Sliding bearing fault diagnosis method and system under data extreme imbalance

Abstract

The invention belongs to the technical field of sliding bearings, and particularly discloses a sliding bearing fault diagnosis method and system under data extreme imbalance, and the method comprises the steps: collecting vibration accelerations and acoustic emission signals at different positions of a sliding bearing in different states, extracting initial features of data feature vectors, and splicing the initial features to obtain fusion features; a multi-source heterogeneous data fusion fault diagnosis model is constructed and split, a multi-source heterogeneous data fusion feature generation model is constructed and trained based on CWGAN-GP, a CWGAN-GP model generator and a discriminator are subjected to confrontation adjustment, features generated by the generator each time are input into a fault diagnosis classification part of the fault diagnosis model to calculate classifier loss, and a fault diagnosis result is obtained. And feeding back to the generator network to adjust parameters until nearly real characteristics are generated, and outputting a bearing fault diagnosis result. By adopting the technical scheme, the characteristic data with higher value density is generated under the condition of fully utilizing the data information to supplement the unbalanced sample, and the fault diagnosis under the extreme imbalance of the sliding bearing data is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of sliding bearings, and relates to a sliding bearing fault diagnosis method and system under extreme data imbalance. Background Art

[0002] As the joint of heavy-duty mechanical transmission system, sliding bearings are widely used in rotating machinery in mining, metallurgy, electric power and other industries. Since they play the role of slewing support, once a failure occurs, it will cause abnormal vibration of the equipment, which will have a great impact on the normal operation and service life of the entire equipment. Therefore, it is very necessary to detect and diagnose the faults. When the project is actually running, most of the time it is in normal state, there is sufficient normal data, missing fault data, and extreme imbalance in the data. Therefore, it is necessary to conduct fault diagnosis research under the state of sliding bearings based on data imbalance.

[0003] At present, the research on multi-source heterogeneous data fusion is mainly divided into data-level fusion, feature-level fusion and decision-level fusion. Most of them choose the latter two for better accuracy and robustness in some cases, while ignoring the complexity and computational consumption of the fusion model.

[0004] Moreover, most current data-level fusions calculate indicators and extract features from time-domain data, ignoring the ability of spectral features to characterize bearing status information. When analyzing data with unbalanced fault categories, the construction of the loss function ignores the impact of the number of labels of different categories on model training. When the fault samples are extremely unbalanced, the generative adversarial network is often used to directly generate data, which will cause information redundancy and loss of some detailed information. Summary of the invention

[0005] The purpose of the present invention is to provide a sliding bearing fault diagnosis method and system under extreme data imbalance, improve the data sample quality, and realize sliding bearing fault diagnosis under extreme data imbalance.

[0006] In order to achieve the above object, the basic scheme of the present invention is: a sliding bearing fault diagnosis method under extreme data imbalance, comprising the following steps:

[0007] S1, collects vibration acceleration and acoustic emission signals at different positions of the sliding bearing in different states;

[0008] S2, performing data preprocessing on the vibration acceleration signal, extracting its frequency spectrum features to form a vibration acceleration feature vector;

[0009] Perform data preprocessing on the acoustic emission signal, calculate its energy characteristics and spectrum characteristics, and obtain the acoustic emission data feature vector;

[0010] Creating data labels for the data feature vectors of the vibration acceleration and the acoustic emission signal, and matching the data labels with the corresponding data feature vectors;

[0011] S3, based on convolutional neural network and attention weighting mechanism, extracts preliminary features of vibration acceleration feature vector and acoustic emission data feature vector and concatenates them to obtain fusion features;

[0012] Taking the fusion features as input, a multi-source heterogeneous data fusion fault diagnosis model based on adaptive focal loss function is constructed;

[0013] The trained multi-source heterogeneous data fusion fault diagnosis model is split into a feature extraction and fusion part and a classifier part, and the model parameters and weights of the two parts are frozen and recorded for data extraction and classification respectively;

[0014] S4, based on CWGAN-GP, construct a multi-source heterogeneous data fusion feature generation model, namely, the CWGAN-GP model, and use the fusion features obtained in step S3 as real features to train the CWGAN-GP model;

[0015] While the generator and discriminator of the CWGAN-GP model are being adjusted adversarially, the features generated by the generator each time are input into the fault diagnosis classification part of the multi-source heterogeneous data fusion fault diagnosis model with frozen parameter weights to calculate the classifier loss, and then fed back to the generator to adjust its parameters until features close to the real ones are generated and the bearing fault diagnosis results are output.

[0016] The working principle and beneficial effects of this basic solution are as follows: This technical solution is based on the fusion of multi-source heterogeneous data, fully considering the expressiveness of the spectral characteristics of vibration acceleration and acoustic emission data, the influence of the category label comparison model under extreme sample imbalance, and the advantages of generating feature data over generating original signals, to achieve the fusion extraction of multi-source heterogeneous data features and improve the quality of data samples. Generate more valuable feature data to supplement the unbalanced samples while making full use of data information, and realize fault diagnosis under extreme unbalanced sliding bearing data.

[0017] Furthermore, in step S2, the vibration acceleration signal is preprocessed to extract its spectrum features to form a vibration acceleration feature vector as follows:

[0018] S201, dividing the multi-channel vibration acceleration signal into vibration data segments of equal acquisition time in the time domain, and ensuring that each segment of data contains at least data of multiple rotation cycles, and ensuring that the spectrum has sufficient frequency resolution;

[0019] S202, using a wavelet threshold denoising algorithm to remove high-frequency noise components of each vibration data segment;

[0020] S203, performing fast Fourier transform on the denoised vibration data to obtain its spectrum, and defining the spectrum feature F of the vibration acceleration signal vb(n1～n2) A is the spectrum area in the corresponding frequency interval (n1~n2) Hz (n1～n2) 、Frequency center of gravity FC (n1～n2) , mean square frequency MSF (n1～n2) And the frequency variance FV (n1～n2) The eigenvector formed by the combination of two adjacent spectral lines on the spectrum graph is defined as the trapezoidal area:

[0021] F vb(n1～n2) =[A (n1～n2) ,FC (n1～n2) ,MSF (n1～n2) ,FV (n1～n2) ]

[0022] A (n1～n2) =(v 1 +v 2 )×(f 2 -f 1 ) / 2

[0023] Among them, f i is the frequency corresponding to spectral line i; v i is the amplitude corresponding to spectral line i;

[0024] The characteristic vector F of the vibration acceleration signal vb It consists of the spectrum amplitude of the signal from 0 to 300 Hz and the spectrum characteristics of 300 to 600 Hz, 600 to 900 Hz and the full spectrum, that is:

[0025] Fvb=[f vb0～300 ,F vb300～600 ,F vb600～900 ,F vbf ]

[0026] Among them, F vbf =[FC f ,MSF f ,FV f ] is the spectrum characteristic of the full spectrum, FC f ,MSF f ,FV f is the frequency center of gravity, mean square frequency and frequency variance of the full spectrum;

[0027] S204, dividing the feature vectors of the vibration acceleration signal of each channel into a model training set, a validation set and a test set in chronological order, performing maximum and minimum normalization on them, and making feature set labels according to the fault category to which the data belongs.

[0028] Preprocess the vibration acceleration signal for subsequent use.

[0029] Further, in step S2, the acoustic emission signal is preprocessed to calculate its energy characteristics and spectrum characteristics, and the steps of obtaining the acoustic emission data feature vector are as follows:

[0030] S2201, combine nft acoustic emission data into a sample group, and use one-dimensional discrete wavelet transform to obtain sample group num 1 Layer detail coefficients and approximate coefficients of the i-th layer, and calculate the corresponding energy values and Calculate the mean energy value of each layer for all samples and find the total energy value E of the signal group:

[0031]

[0032] Among them, E di is the energy value of the detail coefficient, E a4 is the energy value of the approximate coefficient, nf is the code of the sample;

[0033] Calculate the ratio of the mean energy of each layer to the total energy E:

[0034]

[0035] Among them, p di is the ratio of detail coefficient energy value to total energy, p a4 is the ratio of the approximate coefficient energy value to the total energy;

[0036] Calculate the energy entropy E of each sample group n :

[0037]

[0038] The energy characteristic E of each sample group f Composition:

[0039] E f =[p d1 ,p d2 ,...,p dnum1 ,p a4 ,E n ];

[0040] S2202, calculating all sample spectra, deleting sample spectrum components exceeding the maximum frequency, and calculating the spectrum average;

[0041] S2203, extract the frequency spectrum feature F of the acoustic emission data corresponding to the frequency range (n3~n4) kHz from the spectrum average result. ae(n3～n4) :

[0042] F ae(n3～n4) =[mean(n3～n4) ,FC ((n3～n4) ,MSF (n3～n4) ,FV (n3～n4) ,A (n3～n4) ]

[0043] Among them, mean (n3～n4) , A (n3～n4) ,FC ((n3～n4) ,MSF (n3～n4) ,FV (n3～n4) The spectrum mean, spectrum area, frequency center of gravity, mean square frequency and frequency variance in the corresponding frequency interval (n3~n4) Hz;

[0044] Construct the feature vector F of the kth acoustic emission sensor signal k ae :

[0045] F k ae =[F aef ,F ae0～100 ,F ae100～200 ,F ae200～300 ,F ae300～450 ,E f ]

[0046] Among them, F aef is the characteristic of the full spectrum, F ae0～100 、F ae100～200 、F ae200～300 、F ae300～450 The characteristics of the ranges of 0-100, 100-200, 200-300 and 300-450kHz are shown in Figure 1. f is the energy characteristic;

[0047] S2204, concatenate the feature vectors of all acoustic emission signals to obtain a data-level fused acoustic emission signal feature vector F ae :

[0048] F ae =[F ae1 ,F ae2 ,F ae3 ,F ae4 ];

[0049] S2205, divide the feature vector of the acoustic emission signal of each channel into a model training set, a validation set and a test set in chronological order, and perform maximum and minimum normalization on them, share the label data of the vibration acceleration feature vector with the acoustic emission feature vector, and randomly sort the multi-source heterogeneous data feature set with different labels to remove the influence of the label order on the model training, and save the multi-source heterogeneous data feature set after completion.

[0050] Pre-process the acoustic emission signal for subsequent use.

[0051] Furthermore, in step S3, based on the convolutional neural network and the attention weighting mechanism, the preliminary features of the vibration acceleration feature vector and the acoustic emission data feature vector are extracted and spliced ​​to obtain the fusion feature. The specific method is:

[0052] S301, constructing a convolution block for each vibration acceleration signal, each convolution block consists of a one-dimensional convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer, and the vibration acceleration feature vectors of multiple channels are summed after preliminary feature extraction by the convolution block to achieve homogeneous information fusion;

[0053] S302, after the fused features pass through the BN layer, the GELU layer and the maximum pooling layer, they are respectively input into a one-dimensional convolution layer and a convolution block, and residual connection is realized by adding the projection of the features by the one-dimensional convolution layer and the output of the convolution block;

[0054] The summed result is input into the convolution block and the one-dimensional convolution layer again through the activation layer and then residually connected again;

[0055] The multiple channel feature maps obtained by summing up are summed up to obtain the final vibration data fusion feature;

[0056] S303, the acoustic emission data feature vectors are respectively passed through a convolution block and a one-dimensional convolution layer, and then summed up to realize residual connection. The summed result is respectively input into the convolution block and the one-dimensional convolution layer again through a GELU activation layer for residual connection, and then the preliminary extraction features of the acoustic emission are obtained through the GELU activation layer;

[0057] S304, based on the squeeze and excitation network, an attention weighted network is constructed, the acoustic emission is initially extracted, the output feature map after the input convolution block has n channels, the feature map is globally averaged pooled, and the channel average value is extracted as its descriptor to realize data squeezing;

[0058] The mean vector composed of channel descriptors is input into the attention weighted network activation. The attention weighted network hidden layer has only one neuron, and the activation function is set to the rectified linear unit.

[0059] The output layer has n neurons, and the weight of each channel is obtained through the Sigmoid function, and the weight vector of each channel is output;

[0060] S305, multiplying the weight vector of each channel obtained by adaptive attention weighting with the original feature map channel output by the corresponding convolution block to weight the channel dimension, and summing the feature maps before and after weighting to realize residual connection;

[0061] The feature map is reduced in dimension through one-dimensional convolutional layers and flattening layers, which reduces the number of model parameters, improves training speed, and unifies the dimensions of acoustic emission features and vibration data features;

[0062] S306, concatenating the vibration acceleration data feature with the acoustic emission data feature vector to obtain a final fusion feature.

[0063] The spectral features of acoustic emission and vibration data are further extracted through the feature extraction network to achieve the fusion extraction of multi-source heterogeneous data features and improve the quality of data samples.

[0064] Furthermore, in step S3, the fusion features are used as input to construct a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focal loss function as follows:

[0065] S3301, inputting the fused feature vector into a classification diagnosis network in a multi-source heterogeneous data fusion fault diagnosis model, where the classification diagnosis network includes two groups of random dropout layers, a fully connected layer, and an activation layer;

[0066] Some neuron nodes are deleted through the random inactivation layer, L1 regularization is used in the fully connected layer to control the complexity of the model, Gaussian error linear unit is selected as the activation function in activation layer 1, and Softmax activation function is used in activation layer 2 to output the probability estimate of the health state;

[0067] S3302, based on the focal loss function, the square root of the inverse of the category label ratio in the training set is used as a penalty factor to establish an adaptive focal loss function Loss:

[0068]

[0069] Where N is the total number of categories; λ is the focusing parameter, which is adjusted according to the specific application; δ i is the predicted probability of category i; y i is the label of category i; p i is the ratio of the number of labels of category i to the total number of labels of all categories.

[0070] The square root of the reciprocal of the proportion of each category label in the training sample is used as the penalty factor, and the classification loss penalty for the majority class is adaptively adjusted according to the degree of class imbalance to reduce the impact of the category ratio under extreme data imbalance conditions.

[0071] Further, the specific steps of step S4 are as follows:

[0072] S401, due to the extreme imbalance of bearing fault data, a multi-source heterogeneous data fusion feature generation model based on the conditional Wasserstein generative adversarial network with gradient penalty is constructed. The training objectives of this model are:

[0073]

[0074] Among them, x is the real data; y is the label corresponding to the data; z is the noise vector, with the subscript p data (z) is the real data distribution, p z (z) is the noise distribution in the latent space, is the interpolation distribution between the generated samples and the real samples; D(·) is the discriminator output; G(·) is the generator output; λ is the penalty coefficient, and the penalty term is the gradient constraint, is a random distribution between the original sample distribution and the generated sample distribution, and α~U(0,1); Lc is the classification loss, and γ is the weight of Lc;

[0075] During training, the discriminator aims to maximize the output value when real data is input and minimize the output value when generated data is input. The generator aims to maximize the output value of the discriminator for the generated samples and ensure that the generated features correspond correctly to the category labels.

[0076] S402, extracting data and its fault category label information from the fused feature vector as a real sample;

[0077] The generator parameters are fixed, random noise and fault category label information that conform to the normal distribution are extracted from the generator input distribution and input into the generator to generate multi-source heterogeneous data fusion feature fake samples with the same number of real samples;

[0078] The multi-source heterogeneous data are fused with feature false samples and real samples and their respective fault category label information, and input into the discriminator to calculate the discrimination results and gradient penalty items, and update the discriminator parameters;

[0079] S403, input the fake samples generated by the generator into the classifier part of the multi-source heterogeneous data fusion fault diagnosis model built in step S3, calculate the classification loss, and feed the classification loss back to the training loss in the generator network to adjust the generator network parameters;

[0080] S404, fix the discriminator parameters, train the generator, extract random noise and fault category label information that conform to the normal distribution from the generator input distribution, input them into the generator to generate fake samples, calculate the generator loss, and adjust the generator network structure parameters together with the classifier loss calculated in step S403;

[0081] S405, repeating steps S402 to S404, alternately training the discriminator and the generator, and updating the parameters of the generative adversarial network until the accuracy of the samples generated by the generator reaches the expected level in the classifier;

[0082] S406, saving the trained generator network, and then inputting random noise and label information into it to generate multi-source heterogeneous data fusion features close to reality, and combining it with the normal state features in the real samples to form a new training set;

[0083] The training set is used to retrain the fully connected classifier network, and the network diagnosis accuracy is verified using a real multi-source heterogeneous data fusion feature set to perform bearing fault diagnosis under extreme data imbalance.

[0084] The conditional Wasserstein generative adversarial network with gradient penalty is used to realize the fusion feature generation of multi-source heterogeneous data under the condition of extreme class imbalance, and the sliding bearing fault diagnosis is realized based on the original fusion features and the generated features.

[0085] The present invention also provides a sliding bearing fault diagnosis system under extreme data imbalance, comprising a data acquisition module and a processing module, wherein the data acquisition module is used to collect vibration acceleration and acoustic emission signals at different positions of the sliding bearing in different states;

[0086] The input end of the processing module is connected to the output end of the data acquisition module, and the processing module executes the method of the present invention to complete the sliding bearing fault diagnosis under extreme data imbalance.

[0087] This system uses data acquisition module and processing module to develop a sliding bearing fault diagnosis method based on multi-source heterogeneous data fusion, adaptive focal loss function and conditional Wasserstein generative adversarial network in the scenario of data category imbalance.

[0088] Furthermore, the data acquisition module includes a multi-channel vibration acceleration and acoustic emission sensor, which is installed at different measuring points of the sliding bearing to obtain multi-source heterogeneous data under different states of the sliding bearing.

[0089] The data acquisition module has a simple structure and is convenient for collecting required information. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 It is a flow chart of a method for diagnosing sliding bearing faults under extreme data imbalance of the present invention;

[0091] Figure 2 It is a schematic flow chart of the feature extraction and fusion of vibration acceleration data of the sliding bearing fault diagnosis method under extreme data imbalance of the present invention;

[0092] Figure 3 It is a schematic flow chart of acoustic emission data feature extraction and fusion of the sliding bearing fault diagnosis method under extreme data imbalance of the present invention;

[0093] Figure 4It is a structural schematic diagram of an attention weighted network of a sliding bearing fault diagnosis method under extreme data imbalance of the present invention;

[0094] Figure 5 It is a structural schematic diagram of a classifier network of a sliding bearing fault diagnosis method under extreme data imbalance of the present invention;

[0095] Figure 6 It is a structural schematic diagram of a multi-source heterogeneous data fusion feature generation model of a sliding bearing fault diagnosis method under extreme data imbalance of the present invention. DETAILED DESCRIPTION

[0096] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0097] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0098] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0099] The present invention discloses a method for diagnosing sliding bearing faults under extreme data imbalance. The method is based on multi-source heterogeneous data fusion, fully considers the expression ability of spectral characteristics of vibration acceleration and acoustic emission data and the influence of category label comparison model under extreme sample imbalance, and the advantage of generating feature data over generating original signals, generates feature data with higher value density to supplement unbalanced samples, and realizes fault diagnosis of sliding bearings under extreme data imbalance.

[0100] like Figure 1 As shown in the figure, the fault diagnosis method of sliding bearing under extreme data imbalance includes the following steps:

[0101] S1, collects vibration acceleration and acoustic emission signals at different positions of the sliding bearing in different states (such as normal state, bearing wear and abrasive wear, etc.). Since fault conditions are difficult to occur, the data is extremely unbalanced;

[0102] S2, performing data preprocessing on the vibration acceleration signal, extracting its frequency spectrum features to form a vibration acceleration feature vector;

[0103] Perform data preprocessing on the acoustic emission signal, calculate its energy characteristics and spectrum characteristics, and obtain the acoustic emission data feature vector;

[0104] Create data labels for the data feature vectors of vibration acceleration and acoustic emission signals, and match the data labels with the corresponding data feature vectors to achieve data-level multi-source heterogeneous data feature extraction;

[0105] S3, based on convolutional neural network and attention weighting mechanism, extracts preliminary features of vibration acceleration feature vector and acoustic emission data feature vector and concatenates them to obtain fusion features;

[0106] Taking the fusion features as input, a multi-source heterogeneous data fusion fault diagnosis model based on adaptive focal loss function is constructed;

[0107] The trained multi-source heterogeneous data fusion fault diagnosis model is split into a feature extraction and fusion part and a classifier part, and the model parameters and weights of the two parts are frozen and recorded respectively for data extraction and classification of subsequent models;

[0108] S4, based on CWGAN-GP, construct a multi-source heterogeneous data fusion feature generation model, namely, the CWGAN-GP model, and use the fusion features obtained in step S3 as real features to train the CWGAN-GP model;

[0109] While the generator and discriminator of the CWGAN-GP model are being adjusted adversarially, the features generated by the generator each time are input into the fault diagnosis classification part of the multi-source heterogeneous data fusion fault diagnosis model with frozen parameter weights to calculate the classifier loss, and then fed back to the generator to adjust its parameters until features close to the real ones are generated and the bearing fault diagnosis results are output.

[0110] In a preferred embodiment of the present invention, in step S2, the vibration acceleration signal is subjected to data preprocessing, and the method of extracting its spectrum features to form a vibration acceleration feature vector is as follows:

[0111] S201, dividing the multi-channel vibration acceleration signal into vibration data segments of equal acquisition time in the time domain, and ensuring that each segment of data contains at least data of multiple rotation cycles, and ensuring that the spectrum has sufficient frequency resolution;

[0112] S202, using a wavelet threshold denoising algorithm to remove high-frequency noise components of each vibration data segment;

[0113] S203, performing fast Fourier transform on the denoised vibration data to obtain its spectrum, and defining the spectrum feature F of the vibration acceleration signal vb(n1～n2) A is the spectrum area in the corresponding frequency interval (n1~n2) Hz (n1～n2) 、Frequency center of gravity FC (n1～n2) , mean square frequency MSF (n1～n2) And the frequency variance FV (n1～n2) The eigenvector formed by the combination of two adjacent spectral lines on the spectrum graph is defined as the trapezoidal area:

[0114] F vb(n1～n2) =[A (n1～n2) ,FC (n1～n2) ,MSF (n1～n2) ,FV (n1～n2) ]

[0115] A (n1～n2) =(v 1 +v 2 )×(f 2 -f 1 ) / 2

[0116] Among them, f i is the frequency corresponding to spectral line i; v i is the amplitude corresponding to spectral line i;

[0117] The characteristic vector F of the vibration acceleration signal vb It consists of the spectrum amplitude of the signal from 0 to 300 Hz and the spectrum characteristics of 300 to 600 Hz, 600 to 900 Hz and the full spectrum, that is:

[0118] F vb =[f vb0～300 ,F vb300～600 ,F vb600～900 ,F vbf ]

[0119] Among them, F vbf =[FC f ,MSF f ,FV f ] is the spectrum characteristic of the full spectrum, FC f ,MSF f ,FV f is the frequency center of gravity, mean square frequency and frequency variance of the full spectrum;

[0120] S204, dividing the feature vectors of the vibration acceleration signal of each channel into a model training set, a validation set and a test set in chronological order, performing maximum and minimum normalization on them, and making feature set labels according to the fault category to which the data belongs.

[0121] In a preferred embodiment of the present invention, the step of performing data preprocessing on the acoustic emission signal in step S2, calculating its energy characteristics and spectrum characteristics, and obtaining the acoustic emission data feature vector is as follows:

[0122] S2201, considering that the acoustic emission data sample time is short and cannot fully represent the fault information, the nft acoustic emission data are combined into a sample group, and the sample group num is obtained by using one-dimensional discrete wavelet transform 1 Layer detail coefficients and approximate coefficients of the i-th layer, and calculate the corresponding energy values and Calculate the mean energy value of each layer for all samples and find the total energy value E of the signal group:

[0123]

[0124] Among them, E di is the energy value of the detail coefficient, E a4 is the energy value of the approximate coefficient, nf is the code of the sample;

[0125] Calculate the ratio of the mean energy of each layer to the total energy E:

[0126]

[0127] Among them, p di is the ratio of detail coefficient energy value to total energy, p a4 is the ratio of the approximate coefficient energy value to the total energy;

[0128] Calculate the energy entropy E of each sample group n :

[0129]

[0130] The energy characteristic E of each sample group f Composition:

[0131] E f =[p d1 ,p d2 ,...,p dnum1 ,p a4 ,E n ];

[0132] S2202, calculating all sample spectra, deleting sample spectrum components that exceed a maximum frequency (after exceeding this frequency, the spectrum amplitude is close to 0), and calculating the spectrum average;

[0133] S2203, extract the frequency spectrum feature F of the acoustic emission data corresponding to the frequency range (n3~n4) kHz from the spectrum average result. ae(n3～n4) :

[0134] F ae(n3～n4) =[mean (n3～n4) ,FC ((n3～n4) ,MSF (n3～n4) ,FV (n3～n4) ,A (n3～n4) ]

[0135] Among them, mean (n3～n4) , A (n3～n4) ,FC ((n3～n4) ,MSF (n3～n4) ,FV (n3～n4) is the spectrum mean, spectrum area, frequency center of gravity, mean square frequency and frequency variance in the corresponding frequency interval (n3~n4) Hz;

[0136] Construct the feature vector F of the kth acoustic emission sensor signal k ae :

[0137] F k ae =[F aef ,F ae0～100 ,F ae100～200 ,F ae200～300 ,F ae300～450 ,E f ]

[0138] Among them, F aef is the characteristic of the full spectrum, F ae0～100 、F ae100～200 、F ae200～300 、F ae300～450 The characteristics of the ranges of 0-100, 100-200, 200-300 and 300-450kHz are shown in Figure 1. f is the energy characteristic;

[0139] S2204, concatenate the feature vectors of all acoustic emission signals to obtain a data-level fused acoustic emission signal feature vector F ae :

[0140] F ae =[F ae1 ,F ae2 ,F ae3 ,F ae4 ];

[0141] S2205, divide the feature vector of the acoustic emission signal of each channel into a model training set, a validation set and a test set in chronological order, and perform maximum and minimum normalization on them, share the label data of the vibration acceleration feature vector with the acoustic emission feature vector, and randomly sort the multi-source heterogeneous data feature set with different labels to remove the influence of the label order on the model training, and save the multi-source heterogeneous data feature set after completion.

[0142] In a preferred embodiment of the present invention, in step S3, based on a convolutional neural network and an attention weighting mechanism, preliminary features of the vibration acceleration feature vector and the acoustic emission data feature vector are extracted and spliced ​​to obtain fusion features. The specific method is:

[0143] S301, a convolution block is constructed for each vibration acceleration signal. Each convolution block consists of a one-dimensional convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer. The vibration acceleration feature vectors of multiple channels are extracted by the convolution block and then summed to achieve homogeneous information fusion, such as Figure 2 As shown;

[0144] S302, after the fused features pass through the BN layer, the GELU layer and the maximum pooling layer, they are respectively input into a one-dimensional convolution layer and a convolution block, and residual connection is realized by adding the projection of the features by the one-dimensional convolution layer and the output of the convolution block;

[0145] The summed result is input into the convolution block and the one-dimensional convolution layer again through the activation layer and then residually connected again;

[0146] The multiple channel feature maps obtained by summing up are summed up to obtain the final vibration data fusion feature;

[0147] S303, the acoustic emission data feature vectors are respectively passed through a convolution block and a one-dimensional convolution layer, and then summed up to realize residual connection. The summation result is respectively input into the convolution block and the one-dimensional convolution layer again through the GELU activation layer for residual connection. Then, the preliminary extraction features of acoustic emission are obtained through the GELU activation layer, such as Figure 3 As shown;

[0148] S304, based on the squeezing and excitation network, construct an attention weighted network (such as Figure 4 As shown in the figure, the output feature map of acoustic emission after initial feature extraction input into the convolution block has n channels. The feature map is globally averaged and pooled, and the channel average is extracted as its descriptor to achieve data squeezing.

[0149] The mean vector composed of channel descriptors is input into the attention weighted network activation. The attention weighted network hidden layer has only one neuron, and the activation function is set to the rectified linear unit.

[0150] The output layer has n neurons, and the weight of each channel is obtained through the Sigmoid function, and the weight vector of each channel is output;

[0151] S305, multiplying the weight vector of each channel obtained by adaptive attention weighting with the original feature map channel output by the corresponding convolution block to weight the channel dimension, and summing the feature maps before and after weighting to realize residual connection;

[0152] The feature map is reduced in dimension through one-dimensional convolutional layers and flattening layers, which reduces the number of model parameters, improves training speed, and unifies the dimensions of acoustic emission features and vibration data features;

[0153] S306, concatenating the vibration acceleration data feature with the acoustic emission data feature vector to obtain a final fusion feature.

[0154] In a preferred solution of the present invention, the steps of taking the fusion features as input and constructing a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focal loss function in step S3 are as follows:

[0155] S3301, inputting the fused feature vector into a classification diagnosis network in a multi-source heterogeneous data fusion fault diagnosis model, where the classification diagnosis network includes two groups of random dropout layers, a fully connected layer, and an activation layer;

[0156] The random inactivation layer deletes some neuron nodes, thereby reducing the connection between hidden units, improving robustness and alleviating overfitting. The fully connected layer uses L1 regularization to control model complexity and reduce overfitting. Activation layer 1 selects Gaussian error linear unit as the activation function, and activation layer 2 uses Softmax activation function to output the probability estimate of the health state;

[0157] S3302, taking into account the impact of extreme imbalance in the number of data of different categories on the loss function, and avoiding excessive loss of categories with extremely scarce samples, based on the focal loss function, the square root of the inverse of the category label ratio in the training set is used as the penalty factor to establish an adaptive focal loss function Loss:

[0158]

[0159] Where N is the total number of categories; λ is the focusing parameter, which is adjusted according to the specific application; δ i is the predicted probability of category i; y i is the label of category i; p i is the ratio of the number of labels of category i to the total number of labels of all categories.

[0160] According to the adaptive focus loss results, the classifier weights are iteratively adjusted until the final bearing status classification results are output. The trained model is split, and the model parameters and weights of the feature extraction fusion part and the fault diagnosis classification part are frozen and recorded respectively for data extraction and classification of subsequent models.

[0161] In a preferred embodiment of the present invention, the specific steps of step S4 are as follows:

[0162] S401, due to the extreme imbalance of bearing fault data, a multi-source heterogeneous data fusion feature generation model based on the conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP) is constructed (such as Figure 6 As shown), the model training objective is:

[0163]

[0164] Among them, x is the real data; y is the label corresponding to the data; z is the noise vector, with the subscript p data (z) is the real data distribution, p z (z) is the noise distribution in the latent space, is the interpolation distribution between the generated samples and the real samples; D(·) is the discriminator output; G(·) is the generator output; λ is the penalty coefficient, and the penalty term is the gradient constraint, is a random distribution between the original sample distribution and the generated sample distribution, and α~U(0,1); Lc is the classification loss, and γ is the weight of Lc;

[0165] During training, the discriminator aims to maximize the output value when real data is input and minimize the output value when generated data is input. The generator aims to maximize the output value of the discriminator for the generated samples and ensure that the generated features correspond correctly to the category labels.

[0166] S402, extracting data and its fault category label information from the fused feature vector as a real sample;

[0167] The generator parameters are fixed, random noise and fault category label information that conform to the normal distribution are extracted from the generator input distribution and input into the generator to generate multi-source heterogeneous data fusion feature fake samples with the same number of real samples;

[0168] The multi-source heterogeneous data are fused with feature false samples and real samples and their respective fault category label information, and input into the discriminator to calculate the discrimination results and gradient penalty items, and update the discriminator parameters;

[0169] S403, input the fake samples generated by the generator into the classifier part of the multi-source heterogeneous data fusion fault diagnosis model built in step S3, calculate the classification loss, and feed the classification loss back to the training loss in the generator network to adjust the generator network parameters;

[0170] S404, fix the discriminator parameters, train the generator, extract random noise and fault category label information that conform to the normal distribution from the generator input distribution, input them into the generator to generate fake samples, calculate the generator loss, and adjust the generator network structure parameters together with the classifier loss calculated in step S403;

[0171] S405, repeating steps S402 to S404, alternately training the discriminator and the generator, and updating the parameters of the generative adversarial network until the accuracy of the samples generated by the generator reaches the expected level in the classifier;

[0172] S406, saving the trained generator network, and then inputting random noise and label information into it to generate multi-source heterogeneous data fusion features close to reality, and combining it with the normal state features in the real samples to form a new training set;

[0173] Use this training set to retrain the fully connected classifier network (such as Figure 5 As shown in the figure, the network diagnosis accuracy is verified by using the real multi-source heterogeneous data fusion feature set to diagnose bearing faults under extreme data imbalance.

[0174] The present invention also provides a sliding bearing fault diagnosis system under extreme data imbalance, comprising a data acquisition module and a processing module, wherein the data acquisition module is used to collect vibration acceleration and acoustic emission signals at different positions of the sliding bearing in different states.

[0175] The input end of the processing module is electrically connected to the output end of the data acquisition module, and the processing module executes the method of the present invention to complete the fault diagnosis of the sliding bearing under extreme data imbalance.

[0176] This system uses data acquisition module and processing module to develop a sliding bearing fault diagnosis method based on multi-source heterogeneous data fusion, adaptive focal loss function and conditional Wasserstein generative adversarial network in the scenario of data category imbalance.

[0177] In a preferred embodiment of the present invention, the data acquisition module includes a multi-channel vibration acceleration and acoustic emission sensor, which is installed at different measuring points of the sliding bearing to obtain multi-source heterogeneous data under different states of the sliding bearing.

[0178] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0179] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for diagnosing sliding bearing faults under extreme data imbalance, characterized in that: The steps include: S1, collects vibration acceleration and acoustic emission signals at different positions of the sliding bearing in different states; S2, performing data preprocessing on the vibration acceleration signal, extracting its frequency spectrum features to form a vibration acceleration feature vector; Perform data preprocessing on the acoustic emission signal, calculate its energy characteristics and spectrum characteristics, and obtain the acoustic emission data feature vector; Creating data labels for the data feature vectors of the vibration acceleration and the acoustic emission signal, and matching the data labels with the corresponding data feature vectors; S3, based on convolutional neural network and attention weighting mechanism, extracts preliminary features of vibration acceleration feature vector and acoustic emission data feature vector and concatenates them to obtain fusion features; Taking the fusion features as input, a multi-source heterogeneous data fusion fault diagnosis model based on adaptive focal loss function is constructed; The trained multi-source heterogeneous data fusion fault diagnosis model is split into a feature extraction and fusion part and a classifier part, and the model parameters and weights of the two parts are frozen and recorded respectively; S4, based on CWGAN-GP, construct a multi-source heterogeneous data fusion feature generation model, namely, the CWGAN-GP model, and use the fusion features obtained in step S3 as real features to train the CWGAN-GP model; While the generator and discriminator of the CWGAN-GP model are being adjusted adversarially, the features generated by the generator each time are input into the fault diagnosis classification part of the multi-source heterogeneous data fusion fault diagnosis model with frozen parameter weights to calculate the classifier loss, and then fed back to the generator to adjust its parameters until features close to the real ones are generated and the bearing fault diagnosis results are output.

2. The method for diagnosing sliding bearing faults under extreme data imbalance according to claim 1, characterized in that: In step S2, the vibration acceleration signal is preprocessed to extract its spectrum features to form a vibration acceleration feature vector as follows: S201, dividing the multi-channel vibration acceleration signal into vibration data segments of equal acquisition time in the time domain, and ensuring that each data segment contains at least data of multiple rotation cycles; S202, using a wavelet threshold denoising algorithm to remove high-frequency noise components of each vibration data segment; S203, performing fast Fourier transform on the denoised vibration data to obtain its spectrum, and defining the spectrum feature F of the vibration acceleration signal vb(n1～n2) A is the spectrum area in the corresponding frequency interval (n1~n2) Hz (n1～n2) 、Frequency center of gravity FC (n1～n2) , mean square frequency MSF (n1～n2) And the frequency variance FV (n1～n2) The eigenvector formed by the combination of two adjacent spectral lines on the spectrum graph is defined as the trapezoidal area: F vb(n1～n2) =[A (n1～n2) ,FC (n1～n2) ,MSF (n1～n2 ),FV (n1～n2) ] <h2 style=";text-align:left;direction:ltr">A<h2 style=";text-align:left;direction:ltr"> (n1～n2) <h2 style=";text-align:left;direction:ltr"> (v1+v2)×(f2-f1) / 2 Among them, f i is the frequency corresponding to spectral line i; v i is the amplitude corresponding to spectral line i; The characteristic vector F of the vibration acceleration signal vb It consists of the spectrum amplitude of the signal from 0 to 300 Hz and the spectrum characteristics of 300 to 600 Hz, 600 to 900 Hz and the full spectrum, that is: F vb =[f vb0～300 ,F vb300～600 ,F vb600～900 ,F vbf ] Among them, F vbf =[FC f ,MSF f ,FV f ] is the spectrum feature of the full spectrum, FC f ,MSF f ,FV f is the frequency center of gravity, mean square frequency and frequency variance of the full spectrum; S204, dividing the feature vectors of the vibration acceleration signal of each channel into a model training set, a validation set and a test set in chronological order, performing maximum and minimum normalization on them, and making feature set labels according to the fault category to which the data belongs.

3. The method for diagnosing sliding bearing faults under extreme data imbalance according to claim 1, characterized in that: In step S2, the steps of performing data preprocessing on the acoustic emission signal, calculating its energy characteristics and spectrum characteristics, and obtaining the acoustic emission data feature vector are as follows: S2201, combine nft acoustic emission data into a sample group, use one-dimensional discrete wavelet transform to obtain the detail coefficient of the sample group num1 layer and the approximate coefficient of the i-th layer, and calculate the corresponding energy value and Calculate the mean energy value of each layer for all samples and find the total energy value E of the signal group: Among them, E di is the energy value of the detail coefficient, E a4 is the energy value of the approximate coefficient, nf is the code of the sample; Calculate the ratio of the mean energy of each layer to the total energy E: Among them, p di is the ratio of detail coefficient energy value to total energy, p a4 is the ratio of the approximate coefficient energy value to the total energy; Calculate the energy entropy E of each sample group n : The energy characteristic E of each sample group f It is composed of: E f =[p d1 ,p d2 ,...,p dnum1 ,p a4 ,E n ]; S2202, calculating all sample spectra, deleting sample spectrum components exceeding the maximum frequency, and calculating the spectrum average; S2203, extract the frequency spectrum feature F of the acoustic emission data corresponding to the frequency range (n3~n4) kHz from the spectrum average result. ae(n3～n4) : F ae(n3～n4) =[mean (n3～n4) ,FC ((n3～n4) ,MSF (n3～n4) ,FV (n3～n4) ,A (n3～n4) ] Among them, mean (n3～n4) ,A (n3～n4) ,FC ((n3～n4) ,MSF (n3～n4) ,FV (n3～n4) is the spectrum mean, spectrum area, frequency center of gravity, mean square frequency and frequency variance in the corresponding frequency interval (n3~n4) Hz; Construct the feature vector F of the kth acoustic emission sensor signal k ae : F k ae =[F ae f,F ae0～100 ,F ae100～200 ,F ae200～300 ,F ae300～450 ,E f ] Among them, F aef is the characteristic of the full spectrum, F ae0～100 、F ae100～200 、F ae200～300 、F ae300～450 The characteristics of the ranges of 0-100, 100-200, 200-300 and 300-450kHz are shown in Figure 1. f is the energy characteristic; S2204, concatenate the feature vectors of all acoustic emission signals to obtain a data-level fused acoustic emission signal feature vector F ae : F ae =[F ae1 ,F ae2 ,F ae3 ,F ae4 ]; S2205, divide the feature vector of the acoustic emission signal of each channel into a model training set, a validation set and a test set in chronological order, and perform maximum and minimum normalization on them, share the label data of the vibration acceleration feature vector with the acoustic emission feature vector, and randomly sort the multi-source heterogeneous data feature set with different labels to remove the influence of the label order on the model training, and save the multi-source heterogeneous data feature set after completion.

4. The method for diagnosing sliding bearing faults under extreme data imbalance according to claim 1, characterized in that: In step S3, based on the convolutional neural network and the attention weighting mechanism, the preliminary features of the vibration acceleration feature vector and the acoustic emission data feature vector are extracted and spliced ​​to obtain the fusion features. The specific method is as follows: S301, constructing a convolution block for each vibration acceleration signal, each convolution block consists of a one-dimensional convolution layer, a batch normalization layer, an activation layer, and a maximum pooling layer, and the vibration acceleration feature vectors of multiple channels are summed after preliminary feature extraction by the convolution block to achieve homogeneous information fusion; S302, after the fused features pass through the BN layer, the GELU layer and the maximum pooling layer, they are respectively input into a one-dimensional convolution layer and a convolution block, and residual connection is realized by adding the projection of the features by the one-dimensional convolution layer and the output of the convolution block; The summed result is input into the convolution block and the one-dimensional convolution layer again through the activation layer and then residually connected again; The multiple channel feature maps obtained by summing up are summed up to obtain the final vibration data fusion feature; S303, the acoustic emission data feature vectors are respectively passed through a convolution block and a one-dimensional convolution layer, and then summed up to realize residual connection. The summed result is respectively input into the convolution block and the one-dimensional convolution layer again through a GELU activation layer for residual connection, and then the preliminary extraction features of the acoustic emission are obtained through the GELU activation layer; S304, based on the squeeze and excitation network, an attention weighted network is constructed, the acoustic emission is initially extracted, the output feature map after the input convolution block has n channels, the feature map is globally averaged pooled, and the channel average value is extracted as its descriptor to realize data squeezing; The mean vector composed of channel descriptors is input into the attention weighted network activation. The attention weighted network hidden layer has only one neuron, and the activation function is set to the rectified linear unit. The output layer has n neurons, and the weight of each channel is obtained through the Sigmoid function, and the weight vector of each channel is output; S305, multiplying the weight vector of each channel obtained by adaptive attention weighting with the original feature map channel output by the corresponding convolution block to weight the channel dimension, and summing the feature maps before and after weighting to realize residual connection; The feature map is reduced in dimension through one-dimensional convolutional layers and flattening layers, which reduces the number of model parameters, improves training speed, and unifies the dimensions of acoustic emission features and vibration data features; S306, concatenating the vibration acceleration data feature with the acoustic emission data feature vector to obtain a final fusion feature.

5. The method for diagnosing sliding bearing faults under extreme data imbalance according to claim 4, characterized in that: In step S3, the steps of taking the fusion features as input and constructing a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focal loss function are as follows: S3301, inputting the fused feature vector into a classification diagnosis network in a multi-source heterogeneous data fusion fault diagnosis model, where the classification diagnosis network includes two groups of random dropout layers, a fully connected layer, and an activation layer; Some neuron nodes are deleted through the random inactivation layer, L1 regularization is used in the fully connected layer to control the complexity of the model, Gaussian error linear unit is selected as the activation function in activation layer 1, and Softmax activation function is used in activation layer 2 to output the probability estimate of the health state; S3302, based on the focal loss function, the square root of the inverse of the category label ratio in the training set is used as a penalty factor to establish an adaptive focal loss function Loss: Where N is the total number of categories; λ is the focusing parameter, which is adjusted according to the specific application; δ i is the predicted probability of category i; y i is the label of category i; p i is the ratio of the number of labels of category i to the total number of labels of all categories.

6. The method for diagnosing sliding bearing faults under extreme data imbalance according to claim 1, characterized in that: The specific steps of step S4 are as follows: S401, due to the extreme imbalance of bearing fault data, a multi-source heterogeneous data fusion feature generation model based on the conditional Wasserstein generative adversarial network with gradient penalty is constructed. The training objectives of this model are: Among them, x is the real data; y is the label corresponding to the data; z is the noise vector, with the subscript p data (z) is the real data distribution, p z (z) is the noise distribution in the latent space, is the interpolation distribution between the generated samples and the real samples; D(·) is the discriminator output; G(·) is the generator output; λ is the penalty coefficient, and the penalty term is the gradient constraint, is a random distribution between the original sample distribution and the generated sample distribution, and α~U(0,1); Lc is the classification loss, and γ is the weight of Lc; During training, the discriminator aims to maximize the output value when real data is input and minimize the output value when generated data is input. The generator aims to maximize the output value of the discriminator for the generated samples and ensure that the generated features correspond correctly to the category labels. S402, extracting data and its fault category label information from the fused feature vector as a real sample; The generator parameters are fixed, random noise and fault category label information that conform to the normal distribution are extracted from the generator input distribution and input into the generator to generate multi-source heterogeneous data fusion feature fake samples with the same number of real samples; The multi-source heterogeneous data are fused with feature false samples and real samples and their respective fault category label information, and input into the discriminator to calculate the discrimination results and gradient penalty items, and update the discriminator parameters; S403, input the fake samples generated by the generator into the classifier part of the multi-source heterogeneous data fusion fault diagnosis model built in step S3, calculate the classification loss, and feed the classification loss back to the training loss in the generator network to adjust the generator network parameters; S404, fix the discriminator parameters, train the generator, extract random noise and fault category label information that conform to the normal distribution from the generator input distribution, input them into the generator to generate fake samples, calculate the generator loss, and adjust the generator network structure parameters together with the classifier loss calculated in step S403; S405, repeating steps S402 to S404, alternately training the discriminator and the generator, and updating the parameters of the generative adversarial network until the accuracy of the samples generated by the generator reaches the expected level in the classifier; S406, saving the trained generator network, and then inputting random noise and label information into it to generate multi-source heterogeneous data fusion features close to reality, and combining it with the normal state features in the real samples to form a new training set; The training set is used to retrain the fully connected classifier network, and the network diagnosis accuracy is verified using a real multi-source heterogeneous data fusion feature set to perform bearing fault diagnosis under extreme data imbalance.

7. A sliding bearing fault diagnosis system under extreme data imbalance, characterized in that: It includes a data acquisition module and a processing module, wherein the data acquisition module is used to collect vibration acceleration and acoustic emission signals at different positions of the sliding bearing in different states; The input end of the processing module is connected to the output end of the data acquisition module, and the processing module executes the method described in any one of claims 1 to 6 to complete the sliding bearing fault diagnosis under extreme data imbalance.

8. The sliding bearing fault diagnosis system under extreme data imbalance as claimed in claim 7, characterized in that: The data acquisition module includes a multi-channel vibration acceleration and acoustic emission sensor, which is installed at different measuring points of the sliding bearing and is used to obtain multi-source heterogeneous data under different states of the sliding bearing.

Images