A fault diagnosis method for rotating components in phase regulators based on adaptive multi-channel graph neural network

Abstract

The application discloses a method for diagnosing faults of rotating parts of a phase modifier based on an adaptive multi-channel graph neural network, and belongs to the technical field of industrial equipment fault diagnosis.The application solves the problems of low fault diagnosis precision, long training time and high training cost of traditional methods when there is a deviation between the real distribution of actual signals and the distribution of training samples.The vibration signals of the rotating parts of the phase modifier are collected and preprocessed, and then the preprocessed vibration signal segments are input into a double-channel graph convolution network structure constructed in a feature space and a topological space.The double-channel graph convolution network structure models the local features and global topological relations of the signals respectively, can comprehensively extract the local and global information of the signals, and introduces an edge enhancement attention mechanism and an adaptive module fusion strategy to weight and distribute the feature contributions of different modules.Then, the weighted and fused features are input into a classifier to realize fault diagnosis.The method can be applied to the fault diagnosis of the phase modifier.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial equipment fault diagnosis, and in particular relates to a method for diagnosing faults of rotating components in a phase regulator based on an adaptive multi-channel graph neural network. Background Art

[0002] With the rapid development of intelligent manufacturing, rotating machinery such as phase regulators plays a vital role in power grid stability and industrial production. Fault diagnosis directly impacts equipment operational safety and maintenance costs. Traditional methods rely primarily on feature extraction and classification of vibration and electrical signals. However, in complex operating conditions, with noise interference and nonlinear and non-stationary signals, traditional CNN and GCN models often suffer from insufficient feature extraction and insufficient fusion of global and local information.

[0003] In recent years, fault diagnosis methods based on graph neural networks (GNNs) have begun to attract attention. The Graph Attention Network (GAT) has achieved some success in aggregating local features by introducing an attention mechanism, while the extended Edge-Enhanced Graph Attention Network (EGAT) incorporates edge features into weight calculations, effectively capturing complex interactions between nodes. However, a single model cannot account for the heterogeneity of signals in both feature and topological spaces. Therefore, a new method for adaptive multi-channel modeling, which simultaneously considers local features and global topological structures, is urgently needed to improve the accuracy and robustness of fault diagnosis.

[0004] In recent years, scholars have conducted extensive research on graph neural networks to improve diagnostic performance. For example, Velickovic et al. proposed a graph attention network (GAT), which dynamically adjusts the importance of neighboring nodes through an attention mechanism, thereby improving the fixed weight restrictions of traditional graph convolutional networks (GCNs). This approach effectively improves the accuracy of local feature aggregation in graph data, but its ability to capture high-order neighbor features is relatively weak. The MNGCN model proposed by Yuan et al. introduces a multi-scale neighborhood aggregation mechanism that can simultaneously capture the feature relationships of nodes at different scales, significantly improving the performance of clustering and classification tasks. However, multi-scale design increases model complexity and may result in longer training times. The CBAM-ResNet-GCN method proposed by Wang et al. achieves high-precision diagnosis of rotating machinery faults by combining a CBAM residual network and a GCN. This method generates high-quality samples through a WGAN-GP, thus addressing the problem of imbalanced fault samples. However, this method is highly dependent on the quality of GAN-generated samples, and may reduce the accuracy of fault diagnosis when the distribution of generated training samples deviates from the true distribution. The CaEGCN model proposed by Huo et al. uses a cross-attention mechanism to fuse a content autoencoder (CAE) and a graph convolutional autoencoder (GAE), overcoming the oversmoothing problem of traditional GCNs. This method demonstrates excellent robustness in clustering tasks, but the introduction of the cross-attention module may increase the training cost of the model. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of low fault diagnosis accuracy, long training time and high training cost of traditional methods when there is a deviation between the true distribution of actual signals and the distribution of training samples, and to propose a fault diagnosis method for rotating components in phase regulators based on an adaptive multi-channel graph neural network.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: a method for diagnosing faults of rotating components in a phase regulator based on an adaptive multi-channel graph neural network, the method specifically comprising the following steps:

[0007] Step 1: collecting vibration signals of rotating parts in the phase regulator, and then preprocessing the collected vibration signals to obtain preprocessed signal segments;

[0008] Step 2: extract the local features of each preprocessed signal segment based on the feature space graph convolution network, and extract the global topological features of each preprocessed signal segment based on the topological space graph convolution network;

[0009] Step 3: Perform weighted fusion on the local features and global topological features of each pre-processed signal segment to obtain the fused fault features;

[0010] Step 4: Input the fused fault features into the classifier, and output the fault diagnosis result of the phase shifter through the classifier.

[0011] Furthermore, the vibration signal of the rotating component in the phase regulator is collected using a three-axis accelerometer, and the vibration signal of the rotating component is continuously collected using the three-axis accelerometer with a sampling frequency of 10 kHz.

[0012] Furthermore, the collected vibration signal is preprocessed, and the specific process is as follows:

[0013] Step 1: De-noise the collected vibration signal

[0014] A Butterworth bandpass filter is used to filter out high-frequency noise in the vibration signal, and then a Wiener filter is used to eliminate power frequency interference in the vibration signal to obtain a vibration signal after noise reduction.

[0015] Step 1 and 2: Set the sliding window length to 1024 points, use the sliding window to sample the vibration signal after noise reduction, and use the sliding window moving step size to 512. The signal in each sliding window is regarded as a signal segment.

[0016] Step 13: Perform maximum and minimum normalization processing on the signals in each signal segment to obtain each normalized signal segment.

[0017] Furthermore, the local features of each pre-processed signal segment are extracted based on the feature space graph convolution network, and the specific process is as follows:

[0018] Step S1: Generate a feature map based on each preprocessed signal segment , specifically:

[0019] Step S11: Each pre-processed signal segment is used as a feature map A node in the , the signal segment corresponding to each node is used as the feature matrix A row of the signal fragments corresponding to all nodes constitutes the feature matrix ;

[0020] Step S12, calculate the The signal segment corresponding to the node The W distance of the signal segment corresponding to the node, and then The signal segment corresponding to the node The W distance of the signal segments corresponding to the nodes is used as the similarity matrix Middle Rank Elements of a column ;

[0021] Similarly, calculate the The similarity matrix is ​​obtained by calculating the W distance between the signal segment corresponding to each node and the signal segment corresponding to each node. ;

[0022] Step S13: Similarity matrix No. Sort the elements in the row from small to large and convert the similarity matrix No. Assign the first K small elements in the row a value of 1, and All other elements in the row are assigned a value of 0;

[0023] Similarly, for the similarity matrix After processing each row in separately, we get the adjacency matrix ;

[0024] Step S2: feature map As the input of the feature space graph convolutional network, the local features of each signal segment after preprocessing are output through the feature space graph convolutional network; specifically:

[0025] Step S21: Initialization ;

[0026] Step S22: Initialize the number of layers ;

[0027] Step S23, calculate the Node feature matrix output by the layer :

[0028]

[0029] in, Represents the feature space graph convolutional network The weight matrix of the layer;

[0030] represents the adjacency matrix with self-connection, , represents the identity matrix;

[0031] Representation feature map degree matrix of ;

[0032] represents the activation function;

[0033] Step S24: Determine whether , Indicates the total number of layers;

[0034] If satisfied , then As the final local feature extracted ;

[0035] If not satisfied , then let , return to execute step S23.

[0036] Furthermore, the global topological features of each signal segment after preprocessing are extracted based on the topological space graph convolutional network, and the specific process is as follows:

[0037] Step B1: Generate a feature map based on each preprocessed signal segment , specifically:

[0038] Step B11: Use each pre-processed signal segment as a feature map A node in the , the signal segment corresponding to each node is used as the feature matrix A row of the signal fragments corresponding to all nodes constitutes the feature matrix ;

[0039] Step B12: Adjacency matrix based on node label information To optimize:

[0040]

[0041] in, Indicates the The label of each node;

[0042] Indicates the The label of each node;

[0043] represents the reinforcement coefficient, ;

[0044] represents the optimized adjacency matrix;

[0045] Represents the adjacency matrix Middle Rank Elements of the column;

[0046] represents the attenuation coefficient;

[0047] Step B2: Feature map As the input of the topological space graph convolutional network, the global topological features of each signal segment after preprocessing are output through the topological space graph convolutional network; specifically:

[0048] Step B21: Initialization ,initialization ;

[0049] Step B22: Initialize the number of layers ;

[0050] Step B23: Calculate node feature matrix :

[0051]

[0052] in, Represents the topological space graph convolutional network The weight matrix of the layer;

[0053] Representation feature map degree matrix of ;

[0054] Step B24: Node feature matrix Update to get the updated node feature matrix ; For the adjacency matrix Update to get the updated adjacency matrix ;

[0055] Step B25: Determine whether , Indicates the total number of layers;

[0056] If satisfied , then As the final global topological feature extracted ;

[0057] If not satisfied , then let , return to step B23.

[0058] Furthermore, the specific process of step B24 is as follows:

[0059] Step B241, calculate the Node and The attention weight of each node :

[0060]

[0061] in, represents the activation function;

[0062] is the vector used for attention calculation;

[0063] The superscript T indicates transposition;

[0064] represents the weight matrix;

[0065] Represents the node feature matrix Middle The characteristics of each node;

[0066] Represents the node feature matrix Middle The characteristics of each node;

[0067] “||” indicates feature concatenation operation;

[0068] Indicates the Node and Edge features between nodes;

[0069] Indicates the The set of neighbor nodes of a node;

[0070] Indicates the The node neighbor nodes;

[0071] Indicates the Node and Edge features between nodes;

[0072] Represents the node feature matrix Middle The characteristics of each node;

[0073] Step B242: Update node features according to attention weights:

[0074]

[0075] in, Indicates the updated The characteristics of each node;

[0076] σ(·) represents the nonlinear activation function;

[0077] Indicates the Node and The attention weight of each node;

[0078] Step B243: After updating, the features of all nodes form a node feature matrix ;

[0079] Step B244: adjacency matrix To update:

[0080]

[0081] in, express Middle Rank Elements of the column; Represents the updated adjacency matrix Middle Rank The elements of the column, according to Get the updated adjacency matrix .

[0082] Furthermore, the Node and Edge features between nodes for:

[0083]

[0084] in, Indicates calculation of the 2-norm.

[0085] Furthermore, the Node and Edge features between nodes for:

[0086] .

[0087] Furthermore, the Node and Edge features between nodes for:

[0088]

[0089] in, Represents a multilayer perceptron.

[0090] Furthermore, the specific process of step three is:

[0091] Step 31. According to and Generate weights :

[0092]

[0093] in, is the weight matrix;

[0094] represents global average pooling;

[0095] represents the bias term;

[0096] Step 3.2: According to weight right and To perform the fusion:

[0097]

[0098] in, Represents the fused features.

[0099] The beneficial effects of the present invention are:

[0100] The present invention collects and preprocesses the vibration signals of the rotating components of the camera, then inputs the preprocessed vibration signal segments into a constructed dual-channel graph convolutional network structure in feature space and topological space. The dual-channel graph convolutional network structure models the local features and global topological relationships of the signal, respectively, and can comprehensively extract local and global information of the signal. It also introduces an edge-enhanced attention mechanism and an adaptive module fusion strategy to weight the feature contributions of different modules. The weighted fused features are then input into a classifier to perform fault diagnosis, thereby improving the modeling capability and diagnostic accuracy of complex fault modes. Even when the actual distribution of the actual signal deviates from the distribution of the training samples, it can still accurately capture and diagnose the fault state of the camera, improving the system's real-time monitoring and predictive maintenance capabilities.

[0101] Moreover, compared with existing methods, the model proposed in the present invention has a simple structure and requires fewer parameters to be trained, thus reducing training time and training costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 This is a framework diagram of a method for diagnosing faults of rotating components in a phase regulator based on an adaptive multi-channel graph neural network according to the present invention;

[0103] Figure 2(a) is the confusion matrix of the GAT model;

[0104] Figure 2(b) is the confusion matrix of the AM-GCN model;

[0105] Figure 2(c) is the confusion matrix of the ST-GCN model;

[0106] Figure 2(d) is the confusion matrix of the model proposed in this invention. DETAILED DESCRIPTION

[0107] Specific implementation method 1. Combination Figure 1This embodiment describes a method for diagnosing faults of rotating components in a phase regulator based on an adaptive multi-channel graph neural network, and the method specifically includes the following steps:

[0108] Step 1: collecting vibration signals of rotating parts in the phase regulator, and then preprocessing the collected vibration signals to obtain preprocessed signal segments;

[0109] Step 2: extract the local features of each preprocessed signal segment based on the feature space graph convolution network, and extract the global topological features of each preprocessed signal segment based on the topological space graph convolution network;

[0110] Step 3: Perform weighted fusion on the local features and global topological features of each pre-processed signal segment to obtain the fused fault features;

[0111] Step 4: Input the fused fault features into the classifier, and output the fault diagnosis results of the rotating components inside the condenser through the classifier, including whether there is a fault in the rotating components inside the condenser and the specific fault type.

[0112] The network parameters in the present invention are all obtained by training based on a data set with known fault types. The rotating parts can be the shaft, bearing or rotor of the phase shifter. For example, by collecting and processing the vibration signal of the shaft, it can be determined whether the shaft has a fault based on the processing result of the shaft vibration signal; by collecting and processing the vibration signal of the rotor, it can be determined whether the rotor has a fault based on the processing result of the rotor vibration signal; by collecting and processing the vibration signal of the bearing, it can be determined whether the bearing has a fault based on the processing result of the bearing vibration signal.

[0113] Specific embodiment 2: This embodiment is a further limitation of specific embodiment 1. The vibration signal of the rotating component in the phase regulator is collected using a three-axis accelerometer. The vibration signal of the rotating component is continuously collected using the three-axis accelerometer with a sampling frequency of 10kHz.

[0114] Other steps and parameters are the same as those in the first embodiment.

[0115] A three-axis accelerometer is used to collect acceleration signals in the three axis directions, and then the acceleration signals in the three axis directions are synthesized into a combined acceleration signal according to the Pythagorean theorem. The combined acceleration signal is used to represent the total intensity of the acceleration received by the object, and then the combined acceleration signal is preprocessed.

[0116] Specific embodiment three: This embodiment further limits the specific embodiment one, and the collected vibration signal is preprocessed. The specific process is as follows:

[0117] Step 1: De-noise the collected vibration signal

[0118] A Butterworth bandpass filter with a cutoff frequency of 50-1000Hz is used to filter out high-frequency noise in the vibration signal, and a Wiener filter is then used to eliminate power frequency interference in the vibration signal to obtain a noise-reduced vibration signal.

[0119] Step 1 and 2: Set the sliding window length to 1024 points, use the sliding window to sample the vibration signal after noise reduction, and use the sliding window moving step size to 512. The signal in each sliding window is regarded as a signal segment.

[0120] Step 13: Perform maximum and minimum normalization processing on the signals in each signal segment to obtain each normalized signal segment.

[0121] Other steps and parameters are the same as those in the first embodiment.

[0122] Further explanation of steps one and two: Since the length of the sliding window is 1024 points and the overlap rate is set to 50% in the present invention, the moving step size of the sliding window is 512. When the vibration signal after noise reduction is sampled using the sliding window, the first to the 1024th points in the vibration signal after noise reduction are taken as a signal segment, the 513th to the 1536th points are taken as a signal segment, and so on.

[0123] By collecting multiple sets of time-series vibration signal data and performing preprocessing, multiple signal segments can be obtained, and label information is added to the multiple signal segments, where the label information includes whether a fault exists or not.

[0124] Specific embodiment 4: This embodiment is a further limitation of specific embodiment 1. The local features of each pre-processed signal segment are extracted based on the feature space graph convolution network. The specific process is as follows:

[0125] Step S1: Generate a feature map based on each preprocessed signal segment , specifically:

[0126] Step S11: Each pre-processed signal segment is used as a feature map A node in the , the signal segment corresponding to each node is used as the feature matrix A row of the signal fragments corresponding to all nodes constitutes the feature matrix ;

[0127] Step S12, calculate the The signal segment corresponding to the node The W distance (Wasserstein distance) of the signal segment corresponding to the node, and then the The signal segment corresponding to the node The W distance of the signal segments corresponding to the nodes is used as the similarity matrix Middle Rank Elements of a column ;

[0128] Similarly, calculate the The similarity matrix is ​​obtained by calculating the W distance between the signal segment corresponding to each node and the signal segment corresponding to each node. ;

[0129] Step S13: Using K-nearest neighbor strategy (KNN), the similarity matrix No. Sort the elements in the row from small to large and convert the similarity matrix No. Assign the first K small elements in the row a value of 1, and All other elements in the row are assigned a value of 0;

[0130] Similarly, for the similarity matrix After processing each row in separately, we get the adjacency matrix ;

[0131] Step S2: feature map As the input of the feature space graph convolutional network, the local features of each signal segment after preprocessing are output through the feature space graph convolutional network; specifically:

[0132] Step S21: Initialization ;

[0133] Step S22: Initialize the number of layers ;

[0134] Step S23, calculate the Node feature matrix output by the layer :

[0135]

[0136] in, Represents the feature space graph convolutional network The weight matrix of the layer;

[0137] represents the adjacency matrix with self-connection, , represents the identity matrix;

[0138] Representation feature map degree matrix of ;

[0139] represents the activation function;

[0140] Step S24: Determine whether , Indicates the total number of layers;

[0141] If satisfied , then As the final local feature extracted ;

[0142] If not satisfied , then let , return to execute step S23.

[0143] Other steps and parameters are the same as those in the first embodiment.

[0144] This embodiment can extract local fine-grained features of the signal.

[0145] Specific embodiment 5: This embodiment is a further limitation of specific embodiment 4. The global topological features of each pre-processed signal segment are extracted based on the topological space graph convolution network. The specific process is as follows:

[0146] Step B1: Generate a feature map based on each preprocessed signal segment , specifically:

[0147] Step B11: Use each pre-processed signal segment as a feature map A node in the , the signal segment corresponding to each node is used as the feature matrix A row of the signal fragments corresponding to all nodes constitutes the feature matrix ;

[0148] Step B12: If the fault labels of two nodes are the same, then the connection weight between them is strengthened; if the labels of two nodes are different, then the connection weight between them is weakened; then the adjacency matrix is ​​adjusted based on the label information of the nodes. To optimize:

[0149]

[0150] in, Indicates the The label of each node;

[0151] Indicates the The label of each node; the reinforcement coefficient in this invention The value of is taken as 0.5;

[0152] represents the reinforcement coefficient, ;

[0153] represents the optimized adjacency matrix;

[0154] Represents the adjacency matrix Middle Rank Elements of the column;

[0155] Represents the attenuation coefficient; in the present invention, the attenuation coefficient The value of is taken as 0.3;

[0156] Step B2: Feature map As the input of the topological space graph convolutional network, the global topological features of each signal segment after preprocessing are output through the topological space graph convolutional network; specifically:

[0157] Step B21: Initialization ,initialization ;

[0158] Step B22: Initialize the number of layers ;

[0159] Step B23: Calculate node feature matrix :

[0160]

[0161] in, Represents the topological space graph convolutional network The weight matrix of the layer;

[0162] Representation feature map degree matrix of ;

[0163] Step B24: Node feature matrix Update to get the updated node feature matrix ; For the adjacency matrix Update to get the updated adjacency matrix ;

[0164] Step B25: Determine whether , Indicates the total number of layers;

[0165] If satisfied , then As the final global topological feature extracted ;

[0166] If not satisfied , then let , return to step B23.

[0167] Other steps and parameters are the same as those in the fourth embodiment.

[0168] Specific embodiment 6: This embodiment further limits the specific embodiment 5. The specific process of step B24 is as follows:

[0169] Step B241, calculate the Node and The attention weight of each node :

[0170]

[0171] in, represents the activation function;

[0172] is a learnable vector used for attention calculation;

[0173] The superscript T indicates transposition;

[0174] A learnable weight matrix representing the transformation of node features;

[0175] Represents the node feature matrix Middle The characteristics of each node;

[0176] Represents the node feature matrix Middle The characteristics of each node;

[0177] “||” indicates feature concatenation operation;

[0178] Indicates the Node and Edge features between nodes;

[0179] Indicates the The neighbor node set of a node (according to the node feature matrix , calculate the node features of each node and the The Euclidean distance of the features of the nodes, and the nodes corresponding to the first K smallest Euclidean distances are taken as the first neighbor nodes of a node);

[0180] represents the i-th neighbor node of the i-th node;

[0181] represents the edge feature between the i-th node and the j-th node; represents the feature of the i-th node in the node feature matrix

[0182]

[0183] Step B242, updating the node feature according to the attention weight:

[0184]

[0185] wherein, represents the feature of the i-th node after updating; σ(·) represents a nonlinear activation function;

[0186]

[0187] represents the attention weight of the i-th node and the j-th node; Step B243, the features of all the nodes after updating constitute the node feature matrix

[0188]

[0189] Step B244, updating the adjacency matrix

[0190] wherein,

[0191] represents the element in the i-th row and the j-th column of represents the element in the i-th row and the j-th column of the updated adjacency matrix

[0192] The other steps and parameters are the same as those in Embodiment Five.

[0193] ​​​​​​​​​​​​​​​​​​​The node features and edge features are input into the attention calculation layer, the neighborhood transmission weight is dynamically adjusted, the node information is weighted and converged according to the calculated attention weight, the key fault features are highlighted, the capture ability of key feature information is enhanced, and the updating method of the embodiment is named as edge enhanced attention mechanism.

[0194] Specific embodiment seven: this embodiment is a further limitation of specific embodiment six, the edge feature between the first node and the second node is:

[0195]

[0196] Wherein, represents the calculation of 2 norm.

[0197] The other steps and parameters are the same as those in specific embodiment six.

[0198] Specific embodiment eight: this embodiment is a further limitation of specific embodiment six, the edge feature between the first node and the second node is:

[0199]

[0200] The other steps and parameters are the same as those in specific embodiment six.

[0201] Specific embodiment nine: this embodiment is a further limitation of specific embodiment six, the edge feature between the first node and the second node is:

[0202]

[0203] Wherein, represents the multi-layer perception.

[0204] The other steps and parameters are the same as those in specific embodiment six.

[0205] Specific embodiment ten: this embodiment is a further limitation of specific embodiment five, the specific process of step three is:

[0206] Step three one, the weight is generated according to and:

[0207] ​​​​​​​​​​​​​​

[0208] in, is a learnable weight matrix;

[0209] represents global average pooling;

[0210] represents the bias term;

[0211] Step 3.2: According to weight right and To perform the fusion:

[0212]

[0213] in, Represents the fused features.

[0214] Other steps and parameters are the same as those in the fifth embodiment.

[0215] The present invention adaptively generates fusion weights based on the overall feature distribution, so that the model pays more attention to the content information of the node in certain tasks (improving ), while other tasks focus more on the structural information of nodes (reducing ), thereby achieving a dynamic balance in the contribution to the characteristics of different modules.

[0216] Experimental verification part:

[0217] The present invention uses the public bearing dataset provided by CWRU to evaluate the effectiveness of the fault diagnosis method proposed in the present invention.

[0218] The bearing dataset contains three fault types: inner race fault (IF), outer race fault (OF), and ball fault (BF). Each fault type has three different fault sizes: 7 mils, 14 mils, and 21 mils. Therefore, the bearing dataset contains one normal state and nine fault states. The state labels are shown in Table 1:

[0219] Table 1

[0220]

[0221] For the original vibration signals of ten different health states, in order to ensure the stability of the number of different signals and make full use of the entire data set, we use the same downsampling method to sample these ten different health signals; the first 120,000 length signals of each data in the data set are intercepted as the sampling range, with a sampling window size of 125 and an interval of 125, so that 480 samples are obtained for each fault type. To improve the quality of the data set, we apply Wiener filtering to the entire data set to recover the original signal from the noise-interfered signal, and set the filter window size to 3.

[0222] To verify the effectiveness of our proposed method, we compared our model with three other advanced graph neural network diagnostic methods (GAT, AM-GCN, and ST-GCN). Six experiments were conducted for each diagnostic method, and the average results were taken. Prior to the comparative experiments, we performed an ablation experiment on our proposed model. Because our model consists of three modules, we used ablation experiments to verify the effectiveness of each module individually. The experimental results are shown in Table 2.

[0223] Table 2

[0224]

[0225] In order to observe the performance of the model intuitively, the confusion matrix of the model proposed in this invention and the other three models was generated respectively. The confusion matrix is ​​as follows: Figure 2(a) to Figure 2(d) As shown, it can be seen that the performance of the model proposed in the present invention is significantly better than the other three comparison methods.

[0226] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of rotating components in a phase regulator based on an adaptive multi-channel graph neural network, characterized in that: The method specifically comprises the following steps: Step 1: collecting vibration signals of rotating parts in the phase regulator, and then preprocessing the collected vibration signals to obtain preprocessed signal segments; Step 2: extract the local features of each preprocessed signal segment based on the feature space graph convolution network, and extract the global topological features of each preprocessed signal segment based on the topological space graph convolution network; The local features of each signal segment after preprocessing are extracted based on the feature space graph convolution network. The specific process is as follows: Step S1: Generate a feature map based on each preprocessed signal segment , specifically: Step S11: Each pre-processed signal segment is used as a feature map A node in the , the signal segment corresponding to each node is used as the feature matrix A row of the signal fragments corresponding to all nodes constitutes the feature matrix ; Step S12, calculate the The signal segment corresponding to the node The W distance of the signal segment corresponding to the node, and then The signal segment corresponding to the node The W distance of the signal segments corresponding to the nodes is used as the similarity matrix Middle Rank Elements of a column ; Similarly, calculate the The similarity matrix is ​​obtained by calculating the W distance between the signal segment corresponding to each node and the signal segment corresponding to each node. ; Step S13: Similarity matrix No. Sort the elements in the row from small to large and convert the similarity matrix No. Assign the first K small elements in the row a value of 1, and All other elements in the row are assigned a value of 0; Similarly, for the similarity matrix After processing each row in separately, we get the adjacency matrix ; Step S2: feature map As the input of the feature space graph convolutional network, the local features of each signal segment after preprocessing are output through the feature space graph convolutional network; specifically: Step S21: Initialization ; Step S22: Initialize the number of layers ; Step S23, calculate the Node feature matrix output by the layer : in, Represents the feature space graph convolutional network The weight matrix of the layer; represents the adjacency matrix with self-connection, = , represents the identity matrix; Representation feature map degree matrix of ; represents the activation function; Step S24: Determine whether , Indicates the total number of layers; If satisfied , then As the final local feature extracted ; If not satisfied , then let , return to step S23; The global topological features of each signal segment after preprocessing are extracted based on the topological space graph convolutional network. The specific process is as follows: Step B1: Generate a feature map based on each preprocessed signal segment , specifically: Step B11: Use each pre-processed signal segment as a feature map A node in the , the signal segment corresponding to each node is used as the feature matrix A row of the signal fragments corresponding to all nodes constitutes the feature matrix ; Step B12: Adjacency matrix based on node label information To optimize: in, Indicates the The label of each node; Indicates the The label of each node; represents the reinforcement coefficient, ; represents the optimized adjacency matrix; Represents the adjacency matrix Middle Rank Elements of the column; [0,1] represents the attenuation coefficient; Step B2: Feature map As the input of the topological space graph convolutional network, the global topological features of each signal segment after preprocessing are output through the topological space graph convolutional network; specifically: Step B21: Initialization ,initialization ; Step B22: Initialize the number of layers ; Step B23: Calculate node feature matrix : in, Represents the topological space graph convolutional network The weight matrix of the layer; Representation feature map degree matrix of ; Step B24: Node feature matrix Update to get the updated node feature matrix ; For the adjacency matrix Update to get the updated adjacency matrix ; The specific process of step B24 is as follows: Step B241, calculate the Node and The attention weight of each node : in, represents the activation function; is the vector used for attention calculation; The superscript T indicates transposition; represents the weight matrix; Represents the node feature matrix Middle The characteristics of each node; Represents the node feature matrix Middle The characteristics of each node; "||" indicates feature concatenation operation; Indicates the Node and Edge features between nodes; Indicates the The set of neighbor nodes of a node; Indicates the The node neighbor nodes; Indicates the Node and Edge features between nodes; Represents the node feature matrix Middle The characteristics of each node; Step B242: Update node features according to attention weights: in, Indicates the updated The characteristics of each node; represents a nonlinear activation function; Indicates the Node and The attention weight of each node; Step B243: After updating, the features of all nodes form a node feature matrix ; Step B244: adjacency matrix To update: in, express Middle Rank Elements of the column; Represents the updated adjacency matrix Middle Rank The elements of the column, according to Get the updated adjacency matrix ; Step B25: Determine whether , Indicates the total number of layers; If satisfied , then As the final global topological feature extracted ; If not satisfied , then let , return to step B23; Step 3: Perform weighted fusion on the local features and global topological features of each pre-processed signal segment to obtain the fused fault features; Step 4: Input the fused fault features into the classifier, and output the fault diagnosis result of the phase shifter through the classifier.

2. The method for fault diagnosis of rotating components in a phase regulator based on an adaptive multi-channel graph neural network according to claim 1, characterized in that: The vibration signals of the rotating parts in the phase regulator are collected using a three-axis accelerometer, which is used to continuously collect the vibration signals of the rotating parts at a sampling frequency of 10 kHz.

3. The method for fault diagnosis of rotating components in a phase regulator based on an adaptive multi-channel graph neural network according to claim 1, characterized in that: The collected vibration signal is preprocessed, and the specific process is as follows: Step 1: De-noise the collected vibration signal A Butterworth bandpass filter is used to filter out high-frequency noise in the vibration signal, and then a Wiener filter is used to eliminate power frequency interference in the vibration signal to obtain a vibration signal after noise reduction. Step 1 and 2: Set the sliding window length to 1024 points, use the sliding window to sample the vibration signal after noise reduction, and use the sliding window moving step size to 512. The signal in each sliding window is regarded as a signal segment. Step 13: Perform maximum and minimum normalization processing on the signals in each signal segment to obtain each normalized signal segment.

4. The method for fault diagnosis of rotating components in a phase regulator based on an adaptive multi-channel graph neural network according to claim 1, characterized in that: The said Node and Edge features between nodes for: in, Indicates calculation of the 2-norm.

5. The method for fault diagnosis of rotating components in a phase regulator based on an adaptive multi-channel graph neural network according to claim 1, characterized in that: The said Node and Edge features between nodes for: 。 6. The method for fault diagnosis of rotating components in a phase regulator based on an adaptive multi-channel graph neural network according to claim 1, characterized in that: The said Node and Edge features between nodes for: in, Represents a multilayer perceptron.

7. The method for fault diagnosis of rotating components in a phase regulator based on an adaptive multi-channel graph neural network according to claim 1, characterized in that: The specific process of step three is: Step 31. According to and Generate weights : in, is the weight matrix; represents global average pooling; represents the bias term; Step 3.2: According to weight right and To perform the fusion: in, Represents the fused features.

Images