Rotating part fault diagnosis method and device

Abstract

The invention discloses a rotating part fault diagnosis method and device. The method comprises the following steps: acquiring to-be-detected time domain signal data of a rotating part; preprocessing the to-be-detected time domain signal data to generate non-Euclidean space data; performing global feature extraction and local feature extraction on the non-Euclidean space data by adopting an improved fusion graph neural network, and outputting a plurality of probability vectors; adopting a preset maximum value function to generate a fault diagnosis result of the rotating part according to each probability vector; the technical problem that a traditional rotating part fault diagnosis method is difficult to maintain high accuracy under variable working conditions is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method and device for diagnosing faults of rotating parts. Background Art

[0002] High-end precision electronic manufacturing equipment is electronic equipment with high technology content, high added value, and at the high end of the industrial chain, such as communication and navigation equipment, radar, large antennas, high-performance computers, high-end network equipment, etc. These equipment represent the advanced level of today's electronic manufacturing field in terms of hardware, software, core technology, and manufacturing process.

[0003] Rotating parts are key components of high-end precision electronic manufacturing equipment, and their fault detection is of great significance to ensure the stable operation of the overall equipment. However, rotating parts are often affected by changes in working load, speed, temperature and other working conditions in the complex and changeable actual operating environment, resulting in complex correlations between fault data, which are no longer limited to traditional linear or geometric relationships.

[0004] Traditional rotating component fault diagnosis methods are often based on the distance and similarity between data points, which can be measured by simple straight line or curve distances. However, when rotating components operate under variable operating conditions, their fault data will show significant high complexity and dynamic variability, making it difficult for traditional rotating component fault diagnosis methods to maintain high accuracy under variable operating conditions. Summary of the invention

[0005] The present invention provides a rotating component fault diagnosis method and device, which are used to solve the technical problem that the traditional rotating component fault diagnosis method is difficult to maintain high accuracy under variable working conditions.

[0006] A first aspect of the present invention provides a method comprising:

[0007] Acquiring time domain signal data to be detected of the rotating component;

[0008] Preprocessing the time domain signal data to be detected to generate non-Euclidean space data;

[0009] Using an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean spatial data, and output multiple probability vectors;

[0010] A preset maximum value function is used to generate a fault diagnosis result of the rotating component according to each of the probability vectors.

[0011] Optionally, the step of preprocessing the time domain signal data to be detected to generate non-Euclidean space data includes:

[0012] Performing a fast Fourier transform on the time-domain signal data to be detected, and outputting frequency-domain signal data;

[0013] A path composition method is used to perform non-Euclidean data construction on the frequency domain signal data to generate non-Euclidean space data.

[0014] Optionally, the improved fusion graph neural network includes an improved graph convolution module, a multi-head graph attention module, a fully connected layer and a random dropout layer; the step of using the improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean space data and outputting multiple probability vectors includes:

[0015] Iteratively updating the global features of the non-Euclidean space data through an improved graph convolution module to generate a plurality of first node features and a plurality of second node features;

[0016] Using a multi-head graph attention module to perform node fusion on each of the first node features and each of the second node features, and output node fusion normalized features;

[0017] Inputting the node fusion normalized features into the fully connected layer for linear transformation to generate transformed node features;

[0018] Regularizing the transformed node features through a random dropout layer, and outputting regularized node features;

[0019] An improved graph convolution module is used to iteratively update the regularized node features globally to generate multiple probability vectors.

[0020] Optionally, the improved graph convolution module includes an improved graph convolution layer and a normalization layer; the step of iteratively updating the global features of the non-Euclidean space data through the improved graph convolution module to generate a plurality of first node features and a plurality of second node features includes:

[0021] Through an improved graph convolution layer based on a preset symmetry processing mechanism, the non-Euclidean spatial data is subjected to polynomial iterative fusion to generate a linear combination of node features;

[0022] A normalization layer is used to normalize the linear combination of the node features, and a plurality of first node features and a plurality of second node features are output.

[0023] Optionally, the multi-head graph attention module includes a multi-head graph attention submodule and a normalization layer; the step of using the multi-head graph attention module to perform node fusion on each of the first node features and each of the second node features, and outputting node fusion normalized features, includes:

[0024] Perform feature enhancement fusion on each of the first node features and each of the second node features through a multi-head graph attention submodule, and output node fusion transformation features;

[0025] A normalization layer is used to normalize the node fusion transformation features to generate node fusion normalized features.

[0026] Optionally, the multi-head graph attention submodule includes a concatenation layer, a first activation function layer, a second activation function layer, and a third activation function layer; the step of performing feature enhancement fusion on each of the first node features and each of the second node features through the multi-head graph attention submodule and outputting node fusion transformation features includes:

[0027] Performing a multiplication operation on each of the first node features and a preset shared parameter to determine a first node parameter feature corresponding to each of the first node features;

[0028] Performing a multiplication operation on each of the second node features and the preset shared parameter to determine a second node parameter feature corresponding to each of the second node features;

[0029] Splicing each of the first node parameter features and each of the second node parameter features through a splicing layer to generate a plurality of node splicing features;

[0030] Inputting each of the node splicing features into a first activation function layer for nonlinear mapping, and outputting first node mapping features corresponding to each of the node splicing features;

[0031] Inputting each of the first node mapping features into a second activation function layer for normalization, and outputting a plurality of attention coefficients;

[0032] Based on a multi-head graph attention mechanism, weighted fusion is performed on each of the attention coefficients and each of the second node parameter features to generate a node fusion domain feature;

[0033] The node fusion domain features are nonlinearly transformed through the third activation function layer, and the node fusion transformation features are output.

[0034] Optionally, before the step of acquiring the time domain signal data to be detected of the rotating component, the method further comprises:

[0035] Acquire time domain signal data to be trained, and preprocess the time domain signal data to be trained to generate non-Euclidean space data to be trained;

[0036] Based on the back propagation algorithm, an initial improved fusion graph neural network is used to determine the model gradient according to the non-Euclidean space data to be trained;

[0037] Based on a preset stochastic gradient descent optimizer, the model parameters of the initial improved fusion graph neural network are updated through the model gradient to determine an intermediate improved fusion graph neural network;

[0038] Using the intermediate improved fusion graph neural network and the preset loss function to calculate the target loss value according to the non-Euclidean space data to be trained, and judging whether the target loss value converges;

[0039] If the target loss value converges, the intermediate improved fusion graph neural network is used as the trained improved fusion graph neural network.

[0040] A second aspect of the present invention provides a method comprising:

[0041] An acquisition module, used for acquiring time domain signal data to be detected of the rotating component;

[0042] A preprocessing module, used for preprocessing the time domain signal data to be detected to generate non-Euclidean space data;

[0043] A feature extraction module, used for performing global feature extraction and local feature extraction on the non-Euclidean spatial data by using an improved fusion graph neural network, and outputting a plurality of probability vectors;

[0044] The result generating module is used to generate the fault diagnosis result of the rotating component according to each of the probability vectors by using a preset maximum value function.

[0045] A third aspect of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the rotating component fault diagnosis method as described in any one of the above items.

[0046] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the steps of the rotating component fault diagnosis method as described in any one of the above items.

[0047] It can be seen from the above technical solutions that the present invention has the following advantages:

[0048] The above technical scheme of the present invention provides a method for diagnosing faults of rotating components, which first obtains time domain signal data to be detected of the rotating components; then, preprocesses the time domain signal data to be detected to generate non-Euclidean space data; uses an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean space data, and outputs multiple probability vectors; finally, uses a preset maximum value function to generate a fault diagnosis result of the rotating component according to each probability vector; based on the above scheme, an improved fusion graph neural network is used to perform global feature extraction and local feature extraction on the non-Euclidean space data obtained after preprocessing, and generates a fault diagnosis result of the rotating component according to each output probability vector, which can comprehensively extract the global and local fault characteristics of the rotating component, and maintain a high accuracy of fault diagnosis even under variable operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 A flowchart of a method for diagnosing a fault of a rotating component provided by an embodiment of the present invention;

[0051] Figure 2 A flowchart of another method for diagnosing a fault of a rotating component provided by an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of the structure of an initial improved fusion graph neural network training framework provided by an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of the structure of a multi-head graph attention submodule provided in an embodiment of the present invention;

[0054] Figure 5 A structural block diagram of a rotating component fault diagnosis device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The embodiments of the present invention provide a rotating component fault diagnosis method and device, which are used to solve the technical problem that the traditional rotating component fault diagnosis method is difficult to maintain high accuracy under variable working conditions.

[0056] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] See also Figure 1 , Figure 1 A flowchart of the steps of a rotating component fault diagnosis method provided by an embodiment of the present invention.

[0058] The present invention provides a method for diagnosing faults of rotating parts, comprising:

[0059] Step 101: Acquire time domain signal data to be detected of a rotating component.

[0060] In this embodiment, the time domain signal data to be detected of the rotating component is obtained.

[0061] Step 102: pre-process the time domain signal data to be detected to generate non-Euclidean space data.

[0062] Specifically, fast Fourier transform is performed on the time domain signal data to be detected to output frequency domain signal data; non-Euclidean data is constructed on the frequency domain signal data using a path composition method to generate non-Euclidean space data; wherein the path composition method is a Path composition method.

[0063] It is worth mentioning that in order to mine the relational information between data, in the process of data preprocessing, the input time domain signal data to be detected is first processed by fast Fourier transform, the signal is converted from time domain to frequency domain, and the frequency domain signal data is obtained. Then, the Path composition method is used to construct the frequency domain signal data into non-Euclidean space data, so as to better capture the complex correlation between the data.

[0064] In this embodiment, the time domain signal data to be detected is preprocessed to generate non-Euclidean space data.

[0065] Step 103: Use an improved fusion graph neural network to perform global feature extraction and local feature extraction on non-Euclidean spatial data, and output multiple probability vectors.

[0066] The improved fusion graph neural network consists of two improved graph convolution modules, a multi-head graph attention module, a fully connected layer and a random dropout layer.

[0067] Specifically, the global features of non-Euclidean space data are iteratively updated through an improved graph convolution module to generate multiple first node features and multiple second node features; the multi-head graph attention module is used to perform node fusion on each first node feature and each second node feature, and the node fusion normalized features are output; the node fusion normalized features are input into the fully connected layer for linear transformation to generate transformed node features; the transformed node features are regularized through a random inactivation layer, and the regularized node features are output; the improved graph convolution module is used to iteratively update the global features of the regularized node features to generate multiple probability vectors.

[0068] Furthermore, in order to take into account the extraction of global and local features of the data, the present invention uses the Chebyshev graph convolution network (ChebyNet, ChebyshevNetwork), that is, the improved graph convolution module to iteratively update the global node features, and combines the multi-head graph attention mechanism, that is, the multi-head graph attention module (GAT, GraphAttentionNetwork), so that the model can deeply mine and focus on key local features. The improved graph convolution module is used as the first and last layers of the improved fusion graph neural network. The present invention improves the feature extraction process by symmetric processing, so that it can more evenly consider the relationship between nodes and accurately reflect the node characteristics.

[0069] In this embodiment, an improved fusion graph neural network is used to perform global feature extraction and local feature extraction on non-Euclidean spatial data, and output multiple probability vectors.

[0070] Step 104: Generate a fault diagnosis result of the rotating component according to each probability vector using a preset maximum value function.

[0071] The preset maximum value function is the Argmax function.

[0072] It should be noted that the output for each sample is an n-dimensional probability vector (each dimension represents the probability P of the sample being the i-th category). i ). The final fault diagnosis result is the predicted category label. Specifically, the probability vector P i The Argmax function is input for processing. The Argmax function outputs the index of the largest element in the probability vector (i.e., the maximum probability value), and the category label corresponding to the index is output as the final fault diagnosis result.

[0073] In this embodiment, a preset maximum value function is used to generate a fault diagnosis result of a rotating component according to each probability vector.

[0074] In an embodiment of the present invention, the present invention provides a method for diagnosing faults of rotating components, firstly, obtaining time domain signal data to be detected of the rotating components; then, preprocessing the time domain signal data to be detected to generate non-Euclidean space data; using an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean space data, and outputting multiple probability vectors; finally, using a preset maximum value function to generate a fault diagnosis result of the rotating component according to each probability vector; based on the above scheme, using an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean space data obtained after preprocessing, and generating the fault diagnosis result of the rotating component according to each output probability vector, the process can comprehensively extract the global and local fault characteristics of the rotating components, and can maintain a high accuracy of fault diagnosis even under variable operating conditions.

[0075] See also Figure 2 , Figure 2 A flowchart of the steps of another rotating component fault diagnosis method provided by an embodiment of the present invention.

[0076] The present invention provides a method for diagnosing faults of rotating parts, comprising:

[0077] Step 201: Obtain time domain signal data to be trained, and preprocess the time domain signal data to be trained to generate non-Euclidean space data to be trained.

[0078] Please note that Figure 3-Figure 4 The initial improved fusion graph neural network training framework proposed in the present invention includes three parts: data preprocessing, initial improved fusion graph neural network, and multi-task collaborative diagnosis; wherein, the initial improved fusion graph neural network consists of two improved graph convolution modules, a multi-head graph attention module, a fully connected layer (Fully Connected) and a random inactivation layer (Dropout Layer), and the graph convolution module consists of an improved graph convolution layer and a normalization layer (BatchNormalization); the multi-head graph attention module consists of a multi-head graph attention sub-module and a normalization layer; the multi-head graph attention sub-module consists of a concatenation layer (Concat), a first activation function layer (LeakyReLU), a second activation function layer (Softmax) and a third activation function layer (ReLU).

[0079] Furthermore, in the process of data preprocessing, the input time domain signal data to be trained is first processed by fast Fourier transform, the signal is converted from time domain to frequency domain, and the frequency domain signal data to be trained is obtained. Then, the Path composition method is used to construct the frequency domain signal data to be trained into non-Euclidean space data (non-Euclidean space data to be trained) for model training, so as to better capture the complex correlation between the data.

[0080] In this embodiment, time domain signal data to be trained is obtained, and the time domain signal data to be trained is preprocessed to generate non-Euclidean space data to be trained.

[0081] Step 202: Based on the back propagation algorithm, the initial improved fusion graph neural network is used to determine the model gradient according to the non-Euclidean space data to be trained.

[0082] It should be noted that in the process of multi-task collaborative diagnosis, in order to better cope with the complexity and dynamics of data under variable working conditions, the model forms a multi-task collaborative fault diagnosis framework by simultaneously executing node classification tasks and node reconstruction tasks. In the node classification task, the second improved ChebyNet layer is optimized through supervised learning, that is, the node representation of the output of the graph convolution layer 2 is improved, so that the model can more accurately adapt to the classification requirements; in the node reconstruction task, a self-supervised learning mechanism is introduced to map the original input features by constructing a node reconstruction layer to provide additional training signals for the model.

[0083] Furthermore, in the forward propagation, for each sample, the network output is an n-dimensional probability vector (each dimension represents the probability P of the sample being the i-th category). i ) and node reconstruction feature vector reconstructed_x i (The output of the second ChebyNet layer is transformed by using a fully connected layer to make it consistent with the feature dimension of the original input), corresponding to the node classification cross entropy loss loss and the node reconstruction mean square error loss reconstruction_loss, that is, the first loss value and the second loss value respectively.

[0084] Specifically, firstly, an initial improved fusion graph neural network is used to output a linear combination of node features to be trained and multiple probability vectors to be trained according to the non-Euclidean space data to be trained; the linear combination of node features to be trained is transformed in dimension through a fully connected layer to obtain a node reconstruction feature vector; a preset first loss function is used to calculate a first loss value according to multiple probability vectors to be trained; a preset second loss function is used to calculate a second loss value according to the node reconstruction feature vector; based on a back propagation algorithm, the model gradient is determined according to the first loss value and the second loss value; wherein, the linear combination of node features to be trained is the output of the improved graph convolution layer 2, that is, the improved graph convolution layer set at the last layer in the initial improved fusion graph neural network.

[0085] Furthermore, the first loss function is preset, specifically:

[0086]

[0087] Among them, loss is the first loss value, which indicates the difference between the prediction and the true label of each category; n is the number of categories; t i is the value of the i-th element in the true label vector (when the true category of the sample is the i-th category, t i is 1, otherwise 0); P i is the probability vector to be trained.

[0088] Furthermore, the node reconstruction loss reconstruction_loss is calculated using the root mean square error, and the second loss function is preset, specifically:

[0089]

[0090] Among them, reconstruction_loss is the second loss value, which represents the difference between the reconstructed output of the node and the original input; m is the number of feature dimensions; reconstructed_x i Reconstruct the i-th feature dimension in the feature vector for the node; x i is the i-th feature dimension in the non-Euclidean space data to be trained.

[0091] Furthermore, in the back propagation, the previous gradient is first cleared, and the sum of the node classification cross entropy loss loss and the node reconstruction loss reconstruction_loss is the total loss tatal_loss, that is, the target loss value. Finally, the back propagation algorithm is called to calculate the gradient relative to the model parameters, that is, the model gradient; the calculation process of the model gradient is specifically as follows:

[0092]

[0093] Among them, tatal_loss is the target loss value; reconstruction_loss is the second loss value; loss is the first loss value; is the model gradient, which indicates the target loss value with respect to the model parameter θ i The gradient of is the first loss value with respect to the model parameter θ i The gradient of is the second loss value about the model parameter θ i gradient.

[0094] In this embodiment, based on the back propagation algorithm, an initial improved fusion graph neural network is used to determine the model gradient according to the non-Euclidean space data to be trained.

[0095] Step 203: Based on a preset stochastic gradient descent optimizer, the model parameters of the initial improved fusion graph neural network are updated through the model gradient to determine an intermediate improved fusion graph neural network.

[0096] It should be noted that the network training process aims to minimize the loss function. Stochastic Gradient Descent (SGD) is used as the optimizer to update the model parameters in the opposite direction according to the gradient. The updating process of the model parameters can be expressed as:

[0097]

[0098] in, is the model parameter of the intermediate improved fusion graph neural network in the t+1th iteration; are the model parameters of the initial improved fusion graph neural network in the tth iteration; is the model gradient; lr is the learning rate.

[0099] In this embodiment, based on a preset stochastic gradient descent optimizer, the model parameters of the initial improved fusion graph neural network are updated through the model gradient to determine the intermediate improved fusion graph neural network.

[0100] Step 204: Use the intermediate improved fusion graph neural network and the preset loss function to calculate the target loss value according to the non-Euclidean space data to be trained, and determine whether the target loss value converges.

[0101] The preset loss function includes a preset first loss function and a preset first loss function.

[0102] It should be noted that the intermediate improved fusion graph neural network is used to obtain a new node reconstruction feature vector and multiple new probability vectors to be trained according to the non-Euclidean space data to be trained, and then the preset loss function is used to calculate the target loss value according to the new node reconstruction feature vector and the multiple new probability vectors to be trained, and it is judged whether the target loss value converges; if the target loss value does not converge, then based on the back propagation algorithm, the model gradient is determined according to the target loss value, and the model gradient is used to update the model parameters of the intermediate improved fusion graph neural network to determine a new intermediate improved fusion graph neural network, and the new intermediate improved fusion graph neural network and the preset loss function are used to recalculate the target loss value according to the non-Euclidean space data to be trained until the target loss value converges, and the intermediate improved fusion graph neural network determined when the target loss value converges is used as the trained improved fusion graph neural network.

[0103] In this embodiment, an intermediate improved fusion graph neural network and a preset loss function are used to calculate the target loss value according to the non-Euclidean space data to be trained, and to determine whether the target loss value converges.

[0104] Step 205: If the target loss value converges, the intermediate improved fusion graph neural network is used as the trained improved fusion graph neural network.

[0105] In this embodiment, if the target loss value converges, the intermediate improved fusion graph neural network is used as the trained improved fusion graph neural network.

[0106] Step 206: Acquire the time domain signal data to be detected of the rotating component.

[0107] In this embodiment, the time domain signal data to be detected of the rotating component is obtained.

[0108] Step 207: pre-process the time domain signal data to be detected to generate non-Euclidean space data.

[0109] In this embodiment, the time domain signal data to be detected is preprocessed to generate non-Euclidean space data.

[0110] Step 208: Use an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean spatial data, and output multiple probability vectors.

[0111] The improved fusion graph neural network includes an improved graph convolution module, a multi-head graph attention module, a fully connected layer and a random dropout layer.

[0112] Further, step 208 may include the following sub-steps:

[0113] S81. Iteratively update the global features of the non-Euclidean space data through an improved graph convolution module to generate a plurality of first node features and a plurality of second node features;

[0114] The improved graph convolution module includes an improved graph convolution layer and a normalization layer.

[0115] Further, S81 may include the following sub-steps:

[0116] S811, through the improved graph convolution layer based on the preset symmetry processing mechanism, the non-Euclidean spatial data is iteratively fused with polynomials to generate linear combinations of node features;

[0117] S812. Use a normalization layer to normalize the linear combination of node features, and output multiple first node features and multiple second node features.

[0118] It is worth mentioning that each iteration in the original processing integrates the previous and next results, captures multi-order neighbor information, and recursively updates to help node feature propagation and aggregation. The initial node feature information is introduced by subtracting the operation of Z0 to strengthen its propagation effect. Combining recursive updates with original features improves the model's ability to capture complex relationships in the graph. By analogy, the node feature Z obtained by approximating the k-th order Chebyshev polynomial can be obtained. k , the process can be expressed as:

[0119]

[0120] Among them, Z0 is the initial node feature matrix; X is the non-Euclidean space data; Z1 is the node feature after one Chebyshev polynomial iteration; is the scaled and standardized Laplacian matrix; Z2 is the node feature after two Chebyshev polynomial iterations; Z k is the node feature after k Chebyshev polynomial iterations; Z (k-1) is the node feature after k-1 Chebyshev polynomial iterations; Z (k-2) is the node feature after k-2 Chebyshev polynomial iterations.

[0121] Furthermore, since Z0 represents the initial node feature matrix, which is asymmetric, symmetry processing is introduced in the calculation process of Z2 (preset symmetry processing mechanism), and the intermediate matrix Z2 is added to its transpose to take the average value. After the introduction of symmetry processing, Z can be obtained through recursive operation. k '; Finally, introduce the symmetry treatment Z k ' and the weight matrix Θ of the Chebyshev graph convolution layer k The multiplication operation is performed so that each order of node features is multiplied by its corresponding weight matrix. This process effectively integrates Chebyshev polynomials of different orders, thereby accurately reflecting the update process of node features. After a complete k-order Chebyshev polynomial recursion, the linear combination X′ of the updated node features is obtained; the processing process of the linear combination of node features can be expressed as:

[0122]

[0123] Among them, X is non-Euclidean spatial data; Z0 is the initial node feature matrix; Z1 is the node feature after one Chebyshev polynomial iteration; Z2′ is the node feature obtained by adding the intermediate matrix Z2 and its transpose to obtain the average value, that is, the node feature after two Chebyshev polynomial iterations based on the preset symmetry processing mechanism; Zk ′ is the node feature after k Chebyshev polynomial iterations based on the preset symmetry processing mechanism; is the scaled and normalized Laplacian matrix, L is the Laplacian matrix, I is the identity matrix, λ max is the maximum eigenvalue of the Laplacian matrix; Z2 is the node feature after two Chebyshev polynomial iterations; Z (k-1) is the node feature after k-1 Chebyshev polynomial iterations; Z (k-2) is the node feature after k-2 Chebyshev polynomial iterations; Θ k is the weight matrix of the Chebyshev graph convolutional layer; X′ is the linear combination of node features.

[0124] S82, using a multi-head graph attention module to perform node fusion on each first node feature and each second node feature, and output node fusion normalized features;

[0125] The multi-head graph attention module includes a multi-head graph attention sub-module and a normalization layer.

[0126] Further, S82 may include the following sub-steps:

[0127] S821, performing feature enhancement fusion on each first node feature and each second node feature through a multi-head graph attention submodule, and outputting a node fusion transformation feature;

[0128] The multi-head graph attention submodule includes a concatenation layer, a first activation function layer, a second activation function layer, and a third activation function layer.

[0129] Further, S821 may include the following sub-steps:

[0130] S8211, performing a multiplication operation on each first node feature and a preset shared parameter to determine a first node parameter feature corresponding to each first node feature;

[0131] S8212, performing a multiplication operation on each second node feature and a preset shared parameter to determine a second node parameter feature corresponding to each second node feature;

[0132] S8213, splicing each first node parameter feature and each second node parameter feature through a splicing layer to generate a plurality of node splicing features;

[0133] S8214, inputting each node splicing feature into the first activation function layer for nonlinear mapping, and outputting the first node mapping feature corresponding to each node splicing feature;

[0134] S8215, inputting each first node mapping feature into the second activation function layer for normalization, and outputting a plurality of attention coefficients;

[0135] S8216. Based on the multi-head graph attention mechanism, weighted fusion is performed on each attention coefficient and each second node parameter feature to generate a node fusion domain feature;

[0136] S8217. Perform nonlinear transformation on the node fusion domain features through the third activation function layer, and output the node fusion transformation features.

[0137] It should be noted that in order to accurately capture the local features of fault data, the present invention introduces GAT, i.e., multi-head graph attention submodule, which comprehensively considers the neighborhood features of nodes and weight update representation. Its input is the output of the first ChebyNet layer, and its output is used as the input of the second ChebyNet layer. Through this fusion method, both global information and local details are considered, and the deep fault features of rotating parts under variable working conditions are fully captured.

[0138] Specifically, for each node i, its neighbor nodes (j∈N i ) calculates the correlation coefficient with itself one by one, and its linear mapping increases the dimension of the node features to achieve feature enhancement, and splices the transformed features of nodes i and j, that is, splices the first node parameter features and the second node parameter features through the splicing layer to generate multiple node splicing features, and maps the spliced ​​high-dimensional features to a real number e ij Up, e ij is the similarity coefficient between node i and node j. The correlation between nodes i and j is completed by the learnable parameter W, that is, the preset shared parameters and mapping. ij After that, it is normalized by the Softmax activation function (the second activation function layer) to obtain the attention coefficient α ij ; Among them, the processing process of the attention coefficient can be expressed as:

[0139]

[0140] Among them, α ij is the attention coefficient; exp(·) is the Softmax activation function; LeakyReLU(·) is the LeakyReLU activation function; N i is the neighbor node domain; e ik is the similarity coefficient between node i and node k, k is the kth domain (neighbor) node of node i, and k belongs to N i , N i is the number of neighboring nodes of node i, including node i itself; e ijis the similarity coefficient between node i and node j; a(·) is the feature map, which is used to map the concatenated high-dimensional features to a real number e ij (Wh i ||Wh j ) is the node splicing feature; h i is the first node feature, indicating the node feature of node i; h j is the second node feature, representing the node feature of node j; W is the preset shared parameter.

[0141] Furthermore, using α ij The features are weighted and fused to obtain new node features. The present invention adopts a multi-head graph attention mechanism to obtain node features through different attentions and average to obtain new node features. That is, based on the multi-head graph attention mechanism, each attention coefficient and each second node parameter feature are weightedly fused to generate node fusion domain features; the node fusion domain features are nonlinearly transformed through the third activation function layer, and the node fusion transformation features are output; wherein, the processing process of the node fusion transformation features can be expressed as:

[0142]

[0143] Among them, h i ′ is the fused transformation feature of node i; M is the number of heads of multi-head attention; is the attention coefficient of node i and node j calculated by the Mth attention head according to the above attention coefficient calculation steps; W M is the input linear transformation weight matrix corresponding to the Mth attention head; σ(·) is the ReLU activation function; h j is the node feature of node j, node j is the neighbor node of node i, that is, the second node feature; N i is the number of neighboring nodes of node i, including node i itself.

[0144] S822. Use a normalization layer to normalize the node fusion transformation features to generate node fusion normalized features.

[0145] S83, input the node fusion normalized features into the fully connected layer for linear transformation to generate transformed node features;

[0146] S84, regularizing the transformed node features through a random dropout layer, and outputting the regularized node features;

[0147] S85. Use the improved graph convolution module to iteratively update the regularized node features globally and generate multiple probability vectors.

[0148] It should be noted that the principle of the steps of iteratively updating the global features of the regularized node features using the improved graph convolution module is consistent with the principle of iteratively updating the global features of the non-Euclidean space data using the improved graph convolution module, and the present invention will not go into details.

[0149] In this embodiment, an improved fusion graph neural network is used to perform global feature extraction and local feature extraction on non-Euclidean spatial data, and output multiple probability vectors.

[0150] Step 209: Generate a fault diagnosis result of the rotating component according to each probability vector using a preset maximum value function.

[0151] In this embodiment, a preset maximum value function is used to generate a fault diagnosis result of a rotating component according to each probability vector.

[0152] As a comparison of technical effects, we can refer to the existing technology. Traditional fault diagnosis methods are often based on the distance and similarity between data points, which can be measured by simple straight line or curve distances. However, when rotating parts operate under variable operating conditions, their fault data show significant high complexity and dynamic variability. Therefore, traditional fault diagnosis methods often find it difficult to maintain high accuracy and stability under variable operating conditions.

[0153] In view of the above problems, the present invention proposes a method for fault diagnosis of rotating parts. This method designs a fusion graph neural network to comprehensively extract the global features of the fault data of rotating parts under variable working conditions, and deeply mines its crucial local features; adopts a multi-task collaborative diagnosis strategy, and effectively enhances the fault feature extraction ability of the main classification task in a complex environment by constructing auxiliary tasks of self-supervised learning. By establishing a logical relationship of multi-task collaboration, we can deeply understand and respond to the various changes produced under various working conditions. This multi-task collaborative method not only improves the accuracy and stability of fault diagnosis, but also enhances the adaptability and generalization ability of variable working condition data, so that it can perform more robustly and reliably when facing changes in working conditions; in the fusion graph neural network module, the Chebyshev graph convolution and the multi-head graph attention mechanism are integrated to comprehensively extract the global and local fault features of rotating parts. In the multi-task collaborative diagnosis module, the node reconstruction task is introduced. The model attempts to restore the representation of the original node through self-supervised learning. By simultaneously executing multi-task collaborative training of node classification and node reconstruction, the adaptability under variable operating conditions is significantly enhanced, enabling it to more accurately capture the key fault characteristics of rotating components, thereby effectively improving the accuracy and stability of fault diagnosis of rotating components under variable operating conditions.

[0154] In summary, the present invention can deeply extract key fault features in the data and mine the potential structural relationships between the data to cope with the challenges brought by variable operating conditions, ensure the accuracy and stability of rotating component fault diagnosis under variable operating conditions, and at the same time, it can be compatible with multiple types of data, deeply mine intrinsic features and potential structural relationships, significantly improve diagnostic performance, and also show significant feature extraction capabilities, which can accurately distinguish different types of faults. Compared with traditional fault diagnosis models, it shows better classification accuracy and diagnostic effects, and shows excellent performance and strong anti-interference ability under complex and changeable working conditions, always maintaining high accuracy and stability, and ensuring the reliability and effectiveness of fault diagnosis; unique improvements have been made in the Chebyshev graph convolutional neural network, and a symmetry processing mechanism has been introduced in the node feature iteration process, so as to better reflect the node features, significantly enhance the representation ability of the node features, and ensure that the graph convolution operation can more evenly consider the relationship between nodes when the node features are propagated, so that the model is more inclined to capture features with symmetrical properties during the learning process; in the feature extraction part, the improved Chebyshev graph convolution and the multi-head graph attention mechanism are cleverly integrated. The graph convolutional network extracts feature information from a global perspective by iteratively fusing the features of nodes and their neighbors. At the same time, a multi-head graph attention mechanism is introduced to make the model more focused on the local connections between nodes, thereby significantly improving its adaptability to complex data. A node reconstruction layer is constructed based on the characteristics of the graph data structure, and the feature dimension of the output of the improved fusion graph neural network is reconstructed to form a node reconstruction task for multi-task collaborative diagnosis of rotating parts. In the node reconstruction task, a self-supervised learning mechanism is innovatively introduced. This mechanism reconstructs node features to map their original input features, which not only retains the original information but also introduces higher-level abstract features.The node representation is optimized by self-supervised learning, which significantly improves the adaptability of the model in specific diagnostic tasks. The present invention also designs a multi-task collaborative fault diagnosis strategy, which combines the node classification task with the node reconstruction task, provides the model with rich learning signals, effectively copes with the complexity and dynamics of data under variable working conditions, and significantly improves the fault diagnosis performance of the model under variable working conditions. The supervised learning of the node classification task is combined with the self-supervised learning of the node reconstruction task. Through this mode, the diagnostic accuracy and stability of the model under variable working conditions are effectively improved, and the fault diagnosis performance is optimized. By comprehensively optimizing the loss functions of the two tasks, namely the cross entropy loss of the node classification task and the mean square error loss of the node reconstruction task, back propagation optimization is performed by minimizing the total loss, which helps the model to more deeply understand the complex structure and characteristics of the data, thereby improving the performance of fault diagnosis. The Euclidean space data collected by the sensor is modeled into a non-Euclidean space, and the improved fusion graph neural network is used to capture the rich relationship information therein, effectively extract the interdependence between the data, and provide valuable auxiliary information for distinguishing different fault modes under variable working conditions, thereby improving the accuracy and reliability of fault diagnosis.

[0155] In an embodiment of the present invention, the present invention provides a method for diagnosing faults of rotating components, firstly, obtaining time domain signal data to be detected of the rotating components; then, preprocessing the time domain signal data to be detected to generate non-Euclidean space data; using an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean space data, and outputting multiple probability vectors; finally, using a preset maximum value function to generate a fault diagnosis result of the rotating component according to each probability vector; based on the above scheme, using an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean space data obtained after preprocessing, and generating the fault diagnosis result of the rotating component according to each output probability vector, the process can comprehensively extract the global and local fault characteristics of the rotating components, and can maintain a high accuracy of fault diagnosis even under variable operating conditions.

[0156] See also Figure 5 , Figure 5 A structural block diagram of a rotating component fault diagnosis device provided by an embodiment of the present invention.

[0157] The present invention provides a rotating component fault diagnosis device, comprising:

[0158] An acquisition module 501 is used to acquire time domain signal data to be detected of a rotating component;

[0159] A preprocessing module 502, used for preprocessing the time domain signal data to be detected to generate non-Euclidean space data;

[0160] A feature extraction module 503 is used to perform global feature extraction and local feature extraction on non-Euclidean spatial data using an improved fusion graph neural network, and output a plurality of probability vectors;

[0161] The result generating module 504 is used to generate the fault diagnosis result of the rotating component according to each probability vector by using a preset maximum value function.

[0162] Furthermore, the preprocessing module 502 is specifically configured to:

[0163] Perform fast Fourier transform on the time domain signal data to be detected, and output frequency domain signal data;

[0164] The path composition method is used to construct non-Euclidean data of frequency domain signal data and generate non-Euclidean spatial data.

[0165] Furthermore, the improved fusion graph neural network includes an improved graph convolution module, a multi-head graph attention module, a fully connected layer and a random inactivation layer; the feature extraction module 503 includes:

[0166] The first submodule is used to iteratively update the global features of the non-Euclidean space data through an improved graph convolution module to generate a plurality of first node features and a plurality of second node features;

[0167] The second submodule is used to use a multi-head graph attention module to perform node fusion on each first node feature and each second node feature, and output node fusion normalized features;

[0168] The third submodule is used to input the node fusion normalized features into the fully connected layer for linear transformation to generate transformed node features;

[0169] The fourth submodule is used to regularize the transformed node features through a random dropout layer and output the regularized node features;

[0170] The fifth submodule is used to use the improved graph convolution module to iteratively update the global features of the regularized node features to generate multiple probability vectors.

[0171] Furthermore, the improved graph convolution module includes an improved graph convolution layer and a normalization layer; the first submodule includes:

[0172] The first unit is used to perform polynomial iterative fusion on non-Euclidean spatial data through an improved graph convolution layer based on a preset symmetry processing mechanism to generate a linear combination of node features;

[0173] The second unit is used to normalize the linear combination of node features by using a normalization layer, and output a plurality of first node features and a plurality of second node features.

[0174] Furthermore, the multi-head graph attention module includes a multi-head graph attention submodule and a normalization layer; the second submodule includes:

[0175] The third unit is used to perform feature enhancement fusion on each first node feature and each second node feature through a multi-head graph attention submodule, and output a node fusion transformation feature;

[0176] The fourth unit is used to normalize the node fusion transformation features by using a normalization layer to generate node fusion normalized features.

[0177] The multi-head graph attention submodule includes a concatenation layer, a first activation function layer, a second activation function layer, and a third activation function layer; the third unit includes:

[0178] A first subunit is used to perform a multiplication operation on each first node feature and a preset shared parameter to determine a first node parameter feature corresponding to each first node feature;

[0179] A second subunit is used to perform a multiplication operation on each second node feature and a preset shared parameter to determine a second node parameter feature corresponding to each second node feature;

[0180] A third subunit is used to splice each first node parameter feature and each second node parameter feature through a splicing layer to generate a plurality of node splicing features;

[0181] A fourth subunit is used to input the splicing features of each node into the first activation function layer for nonlinear mapping, and output the first node mapping features corresponding to the splicing features of each node;

[0182] A fifth subunit is used to input each first node mapping feature into the second activation function layer for normalization, and output a plurality of attention coefficients;

[0183] The sixth subunit is used to perform weighted fusion of each attention coefficient and each second node parameter feature based on a multi-head graph attention mechanism to generate a node fusion domain feature;

[0184] The seventh subunit is used to perform nonlinear transformation on the node fusion domain features through the third activation function layer, and output the node fusion transformation features.

[0185] In an optional embodiment, the device further comprises:

[0186] The first module is used to obtain the time domain signal data to be trained, and pre-process the time domain signal data to be trained to generate the non-Euclidean space data to be trained;

[0187] The second module is used to determine the model gradient based on the non-Euclidean space data to be trained by using the initial improved fusion graph neural network based on the back propagation algorithm;

[0188] The third module is used to update the model parameters of the initial improved fusion graph neural network through the model gradient based on the preset stochastic gradient descent optimizer, and determine the intermediate improved fusion graph neural network;

[0189] The fourth module is used to calculate the target loss value according to the non-Euclidean space data to be trained by using the intermediate improved fusion graph neural network and the preset loss function, and to determine whether the target loss value converges;

[0190] The fifth module is used to use the intermediate improved fusion graph neural network as the trained improved fusion graph neural network if the target loss value converges.

[0191] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices, modules, sub-modules, units and sub-units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0192] An embodiment of the present invention further provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the rotating component fault diagnosis method of any of the above embodiments.

[0193] An embodiment of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the rotating component fault diagnosis method as in any of the above embodiments are implemented.

[0194] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0195] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0196] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing faults of rotating parts, characterized in that: include: Acquiring time domain signal data to be detected of the rotating component; Preprocessing the time domain signal data to be detected to generate non-Euclidean space data; Using an improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean spatial data, and output multiple probability vectors; A preset maximum value function is used to generate a fault diagnosis result of the rotating component according to each of the probability vectors.

2. The rotating component fault diagnosis method according to claim 1, characterized in that: The step of preprocessing the time domain signal data to be detected to generate non-Euclidean space data includes: Performing a fast Fourier transform on the time-domain signal data to be detected, and outputting frequency-domain signal data; A path composition method is used to perform non-Euclidean data construction on the frequency domain signal data to generate non-Euclidean space data.

3. The rotating component fault diagnosis method according to claim 1, characterized in that: The improved fusion graph neural network includes an improved graph convolution module, a multi-head graph attention module, a fully connected layer and a random inactivation layer; the step of using the improved fusion graph neural network to perform global feature extraction and local feature extraction on the non-Euclidean space data and output multiple probability vectors includes: Iteratively updating the global features of the non-Euclidean space data through an improved graph convolution module to generate a plurality of first node features and a plurality of second node features; Using a multi-head graph attention module to perform node fusion on each of the first node features and each of the second node features, and output node fusion normalized features; Inputting the node fusion normalized features into the fully connected layer for linear transformation to generate transformed node features; Regularizing the transformed node features through a random dropout layer, and outputting regularized node features; An improved graph convolution module is used to iteratively update the regularized node features globally to generate multiple probability vectors.

4. The rotating component fault diagnosis method according to claim 3, characterized in that: The improved graph convolution module includes an improved graph convolution layer and a normalization layer; the step of iteratively updating the global features of the non-Euclidean space data through the improved graph convolution module to generate a plurality of first node features and a plurality of second node features includes: Through an improved graph convolution layer based on a preset symmetry processing mechanism, the non-Euclidean spatial data is subjected to polynomial iterative fusion to generate a linear combination of node features; A normalization layer is used to normalize the linear combination of the node features, and a plurality of first node features and a plurality of second node features are output.

5. The rotating component fault diagnosis method according to claim 3, characterized in that: The multi-head graph attention module includes a multi-head graph attention submodule and a normalization layer; the step of using the multi-head graph attention module to perform node fusion on each of the first node features and each of the second node features, and outputting the node fusion normalized features, includes: Perform feature enhancement fusion on each of the first node features and each of the second node features through a multi-head graph attention submodule, and output node fusion transformation features; A normalization layer is used to normalize the node fusion transformation features to generate node fusion normalized features.

6. The rotating component fault diagnosis method according to claim 5, characterized in that: The multi-head graph attention submodule includes a splicing layer, a first activation function layer, a second activation function layer, and a third activation function layer; the step of performing feature enhancement fusion on each of the first node features and each of the second node features through the multi-head graph attention submodule and outputting node fusion transformation features includes: Performing a multiplication operation on each of the first node features and a preset shared parameter to determine a first node parameter feature corresponding to each of the first node features; Performing a multiplication operation on each of the second node features and the preset shared parameter to determine a second node parameter feature corresponding to each of the second node features; Splicing each of the first node parameter features and each of the second node parameter features through a splicing layer to generate a plurality of node splicing features; Input each of the node splicing features into a first activation function layer for nonlinear mapping, and output a first node mapping feature corresponding to each of the node splicing features; Inputting each of the first node mapping features into a second activation function layer for normalization, and outputting a plurality of attention coefficients; Based on a multi-head graph attention mechanism, weighted fusion is performed on each of the attention coefficients and each of the second node parameter features to generate a node fusion domain feature; The node fusion domain features are nonlinearly transformed through the third activation function layer, and the node fusion transformation features are output.

7. The rotating component fault diagnosis method according to claim 1, characterized in that: Before the step of obtaining the time domain signal data to be detected of the rotating component, the method includes: Acquire time domain signal data to be trained, and preprocess the time domain signal data to be trained to generate non-Euclidean space data to be trained; Based on the back propagation algorithm, an initial improved fusion graph neural network is used to determine the model gradient according to the non-Euclidean space data to be trained; Based on a preset stochastic gradient descent optimizer, the model parameters of the initial improved fusion graph neural network are updated through the model gradient to determine an intermediate improved fusion graph neural network; Using the intermediate improved fusion graph neural network and the preset loss function to calculate the target loss value according to the non-Euclidean space data to be trained, and judging whether the target loss value converges; If the target loss value converges, the intermediate improved fusion graph neural network is used as the trained improved fusion graph neural network.

8. A rotating component fault diagnosis device, characterized in that: include: An acquisition module, used for acquiring time domain signal data to be detected of the rotating component; A preprocessing module, used for preprocessing the time domain signal data to be detected to generate non-Euclidean space data; A feature extraction module, used for performing global feature extraction and local feature extraction on the non-Euclidean spatial data by using an improved fusion graph neural network, and outputting a plurality of probability vectors; The result generating module is used to generate the fault diagnosis result of the rotating component according to each of the probability vectors by using a preset maximum value function.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the rotating component fault diagnosis method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the rotating component fault diagnosis method as described in any one of claims 1-7 are implemented.

Images