Bearing fault diagnosis method and system based on self-supervised migration of physical information

Abstract

The invention relates to a bearing fault diagnosis method and system based on physical information self-supervised migration, and belongs to the technical field of bearing fault diagnosis. Vibration signals of the bearing are collected to obtain a data set, the data set is analyzed to obtain a physical label, a pseudo label and a data label, and the pseudo label is dynamically updated; obtaining data loss and physical loss according to the cross entropy of the pseudo tag and the data tag and the cross entropy of the physical tag and the data tag, and dynamically allocating weights and fusing the two to obtain a loss function; wherein the weight of the physical loss is dynamically adjusted according to the reciprocal of the variance of the data loss; inputting the target data of the to-be-detected bearing into a diagnosis model obtained by training the loss function, and outputting target features; and inputting the features into the physical model and the clustering model to obtain a target physical tag and a target pseudo tag, and aligning the target physical tag and the target pseudo tag to obtain a fault diagnosis result. According to the method, the generalization ability and the interpretability of the diagnosis model are improved, and the method does not depend on a tag.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method and system based on physical information self-supervision migration. Background Art

[0002] Intelligent fault diagnosis is an important research direction in the field of industrial reliability. As a common and critical component in industrial systems, the health status of bearings directly affects the operating efficiency and safety of the entire system. Therefore, intelligent fault diagnosis technology for bearings is not only an important way to improve the overall reliability of industrial systems, but also one of the key points of current technological innovation and breakthroughs.

[0003] In the prior art, the training effect of the transfer learning model depends on the label information of the source domain data. However, in actual industrial monitoring scenarios, the acquisition of accurate and complete data label information often consumes a lot of manpower and material resources, and when the adaptability and dependence of the diagnostic model on the training data and labels are too high, the generalization performance of the diagnostic model will be affected. In addition, the purely data-driven deep learning model relies on data mining and ignores the physical knowledge contained in the fault mechanism, resulting in poor interpretability. The physical model is consistent based on the fault damage mechanism but has poor fitting ability. The presence of noise in the industrial field monitoring data makes the physical model unable to fit all the data, and its diagnostic accuracy is not high. Therefore, the development of existing intelligent fault diagnosis technology is constrained by label annotation, model generalization performance improvement under data change distribution, and model interpretability. Summary of the invention

[0004] To this end, the technical problem to be solved by the present invention is to overcome the problems in the prior art that the performance of the diagnostic model is greatly affected by the labels, the model generalization ability is low and the interpretability is poor, and to provide a bearing fault diagnosis method and system based on self-supervised migration of physical information, which not only does not rely on manual labeling, can achieve true end-to-end diagnosis, but also improves the generalization ability and interpretability of the diagnostic model.

[0005] In a first aspect, in order to solve the above technical problems, the present invention provides a bearing fault diagnosis method based on self-supervisory migration of physical information, comprising:

[0006] Collect and process the vibration signals of the bearing to obtain a data set;

[0007] Processing and analyzing the data set to obtain physical labels, pseudo labels and data labels; wherein, combining the consistency judgment of the physical labels, the pseudo labels and the data labels, the pseudo labels are dynamically updated using similarity measurement and gradient mapping;

[0008] Obtaining data loss and physical loss according to the cross entropy between the pseudo label and the data label, and the cross entropy between the physical label and the data label, respectively;

[0009] Dynamically assigning weights to the data loss and the physical loss to obtain a loss function; wherein the weight of the physical loss is dynamically adjusted according to the inverse of the variance of the data loss;

[0010] Using the loss function to train the neural network to obtain a diagnostic model;

[0011] Inputting target data obtained by collecting the bearing to be inspected into the diagnostic model to obtain target features;

[0012] The target features are input into a physical model and a clustering model to obtain a target physical label and a target pseudo label; the target physical label and the target pseudo label are aligned to obtain a fault diagnosis result of the detected bearing.

[0013] In one embodiment of the present invention, the method for obtaining the physical tag is a physical kinetic model and envelope spectrum analysis, comprising:

[0014] The envelope spectrum of the vibration signal is obtained according to Hilbert transform and fast Fourier transform, and the calculation formula is:

[0015]

[0016] The sum of the amplitudes within k1 to k2 is calculated according to the envelope spectrum, and the calculation formula is:

[0017]

[0018] The amplitude sum calculation formula is used to calculate the amplitude sum of multiple key information frequency bands to obtain the key information frequency band amplitude sum, and the calculation formula is:

[0019]

[0020] The key information frequency band is obtained according to the state characteristic frequency, and the calculation formula of the state characteristic frequency is:

[0021]

[0022] The sum of the amplitudes of the key information frequency bands is normalized, and the calculation formula is:

[0023]

[0024] Where N is the signal length, f i Representative basis f outer 、f inner 、fball and f cage The calculated theoretical fault characteristic frequency value of the i-th type bearing state, Δ represents the 1 / 2 width of the key information frequency band, βf i represents the β-fold frequency of the characteristic frequency, Am i represents the sum of the amplitudes of the fault characteristic frequency and its multiples in the information band under the i-th bearing state, L is the total number of basic state categories of the bearing signal, R q,i is the ratio of the sum of the amplitudes of the qth sample in state i to the sum of the amplitudes of all states, n is the sample point index of the signal, is an imaginary unit, z is the number of rolling elements, f is the speed frequency, d is the rolling element diameter, D is the raceway pitch diameter, and α is the bearing contact angle.

[0025] In one embodiment of the present invention, the loss function includes the data loss and the physical loss, and the calculation formula of the data loss is:

[0026]

[0027] The calculation formula of the physical loss is:

[0028]

[0029] The calculation formula of the loss function is:

[0030] loss=(1-λ)*loss data +λ*loss phy ;

[0031] Among them, y (q) The data-driven prediction label for the qth sample, is the unsupervised clustering pseudo label of the qth sample, Predict the physical driver label for the qth sample, y phy =[R q,1 ,R q,2 ,...,R q,L ], L is the total number of basic state categories of bearing signals, R q,i is the ratio of the sum of the amplitudes of the qth sample in state i to the sum of the amplitudes of all states, w (m) is the weight matrix of the mth layer, Q is the number of samples, M is the total number of layers of the weight matrix, μ is the regularization coefficient, and λ is the inverse of the variance of the data loss.

[0032] In one embodiment of the present invention, the λ is the inverse of the variance of the data loss, and the calculation formula is:

[0033]

[0034] in, The variance of the data loss over all iterations of the previous epoch for model training.

[0035] In one embodiment of the present invention, combining the consistency judgment of the physical label, the pseudo label and the data label, dynamically updating the pseudo label by using similarity measurement and gradient mapping includes:

[0036] Determining whether the pseudo-label is a low-confidence pseudo-label;

[0037] Performing similarity measurement on the physical label, the pseudo label and the data label, and using a sigmoid function to guide the update amplitude of the low-confidence pseudo label;

[0038] The low confidence pseudo label is dynamically updated and normalized according to the update amplitude, and the calculation formula is:

[0039] H(x)=ω*D1*(1-1 / 1+e -10*(x-0.5) ),

[0040] PL' max =|PL max -H(PL max )|,

[0041] PL' i =PL i +H(PL i )(i=0,1,...,n1-1);

[0042] Among them, ω is the regularization coefficient, D1 is the Euclidean distance of multiple multidimensional coordinates, n1 is the total number of target domain categories, PL max is the current maximum probability value of the pseudo-label, PL i is the current non-maximum probability value of the pseudo-label, PL' max The updated value adjusted for the maximum probability value, PL' i The updated value after adjusting the non-maximum probability value, f(x)=1-1 / 1+e -10*(x-0.5) To improve the sigmoid function.

[0043] In one embodiment of the present invention, the low-confidence pseudo-label is a sample label in which the onehot labels of the physical label, the pseudo-label and the data label are all different.

[0044] In one embodiment of the present invention, the method for acquiring the target pseudo label includes:

[0045] Calculating the similarity between the target features to obtain a similarity matrix;

[0046] According to the similarity matrix, a Laplace matrix is ​​obtained;

[0047] Performing dimensionality reduction processing on the eigenvector of the Laplace matrix to obtain a reduced-dimensional eigenvector;

[0048] Inputting the dimension-reduced feature vector into a fuzzy C-means clustering algorithm for clustering, and obtaining cluster centers using the objective function of the C-means clustering algorithm;

[0049] According to the updating rule of the fuzzy C-means clustering algorithm, the membership matrix and the cluster centers are iteratively updated to obtain the fuzzy clustering result.

[0050] In one embodiment of the present invention, the calculation formula of the similarity matrix is:

[0051] W(v i ,v j )=exp(-||v i -v j || 2 / 2s 2 );

[0052] Among them, W is the similarity matrix, v i and v j are two features in the feature matrix, and s is the scale parameter of the Gaussian kernel function.

[0053] In one embodiment of the present invention, the objective function is:

[0054]

[0055] in, For sample v i Belongs to cluster center c j The membership degree of is the fuzzy parameter, N2 is the total number of samples, 1≤i,j≤N2, and i≠j, i and j are natural numbers.

[0056] In a second aspect, in order to solve the above technical problems, the present invention provides a bearing fault diagnosis system based on self-supervisory migration of physical information, comprising:

[0057] A signal acquisition module is used to collect and process the vibration signal of the bearing to obtain a data set;

[0058] A label acquisition module, used to process and analyze the data set to obtain physical labels, pseudo labels and data labels; wherein, in combination with the consistency judgment of the physical labels, the pseudo labels and the data labels, the pseudo labels are dynamically updated using similarity measurement and gradient mapping;

[0059] A loss acquisition module, used to obtain data loss and physical loss according to the cross entropy between the pseudo label and the data label, and the cross entropy between the physical label and the data label, respectively;

[0060] A weight allocation module, used to dynamically allocate weights to the data loss and the physical loss to obtain a loss function; wherein the weight of the physical loss is dynamically adjusted according to the inverse of the variance of the data loss;

[0061] A model training module, used to train the neural network using the loss function to obtain a diagnostic model;

[0062] A target extraction module is used to input target data into the diagnostic model to obtain target features; the target data is obtained by collecting the bearing to be tested;

[0063] The fault diagnosis module is used to input the target features into the physical model and the clustering model to obtain the target physical label and the target pseudo label; align the target physical label and the target pseudo label to obtain the fault diagnosis result of the detected bearing.

[0064] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0065] The bearing fault diagnosis method and system based on the self-supervised migration of physical information described in the present invention realizes the stability of data loss and the association of embedded physical loss, and the optimal fusion of data information and physical information by dynamically adjusting the weight ratio of data loss and physical loss, thereby enhancing the generalization ability and diagnostic accuracy of the diagnosis model in cross-domain application scenarios. At the same time, the embedding of physical loss can transform domain difference factors into domain invariant factors, which not only enhances the interpretability of the model, but also broadens its generalization boundary. In addition, the idea of ​​self-supervised learning is adopted to realize the training of self-supervised models with uncertainty awareness, providing the model with uncertainty probability pseudo-labels, and as the network training pseudo-labels continuously improve the confidence, this design reduces the model's dependence on source domain data and labels, can save manpower and material resources for label annotation, and thus greatly improves the efficiency and intelligence level of industrial field monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein

[0067] Figure 1 A flowchart of a bearing fault diagnosis method based on self-supervisory migration of physical information in a preferred embodiment of the present invention;

[0068] Figure 2Partial two-dimensional time-frequency diagram of the normal state and faulty state of the bearing in the preferred embodiment of the present invention;

[0069] Figure 3 A flow chart of physical label generation in a preferred embodiment of the present invention;

[0070] Figure 4 This is a flowchart of dynamic updating of pseudo labels in a preferred embodiment of the present invention;

[0071] Figure 5 It is a schematic diagram of the alignment of physical labels and cluster pseudo labels in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0072] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0073] Reference Figure 1 As shown, an embodiment of the present invention provides a bearing fault diagnosis method based on physical information self-supervision migration, including but not limited to the following steps:

[0074] Collect and process the vibration signals of the bearing to obtain a data set;

[0075] The data set is processed and analyzed to obtain physical labels, pseudo labels and data labels; wherein, the pseudo labels are dynamically updated by using similarity measurement and gradient mapping in combination with the consistency judgment of the physical labels, pseudo labels and data labels;

[0076] The data loss and physical loss are obtained according to the cross entropy between the pseudo label and the data label, and the cross entropy between the physical label and the data label respectively;

[0077] Dynamically assign weights to data loss and physical loss to obtain a loss function; the weight of physical loss is dynamically adjusted according to the inverse of the variance of data loss;

[0078] The loss function is used to train the neural network to obtain a diagnostic model;

[0079] Inputting the target data obtained by collecting the bearing to be inspected into the diagnosis model to obtain the target features;

[0080] The target features are input into the physical model and the clustering model to obtain the target physical label and the target pseudo label; the target physical label and the target pseudo label are aligned to obtain the fault diagnosis result of the detected bearing.

[0081] The bearing fault diagnosis method based on self-supervised migration of physical information described in the present invention realizes the stability of data loss and the association of physical loss embedding by dynamically adjusting the weight ratio of data loss and physical loss, thereby realizing the optimal fusion of data information and physical information, and enhancing the generalization ability and diagnostic accuracy of the diagnosis model in cross-domain application scenarios. At the same time, the physical model contains domain difference factors such as working conditions. The embedding of physical loss can transform domain difference factors into domain invariant factors, which not only enhances the interpretability of the model, but also broadens its generalization boundary. In addition, the use of self-supervised learning training methods improves the robustness of the model and reduces the manpower and resource consumption required for source domain data labeling, thereby greatly improving the efficiency and intelligence level of industrial field monitoring.

[0082] Specifically, the method for collecting bearing vibration signals can be determined according to the specific application environment and requirements. In this embodiment, an acceleration sensor is preferably used to collect bearing vibration signals, and a sliding window is used for data segmentation. The segmented signal is subjected to continuous wavelet transform to obtain a two-dimensional time-frequency diagram of the signal. The bearing data set (which may be referred to as the CWRU data set) provided by Case Western Reserve University (CWRU) is preferably used as an example. This data set uses electrospark machining (EDM) technology to artificially create faults on motor bearings. The fault fault diameter range is controlled between 0.007 inches and 0.040 inches, and is distributed in the key parts of the inner raceway, rolling element (i.e., ball) and outer raceway of the bearing. These bearings with preset faults are reinstalled into the test motor to simulate the actual working environment. In the process of gradually increasing the motor load from 0 to 3 horsepower (while keeping the motor speed in a stable range of 1720 to 1797RPM), the bearing vibration data is fully recorded. This data set contains drive end bearing and fan end bearing data sets. In this embodiment, the drive end bearing data set is preferably selected, and the sampling frequency is 12 khz. Three sub-data sets of different working conditions are constructed, which can be referred to Table 1. Taking the 1797 rpm working condition data in the CWRU data set as an example, refer to Figure 2 , and obtain the partial two-dimensional time-frequency diagram of the normal state and the faulty state of the bearing. In this embodiment, six groups of migration experiments are constructed to verify the method, including: C1→C2, C1→C3, C2→C1, C2→C3, C3→C1, C3→C2.

[0083] Table 1 Working conditions of sub-datasets

[0084]

[0085] Combining the physical dynamics model and envelope spectrum analysis to obtain the physical labels of the unlabeled data set can, on the one hand, provide the clustering model and deep learning model with prior information about the label space, and on the other hand, construct the physical loss and embed it into the model training process to improve its interpretability and generalization. Figure 3 , the specific steps include:

[0086] In step S21, the envelope spectrum of the vibration signal is obtained according to the Hilbert transform and the fast Fourier transform (FFT), and the calculation formula is:

[0087]

[0088] Where N is the signal length, n is the sample point index of the signal, Is an imaginary unit.

[0089] In step S22, the sum of the amplitudes within k1 to k2 is calculated according to the envelope spectrum, and the calculation formula is:

[0090]

[0091] In step S23, the amplitude sum calculation formula is used to calculate the amplitude sum of multiple potential key information frequency bands to obtain the key information frequency band amplitude sum, and the calculation formula is:

[0092]

[0093] Multiple potential key information frequency bands of the signal can be located according to the state characteristic frequency of the bearing, and the center of the key information frequency band is the state characteristic frequency and its multiples. The state characteristic frequency calculation formula is:

[0094]

[0095]

[0096] Among them, f i Representative basis f outer 、f inner 、f ball and f cage The calculated theoretical fault characteristic frequency value of the i-th type bearing state, Δ represents the 1 / 2 width of the key information frequency band, βf i Indicates the characteristic frequency β times, Am i It represents the sum of the amplitudes of the information frequency band where the fault characteristic frequency and its multiples are located under the i-th type of bearing state, z is the number of rolling elements, f is the rotation frequency, d is the rolling element diameter, D is the raceway pitch diameter, and α is the bearing contact angle.

[0097] In step S24, the sum of the amplitudes of the key information frequency bands is normalized, and the calculation formula is:

[0098]

[0099] Where L is the total number of basic status categories of bearing signals, R q,i is the ratio of the sum of the amplitudes of the qth sample in state i to the sum of the amplitudes of all states. Through normalization, the amplitude ratio can be obtained to characterize the probability of the sample belonging to each state, and then the physical label can be obtained.

[0100] After obtaining the physical labels of the data set, the data labels of the data set are obtained based on the AlexNet deep learning model, and the pseudo labels of the data set are obtained based on spectral clustering. Since pseudo labels are often accompanied by noise problems, effective noise reduction processing becomes the key. This example introduces a strategy for dynamically updating pseudo labels to improve confidence. Specifically, low-confidence pseudo labels are dynamically updated to increase the probability of confidence improvement. Combining consistency judgment under multiple labels, similarity measurement and gradient mapping to dynamically update pseudo labels, please refer to Figure 4 , the specific steps include:

[0101] In step S41, a consistency check is performed on multiple prediction results (ie, physical labels, data labels, and pseudo labels). The labels of samples whose onehot labels of multiple prediction results are all different are considered to be low-confidence pseudo labels.

[0102] In step S42, a plurality of prediction results are measured for similarity. The greater the similarity of the probability labels of the plurality of prediction results, the smaller the update amplitude of the pseudo label, and vice versa, the larger the update amplitude of the pseudo label. In this embodiment, the Euclidean distance is preferably used as the similarity measurement criterion. Since the smaller the Euclidean distance value, the greater the similarity, the Euclidean distance and the pseudo label update amplitude are positively correlated.

[0103] Specifically, suppose there are N1 1*n1-dimensional labels, which are regarded as N1 n1-dimensional coordinates Calculate the Euclidean distance between two pairs. The calculation formula is:

[0104]

[0105] Among them, 1≤i,j≤N1 and i≠j, i and j are natural numbers.

[0106] The Euclidean distance calculation formula for multiple multidimensional coordinates is:

[0107]

[0108] In step S43, in order to fit the results of the unsupervised clustering model itself, the pseudo-label is prevented from being updated in the direction of lower confidence. It is expected that when the current maximum value is large, the update amplitude is small. Since the sum of the probability values ​​of the pseudo-probability label adds up to 1. Each probability value is less than or equal to 1. Therefore, the sigmoid function is very suitable for guiding the update amplitude. The sigmoid function is improved to have a greater convergence trend in the interval [0,1]. At the same time, the newly constructed sigmoid-like function can achieve the effect of obtaining a smaller output with a larger input.

[0109] In step S44, the update rule of the pseudo-label is to reduce the current maximum value in the pseudo-label to a certain extent, and increase other values ​​to a corresponding extent. The calculation formula for the dynamic update amplitude of the pseudo-label is:

[0110] H(x)=ω*D1*(1-1 / 1+e -10*(x-0.5) ),

[0111] The current maximum probability value PL of the pseudo label max The calculation formula for adjustment update is:

[0112] PL' max =|PL max -H(PL max )|,

[0113] The remaining values ​​PL i (i.e. the current non-maximum probability value of the pseudo-label) The calculation formula for adjusting and updating is:

[0114] PL' i =PL i +H(PL i )(i=0,1,...,n1-1),

[0115] Among them, ω is the regularization coefficient, D1 is the Euclidean distance of multiple multidimensional coordinates, n1 is the total number of target domain categories, PL max is the maximum probability value of the pseudo-label at present, PL i is the current non-maximum probability value of the pseudo-label, PL' max Adjust the updated value for the maximum probability value, PL' i Adjust the updated value for the non-maximum probability value, f(x) = 1-1 / 1+e -10*(x-0.5) It is an improved sigmoid function.

[0116] In step S45, each probability value of the adjusted pseudo-label is normalized, and the probability value of the clustering metric pseudo-label is updated and replaced.

[0117] The pseudo-label dynamic update strategy can provide the model with uncertain but gradually increasing pseudo-labels for progressive self-supervised learning, thereby improving the model's ability to cope with data of different distributions and thus improving the diagnostic performance of the model's feature extraction capability.

[0118] In order to optimize the overall performance of the diagnostic model, the data loss of the diagnostic model is obtained according to the cross entropy of the pseudo label and the data label, and the physical loss of the model is obtained according to the cross entropy of the physical label and the data label. Based on the data loss and physical loss, the deep learning network loss function is reconstructed. The specific steps include:

[0119] In step S31, the data loss of the deep neural network is calculated using the following formula:

[0120]

[0121] The physical loss of the deep neural network is calculated using the following formula:

[0122]

[0123] Among them, loss data is the data loss, loss phy is the physical loss, y (q) The data-driven prediction label for the qth sample, is the unsupervised clustering pseudo label of the qth sample, Predict the physical driver label for the qth sample, w (m) is the weight matrix of the mth layer, Q is the number of samples, M is the total number of layers of the weight matrix, μ is the regularization coefficient, and y phy =[R q,1 ,R q,2 ,...,R q,L ].

[0124] In step S32, the weights of physical loss and data loss in the total loss function are dynamically allocated to obtain the optimal function of the model. Specifically, the inverse of the data loss variance is used as the weight of the physical loss to achieve the association between the stability of the data loss and the embedding of the physical loss, thereby achieving the optimal fusion of data information and physical information, and promoting the balance between the accuracy and generalization performance of the model. The calculation formula of the deep learning network loss function (i.e., the total loss function) is:

[0125] loss=(1-λ)*loss data +λ*loss phy ;

[0126] The weight calculation formula of physical loss is:

[0127]

[0128] Among them, λ is the inverse of the variance of data loss, that is, the weight of physical loss; The variance of the data loss for all iterations in the previous cycle (i.e. the previous epoch) of model training.

[0129] The weight of physical loss is assigned based on the method of quantifying the variance of data loss to achieve effective association between data loss and physical loss. The variance of data loss is used to constrain the embedding of physical loss. The amount of embedded physical loss will affect the stability of data loss. By dynamically assigning the weights of data loss function and physical loss function, the generalization performance of the model is improved as much as possible while ensuring the diagnostic accuracy of the model, achieving the optimal fusion drive of data-physics.

[0130] In view of the many advantages of the back propagation method in neural network training, such as automated processing, efficient optimization, wide applicability, and good support for complex network structures, the embodiment of the present invention preferably uses the back propagation method to train the neural network. Specifically, on the source domain data set, the loss between the network prediction result and the true label is calculated iteratively, and the weight of the network is updated by using the back propagation method to minimize the loss, thereby obtaining a trained diagnostic model.

[0131] After using the feature extractor trained with the source domain dataset (i.e., the feature extractor of the diagnostic model), the feature extractor is used to extract features from the target domain dataset, and the diagnosis of the target domain task is finally achieved by combining the physical model and the clustering model. The specific steps include:

[0132] In step one, in the target domain diagnosis stage, the classification identifier in the source domain model is removed, and a feature extractor with good generalization performance is used to extract target features from target domain samples.

[0133] In step 2, the similarity between the target features is calculated. The Gaussian kernel function calculates the similarity between features with nonlinear mapping capability, smoothness and robustness, flexibility in parameter adjustment, wide applicability and high computational efficiency. Therefore, the embodiment of the present invention preferably uses the Gaussian kernel function to calculate the similarity, and the calculation formula is:

[0134] W(v i ,v j )=exp(-||v i -v j || 2 / 2s 2 );

[0135] Among them, W is the similarity matrix composed of the similarity of all samples, the size is N2×N2, N2 is the number of samples, v i and v jare two features in the feature matrix respectively, s is the scale parameter of the Gaussian kernel function, 1≤i,j≤N2, and i≠j, i and j are natural numbers.

[0136] In step 3, the Laplace matrix L is constructed based on the similarity matrix W. The standardized Laplace matrix is ​​used and the calculation formula is:

[0137] L=D2 -1 / 2 *(D2-W)*D2 -1 / 2 ;

[0138] Among them, D2 is the degree matrix, which is a diagonal matrix whose diagonal elements are the sum of the similarities of each sample.

[0139] In step 4, the eigenvector of the Laplace matrix L is subjected to dimensionality reduction processing to obtain a reduced eigenvector T. The reduced eigenvector T is input into the fuzzy C-means clustering algorithm for clustering. Fuzzy C-means clustering obtains the cluster center by minimizing the objective function. The objective function formula is:

[0140]

[0141] in, For sample v i Belongs to cluster center c j The membership degree of is a fuzzy parameter (usually greater than 1).

[0142] In step 5, according to the update rule of the fuzzy C-means clustering algorithm, the membership matrix and cluster centers are iteratively updated until the stop condition is met to obtain the final fuzzy clustering result (ie, the target pseudo label). Steps 2 to 5 are methods for obtaining the target pseudo label.

[0143] In step six, the membership of each sample to each cluster is determined, and the sample can be assigned to the cluster with the highest membership value based on the membership value.

[0144] In step seven, the method of obtaining the physical tag of the target data (ie, the target physical tag) is similar to steps S21 to S24 and will not be repeated here.

[0145] In step 8, the physical labels and clustering results of the target data are aligned to obtain the classification results of the target domain data set and realize fault diagnosis. Figure 5 .

[0146] Comparative Example 1

[0147] In order to verify the performance of the bearing fault diagnosis method based on self-supervised migration of physical information described in the present invention in cross-domain diagnosis, this embodiment designed a total of 6 migration diagnosis tasks for experimental verification.

[0148] The embodiment of the present invention compares the diagnostic model with the selected domain adversarial neural network (DANN) method, the causal decoupling neural network (CDN) method, and the multi-domain adversarial coding neural network (MDCC) method, and conducts 6 migration experiments respectively. The obtained accuracy comparison results are shown in Table 2.

[0149] Table 2 Comparison of accuracy of different methods

[0150]

[0151] Referring to Table 2, the method of the present invention has significantly improved the fault diagnosis accuracy compared with the DANN method, the CDN method, and the MDCC method.

[0152] Comparative Example 2

[0153] To further analyze the effectiveness and contribution of each strategy of the method described in the present invention, this embodiment designs an ablation experiment. The specific method of setting up the ablation experiment includes:

[0154] A1: The method described in the present invention uses a supervised approach for training, and the data loss is obtained by the real label and the predicted label.

[0155] A2: The method described in the present invention uses a semi-supervised approach for training, some labeled samples and unlabeled samples are used, and the pseudo labels of the unlabeled samples are generated and updated using the proposed method.

[0156] A3: No physical loss is embedded in the method of the present invention.

[0157] A4: The embedding weight of the physical loss of the method described in the present invention increases linearly.

[0158] A5: The embedding weight of the physical loss in the method described in the present invention increases in a sigmoid function-like trend.

[0159] Experiments were conducted according to five cases A1 to A5, and the experimental results are shown in Table 3. Referring to Table 3, the pseudo-label generation and dynamic update strategy described in the present invention and the data-physics optimization fusion strategy of controlling the physical loss weight with data loss variance both promote the improvement of the generalization performance of the model in cross-domain tasks. The method described in the present invention gives full play to the data mining capabilities of the data-driven model and the consistency advantages of the physical model. Compared with traditional fault diagnosis methods, it can train a diagnostic model with the optimal balance of accuracy and generalization performance in cross-domain diagnosis tasks, thereby improving the cross-domain diagnosis effect.

[0160] Table 3 Comparison of diagnostic accuracy of ablation experiments

[0161]

[0162] Embodiment 2

[0163] Based on the same inventive concept, this embodiment provides a bearing fault diagnosis system based on self-supervisory migration of physical information. The principle of solving the problem is similar to that of the bearing fault diagnosis method based on self-supervisory migration of physical information, and the repeated parts will not be repeated.

[0164] The embodiment of the present invention provides a bearing fault diagnosis system based on physical information self-supervision migration, including:

[0165] A signal acquisition module is used to collect and process the vibration signal of the bearing to obtain a data set;

[0166] The label acquisition module is used to process and analyze the data set to obtain physical labels, pseudo labels and data labels; wherein, the pseudo labels are dynamically updated by combining the consistency judgment of the physical labels, pseudo labels and data labels and using similarity measurement and gradient mapping;

[0167] A loss acquisition module, used to obtain data loss and physical loss according to the cross entropy between pseudo labels and data labels, and the cross entropy between physical labels and data labels, respectively;

[0168] The weight allocation module is used to dynamically allocate weights to data loss and physical loss to obtain a loss function; the weight of the physical loss is dynamically adjusted according to the inverse of the variance of the data loss;

[0169] A model training module is used to train the neural network using a loss function to obtain a diagnostic model;

[0170] The target extraction module is used to input the target data into the diagnostic model to obtain the target features; the target data is obtained by collecting the bearing to be tested;

[0171] The fault diagnosis module is used to input the target features into the physical model and the clustering model to obtain the target physical label and the target pseudo label; align the target physical label and the target pseudo label to obtain the fault diagnosis result of the detected bearing.

[0172] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0173] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0174] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0175] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0176] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.

Claims

1. A bearing fault diagnosis method based on self-supervised migration of physical information, characterized in that: include: Collect and process the vibration signals of the bearing to obtain a data set; Processing and analyzing the data set to obtain physical labels, pseudo labels and data labels; wherein, combining the consistency judgment of the physical labels, the pseudo labels and the data labels, the pseudo labels are dynamically updated using similarity measurement and gradient mapping; Obtaining data loss and physical loss according to the cross entropy between the pseudo label and the data label, and the cross entropy between the physical label and the data label, respectively; Dynamically assigning weights to the data loss and the physical loss to obtain a loss function; wherein the weight of the physical loss is dynamically adjusted according to the inverse of the variance of the data loss; Using the loss function to train the neural network to obtain a diagnostic model; Inputting target data obtained by collecting the bearing to be inspected into the diagnostic model to obtain target features; The target features are input into a physical model and a clustering model to obtain a target physical label and a target pseudo label; the target physical label and the target pseudo label are aligned to obtain a fault diagnosis result of the detected bearing.

2. A bearing fault diagnosis method based on physical information self-supervision migration according to claim 1, characterized in that: The method for obtaining the physical label is a physical kinetic model and envelope spectrum analysis, including: The envelope spectrum of the vibration signal is obtained according to Hilbert transform and fast Fourier transform, and the calculation formula is: The sum of the amplitudes within k1 to k2 is calculated according to the envelope spectrum, and the calculation formula is: The amplitude sum calculation formula is used to calculate the amplitude sum of multiple key information frequency bands to obtain the key information frequency band amplitude sum, and the calculation formula is: The key information frequency band is obtained according to the state characteristic frequency, and the calculation formula of the state characteristic frequency is: The sum of the amplitudes of the key information frequency bands is normalized, and the calculation formula is: Where N is the signal length, f i Representative basis f outer 、f inner 、f ball and f cage The calculated theoretical fault characteristic frequency value of the i-th type bearing state, Δ represents the 1 / 2 width of the key information frequency band, βf i represents the β-fold frequency of the characteristic frequency, Am i represents the sum of the amplitudes of the fault characteristic frequency and its multiples in the information band under the i-th bearing state, L is the total number of basic state categories of the bearing signal, R q,i is the ratio of the sum of the amplitudes of the qth sample in state i to the sum of the amplitudes of all states, n is the sample point index of the signal, is an imaginary unit, z is the number of rolling elements, f is the speed frequency, d is the rolling element diameter, D is the raceway pitch diameter, and α is the bearing contact angle.

3. A bearing fault diagnosis method based on physical information self-supervision migration according to claim 1, characterized in that: The loss function includes the data loss and the physical loss, and the calculation formula of the data loss is: The calculation formula of the physical loss is: The calculation formula of the loss function is: loss=(1-λ)*loss data +λ*loss phy ; Among them, y (q) The data-driven prediction label for the qth sample, is the unsupervised clustering pseudo label of the qth sample, Predict the physical driver label for the qth sample, y phy =[R q,1 ,R q,2 ,...,R q,L ], L is the total number of basic state categories of bearing signals, R q,i is the ratio of the sum of the amplitudes of the qth sample in state i to the sum of the amplitudes of all states, w (m) is the weight matrix of the mth layer, Q is the number of samples, M is the total number of layers of the weight matrix, μ is the regularization coefficient, and λ is the inverse of the variance of the data loss.

4. A bearing fault diagnosis method based on physical information self-supervision migration according to claim 3, characterized in that: The λ is the inverse of the variance of the data loss, and the calculation formula is: in, The variance of the data loss over all iterations of the previous epoch for model training.

5. The bearing fault diagnosis method based on physical information self-supervision migration according to claim 1 is characterized in that: Combining the consistency judgment of the physical label, the pseudo label and the data label, dynamically updating the pseudo label by using similarity measurement and gradient mapping, including: Determining whether the pseudo-label is a low-confidence pseudo-label; Performing similarity measurement on the physical label, the pseudo label and the data label, and using a sigmoid function to guide the update amplitude of the low-confidence pseudo label; The low confidence pseudo label is dynamically updated and normalized according to the update amplitude, and the calculation formula is: H(x)=ω*D1*(1-1 / 1+e -10*(x-0.5) ), PL' max =|PL max -H(PL max )|, PL' i =PL i +H(PL i )(i=0,1,...,n1-1); Among them, ω is the regularization coefficient, D1 is the Euclidean distance of multiple multidimensional coordinates, n1 is the total number of target domain categories, PL max is the current maximum probability value of the pseudo-label, PL i is the current non-maximum probability value of the pseudo-label, PL' max The updated value adjusted for the maximum probability value, PL' i The updated value is adjusted for the non-maximum probability value, f(x)=1-1 / 1+e -10*(x-0.5) To improve the sigmoid function.

6. A bearing fault diagnosis method based on physical information self-supervision migration according to claim 5, characterized in that: The low-confidence pseudo-label is a sample label whose one-hot labels of the physical label, the pseudo-label, and the data label are all different.

7. A bearing fault diagnosis method based on physical information self-supervision migration according to claim 1, characterized in that: The method for obtaining the target pseudo label includes: Calculating the similarity between the target features to obtain a similarity matrix; According to the similarity matrix, a Laplace matrix is ​​obtained; Performing dimensionality reduction processing on the eigenvector of the Laplace matrix to obtain a reduced-dimensional eigenvector; Inputting the dimension-reduced feature vector into a fuzzy C-means clustering algorithm for clustering, and obtaining cluster centers using the objective function of the C-means clustering algorithm; According to the updating rule of the fuzzy C-means clustering algorithm, the membership matrix and the cluster centers are iteratively updated to obtain the fuzzy clustering result.

8. A bearing fault diagnosis method based on physical information self-supervision migration according to claim 7, characterized in that: The calculation formula of the similarity matrix is: W(v i ,v j )=exp(-||v i -v j || 2 / 2s 2 ); Among them, W is the similarity matrix, v i and v j are two features in the feature matrix, and s is the scale parameter of the Gaussian kernel function.

9. A bearing fault diagnosis method based on physical information self-supervision migration according to claim 7, characterized in that: The objective function is: in, For sample v i Belongs to cluster center c j The membership degree of is the fuzzy parameter, N2 is the total number of samples, 1≤i,j≤N2, and i≠j, i and j are natural numbers.

10. A bearing fault diagnosis system based on self-supervised migration of physical information, characterized in that ,include: A signal acquisition module is used to collect and process the vibration signal of the bearing to obtain a data set; A label acquisition module, used to process and analyze the data set to obtain physical labels, pseudo labels and data labels; wherein, in combination with the consistency judgment of the physical labels, the pseudo labels and the data labels, the pseudo labels are dynamically updated using similarity measurement and gradient mapping; A loss acquisition module, used to obtain data loss and physical loss according to the cross entropy between the pseudo label and the data label, and the cross entropy between the physical label and the data label, respectively; A weight allocation module, used to dynamically allocate weights to the data loss and the physical loss to obtain a loss function; wherein the weight of the physical loss is dynamically adjusted according to the inverse of the variance of the data loss; A model training module, used to train the neural network using the loss function to obtain a diagnostic model; A target extraction module, used for inputting target data into the diagnostic model to obtain target features; The target data is obtained by collecting the bearing to be tested; The fault diagnosis module is used to input the target features into the physical model and the clustering model to obtain the target physical label and the target pseudo label; align the target physical label and the target pseudo label to obtain the fault diagnosis result of the detected bearing.

Images