Planetary gearbox fault diagnosis method based on semi-supervised graph convolution

Abstract

The present invention discloses a planetary gearbox fault diagnosis method based on semi-supervised graph convolution, and the present invention relates to the field of rotating machinery fault diagnosis methods. The present invention aims to solve the problem of low accuracy of existing rotating machinery fault diagnosis methods. The present invention uses an autoencoder denoising network to perform denoising on the original data; performs feature extraction on the denoised data to obtain a feature matrix, and performs dimensionality reduction on the feature matrix; obtains an undirected graph between each data in the feature matrix after dimensionality reduction, and inputs the matrix after dimensionality reduction and the undirected graph between the data into a semi-supervised graph convolutional neural network fault diagnosis model to obtain a diagnosis result. The present invention can accurately diagnose planetary gearbox faults.

Description

Technical Field

[0001] The present invention relates to the field of rotating machinery fault diagnosis methods, and in particular to a planetary gearbox fault diagnosis method based on semi-supervised graph convolution. Background Art

[0002] Planetary gearboxes have compact structures and large transmission ratios and are used in mechanical equipment in many industries. Especially in aviation, aerospace and some high-risk industries, the health of planetary gearboxes is closely related to people's lives and property, and these fields have a series of problems such as complex working conditions, long working hours and harsh working environments, including different types of failures of gears, bearings and shafts under various loads and wear. The vibration response of the planetary gearbox system has its own motion characteristics and kinematic characteristics. Its vibration response mainly includes its own motion frequency and its frequency multiples, including the amplitude modulation and frequency modulation components of other internal parts, accompanied by noise in the environment, and interference noise caused by the vibration of other equipment in the same plant. The simultaneous existence of these factors causes subsequent fault diagnosis, interferes with the accuracy of feature extraction, and cannot accurately see the characteristic frequency in spectrum analysis. At present, the noise reduction algorithms in the field of rotating machinery fault diagnosis can be roughly divided into three categories: wavelet noise reduction, adaptive mode decomposition, atomic decomposition and sparse representation. As a kind of sparse representation signal, wavelet transform can concentrate most of the signal energy in a few sparse in the wavelet domain. However, wavelet analysis denoising methods need to select wavelet basis functions according to some characteristics of the signal itself, and also need to set corresponding threshold functions. The quality of its denoising effect depends to a certain extent on the influence of human prior knowledge, and often has poor adaptability. Different types of signals are often poorly generalized when brought into the same wavelet denoising algorithm. Adaptive modal extraction denoising lacks a mathematical theoretical basis, and it is a recursive way to extract modalities, which cannot correct the errors in the extracted modalities, resulting in modal aliasing. The atomic decomposition domain sparse representation denoising algorithm still has the problems of weak signal expression ability and poor adaptability.

[0003] The characteristic of intelligent diagnosis is that the designed diagnostic model is trained based on a large amount of data, and finally a mathematical model that can be used to judge specific faults can be obtained, and its diagnostic accuracy can generally reach the required range of work. This type of data-driven model does not require a lot of prior knowledge like traditional mechanism models. Intelligent diagnosis algorithms are divided into supervised learning, unsupervised learning and semi-supervised learning. Common supervised learning algorithms include support vector machine-based diagnosis algorithms and neural network-based diagnosis algorithms. Although supervised learning algorithms have the ability to feedback and optimize models during the learning process, they often require a lot of manual measurement costs in obtaining labels. Unsupervised learning algorithms mainly perform fault diagnosis by mining the internal structure of unlabeled data and clustering each type of fault. However, unsupervised classes are often classified in the form of clustering, and the learning model does not have feedback, so its diagnostic accuracy is often not ideal; semi-supervised learning fault diagnosis algorithms use both labeled and unlabeled data to improve the diagnostic accuracy of the classification model, in which a small number of labels are used to adjust model parameters, and a large amount of unlabeled data is used to reduce the manual cost of labeling. However, most of the existing fault diagnosis methods take data points as isolated data as input, ignoring the connection between data. However, the data contains a lot of information which is of great significance for diagnosis. Summary of the invention

[0004] In view of the above-mentioned defects in the prior art, the technical problem to be solved by the present invention is to provide a planetary gearbox fault diagnosis method with higher accuracy.

[0005] To achieve the above object, the present invention provides a planetary gearbox fault diagnosis method based on semi-supervised graph convolution, comprising the following steps:

[0006] S1. Acquire raw data of a planetary gearbox, and use an autoencoder denoising network to perform denoising on the raw data;

[0007] S2, extracting features from the denoised data to obtain a feature matrix, and performing dimensionality reduction processing on the feature matrix;

[0008] S3. Obtain an undirected graph between each data in the feature matrix after the dimensionality reduction processing, and input the undirected graph between the matrix after dimensionality reduction and the data into a semi-supervised graph convolutional neural network fault diagnosis model to obtain a diagnosis result.

[0009] Furthermore, the method for performing noise reduction processing on the original data includes:

[0010] S11. Establish an autoencoder denoising network and initialize the dictionaries of each layer of the autoencoder denoising network and allowable error , where i=1, 2, …, v, represents the number of layers of the autoencoder denoising network;

[0011] S12. Use the orthogonal matching pursuit algorithm to calculate the sparse coefficients of each layer , i=1,2,…,v;

[0012] S13. Calculate reconstruction error , when the reconstruction error Is it less than the allowable error? If yes, execute step S15; if no, execute step S14. The reconstruction error for:

[0013] ;

[0014] in, is the original data, is the data after noise reduction;

[0015] S14. Update the dictionary of each layer by K-SVD method , and determine whether the update deadline condition is met, if so, execute step S15, if not, execute step S12;

[0016] S15, the current dictionaries of each layer The v-th layer sparse coefficient , get the noise-reduced data , .

[0017] Furthermore, the original data Split into several matrix blocks, forming the following matrix ,matrix The matrix blocks in the denoising process are processed in batches to obtain the corresponding processed data blocks, and then the denoised signal is obtained. :

[0018] ; .

[0019] Furthermore, the number of atoms in each layer of the dictionary in the autoencoder denoising network first increases and then decreases in a layer-by-layer order.

[0020] Further, step S12 includes:

[0021] S121, initializing the first layer dictionary D1, and initializing the first layer sparse coefficient X1 using the orthogonal matching pursuit algorithm;

[0022] S122, using the first layer sparse coefficient X1 as the input signal of the second layer, and obtaining the second layer sparse coefficient X2 by using the orthogonal matching pursuit algorithm with the second layer dictionary D2;

[0023] S123, repeat step S122 until the v-1th layer sparse coefficient X v-1 As the input signal of the vth layer, and the vth layer dictionary D v Use the orthogonal matching pursuit algorithm to get the v-th layer sparse coefficient X v .

[0024] Furthermore, the dictionary of each layer is updated by the K-SVD method. The methods include:

[0025] S141, updating the first layer dictionary D1 according to the original data, the dictionary D1 and the sparse coefficient X1 according to the K-SVD method;

[0026] S142, continue to use the sparse coefficient X1 as the input signal, and use the K-SVD method together with the dictionary D2 to update the second-layer dictionary D2;

[0027] S143, repeat step S142 until the sparse coefficient X v-1 As the input signal, the dictionary D5 and the sparse coefficient X5 use the K-SVD method to update the v-th layer dictionary D v .

[0028] Further, step S2 includes:

[0029] S21, extracting the memory features of the data after noise reduction, obtaining time domain features and frequency domain features, and forming a feature matrix;

[0030] S22. Determine the edge weight value W using the Gaussian kernel function ij , and then get the adjacency graph W;

[0031] S23, calculating the Laplace matrix L=DW according to the adjacency graph W, and setting the dimension of the reduced matrix to e;

[0032] S24, select the eigenvectors corresponding to the first e smallest non-zero eigenvalues, and arrange them in ascending order according to the eigenvalues ​​to obtain the feature matrix V after feature dimension reduction = (v (1) v (2) … v (e) ).

[0033] Furthermore, the method for establishing the semi-supervised graph convolutional neural network fault diagnosis model includes:

[0034] S31. Build a semi-supervised graph convolutional neural network fault diagnosis model;

[0035] S32, input training data, the training data is the feature matrix after reduction, the training data is divided into test set V train , validation set V val and the test set Vtest , where the test set V train Including labeled data and unlabeled data, validation set V val and the test set V test All are labeled data, and the adjacency matrix A is established train , A val and A test ;

[0036] S33, initializing the hyperparameters of the semi-supervised graph convolutional neural network fault diagnosis model and the convolution kernel parameters of each layer of the network;

[0037] S34, V train and the adjacency matrix A train Perform forward propagation in and extract the predicted category ;

[0038] S35. According to the predicted category and the actual category Y of the labeled data Li Calculate the loss function L;

[0039] S36, back-propagating the loss function, updating the convolution kernel parameters of each layer, and obtaining a semi-supervised graph convolutional neural network fault diagnosis model;

[0040] S37, the verification set V val and the adjacency matrix A val Input the semi-supervised graph convolutional neural network fault diagnosis model obtained in step S36 for verification, and output the verification set loss value and the verification set accuracy;

[0041] S38, determine whether the number of iterations is less than the maximum number of iterations n, if so, repeat steps S34-S37, if not, end the training and output the final semi-supervised graph convolutional neural network fault diagnosis model;

[0042] S39, test set V test and the adjacency matrix A test Input the semi-supervised graph convolutional neural network fault diagnosis model obtained by S38, and output the recognition accuracy and loss value of the planetary gearbox fault diagnosis test set.

[0043] Furthermore, the semi-supervised graph convolutional neural network fault diagnosis model includes a forward propagation module and a back propagation module, the forward propagation module includes a graph convolutional filter layer and a nonlinear function layer Relu, and the back propagation module includes a graph convolutional filter layer and a softmax classification layer.

[0044] Furthermore, the labeled data accounts for 10% of the total data set.

[0045] Compared with the prior art, the present invention has the following technical effects:

[0046] 1. This application uses a K-SVD autoencoder for signal denoising. The adaptive denoising ability of sparse representation signal denoising algorithms is limited, and the effect of deep learning denoising algorithms is easily affected by parameter initialization. This method uses K-SVD as a computational solver for each layer of denoising network. To improve the signal denoising ability of this method, the number of atoms in each layer is set according to the autoencoder rules, and a K-SVD autoencoder denoising network is built. Compared with other sparse representation denoising algorithms, this method has a higher sensitivity to denoised signals and is better than traditional sparse representation denoising algorithms in denoising effect.

[0047] 2. This application adopts a planetary gearbox fault diagnosis method based on semi-supervised graph convolution. In view of the fact that the current fault diagnosis method requires a large amount of supervised learning with faulty labeled data and consumes a lot of manpower and material costs in acquiring fault labels, this method constrains the relationship between data by using Euclidean distance to establish an undirected graph between features after dimensionality reduction, sets the solution mode of each layer of the network to graph convolution, and uses 10% faulty labels and 90% non-faulty labels to jointly train the network. The accuracy of the diagnostic effect test set is higher than that of the comparative fault diagnosis method, and the fault diagnosis accuracy is higher. The concept, specific structure and technical effects of the present invention will be further explained in conjunction with the accompanying drawings below to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flow chart of a planetary gear fault diagnosis method according to a specific embodiment of the present invention;

[0049] Figure 2 A denoising flow chart of an autoencoder denoising network according to a specific embodiment of the present invention;

[0050] Figure 3 This is a flow chart of a dimensionality reduction method according to a specific embodiment of the present invention;

[0051] Figure 4 A structural principle diagram of a semi-supervised graph convolution fault diagnosis model according to a specific embodiment of the present invention;

[0052] Figure 5 A flowchart of a method for constructing a semi-supervised graph convolution fault diagnosis model according to a specific embodiment of the present invention;

[0053] Figure 6 This is a frequency domain diagram of a sun gear fault before denoising using the OMP method, K-SVD and K-SVD autoencoder denoising network in a specific embodiment of the present invention. Figure 6 a is the frequency domain diagram of the original signal sun gear fault, Figure 6 b is the frequency domain diagram of the OMP noise reduction sun gear fault. Figure 6 c is the frequency domain diagram of sun gear fault with K-SVD noise reduction, Figure 6 d is the frequency domain diagram of the sun gear fault denoised by the K-SVD autoencoder denoising network;

[0054] Figure 7 This is a time-frequency domain diagram of the outer gear ring fault before and after denoising using the OMP method, K-SVD and K-SVD autoencoder denoising network in a specific embodiment of the present invention. Figure 7 a is the frequency domain diagram of the outer ring fault of the original signal, Figure 7 b is the frequency domain diagram of the OMP noise reduction outer gear ring fault. Figure 7 c is the frequency domain diagram of the outer gear ring fault with K-SVD noise reduction. Figure 7 c is the frequency domain diagram of the outer ring fault denoised by the K-SVD autoencoder denoising network, Figure 7 d is the frequency domain diagram of the outer gear ring fault after denoising by the K-SVD autoencoder denoising network;

[0055] Figure 8 This is a time-frequency domain diagram of a normal signal before and after denoising using the OMP method, K-SVD and K-SVD type autoencoder denoising network in a specific embodiment of the present invention. Figure 8 a is the original normal signal frequency domain diagram, Figure 8 b is the frequency domain diagram of the normal signal after OMP denoising. Figure 8 c is the frequency domain diagram of the normal signal denoised by K-SVD, Figure 8 d is the frequency domain diagram of the normal signal after the K-SVD autoencoder denoising network denoises the normal signal;

[0056] Fig. 9 This is a result diagram of a specific embodiment of the present invention using the PCA dimensionality reduction method, the KPCA dimensionality reduction method and the LE dimensionality reduction method of the present application. Fig. 9 a is the effect diagram of PCA dimensionality reduction method, Fig. 9 b is the effect diagram of KPCA dimensionality reduction method. Fig. 9 c is the effect diagram of LE dimensionality reduction method;

[0057] Fig.10 This is the accuracy change curve of the BP neural network, CNN and semi-supervised graph convolution fault diagnosis model used in a specific embodiment of the present invention. Fig.10 a is the accuracy change curve of BP neural network, Fig.10 b is the CNN accuracy change curve, Fig.10 c is the accuracy change curve of the semi-supervised graph convolution fault diagnosis model;

[0058] Fig.11 The confusion matrix of the BP neural network, CNN and semi-supervised graph convolution fault diagnosis model used in a specific embodiment of the present invention is shown in FIG. Fig.11 a is the confusion matrix of BP neural network, Fig.11 b is the confusion matrix of CNN, Fig.11c is the confusion matrix of the semi-supervised graph convolutional fault diagnosis model. DETAILED DESCRIPTION

[0059] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0060] It should be noted that the illustrations provided in the following embodiments are only used to illustrate the basic concept of the present invention in a schematic manner, and thus the illustrations only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0061] While some exemplary embodiments of the present invention have been described for the purpose of illustration, it should be understood that the present invention may be implemented in other ways not specifically shown in the drawings.

[0062] like Figure 1 As shown, the planetary gearbox fault diagnosis method based on semi-supervised graph convolution includes the following steps:

[0063] S1. Obtain the original data of the planetary gearbox, and use the autoencoder denoising network to perform denoising on the original data. During the fault diagnosis process, since the collected signals are often affected by interference between devices and environmental noise, the real measured signals are often masked by noise, affecting the accuracy and reliability of fault diagnosis. Therefore, it is necessary to perform denoising on the original vibration signal. This embodiment uses the K-SVD autoencoder denoising network for denoising. When the number of sampling points of the original signal is large, the number of atoms of the K-SVD autoencoder increases accordingly, which will reduce the efficiency of the algorithm. This embodiment processes the original signal set in batches. The original signal , the total amount of data is N, the original data set Split into several matrix blocks to form the following matrix:

[0064] .

[0065] By using the above method to perform denoising on the original planetary gear fault data set in batches, parallel computing can be achieved on the computer. For the planetary gearbox fault data with a large number of sampling points and a large sample size, the computing speed of the K-SVD denoising autoencoder can be significantly improved. In this embodiment, each group of sampling points Q, each batch of planetary gearbox vibration signal set processing , q is the number of signals processed in each batch.

[0066] This embodiment uses a K-SVD autoencoder denoising network to perform denoising processing. Figure 2 As shown, specifically including:

[0067] S11. Establish a K-SVD autoencoder denoising network, set parameters and initialize the dictionaries of each layer of the denoising network and allowable error , where i=1, 2, …, v, represents the number of layers of the autoencoder denoising network;

[0068] The number of iterations is set to M, and the sparse representation result tolerance is , set the K-SVD autoencoder denoising network to include 5 layers. The dictionary of each layer (i=1, 2, …, 5), where , , , , , where the right side of the equation represents a two-dimensional matrix, is the number of sampling points of dictionary atomic signal, is the number of sparse atoms. In this embodiment, the number of atoms in each layer of the dictionary increases first and then decreases in a layer-by-layer order.

[0069] In this embodiment, the network dictionary of each layer of the K-SVD autoencoder denoising network The Fourier dictionary is used as the initialization method. The construction method of the Fourier dictionary is:

[0070] ;

[0071] Where O is the discrete value sequence number in the time domain.

[0072] The dictionary is initialized as:

[0073] ;

[0074] in, .

[0075] S12. Use the orthogonal matching pursuit algorithm to calculate the sparse coefficients of each layer , i=1, 2, ..., 5, the principle of the orthogonal matching pursuit algorithm is:

[0076] Step 1. Initialize sparse coefficients , let the residual , the support of the initial solution , ;

[0077] Step 2, select variables. , find the residual Atom with maximum correlation:

[0078]

[0079] Among them, atoms A column that is a dictionary.

[0080] Step 3, update the variables. , , the residual is:

[0081] ;

[0082] Step 4. If the current residual is unrelated to all atoms, or , then return ; Otherwise, jump to stpe2.

[0083] In this embodiment, according to the initialization dictionary D i Use the orthogonal matching pursuit algorithm (OMP) for forward propagation to solve the sparse coefficient X i , and its forward propagation principle is: , , …, , including:

[0084] S121, initializing the first layer dictionary D1, and initializing the first layer sparse coefficient X1 using the orthogonal matching pursuit algorithm;

[0085] S122, using the first layer sparse coefficient X1 as the input signal of the second layer, and obtaining the second layer sparse coefficient X2 by using the orthogonal matching pursuit algorithm with the second layer dictionary D2;

[0086] S123, repeat step S122 until the 4th layer sparse coefficient X4 is used as the input signal of the 5th layer, and the 5th layer sparse coefficient X5 is obtained by using the orthogonal matching pursuit algorithm with the 5th layer dictionary D5.

[0087] S13, based on the updated dictionaries of each layer (i=1, 2, ..., 5) and sparse coefficient X5 to calculate the reconstruction error , to determine whether the residual error reaches the termination condition, when the reconstruction error Is it less than the allowable error? If yes, execute step S15; if no, execute step S14. The reconstruction error for:

[0088] ;

[0089] in, is the original data, is the data after noise reduction;

[0090] S14. Update the dictionary of each layer by K-SVD method , and determine whether the update deadline condition is met, if so, execute step S15, if not, execute step S12;

[0091] The principle process of the K-SVD dictionary update includes:

[0092] Using OMP in the previous section to obtain the sparse coefficient set X of the sample signal set Y, K-SVD only updates one atom in the dictionary D in each iteration. Suppose we update the wth (i=1,2,…,w)th atom d w , as follows:

[0093] ;

[0094] In the formula, is the jth row of the sparse coefficient X, let is the reconstruction error of all atoms except the w-th atom, we can get:

[0095] ;

[0096] make Indicates the use of atom d w The index of the sample signal Y, for The matrix of is a non-zero value, and we get:

[0097] ;

[0098] In the formula, represent Remove the row vectors with zero values, Indicates the use of atoms in sparse coefficients The error column for Perform singular value decomposition to obtain , let U update the original dictionary , Update x j ; Until all atoms W in the dictionary are updated and the updated dictionary D is output, the iteration ends.

[0099] This embodiment uses K-SVD to perform back propagation to update the dictionaries of each layer. The back propagation principle is:

[0100] , , …, , the updating process of each layer dictionary in this embodiment includes:

[0101] S141, according to the original data Y A , dictionary D1 and sparse coefficient X1, update the first layer dictionary D1 according to the K-SVD method;

[0102] S142, continue to use the sparse coefficient X1 as the input signal, and use the K-SVD method together with the dictionary D2 to update the second-layer dictionary D2;

[0103] S143, repeat step S142 until the sparse coefficient X v-1 As the input signal, the dictionary D5 and the sparse coefficient X5 use the K-SVD method to update the v-th layer dictionary D v .

[0104] S15, the number of iterations t=t+1, when t>M, the iteration is terminated and the process goes to S16; otherwise, step S12 is repeated;

[0105] S16: Output the final dictionary of each layer obtained by updating the above iterative process Layer 5 sparse coefficients , get the noise-reduced data , .

[0106] S17, repeat steps S12-S16 to complete the original data set All batches of data in the denoised data set are obtained to obtain the denoised data blocks corresponding to each batch of original data, and then the denoised signal set is obtained. for:

[0107] .

[0108] S2, extracting features from the denoised data to obtain a feature matrix, performing dimensionality reduction processing on the feature matrix, and arranging the eigenvalues ​​after dimensionality reduction into a feature matrix to achieve the effect of dimensionality reduction processing; Figure 3 As shown, specifically including:

[0109] S21, extracting the memory features of the data after noise reduction, obtaining time domain features and frequency domain features, and forming a feature matrix;

[0110] This embodiment uses time domain indicators and frequency domain indicators to decompose the noise-reduced vibration signal, and then obtains time domain characteristic signals and frequency domain characteristic signals. In this embodiment, the time domain indicators are used to obtain p1-p 11 There are 11 time domain characteristic signals in total, which are decomposed into p by frequency domain characteristic index. 12 -p 24 There are 12 frequency domain feature signals in total.

[0111] The time domain characteristics are:

[0112]

[0113] The frequency domain characteristic index is:

[0114]

[0115] In this embodiment, the signal set after noise reduction is subjected to time domain and frequency domain feature extraction to obtain a feature matrix P with a total of m data, and n features are extracted from each data to form a The characteristic matrix P is:

[0116] ;

[0117] The number of features of the data after feature dimensionality reduction is e (e < n), forming a The mapping matrix.

[0118] ;

[0119] This embodiment uses Laplacian Eigenmaps (LE) to perform dimensionality reduction processing, which is specifically achieved through the following steps.

[0120] S22. Determine the edge weight value W using the Gaussian kernel function ij , and then get the adjacency graph W, the edge weight value W ij for:

[0121] ;

[0122] Where i, j = 1, 2, …, m;

[0123] S23, calculating the Laplace matrix L=DW according to the adjacency graph W, and setting the dimension of the reduced matrix to e;

[0124] In order to find the mapping from high-dimensional space to low-dimensional space , so that adjacent points are as close as possible, and the optimization objective function is established as:

[0125] ;

[0126] The optimization objective function is equivalent to:

[0127] ;

[0128] In the formula, the Lablax matrix L calculated according to the adjacency graph W is: L=DW, D is a diagonal matrix, and the elements are the sum of the columns of the adjacency graph W.

[0129] In order to eliminate the influence of the scale factor, the constraint is imposed , and then transform the above optimization objective function into:

[0130] ;

[0131] Using the Laarrange multiplier method to solve the generalized eigenvalue problem, we get:

[0132] ;

[0133] ;

[0134] Therefore, the optimization objective function is converted into a quadratic optimization function:

[0135] .

[0136] S24. Learn the low-dimensional embedding V through the LE algorithm as the eigenvector corresponding to the first e smallest non-zero eigenvalues ​​in the generalized eigendecomposition of (L, D). Arrange the eigenvalues ​​from small to large to obtain the feature matrix V after feature dimensionality reduction = (v (1) v (2) …v (e) ).

[0137] S3. Obtain an undirected graph between each data in the feature matrix after the dimensionality reduction processing, and input the undirected graph between the matrix after dimensionality reduction and the data into a semi-supervised graph convolutional neural network fault diagnosis model to obtain a diagnosis result.

[0138] The traditional intelligent fault diagnosis algorithm is a supervised learning method, which has a high cost in obtaining labels. In this embodiment, a semi-supervised graph convolutional neural network (Semi-GCN) fault diagnosis model is used to diagnose planetary gearbox faults.

[0139] like Figure 5 As shown, the method for establishing the semi-supervised graph convolutional neural network fault diagnosis model in this embodiment includes:

[0140] S31. Build a semi-supervised graph convolutional neural network fault diagnosis model;

[0141] In this embodiment, for the dataset V after LE dimension reduction, in order to express the relationship between its data, the concept of graph is introduced to represent the fault data itself and reflect the relationship between them. The graph is defined as a set G = (V, E, A), where V represents the set of planetary gear fault data V = (v1, v2, ..., v n ), also called node set V; E represents the link node v i (i=1,2,…,n) and node v j The set of edges between (j=1,2,…,n) is E=(e 11 , e 12 , e ij ,…,e nn ); A represents the weight of the edge connecting any two nodes in the node set V, expressed in the form of an adjacency matrix. ij The Euclidean distance is calculated as shown below. According to the actual working conditions of fault diagnosis, the data point v i With v j The presence or absence of direction has no effect on the event itself, so the graph is defined as an undirected graph.

[0142] ;

[0143] like Figure 4 As shown, the Semi-GCN fault diagnosis model of this embodiment includes two parts, namely a forward propagation module and a back propagation module. The forward propagation module includes a graph convolution filter layer and a nonlinear function layer Relu, and the back propagation module includes a graph convolution filter layer and a softmax classification layer. The forward propagation outputs a predicted value, and the network model parameters are updated by back propagation according to the predicted value.

[0144] The forward propagation function of the fault diagnosis model is:

[0145] ;

[0146] in, is the training set divided by the feature set V after LE dimensionality reduction processing, express The constructed adjacency matrix represents the training set An undirected graph of

[0147] Test Set Contains labeled fault data With unlabeled fault data ,Will Divide each training batch ,Will Substitute it into the forward propagation function and find it Prediction Type , Contains the predicted values ​​of labeled fault data Prediction value of unlabeled fault data However, due to the semi-supervised nature of the algorithm, the participation loss function L only uses the actual value of the labeled data With the predicted value Calculate, the loss function of the fault diagnosis model is:

[0148] ;

[0149] The model parameter updating method adopted in this embodiment is the stochastic gradient descent method. Since this updating method is an existing method, it will not be described in detail here.

[0150] S32, input training data, the training data is the feature matrix after reduction, the training data is divided into test set V train , validation set V val and the test set V test , where the test set V train Including labeled data and unlabeled data, validation set V val and the test set V test All are labeled data, and the adjacency matrix A is established train , A val and A test In this embodiment, the labeled data accounts for 10% of the total data set, and only 10% of the labeled data is used in the network feedback calculation, while most of the unlabeled data only participates in the forward propagation of the prediction of the fault category, thereby realizing the supervised calculation of a small amount of labeled data and realizing semi-supervised learning.

[0151] S33, initializing the hyperparameters of the semi-supervised graph convolutional neural network fault diagnosis model and the convolution kernel parameters of each layer of the network;

[0152] S34, V train and the adjacency matrix A train Perform forward propagation in and extract the predicted category ;

[0153] S35. According to the predicted category and the actual categories of the labeled data Calculate the loss function L;

[0154] S36, back-propagating the loss function, updating the convolution kernel parameters of each layer, and obtaining a semi-supervised graph convolutional neural network fault diagnosis model;

[0155] S37, the verification set V val and the adjacency matrix A valInput the semi-supervised graph convolutional neural network fault diagnosis model obtained in step S36 for verification, and output the verification set loss value and the verification set accuracy;

[0156] S38, determine whether the number of iterations is less than the maximum number of iterations n, if so, repeat steps S34-S37, if not, end the training and output the final semi-supervised graph convolutional neural network fault diagnosis model;

[0157] S39, test set V test and the adjacency matrix A test Input the semi-supervised graph convolutional neural network fault diagnosis model obtained by S38, and output the recognition accuracy and loss value of the planetary gearbox fault diagnosis test set.

[0158] In order to further verify the effect of the present invention, in a specific embodiment, the simulation signal is denoised by the OMP algorithm denoising in the prior art, the K-SVD algorithm and the K-SVD autoencoder denoising network of the present invention, and the relevant parameter settings of the three denoising methods are shown in Table 1, Table 2 and Table 3. The simulation signal includes a local fault signal of the outer ring gear of the planetary gear system, a local fault signal of the sun gear, and a normal planetary gear signal. The signal set , the number of sampling points for each signal is 5000, the sampling frequency is 2500, 2000 groups of each type of signal are taken, and the number of noise reduction signals in each batch is .

[0159] Table 1: OMP algorithm parameters

[0160]

[0161] Table 2: K-SVD algorithm parameters

[0162]

[0163] Table 3: K-SVD autoencoder algorithm parameters

[0164]

[0165] The time-frequency domain images after denoising are obtained by three denoising methods. Figure 6 , Figure 7 and Figure 8 As shown, Figure 6 As shown in the figure, the noise reduction results of the local fault signal of the sun gear are obtained by three different methods. The peak of the modulation frequency appears in the local fault of the nearby planetary gear and sun gear, indicating that the three methods can better preserve the characteristic information of the local fault of the sun gear after denoising. However, the frequency amplitude is 0.6 before denoising, 0.42 after OMP denoising, 0.54 after K-SVD denoising, and 0.04 after K-SVD autoencoder denoising, indicating that the K-SVD autoencoder of this embodiment has a stronger ability in suppressing noise energy, proving the effectiveness of denoising by this method.

[0166] Figure 7 The noise reduction results of the three methods for the outer gear ring fault signal are shown in the figure. The modulation characteristic frequency of the local fault of the outer ring of the planetary gear system has a peak at the frequency. Therefore, the three methods can better preserve the characteristic information of the local fault of the outer ring after denoising; at the same time, the frequency amplitude is 0.65 before denoising, the frequency amplitude is 0.54 after OMP denoising, the frequency amplitude is 0.64 after K-SVD denoising, and the frequency amplitude is 0.35 after K-SVD autoencoder denoising, which proves the effectiveness of K-SVD autoencoder denoising.

[0167] Figure 8 The following are the results of three methods of noise reduction for normal signals. As can be seen from the figure, The meshing characteristic frequency of the normal signal of the planetary gear appears at a peak near the frequency, indicating that the three denoising methods can better obtain the characteristic information of the normal signal of the planetary gear after denoising; at the same time, the frequency amplitude is 0.87 before denoising, 0.83 after OMP denoising, 0.87 after K-SVD denoising, and 0.53 after K-SVD autoencoder denoising, which proves the effectiveness of K-SVD autoencoder denoising.

[0168] In order to further compare the effectiveness of different methods in noise reduction, the performance evaluation indicators of different noise reduction methods, namely, root mean square error (RMSE) and signal-to-noise ratio (SNR), are calculated, as shown in Table 4 and Figure 5 As shown, it can be seen that the K-SVD autoencoder of this embodiment has a larger SNR and a smaller RMSE after denoising, and has better superiority.

[0169] Table 4: SNR indicators of different methods

[0170]

[0171] Table 5: RMSE indicators of different methods

[0172]

[0173] The new scene after denoising is subjected to feature extraction to obtain a feature matrix P. The PCA method, KPCA method and LE method are used to perform feature dimension reduction on the feature matrix, and the dimension reduction effects of the three methods are compared. The experimental parameters of the three methods are shown in Tables 6, 7 and 8.

[0174] Table 6: PCA algorithm experimental parameters

[0175]

[0176] Table 7: KPCA algorithm experimental parameters

[0177]

[0178] Table 8: LE algorithm experimental parameters

[0179]

[0180] According to the above setting parameters, the dimensionality reduction results of the three methods are as follows Fig. 9 shown. Fig. 9 a, 9b, 9c The effect diagrams of three feature dimensionality reduction methods. It can be seen from the figure that PCA does not form a good distinction for different categories of data after dimensionality reduction, and KPCA does not form a good distinction between sun gear fault and outer ring fault after dimensionality reduction, while LE can clearly show that all three signals have formed a good distinction. And by comparing the performance index TCSS of the three feature dimensionality reduction methods in the figure, the LE method is 4.24, which is lower than 5.65 of KPCA and PCA; by comparing the inter-class performance index TDBC of the three feature dimensionality reduction methods, the LE method is 20.53, which is higher than 17.31 of KPCA and 2.29 of PCA. From the intra-class index and inter-class index, it can be proved that the LE algorithm is better than the two algorithms in dimensionality reduction effect, reflecting the effectiveness and superiority of the LE dimensionality reduction algorithm used.

[0181] The fault data V after dimensionality reduction of the sampled LE is used for fault diagnosis using BP neural network, CNN neural network and Semi-GCN fault diagnosis model. The data set is divided into 60% training set, 20% validation set and 20% test set. The entire data set is set according to the previous fault data volume: 3600 sets of training set V train data, 1200 sets of validation set Vval and 1200 sets of test set V test. The number of training times is 160 times, the learning rate is 0.01, the function is Relu function, the loss function is cross entropy function, the optimization function is stochastic gradient descent, and the batch data volume is 1200 sets.

[0182] The BP neural network is a 4-layer neural network, the first layer is a fully connected layer, the second layer is a fully connected layer, the third layer is a fully connected layer, and the fourth layer is a fully connected layer. The BP neural network parameter settings are shown in Table 9.

[0183] Table 9: BP neural network model parameters

[0184]

[0185] The CNN is a 5-layer neural network, the first layer is a fully connected layer, the second layer is a fully connected layer, the third layer is a convolutional layer with a step size of 1; the fourth layer is a convolutional layer with a step size of 1; the fifth layer is a pooling layer with a step size of 2. The CNN parameter settings are shown in Table 10;

[0186] Table 10: CNN model parameter table:

[0187]

[0188] Table 11: Semi-GCN model parameter table

[0189]

[0190] According to the above three diagnostic methods, Fig.10 The accuracy curves of the three diagnostic networks shown in the figure show that after 160 training iterations, the accuracy of the validation set of BP neural network, CNN and Semi-Semi-GCN has reached about 90%, and they are all in a convergence state. In addition, the loss value of Semi-GCN is lower than that of BP and CNN, indicating that the diagnostic results predicted by GCN are more consistent with the actual situation. Therefore, the effectiveness of the Semi-GCN method proposed in this paper can be demonstrated. The classification confusion matrix of the three fault methods is obtained through test set classification as shown in the figure below. Fig.11 As shown, the accuracy and precision of the three fault diagnosis methods are calculated based on the confusion matrix.

[0191] Table 12: Test set diagnostic accuracy of different algorithms

[0192]

[0193] Table 13: Test set diagnostic accuracy of different algorithms

[0194]

[0195] It can be seen from Tables 12 and 13 that the accuracy of the Semi-GCN test set is the highest. The Semi-GCN fault diagnosis network has the highest accuracy for normal signals, sun gear faults, and outer gear ring faults. It can be seen that for each type of signal, Semi-GCN has a higher accuracy, and the diagnosis results have a higher credibility. The accuracy and precision indicators prove that Semi-GCN is superior.

[0196] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A planetary gearbox fault diagnosis method based on semi-supervised graph convolution, characterized in that: The steps include: S1. Acquire raw data of a planetary gearbox, and use an autoencoder denoising network to perform denoising on the raw data; S2, extracting features from the denoised data to obtain a feature matrix, and performing dimensionality reduction processing on the feature matrix; S3, obtaining an undirected graph between each data in the feature matrix after the dimensionality reduction processing, inputting the matrix after the dimensionality reduction and the undirected graph between the data into a semi-supervised graph convolutional neural network fault diagnosis model to obtain a diagnosis result; The method for establishing the semi-supervised graph convolutional neural network fault diagnosis model includes: S31. Build a semi-supervised graph convolutional neural network fault diagnosis model; S32, input training data, the training data is the feature matrix after reduction, the training data is divided into test set V train , validation set V val and the test set V test , where the test set V train Including labeled data and unlabeled data, validation set V val and the test set V test All are labeled data, and the adjacency matrix A is established train , A val and A test ; S33, initializing the hyperparameters of the semi-supervised graph convolutional neural network fault diagnosis model and the convolution kernel parameters of each layer of the network; S34, V train and the adjacency matrix A train Perform forward propagation in and extract the predicted category ; S35. According to the predicted category and the actual category Y of the labeled data Li Calculate the loss function L; S36, back-propagating the loss function, updating the convolution kernel parameters of each layer, and obtaining a semi-supervised graph convolutional neural network fault diagnosis model; S37, the verification set V val and the adjacency matrix A val Input the semi-supervised graph convolutional neural network fault diagnosis model obtained in step S36 for verification, and output the verification set loss value and the verification set accuracy; S38, determine whether the number of iterations is less than the maximum number of iterations n, if so, repeat steps S34-S37, if not, end the training and output the final semi-supervised graph convolutional neural network fault diagnosis model; S39, test set V test and the adjacency matrix A test Input the semi-supervised graph convolutional neural network fault diagnosis model obtained by S38, and output the recognition accuracy and loss value of the planetary gearbox fault diagnosis test set.

2. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 1 is characterized in that: Methods for denoising raw data include: S11. Establish an autoencoder denoising network and initialize the dictionaries of each layer of the autoencoder denoising network and allowable error , where i=1, 2, …, v, represents the number of layers of the autoencoder denoising network; S12. Use the orthogonal matching pursuit algorithm to calculate the sparse coefficients of each layer , i=1,2,…,v; S13. Calculate reconstruction error , when the reconstruction error Is it less than the allowable error? If yes, execute step S15; if no, execute step S14. The reconstruction error for: ； in, is the original data, is the data after noise reduction; S14. Update the dictionary of each layer by K-SVD method , and determine whether the update deadline condition is met, if so, execute step S15, if not, execute step S12; S15, the current dictionaries of each layer The v-th layer sparse coefficient , get the noise-reduced data , .

3. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 2 is characterized in that: The original data Split into several matrix blocks to form the following matrix: ； matrix The matrix blocks in the denoising process are processed in batches to obtain the corresponding processed data blocks, and then the denoised signal is obtained. for: 。 4. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 2 is characterized in that: The number of atoms in each layer of the dictionary in the autoencoder denoising network first increases and then decreases in a layer-by-layer order.

5. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 2 is characterized in that: Step S12 includes: S121, initializing the first layer dictionary D1, and initializing the first layer sparse coefficient X1 using the orthogonal matching pursuit algorithm; S122, using the first layer sparse coefficient X1 as the input signal of the second layer, and obtaining the second layer sparse coefficient X2 by using the orthogonal matching pursuit algorithm with the second layer dictionary D2; S123, repeat step S122 until the v-1th layer sparse coefficient X v-1 As the input signal of the vth layer, and the vth layer dictionary D v Use the orthogonal matching pursuit algorithm to get the v-th layer sparse coefficient X v .

6. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 5, characterized in that: The K-SVD method is used to update the dictionaries of each layer The methods include: S141, updating the first layer dictionary D1 according to the original data, the dictionary D1 and the sparse coefficient X1 according to the K-SVD method; S142, continue to use the sparse coefficient X1 as the input signal, and use the K-SVD method together with the dictionary D2 to update the second-layer dictionary D2; S143, repeat step S142 until the sparse coefficient X v-1 As the input signal, the dictionary D5 and the sparse coefficient X5 use the K-SVD method to update the v-th layer dictionary D v .

7. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 1 is characterized in that: Step S2 includes: S21, extracting features from the denoised data to obtain time domain features and frequency domain features to form a feature matrix; S22. Determine the edge weight value W using the Gaussian kernel function ij , and then get the adjacency graph W; S23, calculating the Laplace matrix L=DW according to the adjacency graph W, and setting the dimension of the reduced matrix to e; S24, select the eigenvectors corresponding to the first e smallest non-zero eigenvalues, and arrange them in ascending order according to the eigenvalues ​​to obtain the feature matrix V after feature dimension reduction = (v (1) v (2) … v (e) ).

8. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 1, characterized in that: The semi-supervised graph convolutional neural network fault diagnosis model includes a forward propagation module and a back propagation module. The forward propagation module includes a graph convolutional filter layer and a nonlinear function layer Relu, and the back propagation module includes a graph convolutional filter layer and a softmax classification layer.

9. The planetary gearbox fault diagnosis method based on semi-supervised graph convolution according to claim 1, characterized in that: The labeled data accounts for 10% of the total dataset.

Images