration signal Divided into m types of fuel and gas distribution system fault types, and set corresponding fault label Y={y 1 ,y 2 ,…,y m}, where x m The vibration signal representing the mth fault type, y m A label vector representing the mth fault type.

[0083] The constructed vibration data can be expressed as follows:

[0084]

[0085] in, is the time domain signal set The corresponding time frequency represents, c represents a collection of all fault classes, and the lth fault class It is the jth time domain signal and its time frequency represent vectors under the lth fault class, and N represents the number of data.

[0086] S2. Enter the vibration data into the hypergraph model, and build the hypergraph using the time domain signal and its time frequency as graph nodes to obtain two hypergraphs under each type of fault type.

[0087] The hypergraph of time domain signal construction under the L-type fault type can be represented as represents the set of nodes of the hypergraph, represents the set of directed hyperedges, Represents the weight matrix assigned to the hyperedge. The larger the value in the weight matrix, the greater the influence of the corresponding edge. At the same time, for any hyper edge, it is a collection of nodes A subset of . The hypergraph constructed by the time-frequency representation of the corresponding time domain signal is

[0088] As a specific implementation, the construction of the hypergraph includes:

[0089] (1) Calculate the degree of the node:

[0090]

[0091] Among them, v j ,e j are the jth node and hyperedge in the hypergraph, respectively, w(e j ) represents the hyperedge e j The average value of the hot core distance between any two vertices on it;

[0092] (2) Calculate the degree of the hyperedge:

[0093]

[0094] H is dimensional point-edge association matrix, as can be seen above, ε is the node set and hyperedge set of the hypergraph. In the node-edge association matrix, if a node is on a hyperedge, then H(v,e) = 1, otherwise H(v,e) = 0.

[0095] (3) Set the diagonal matrix D v , represents the degree matrix of the node, and the elements on the diagonal are the degree d(v) of each node v in the hypergraph; set the diagonal matrix D e , represents the degree matrix of the hyperedge, and the elements on the diagonal are the degree δ(e) of each hyperedge e in the hypergraph.

[0096] Compute the adjacency matrix of the hypergraph representing the connections between each node From the above, we can know that it is the weight matrix assigned to the hyperedge; the value a of each element in the adjacency matrix i,j =w(e n ), w(e n ) represents the nth hyperedge e n Any two vertices on the node v i ,v j The average value of the hot core distance is:

[0097]

[0098] Among them, x i ,x j and v i ,v j The corresponding time-frequency representation or time domain signal, u is the Gaussian kernel size.

[0099] (4) The optimization of the hypergraph is transformed into solving the eigenvector corresponding to the minimum eigenvalue of the matrix bundle (L, A) to achieve graph segmentation.

[0100] Among them, the Laplacian matrix of the hypergraph The minimum non-negative eigenvector Δ can be calculated using the identity matrix I:

[0101]

[0102] (5) Information transfer of the graph convolutional neural network based on which the hypergraph model is calculated:

[0103]

[0104] Among them, X l and X l-1 represents the input and output of the graph convolutional neural network, σ(·) represents the Recti-fied Linear Units (ReLU) activation function, and θ l-1 is a network parameter.

[0105] S3. Using the weight fuzzy module, data fusion is performed on the nodes of the two hypergraphs under each type of fault, so as to achieve fusion of the features of the two hypergraphs and obtain a fused hypergraph for representing the information cascade.

[0106] The specific steps include:

[0107] Initialize the connection between the nodes of the two hypergraphs;

[0108] Calculate the node fuzzy entropy between the two hypergraphs in turn:

[0109]

[0110] Among them, a is the adjustment factor, v i , The i-th node in the hypergraph constructed for the time domain signal and the j-th node in the hypergraph constructed for the time-frequency representation respectively;

[0111] The node fuzzy entropy is used as the fuzzy membership to calculate the fusion weight:

[0112]

[0113] Among them, m and n are the total number of nodes in the hypergraph constructed by time domain signal and the hypergraph constructed by time-frequency representation, respectively;

[0114] Based on the fusion weight I, the node relationship in the two hypergraphs is determined according to the following fuzzy rules:

[0115]

[0116] Among them, 1 means a strong relationship, 0 means a slight relationship, and ξ is a predefined threshold;

[0117] If the relationship is strong, the nodes are connected by hyperedges, otherwise no connection is required. By connecting two or more nodes with strong relationships by hyperedges, the node initialization connection is replaced, so that the most critical semantics of the two hypergraphs are transferred to the fused hypergraph.

[0118] This embodiment uses a fuzzy weight module to focus on the node entropy of the two hypergraphs, that is, their semantic degrees, so as to transfer the most critical semantics in the two hypergraphs to the final constructed hypergraph (the fused hypergraph).

[0119] like Figure 3 As shown, each hyper-edge can establish a connection between two or more nodes, thereby enabling a many-to-many relationship representation, e1 and e2 in the figure. For the convenience of description, Figure 3 An incidence matrix is ​​used in which each row (A→J) represents a node and each column (e1→e3) represents a hyperedge. If a node is part of a particular hyperedge, the corresponding element of the incidence matrix position is set to 1, otherwise it is set to 0.

[0120] S4. The fused hypergraph features are input into the multi-layer perception hybrid module to suppress knowledge divergence in the hypergraph structure.

[0121] like Figure 1 As shown, the structure of the multi-layer perception hybrid module includes a first graph convolution layer (i.e., graph convolution layer 1), a self-attention mechanism, a second graph convolution layer (i.e., graph convolution layer 2), a first multi-layer perceptron (i.e., multi-layer perceptron 1) and a second multi-layer perceptron (i.e., multi-layer perceptron 2), a third graph convolution layer (i.e., graph convolution layer 3) and an embedding vector;

[0122] The first multilayer perceptron and the second multilayer perceptron are composed of multiple fully connected layers of the same size. As a preferred embodiment, each multilayer perceptron includes a fully connected layer 1, a layer regularization, and a fully connected layer 2.

[0123] The features output by the first convolutional layer Construct the dependency relationship of the self-attention mechanism:

[0124]

[0125] Among them, ψ1 is the output feature of the self-attention mechanism, conv() represents the convolution operation, and its subscript represents the corresponding graph convolution layer;

[0126] The features of the second layer hypergraph convolution output After linear transformation and mapping in the first multilayer perceptron:

[0127]

[0128] Where ψ2 is the output feature of the first multilayer perceptron, w1 and w2 are the weights of the first and second fully connected layers in the first multilayer perceptron, σ(·) is the unit nonlinearity Gaussian error linear unit (GeLU); LayerNorm() is the layer regularization operation;

[0129] The second multilayer perceptron And ψ2 are mapped:

[0130]

[0131] Among them, ψ3 is the output characteristics of the second multi-layer perceptron, w3 and w4 are the weights of the first and second fully connected layers in the second multi-layer perceptron, respectively;

[0132] Make ψ1, ψ2 and ψ3 dimensionally aligned, and then the embedding vector is completed through feedforward combination The embedded vector is also used as input to the third graph convolution layer to realize the accumulation of the first layer of knowledge in the last layer.

[0133] S5, through the output fully connected layer (i.e. Figure 1 The full connection layer 4) and softmax classification function calculate the probability of all fault classes and calculate the loss function of fault evaluation to realize the training, optimization and update of hypergraph convolutional fuzzy network.

[0134] The hypergraph convolutional neural fuzzy network takes the time-frequency representation and time-domain signal of vibration data as input, constructs the hypergraph fault features obtained by the vibration signal and its time-frequency representation into respective graph nodes, and uses the weight fuzzy module to complete the data fusion of the graph nodes, that is, the fusion of two hypergraph fault features, realizes the extraction of key fault information from the two inputs, and reconstructs them into fused hypergraph features. The calculation method is as follows:

[0135]

[0136] Among them, X represents the hypergraph fault feature extracted by the vibration signal in the graph structure, It represents the time frequency to represent the hypergraph fault characteristics extracted in the graph structure. Represents hypergraph feature fusion.

[0137] Among them, the update of the hypergraph convolutional fuzzy network is designed to learn a classifier from the marked vibration data set. By calculating the label probability of all classes with a softmax function, the calculation process is as follows:

[0138]

[0139] in, is the predicted value of the label, w C and b C Represents the learnable parameters of the classifier, and F represents the final feature representation

[0140] As a specific embodiment, a cross entropy function is used as a loss function for calculating the fault evaluation:

[0141]

[0142] in, Represents the loss function value, y i is the label value of the i-th sample, is the predicted value of the i-th sample, and N is the number of samples.

[0143] S6. Use the trained hypergraph convolutional fuzzy network to perform a health assessment of the rotating mechanical system.

[0144] As an improvement of the above-mentioned embodiment, before step S4, it is further included:

[0145] Repeat S1 to S3, use more vibration data to iteratively train the hypergraph model, expand the number of graph nodes and layers of the fused hypergraph to improve the graph structure's ability to extract fault features, and construct the hypergraph model into a deep neural network.

[0146] As a specific implementation, in the iterative training of the hypergraph model, two types of input data are input into their respective graph structures with the same batch size, so as to extract the hypergraph features of both. Each iterative training uses vibration data of different time batches.

[0147] The following is a specific experiment to verify the effectiveness of the method of this embodiment.

[0148] The method of this embodiment is used to identify multiple faults on the fuel and gas distribution system of a certain intelligent production line rotary machinery system.

[0149] The following steps are involved:

[0150] Step 1: Use acceleration sensors and photoelectric speed sensors to collect vibration signals of different fault types of the experimental system, with a sampling frequency of 25kHz, including intake valve clearance 0.1 mm larger, intake valve clearance 0.2 mm larger, exhaust valve clearance 0.1 mm larger, exhaust valve clearance 0.2 mm larger, intake filter blockage, exhaust pipe blockage, lubrication advance angle reduction of 3°, lubrication advance angle increase of 3°, abnormal cylinder combustion, cylinder wear, serious cylinder failure and normal, a total of 12 different types of vibration signal data, each of which has 500 samples, and each sample is labeled. On this basis, the time-frequency representation of each sample vibration data is extracted, and the corresponding labels are marked to construct a data set. Among them, the collected vibration data is as follows Figure 4 As shown, the lower part in the figure is a time domain signal, and the upper part is a corresponding time frequency representation. The system failure type and sample number are shown in Table 1:

[0151] Table 1 System failure types and sample numbers

[0152] Fault type Sample size Intake valve clearance is 0.1 mm larger 500 Intake valve clearance is 0.2 mm larger 500 Exhaust valve clearance is 0.1 mm larger 500 Exhaust valve clearance 0.2 mm larger 500 Air intake filter clogged 500 Exhaust pipe blocked 500 Lubrication advance angle reduced by 3° 500 Lubrication advance angle increased by 3° 500 Abnormal cylinder combustion 500 Cylinder working wear 500 Cylinder has serious fault 500 normal 500

[0153] Step 2: The data set is randomly divided into a training set and a test set in a ratio of 8:2. On this basis, the vibration signal and time-frequency representation in the training set are input into the hypergraph model in parallel to construct their respective graph nodes, complete the graph construction, and output their respective hypergraph features with fault information.

[0154] Step 3: Through the weight fuzzy module, the vibration signal in the hypergraph feature, i.e. the graph structure of the vibration signal, and the hypergraph feature in time-frequency representation, i.e. the graph structure of the time-frequency signal, are fused to obtain the fused hypergraph feature.

[0155] In the fourth step, repeat the second and third steps to input all batches of vibration signals and time-frequency representations into the hypergraph model, thereby deepening the number of graph nodes and layers and improving the hypergraph model's training and learning of fault features.

[0156] Step 5: Combine the hypergraph model, weight fuzzy module, multi-layer perception hybrid module, fully connected layer and Softmax classification layer to build a hypergraph convolutional fuzzy network, and input the fused hypergraph features into the multi-layer perception hybrid module.

[0157] Step 6: Construct the dependency relationship of the features of the graph convolution output of the first layer in the multi-layer perception hybrid module through the self-attention mechanism, and realize the knowledge divergence suppression in the graph structure under the hybrid embedding of multi-layer perceptron and embedding vector.

[0158] Step 7: Use the output of the multi-layer perception hybrid module as the input of the full connection. Finally, calculate the probabilities of all fault classes in the Softmax classification layer, and use the cross entropy function as the loss function for calculating fault assessment to achieve the training, optimization, and update of the overall model.

[0159] Step 8: Save the hypergraph convolutional fuzzy network model with optimal parameters to perform a healthy evaluation of the test set of the rotating mechanical system.

[0160] The multi-failure identification of fuel and gas distribution systems is performed using hypergraph convolutional fuzzy network model, and the prediction results are as follows Figure 5 Shown. From the figure, it can be seen that the predicted value successfully realizes the complete classification of the faults, verifying the effectiveness of the method proposed in this embodiment in the health evaluation of the intelligent production line rotary machinery system.

[0161] Those skilled in the art can understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for health assessment of rotating machinery in intelligent production lines based on hypergraph convolutional fuzzy networks, characterized in that: The hypergraph convolutional fuzzy network includes a hypergraph model, a weight fuzzy module, a multi-layer perception hybrid module, an output fully connected layer and a classification function; The method comprises: S1. Collect time domain signals generated by various faults during the operation of the rotating machinery system, divide the time domain signals according to the fault type and add labels, obtain the time-frequency representation of each time domain signal, and mark the corresponding labels to form vibration data; S2, inputting the vibration data into the hypergraph model, constructing a hypergraph with the time domain signal and its time-frequency representation as graph nodes, and obtaining two hypergraphs for each type of fault; S3, using the weight fuzzy module to perform data fusion on the nodes of the two hypergraphs under each type of fault, to achieve fusion of the features of the two hypergraphs, and to obtain a fused hypergraph; S4, input the fused hypergraph features into the multi-layer perception hybrid module to suppress knowledge divergence in the graph structure; S5. Calculate the probabilities of all fault classes through the output fully connected layer and the classification function, and calculate the loss function of fault assessment to achieve training, optimization and updating of the hypergraph convolutional fuzzy network; S6. Use the trained hypergraph convolutional fuzzy network to perform health assessment on the rotating machinery system.

2. The method according to claim 1, characterized in that In step S3, the weight fuzzy module is used to perform data fusion on the nodes of the two hypergraphs under each type of fault, including: Initialize the connection between the nodes of the two hypergraphs; Calculate the node fuzzy entropy between the two hypergraphs in turn: Among them, a is the adjustment factor, v i , The i-th node in the hypergraph constructed for the time domain signal and the j-th node in the hypergraph constructed for the time-frequency representation respectively; The node fuzzy entropy is used as the fuzzy membership to calculate the fusion weight: Among them, m and n are the total number of nodes in the hypergraph constructed by time domain signal and the hypergraph constructed by time-frequency representation, respectively; Based on the fusion weight I, the node relationship in the two hypergraphs is determined according to the following fuzzy rules: Among them, 1 indicates a strong relationship, 0 indicates a slight relationship, and ξ is a predefined threshold; Two or more nodes with a strong relationship are connected through a hyperedge to replace the node initialization connection, so as to transfer the most critical semantics of the two hypergraphs to the fused hypergraph.

3. The method according to claim 1, characterized in that: The structure of the multi-layer perception hybrid module includes a first graph convolution layer, a self-attention mechanism, a second graph convolution layer, a first multi-layer perceptron and a second multi-layer perceptron, a third graph convolution layer and an embedding vector; The first multilayer perceptron and the second multilayer perceptron are composed of multiple fully connected layers of the same size; The features output by the first convolutional layer Carry out the dependency construction of the self-attention mechanism: Among them, ψ1 is the output feature of the self-attention mechanism, conv() represents the convolution operation, and its subscript represents the corresponding graph convolution layer; The features of the second layer hypergraph convolution output After linear transformation, it is mapped in the first multi-layer perceptron: Where ψ2 is the output feature of the first multilayer perceptron, w1 and w2 are the weights of the first and second fully connected layers in the first multilayer perceptron, σ(·) is the unit nonlinear Gaussian error linear unit; LayerNorm() is the layer regularization operation; The second multilayer perceptron And ψ2 are mapped: Where ψ3 is the output feature of the second multilayer perceptron, w3 and w4 are the weights of the first and second fully connected layers in the second multilayer perceptron respectively; Align the dimensions of ψ1, ψ2 and ψ3, and complete the embedding vector through feed-forward combination The embedding vector is also used as the input of the third graph convolutional layer to achieve the accumulation of the first layer of knowledge in the last layer.

4. The method according to claim 1, characterized in that: In step S2, the information transmission mechanism of the graph convolutional neural network used to construct the hypergraph is as follows: In the above formula: X l and X l-1 represents the input and output of the graph convolutional neural network, σ(·) represents the ReLU activation function, and θ l-1 are network parameters; is the smallest non-negative eigenvector, which can be calculated by the identity matrix I, where: D v is a diagonal matrix, representing the degree matrix of the node, and the elements on the diagonal are the degree d(v) of each node v in the hypergraph; D e is a diagonal matrix, representing the degree matrix of the hyperedge, and the elements on the diagonal are the degree δ(e) of each hyperedge e in the hypergraph; H is dimensional point-edge incidence matrix, ε、 They are respectively the node set, hyperedge set, and weight set assigned to the hyperedges of the hypergraph.

5. The method according to claim 4, characterized in that The degree of a node is calculated as follows: Among them, vj, ej are the jth node and hyperedge in the hypergraph respectively, w(e j ) represents the hyperedge e j The average value of the hot core distance between any two vertices on it; The degree of a hyperedge is calculated as follows:

6. The method according to claim 4, characterized in that In step S2, the optimization of the hypergraph is transformed into solving the eigenvector corresponding to the minimum eigenvalue of the matrix bundle (L, A) to achieve the segmentation of the graph; L is the Laplacian matrix of the hypergraph, A is the adjacency matrix of the hypergraph, The value a of each element in the adjacency matrix i,j =w(e n ), w(e n ) represents the nth hyperedge e n Any two vertices on the node v i ,v j The average value of the hot core distance is: x i ,x j and v i ,v j The corresponding time-frequency representation or time domain signal, u is the Gaussian kernel size.

7. The method according to claim 1, characterized in that Before step S4, the method further includes: Repeat S1 to S3, use more vibration data to iteratively train the hypergraph model, expand the number of graph nodes and layers of the fused hypergraph, and construct the hypergraph model into a deep neural network with high fault feature extraction capability.

8. The method according to claim 1, characterized in that The vibration data obtained in step S1 is expressed as follows: in, is the time domain signal set The corresponding time-frequency representation, c represents the set of all fault classes, the lth fault class is the jth time domain signal under the lth fault type and its time-frequency representation vector, and N represents the number of data.

9. The method according to claim 1, characterized in that: The time domain signals generated by various faults during the operation of the rotating mechanical system are collected, including: The acceleration sensor is installed on the cylinder head of the fuel and valve system of the rotating mechanical system, and the reference standard reflective label of the dead point signal measured by the photoelectric speed sensor is pasted on the main shaft end of the rotating mechanical system. According to the dead point signal, the time domain signal with the same time length is collected.

10. The method according to claim 1, characterized in that The cross entropy function is used as the loss function for calculating fault assessment to achieve the training, optimization and updating of the hypergraph convolutional fuzzy network.

Images