Methods and systems for diagnosing sliding bearing faults under extreme data imbalance

Abstract

This invention belongs to the field of sliding bearing technology, specifically disclosing a method and system for fault diagnosis of sliding bearings under extreme data imbalance. The method collects vibration acceleration and acoustic emission signals at different locations under different states of the sliding bearing, extracts preliminary features from the data feature vectors, and concatenates them to obtain fused features. A multi-source heterogeneous data fusion fault diagnosis model is constructed and decomposed. A multi-source heterogeneous data fusion feature generation model is built and trained based on CWGAN-GP. The CWGAN-GP model generator and discriminator are adjusted adversarially. The features generated by the generator each time are input into the fault diagnosis classification part of the fault diagnosis model to calculate the classifier loss, and fed back to the generator network to adjust its parameters until near-realistic features are generated, outputting the bearing fault diagnosis result. This technical solution fully utilizes data information to generate more valuable and dense feature data to supplement imbalanced samples, achieving fault diagnosis of sliding bearings under extreme data imbalance.

Description

Technical Field

[0001] This invention belongs to the field of sliding bearing technology and relates to a method and system for diagnosing sliding bearing faults under extreme data imbalance. Background Technology

[0002] Sliding bearings, as joints in heavy-duty mechanical transmission systems, are widely used in rotating machinery in industries such as mining, metallurgy, and power. Because they act as rotational supports, a failure can lead to abnormal vibrations, significantly impacting the normal operation and lifespan of the entire equipment. Therefore, condition monitoring and fault diagnosis are essential. In actual engineering operation, the bearings are mostly in a normal state, resulting in abundant normal data but missing fault data, and extreme data imbalances. Therefore, research on fault diagnosis of sliding bearings under unbalanced conditions is necessary.

[0003] Currently, research on multi-source heterogeneous data fusion is mainly divided into data-level fusion, feature-level fusion, and decision-level fusion. Most studies choose the latter two for better accuracy and robustness in certain situations, while ignoring the complexity and computational cost of the fusion model.

[0004] Furthermore, current data-level fusion methods mostly rely on time-domain data for index calculation and feature extraction, neglecting the ability of spectral features to represent bearing condition information. When analyzing data with imbalanced fault categories, the loss function ignores the impact of the ratio of different category labels on model training. When using generative adversarial networks (GANs) to handle extremely imbalanced fault samples, directly generating data often leads to information redundancy and the loss of some detailed information. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for diagnosing sliding bearing faults under extreme data imbalance, thereby improving the quality of data samples and enabling fault diagnosis of sliding bearings under extreme data imbalance.

[0006] To achieve the above objectives, the basic solution of the present invention is: a method for diagnosing sliding bearing faults under extreme data imbalance, comprising the following steps:

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

[0008] S2, perform data preprocessing on the vibration acceleration signal and extract its spectral features to form a vibration acceleration feature vector;

[0009] The acoustic emission signal is preprocessed to calculate its energy characteristics and spectral characteristics, thus obtaining the acoustic emission data feature vector.

[0010] Create data labels for the feature vectors of vibration acceleration and acoustic emission signals, and match the data labels with the corresponding feature vectors;

[0011] S3, based on convolutional neural networks and attention weighting mechanisms, extracts preliminary features from vibration acceleration feature vectors and acoustic emission data feature vectors and concatenates them to obtain fused features;

[0012] Using the fused features as input, a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focus 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. The model parameters and weights of the two parts are frozen and recorded for data extraction and classification.

[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 against each other, 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 reality are generated and the bearing fault diagnosis results are output.

[0016] The working principle and beneficial effects of this basic scheme are as follows: This technical scheme is based on the fusion of multi-source heterogeneous data, fully considering the expressive power of the spectral characteristics of vibration acceleration and acoustic emission data, the impact of class label comparison models under extreme sample imbalance, and the advantages of generating feature data compared to generating the original signal. It achieves the fusion and extraction of features from multi-source heterogeneous data, improving the quality of data samples. By fully utilizing data information, it generates more valuable and denser feature data to supplement imbalanced samples, enabling fault diagnosis under extreme imbalance of sliding bearing data.

[0017] Furthermore, in step S2, the method for preprocessing the vibration acceleration signal and extracting its spectral features to construct the vibration acceleration feature vector is as follows:

[0018] S201 divides the multi-channel vibration acceleration signal into vibration data segments with equal acquisition duration in the time domain, and ensures that each data segment contains data of at least several rotation cycles, and ensures that the spectrum has sufficient frequency resolution.

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

[0020] S203, perform a Fast Fourier Transform on the denoised vibration data to obtain its spectrum, and define the spectral characteristics F of the vibration acceleration signal. vb(n1～n2) Let A be the spectral area within the corresponding frequency range (n1~n2) Hz. (n1～n2) , frequency centroid FC (n1～n2) Mean Square Frequency (MSF) (n1～n2) and frequency variance FV (n1～n2) The combined eigenvectors; where the spectral area is defined as the area of ​​the trapezoid formed by two adjacent spectral lines on the spectrum:

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

[0022] A (n1～n2) = (v1+v2)×(f2-f1) / 2

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

[0024] The eigenvector F of the vibration acceleration signal vb It consists of the spectral amplitude of the signal from 0 to 300 Hz and the spectral characteristics of 300 to 600 Hz, 600 to 900 Hz, and the entire 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 [ ] represents the spectral characteristics of the entire spectrum, FC f MSF f ,FV f The frequency centroid, mean square frequency, and frequency variance of the entire spectrum;

[0027] S204 divides the feature vector of each channel vibration acceleration signal into a model training set, a validation set, and a test set in chronological order, performs max-min normalization on it, and creates feature set labels according to the fault category to which the data belongs.

[0028] The vibration acceleration signal is preprocessed to facilitate subsequent use.

[0029] Furthermore, in step S2, the acoustic emission signal is preprocessed to calculate its energy characteristics and spectral characteristics, thereby obtaining the acoustic emission data feature vector.

[0030] S2201, combine nft acoustic emission data into a sample group, use one-dimensional discrete wavelet transform to obtain the detail coefficients of the num1 layer and the approximation coefficients of the i-th layer of the sample group, and calculate the corresponding energy values. and Calculate the mean energy value of each layer for all samples, and then calculate the total energy value E of the signal group:

[0031]

[0032] Among them, E di E represents the energy value for the detail factor. a4 is the energy value of the approximation coefficient, and nf is the sample code;

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

[0034]

[0035] Where, p di p represents the ratio of the detail factor energy value to the total energy. a4 It is the ratio of the approximate coefficient energy value to the total energy;

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

[0037]

[0038] Energy characteristics E of each sample group f The composition is as follows:

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

[0040] S2202, calculate the spectrum of all samples, remove the spectral components of samples that exceed the maximum frequency, and calculate the spectral average;

[0041] S2203, extract the kHz spectral characteristics of acoustic emission data corresponding to the frequency range (n3~n4) from the spectral averaging results. ae(n3～n4) :

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

[0043] Where, mean (n3～n4) A (n3～n4) FC ((n3～n4) MSF (n3～n4) ,FV (n3～n4) The mean, area, centroid, mean square frequency, and variance of the spectrum within the corresponding frequency range (n3~n4) Hz;

[0044] Construct the feature vector F of the k-th 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 For the characteristics of the full spectrum, F ae0～100 F ae100～200 F ae200～300 F ae300～450 The characteristics are respectively in the ranges of 0–100, 100–200, 200–300, and 300–450 kHz, E f Energy characteristics;

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

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

[0049] S2205 divides 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 performs max-min normalization on them. The label data of the vibration acceleration feature vector is shared with the acoustic emission feature vector. The multi-source heterogeneous data feature sets with different labels are randomly sorted to remove the influence of the label order on the model training. After completion, the multi-source heterogeneous data feature sets are saved.

[0050] Preprocessing the acoustic emission signal facilitates its subsequent use.

[0051] Furthermore, in step S3, preliminary features of the vibration acceleration feature vector and acoustic emission data feature vector are extracted based on convolutional neural networks and attention weighting mechanisms, and then concatenated to obtain fused features. The specific method is as follows:

[0052] S301, a convolutional block is constructed for each vibration acceleration signal. Each convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer. The vibration acceleration feature vectors of multiple channels are summed after preliminary feature extraction by the convolutional block to achieve homogeneous information fusion.

[0053] S302, the fused features are passed through a BN layer, a GELU layer and a max pooling layer, and then input into a one-dimensional convolutional layer and a convolutional block respectively. The residual connection is achieved by adding the projection of the features by the one-dimensional convolutional layer and the output of the convolutional block.

[0054] The summed result is fed into the activation layer and then into the convolutional block and the one-dimensional convolutional layer respectively, and then residually connected again;

[0055] The summation of multiple channel feature maps is used to obtain the final vibration data fusion feature;

[0056] S303, the acoustic emission data feature vectors are summed after passing through a convolutional block and a one-dimensional convolutional layer to achieve residual connection. The summation result is then fed into the convolutional block and the one-dimensional convolutional layer again through the GELU activation layer for residual connection. Finally, the acoustic emission features are extracted through the GELU activation layer.

[0057] S304, based on the squeezing and excitation network, constructs an attention weighted network. The output feature map after the acoustic emission preliminary feature extraction input convolution block has n channels. Global average pooling is performed on the feature map, and the channel average value is extracted as its descriptor to achieve data squeezing.

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

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

[0060] S305: The channel weight vectors obtained by adaptive attention weighting are multiplied by the original feature map channels output by the corresponding convolutional block to perform channel-dimensional weighting, and the feature maps before and after weighting are summed to achieve residual connection.

[0061] The feature map is reduced in dimensionality by using one-dimensional convolutional layers and flattening layers, which reduces the number of model parameters, improves training speed, and unifies the dimensionality of acoustic emission features with vibration data features.

[0062] S306, the vibration acceleration data features and acoustic emission data feature vectors are concatenated to obtain the final fused features.

[0063] By using a feature extraction network, the spectral features of acoustic emission and vibration data are further extracted, enabling the fusion extraction of features from multi-source heterogeneous data and improving the quality of data samples.

[0064] Furthermore, in step S3, the steps for constructing a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focus loss function, using the fused features as input, are as follows:

[0065] S3301, the fused feature vector is input into the classification diagnosis network in the multi-source heterogeneous data fusion fault diagnosis model. The classification diagnosis network includes two sets of random deactivation layers, fully connected layers and activation layers.

[0066] By deleting some neuron nodes through a random deactivation layer, using L1 regularization to control model complexity in the fully connected layer, selecting Gaussian error linear units as the activation function for activation layer 1, and using the Softmax activation function for activation layer 2, the output is a probability estimate of the healthy state.

[0067] S3302, based on the focus loss function, uses the square root of the inverse of the proportion of class labels in the training set as the penalty factor to establish an adaptive focus 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 y represents the predicted probability for category i; i The label for category i; p i This represents the ratio of the number of tags in category i to the total number of tags in all categories.

[0070] The penalty factor is the square root of the inverse of the proportion of each class label in the training samples. The penalty for the classification loss of the majority class is adaptively adjusted according to the degree of class imbalance, so as to reduce the impact of class proportion under extreme data imbalance.

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

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

[0073]

[0074] Where x represents the real data; y represents the label corresponding to that data; and z represents the noise vector, with subscript p.data (z) represents the true data distribution, p z (z) represents the noise distribution in the potential space. 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... For gradient constraints, It 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 its output value when given real data input and minimize its output value when given generated data input. The generator aims to maximize the discriminator's output value for generated samples and ensure that the generated features correspond correctly to the category labels.

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

[0077] With fixed generator parameters, random noise and fault category label information conforming to a 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] Multi-source heterogeneous data is fused with feature fake samples and real samples and their respective fault category label information, and then input into the discriminator to calculate the discrimination result and gradient penalty term, 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 and 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, repeat steps S402 to S404, alternately train the discriminator and generator, update the parameters of the generative adversarial network until the accuracy of the samples generated by the generator in the classifier reaches the expected level.

[0082] S406: Save the generator network after training, then input random noise and label information into it to generate near-realistic multi-source heterogeneous data fusion features, and combine them with normal state features in real samples to form a new training set.

[0083] The fully connected classifier network was retrained using this training set, and the network's diagnostic accuracy was verified by fusing feature sets with real multi-source heterogeneous data to perform bearing fault diagnosis under extreme data imbalance.

[0084] Based on a conditional Wasserstein generative adversarial network with gradient penalty, this method is used to generate fusion features from multi-source heterogeneous data under extreme class imbalance, and to realize sliding bearing fault diagnosis based on the original fusion features and generated features.

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

[0086] The input terminal of the processing module is connected to the output terminal of the data acquisition module. The processing module executes the method described in this invention to complete the fault diagnosis of the sliding bearing under extreme data imbalance.

[0087] This system utilizes data acquisition and processing modules to develop a sliding bearing fault diagnosis method based on multi-source heterogeneous data fusion, adaptive focus loss function, and conditional Wasserstein generative adversarial network in scenarios with imbalanced data categories.

[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 acquire multi-source heterogeneous data of the sliding bearing under different conditions.

[0089] The data acquisition module has a simple structure, making it easy to collect the required information. Attached Figure Description

[0090] Figure 1 This is a flowchart illustrating the sliding bearing fault diagnosis method under extreme data imbalance according to the present invention.

[0091] Figure 2 This is a schematic diagram of the vibration acceleration data feature extraction and fusion process in the sliding bearing fault diagnosis method under extreme data imbalance of the present invention.

[0092] Figure 3 This is a schematic diagram of the acoustic emission data feature extraction and fusion process in the sliding bearing fault diagnosis method under extreme data imbalance of the present invention.

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

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

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

[0096] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0097] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0098] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0099] This invention discloses a method for fault diagnosis of sliding bearings under extreme data imbalance. Based on the fusion of multi-source heterogeneous data, it fully considers the expressive power of the spectral features of vibration acceleration and acoustic emission data, the impact of the class label comparison model under extreme sample imbalance, and the advantages of generating feature data compared with generating the original signal. It generates feature data with higher value density to supplement the imbalanced samples, thereby realizing fault diagnosis of sliding bearings under extreme data imbalance.

[0100] like Figure 1 As shown, the method for diagnosing sliding bearing faults under extreme data imbalance includes the following steps:

[0101] S1 collects vibration acceleration and acoustic emission signals at different locations of the sliding bearing under different conditions (such as normal condition, bearing wear and abrasive wear, etc.). Due to the difficulty in occurrence of fault conditions, the data exhibits extreme imbalance.

[0102] S2, perform data preprocessing on the vibration acceleration signal and extract its spectral features to form a vibration acceleration feature vector;

[0103] The acoustic emission signal is preprocessed to calculate its energy characteristics and spectral characteristics, thus obtaining the acoustic emission data feature vector.

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

[0105] S3, based on convolutional neural networks and attention weighting mechanisms, extracts preliminary features from vibration acceleration feature vectors and acoustic emission data feature vectors and concatenates them to obtain fused features;

[0106] Using the fused features as input, a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focus 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. The model parameters and weights of the two parts are frozen and recorded respectively for subsequent data extraction and classification.

[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 against each other, 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 reality are generated and the bearing fault diagnosis results are output.

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

[0111] S201 divides the multi-channel vibration acceleration signal into vibration data segments with equal acquisition duration in the time domain, and ensures that each data segment contains data of at least several rotation cycles, and ensures that the spectrum has sufficient frequency resolution.

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

[0113] S203, perform a Fast Fourier Transform on the denoised vibration data to obtain its spectrum, and define the spectral characteristics F of the vibration acceleration signal. vb(n1～n2)Let A be the spectral area within the corresponding frequency range (n1~n2) Hz. (n1～n2) , frequency centroid FC (n1～n2) Mean Square Frequency (MSF) (n1～n2) and frequency variance FV (n1～n2) The combined eigenvectors; where the spectral area is defined as the area of ​​the trapezoid formed by two adjacent spectral lines on the spectrum:

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

[0115] A (n1～n2) = (v1+v2)×(f2-f1) / 2

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

[0117] The eigenvector F of the vibration acceleration signal vb It consists of the spectral amplitude of the signal from 0 to 300 Hz and the spectral characteristics of 300 to 600 Hz, 600 to 900 Hz, and the entire 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 [ ] represents the spectral characteristics of the entire spectrum, FC f MSF f ,FV f The frequency centroid, mean square frequency, and frequency variance of the entire spectrum;

[0120] S204 divides the feature vector of each channel vibration acceleration signal into a model training set, a validation set, and a test set in chronological order, performs max-min normalization on it, and creates feature set labels according to the fault category to which the data belongs.

[0121] In a preferred embodiment of the present invention, step S2, which involves preprocessing the acoustic emission signal to calculate its energy characteristics and spectral characteristics to obtain the acoustic emission data feature vector, is as follows:

[0122] S2201, considering that the acoustic emission data sample time is too short to completely represent the fault information, nft acoustic emission data are combined into a sample group. The detail coefficients of the num1 layer and the approximation coefficients of the i-th layer of the sample group are obtained by one-dimensional discrete wavelet transform, and the corresponding energy values ​​are calculated. and Calculate the mean energy value of each layer for all samples, and then calculate the total energy value E of the signal group:

[0123]

[0124] Among them, E di E represents the energy value for the detail factor. a4 is the energy value of the approximation coefficient, and nf is the sample code;

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

[0126]

[0127] Where, p di p represents the ratio of the detail factor energy value to the total energy. a4 It is the ratio of the approximate coefficient energy value to the total energy;

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

[0129]

[0130] Energy characteristics E of each sample group f The composition is as follows:

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

[0132] S2202, calculate the spectrum of all samples, delete the spectral components of samples that exceed the maximum frequency (the spectral amplitude is close to 0 after exceeding this frequency), and calculate the spectral average.

[0133] S2203, extract the kHz spectral characteristics of acoustic emission data corresponding to the frequency range (n3~n4) from the spectral averaging results. ae(n3～n4) :

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

[0135] Where, mean (n3～n4) A (n3～n4) FC ((n3～n4) MSF (n3～n4) ,FV (n3～n4) The mean, area, centroid, mean square frequency, and variance of the spectrum within the corresponding frequency range (n3~n4) Hz;

[0136] Construct the feature vector F of the k-th 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 For the characteristics of the full spectrum, F ae0～100 F ae100～200 F ae200～300 F ae300～450 The characteristics are respectively in the ranges of 0–100, 100–200, 200–300, and 300–450 kHz, E f Energy characteristics;

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

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

[0141] S2205 divides 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 performs max-min normalization on them. The label data of the vibration acceleration feature vector is shared with the acoustic emission feature vector. The multi-source heterogeneous data feature sets with different labels are randomly sorted to remove the influence of the label order on the model training. After completion, the multi-source heterogeneous data feature sets are saved.

[0142] In a preferred embodiment of the present invention, step S3 involves extracting preliminary features from the vibration acceleration feature vector and the acoustic emission data feature vector based on a convolutional neural network and an attention weighting mechanism, and then concatenating them to obtain fused features. The specific method is as follows:

[0143] S301, a convolutional block is constructed for each vibration acceleration signal. Each convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer. Vibration acceleration feature vectors from multiple channels are initially extracted through the convolutional block and then summed to achieve homogeneous information fusion. Figure 2 As shown;

[0144] S302, the fused features are passed through a BN layer, a GELU layer and a max pooling layer, and then input into a one-dimensional convolutional layer and a convolutional block respectively. The residual connection is achieved by adding the projection of the features by the one-dimensional convolutional layer and the output of the convolutional block.

[0145] The summed result is fed into the activation layer and then into the convolutional block and the one-dimensional convolutional layer respectively, and then residually connected again;

[0146] The summation of multiple channel feature maps is used to obtain the final vibration data fusion feature;

[0147] S303, the acoustic emission data feature vectors are summed after passing through a convolutional block and a one-dimensional convolutional layer to achieve residual concatenation. The summation result is then fed back into the convolutional block and the one-dimensional convolutional layer through a GELU activation layer for residual concatenation. Finally, the acoustic emission features are preliminarily extracted through the GELU activation layer, such as... Figure 3 As shown;

[0148] S304, based on squeezing and incentive networks, constructs an attention-weighted network (such as...). Figure 4 As shown in the figure, the output feature map after the acoustic emission feature is initially extracted and input into the convolutional block has n channels. Global average pooling is performed on the feature map, and the average value of each channel is extracted as its descriptor to achieve data compression.

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

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

[0151] S305: The channel weight vectors obtained by adaptive attention weighting are multiplied by the original feature map channels output by the corresponding convolutional block to perform channel-dimensional weighting, and the feature maps before and after weighting are summed to achieve residual connection.

[0152] The feature map is reduced in dimensionality by using one-dimensional convolutional layers and flattening layers, which reduces the number of model parameters, improves training speed, and unifies the dimensionality of acoustic emission features with vibration data features.

[0153] S306, the vibration acceleration data features and acoustic emission data feature vectors are concatenated to obtain the final fused features.

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

[0155] S3301, the fused feature vector is input into the classification diagnosis network in the multi-source heterogeneous data fusion fault diagnosis model. The classification diagnosis network includes two sets of random deactivation layers, fully connected layers and activation layers.

[0156] By randomly deactivating some neurons, connections between hidden units are reduced, improving robustness and mitigating overfitting. L1 regularization is used in the fully connected layers to control model complexity and reduce overfitting. Activation layer 1 selects Gaussian error linear units as the activation function, and activation layer 2 uses the Softmax activation function to output a probability estimate of the healthy state.

[0157] S3302, considering the impact of extreme imbalance in the number of data points across different categories on the loss function, and to avoid excessive loss for categories due to extreme scarcity of samples, an adaptive focus loss function, Loss, is established based on the focus loss function. This function uses the square root of the inverse of the proportion of class labels in the training set as the penalty factor.

[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 y represents the predicted probability for category i; i The label for category i; p i This represents the ratio of the number of tags in category i to the total number of tags in all categories.

[0160] Based on the adaptive focus loss result, the classifier weights are iteratively adjusted until the final bearing state classification result is output. The trained model is then split, and the model parameters and weights of the feature extraction and fusion part and the fault diagnosis and classification part are frozen and recorded separately for subsequent model data extraction and classification.

[0161] In a preferred embodiment of the present invention, step S4 is specifically performed as follows:

[0162] S401, due to the extreme imbalance in bearing fault data, a multi-source heterogeneous data fusion feature generation model based on a Conditional Waserstein Generative Adversarial Network (CWGAN-GP) with gradient penalty is constructed (e.g., Figure 6 As shown), the training objective of this model is:

[0163]

[0164] Where x represents the real data; y represents the label corresponding to that data; and z represents the noise vector, with subscript p. data (z) represents the true data distribution, p z (z) represents the noise distribution in the potential space. G represents 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... For gradient constraints, It 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 its output value when given real data input and minimize its output value when given generated data input. The generator aims to maximize the discriminator's output value for generated samples and ensure that the generated features correspond correctly to the category labels.

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

[0167] With fixed generator parameters, random noise and fault category label information conforming to a 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] Multi-source heterogeneous data is fused with feature fake samples and real samples and their respective fault category label information, and then input into the discriminator to calculate the discrimination result and gradient penalty term, 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 and 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, repeat steps S402 to S404, alternately train the discriminator and generator, update the parameters of the generative adversarial network until the accuracy of the samples generated by the generator in the classifier reaches the expected level.

[0172] S406: Save the generator network after training, then input random noise and label information into it to generate near-realistic multi-source heterogeneous data fusion features, and combine them with normal state features in real samples to form a new training set.

[0173] Retrain the fully connected classifier network using this training set (e.g.) Figure 5 As shown, the accuracy of network diagnosis is verified by fusing feature sets from real multi-source heterogeneous data, and bearing fault diagnosis is performed under extreme data imbalance.

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

[0175] The input terminal of the processing module is electrically connected to the output terminal of the data acquisition module. The processing module executes the method described in this invention to complete the fault diagnosis of the sliding bearing under extreme data imbalance.

[0176] This system utilizes data acquisition and processing modules to develop a sliding bearing fault diagnosis method based on multi-source heterogeneous data fusion, adaptive focus loss function, and conditional Wasserstein generative adversarial network in scenarios with imbalanced data categories.

[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 acquire multi-source heterogeneous data of the sliding bearing under different states.

[0178] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

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

Claims

1. A method for fault diagnosis of a sliding bearing under data extreme imbalance, characterized in that, The method comprises the following steps: S1, collecting vibration acceleration and acoustic emission signals at different positions in different states of the sliding bearing; S2, data preprocessing of the vibration acceleration signals, extracting the frequency spectrum features to form a vibration acceleration feature vector; data preprocessing of the acoustic emission signals, calculating the energy features and the frequency spectrum features to obtain an acoustic emission data feature vector; making data labels of the vibration acceleration and acoustic emission signal data feature vectors, and matching the data labels with the corresponding data feature vectors; S3, based on a convolutional neural network and an attention weighting mechanism, extracting preliminary features of the vibration acceleration feature vector and the acoustic emission data feature vector and splicing to obtain a fusion feature; taking the fusion feature as input, constructing a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focal loss function; According to the adaptive focal loss result, the classifier weight is adjusted in a loop until the final bearing state classification result is output. The trained multi-source heterogeneous data fusion fault diagnosis model is divided into a feature extraction fusion part and a classifier part, and the model parameters and weights of the two parts are respectively frozen and recorded; S4, based on CWGAN-GP, a multi-source heterogeneous data fusion feature generation model, namely a CWGAN-GP model, is constructed, and the fusion feature obtained in step S3 is taken as a real feature to train the CWGAN-GP model; While the generator and the discriminator of the CWGAN-GP model are adjusted in an antagonistic manner, the feature generated by the generator each time is 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, which is fed back to the generator to adjust its parameters until the generated feature is close to the real feature, and the bearing fault diagnosis result is output; 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, and the specific method is as follows: S301, a convolution block is constructed for each vibration acceleration signal, each convolution block comprises 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 realize 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 the one-dimensional convolution layer and a convolution block, and the projection of the one-dimensional convolution layer on the features and the output of the convolution block are added to realize residual connection; the summed result is input into the convolution block and the one-dimensional convolution layer again through the activation layer and residual connection again; the multiple channel feature maps obtained by summing are summed to obtain the final vibration data fusion feature; S303, the acoustic emission data feature vector is respectively input into a convolution block and a one-dimensional convolution layer, and then summed to realize residual connection, and the summed result is input into the convolution block and the one-dimensional convolution layer again through the GELU activation layer to realize residual connection, and then the acoustic emission preliminary extraction features are obtained through the GELU activation layer; S304, based on the squeezing and excitation network, an attention weighted network is constructed, the output feature map of the acoustic emission preliminary feature extraction after the convolution block has n channels, the global average pooling is performed on the feature map, and the channel average value is extracted as the descriptor to realize data squeezing; The mean vector composed of the channel descriptor is input into the attention weighted network for activation, the hidden layer of the attention weighted network has only one neuron, and the activation function is set to a rectified linear unit; The output layer has n neurons, each channel weight is obtained through a Sigmoid function, and an output channel weight vector is obtained; S305, the channel weight vector obtained through the adaptive attention weighting is multiplied by the original feature map channel output by the corresponding convolution block for channel dimension weighting, and the weighted feature map is summed to realize residual connection; The feature map is reduced in dimension through a one-dimensional convolution layer and a flattening layer, the model parameter quantity is reduced to improve the training speed, and the acoustic emission feature and the vibration data feature are unified in dimension; S306, the vibration acceleration data feature and the acoustic emission data feature vector are spliced to obtain the final fusion feature.

2. The data extreme unbalance slide bearing fault diagnostic method of claim 1, wherein, In step S2, the vibration acceleration signal is preprocessed, the frequency spectrum feature is extracted to form a vibration acceleration feature vector, a data label of the vibration acceleration data feature vector is made, and the data label is matched with the corresponding data feature vector. The method is as follows: S201, the multi-channel vibration acceleration signal is divided into vibration data segments with equal collection time length in the time domain, and it is ensured that each data segment contains at least multiple rotation period data; S202, the wavelet threshold denoising algorithm is used to remove the high-frequency noise component of each vibration data segment; S203, the vibration data after denoising is subjected to fast Fourier transform to obtain its frequency spectrum, and the frequency spectrum features F of the vibration acceleration signal are defined vb（n1~n2） The frequency spectrum area A in the corresponding frequency interval (n1~n2) Hz （n1~n2） , the frequency center of gravity FC （n1~n2） , the mean square frequency MSF （n1~n2） , and the frequency variance FV （n1~n2） are combined to form a feature vector; wherein the frequency spectrum area is defined as the trapezoidal area formed by two adjacent spectrum lines on the spectrum diagram: F vb（n1~n2） = [A （n1~n2） , FC （n1~n2） , MSF （n1~n2） , FV （n1~n2） ], A （n1~n2） = (v1 + v2) x (f2 - f1) / 2, wherein, f i is the frequency corresponding to the spectral line i; v i is the amplitude corresponding to the spectral line i; Feature vector F of the vibration acceleration signal vb The feature vector F consists of the spectral amplitudes of the signal from 0 to 300 Hz and the spectral features of 300 to 600 Hz, 600 to 900 Hz and the full spectrum, i.e.: F 𝑣𝑏 = [f 𝑣𝑏0~300 , F 𝑣𝑏300~600 , F 𝑣𝑏600~900 , F 𝑣𝑏𝑓 ], where F vbf = [FC f , MSF f , FV f ] are the spectral features of the full spectrum, FC f , MSF f , FV f are the frequency center of mass, mean square frequency and frequency variance of the full spectrum; S204, the feature vector of each channel vibration acceleration signal is divided into a model training set, a validation set and a test set in time sequence, and is subjected to maximum and minimum normalization, and a feature set label is made according to the fault category to which the data belongs.

3. The data extreme unbalance slide bearing fault diagnostic method of claim 1, wherein, In step S2, the acoustic emission signal is preprocessed, the energy feature and the frequency spectrum feature are calculated, the acoustic emission data feature vector is obtained, the data label of the acoustic emission signal data feature vector is made, and the data label is matched with the corresponding data feature vector. The steps are as follows: S2201, combine nft acoustic emission data groups into a sample group, obtain the num1 layer detail coefficients and the i layer approximate coefficients of the sample group by using one-dimensional discrete wavelet transform, and calculate the corresponding energy values and calculate the mean value of each layer energy value for all samples, and obtain the total energy value E of the signal group: ， wherein, is the energy value of the detail coefficient, is the energy value of the approximation coefficient, nf is the index of the sample; The ratio of the energy mean value of each layer to the total energy E is calculated: ， wherein is the ratio of the detail coefficient energy value to the total energy, is the ratio of the approximation coefficient energy value to the total energy; calculating the energy entropy E for each sample group n : ， Energy feature E of each sample group f is configured to: ； S2202, the spectrum of all samples is calculated, the sample spectrum component exceeding the maximum frequency is deleted, and the spectral average is calculated; S2203, extracting the frequency interval (n3~n4) kHz spectrum feature F corresponding to the acoustic emission data 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） ], wherein mean （n3~n4） , A （n3~n4） , FC （（n3~n4） , MSF （n3~n4） , FV （n3~n4） are the spectral mean, the spectral area, the frequency center of gravity, the mean square frequency and the frequency variance in the corresponding frequency interval (n3~n4) Hz; constructing a feature vector F for the kth acoustic emission sensor signal k 𝑎𝑒 : F k 𝑎𝑒 =[F 𝑎𝑒𝑓 , F 𝑎𝑒0~100 , F 𝑎𝑒100~200 , F 𝑎𝑒200~300 , F 𝑎𝑒300~450 , E 𝑓 ], wherein, F 𝑎𝑒𝑓 is the feature of the full spectrum, F 𝑎𝑒0~100 , F 𝑎𝑒100~200 , F 𝑎𝑒200~300 , F 𝑎𝑒300~450 are the features in the 0~100, 100~200, 200~300 and 300~450 kHz interval respectively, E f is the energy feature; S2204, splice 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, the feature vector of the acoustic emission signal of each channel is divided into a model training set, a validation set and a test set in time sequence, and is subjected to maximum and minimum normalization, the label data of the vibration acceleration feature vector is shared with the acoustic emission feature vector, and the multi-source heterogeneous data feature sets with different labels are randomly sorted to remove the influence of label sequence on model training, and the multi-source heterogeneous data feature sets are saved after completion.

4. The data extreme unbalance slide bearing fault diagnostic method of claim 1, wherein, In step S3, the fusion feature is taken as input to construct a multi-source heterogeneous data fusion fault diagnosis model based on an adaptive focal loss function. The steps are as follows: S3301, the fusion feature vector is input into the classification diagnosis network in the multi-source heterogeneous data fusion fault diagnosis model, and the classification diagnosis network includes two groups of random inactivation layers, fully connected layers and activation layers; Part of the neuron nodes are deleted by random inactivation layer, the model complexity is controlled by L1 regularization in full connection layer, Gaussian error linear unit is selected as the activation function in activation layer 1, and the Softmax activation function is used in activation layer 2 to output the probability estimation of the health state; S3302, based on the focal loss function, the square root of the reciprocal of the category label proportion 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 classes; λ is a focusing parameter, which is adjusted according to specific application; δ i is the predicted probability of class i; y i is the label of class i; p i is the ratio of the number of class i labels to the total number of labels of all classes.

5. The data extreme imbalance sliding bearing fault diagnostic method of claim 1, wherein, The specific steps of step S4 are as follows: S401, since the bearing fault data is extremely unbalanced, a multi-source heterogeneous data fusion feature generation model based on a conditional Wasserstein generative adversarial network with gradient penalty is constructed, and the training target of the model is: ， where x is the real data; y is the label corresponding to the data; z is the noise vector, subscript p data (z) is the real data distribution, p z (z) is the noise distribution in the latent space, p 𝑥̂ (𝑥̂) is the interpolation distribution between the generated sample and the real sample; D(·) is the discriminator output; G(·) is the generator output; λ is the penalty coefficient, the penalty term is the gradient constraint, 𝑥̂ = 𝑥 + 𝛼 ∙ (𝐺(𝑧) − 𝑥), 𝑥̂ 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 the real data is input and minimize the output value when the generated data is input, and the generator aims to maximize the output value of the discriminator for the generated sample and correctly correspond the generated feature and the category label; S402, the data and its fault category label information are extracted from the fusion feature vector as real samples; The generator parameters are fixed, random noise conforming to the normal distribution and fault category label information are extracted from the generator input distribution and input into the generator to generate a plurality of source heterogeneous data fusion feature fake samples with the same number of real samples; The multi-source heterogeneous data fusion feature fake samples and real samples and their respective fault category label information are input into the discriminator to calculate the discrimination result and the gradient penalty term, and the discriminator parameters are updated; S403, the fake samples generated by the generator are input into the classifier part of the multi-source heterogeneous data fusion fault diagnosis model built in step S3 to calculate the classification loss, and the classification loss is fed back to the training loss in the generator network to adjust the generator network parameters; S404, the discriminator parameters are fixed, the generator is trained, random noise conforming to the normal distribution and fault category label information are extracted from the generator input distribution and input into the generator to generate fake samples, and the generator loss is calculated together with the classifier loss calculated in step S403 to adjust the generator network structure parameters; S405, steps S402-S404 are repeated, and the training of the discriminator and the generator is alternately performed to update the generative adversarial network parameters until the accuracy of the samples generated by the generator in the classifier reaches the expectation; S406, the generator network after training is saved, and then random noise and label information are input into it to generate multi-source heterogeneous data fusion features close to the real ones, which are combined with the normal state features in the real samples to form a new training set; The fully connected classifier network is retrained using the training set, and the network diagnosis accuracy is verified using the real multi-source heterogeneous data fusion feature set to perform bearing fault diagnosis under data extreme imbalance.

6. A data extreme unbalance sliding bearing fault diagnosis system characterized by, The data acquisition module is used for acquiring 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 with the output end of the data acquisition module, the processing module executes the method in any one of claims 1-5, and the sliding bearing fault diagnosis under data extreme imbalance is completed.

7. The data extreme imbalance sliding bearing fault diagnostic system of claim 6, wherein, The data acquisition module comprises a multi-channel vibration acceleration and acoustic emission sensor, which is installed at different measuring points of the sliding bearing and used for acquiring multi-source heterogeneous data under different states of the sliding bearing.

Images