Bearing fault classification system and method based on multi-scale domain adaptive network

Abstract

The application provides a bearing fault classification system and method based on a multi-scale field adaptive network. In view of the situation that label information is missing and data distribution is deviated in actual industry, field knowledge is combined into the model, the existing label data resources can be fully utilized, the multi-device multi-condition label data knowledge can be migrated to the device to be diagnosed when the device data distribution is deviated due to the working environment, the problem of label missing in real industrial production is greatly made up, the use value of the label is greatly improved, and the efficiency of bearing fault diagnosis is also improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of automation, and relates to bearing fault diagnosis, in particular to a bearing fault classification system and method based on a multi-scale domain adaptive network. BACKGROUND

[0002] With the continuous development of modern scientific and technological productivity and the continuous progress of information technology, the demand for low-cost, high-quality products and production safety in modern industry is increasing. The maintenance of rotating equipment in industrial production is rapidly shifting from preventive maintenance to real-time monitoring and intelligent fault diagnosis based on the state of rotating equipment. Rotating machinery is one of the most widely used types of rotating equipment and is very important in industrial production, playing an irreplaceable role. At the same time, rolling bearings are one of the most important components of rotating machinery, which have an important influence on the overall performance. In actual industry, production equipment is often in harsh working conditions, and rolling bearing failures occur frequently, resulting in unexpected downtime and causing huge economic losses. Therefore, based on the state monitoring and fault maintenance of bearings in rotating machinery, the number of unexpected equipment downtime can be effectively reduced, which is of great significance for the safe and efficient operation of rotating equipment and the saving of resource costs.

[0003] In actual industrial production, the main methods for diagnosing the faults of rotating bearings include mechanism-based fault diagnosis and data-driven machine learning intelligent algorithm models:

[0004] 1) Mechanism-based fault diagnosis method. At present, in actual fault diagnosis analysis, the main method is to analyze the vibration data collected by an acceleration sensor, use time-frequency domain analysis methods such as short-time Fourier transform and wavelet analysis, and use the time domain waveform, frequency spectrum data and envelope spectrum fields to diagnose the faults of rotating equipment bearings by engineering technical diagnosis experts. Specifically, based on frequency spectrum and envelope spectrum image analysis, using frequency multiplication and center frequency and sideband characteristics, combined with rotation frequency information and the inherent properties of the equipment, the fault frequency of the component is determined to determine the type of fault, such as cage fault, rolling element fault and inner and outer ring fault; based on frequency spectrum image analysis, using energy area and the number of harmonics to determine whether the bearing has a bearing looseness fault.

[0005] 2) Data-driven intelligent algorithm model. With the development of information technology, intelligent algorithm classification model is introduced into equipment fault diagnosis. For some traditional machine learning algorithms, such as envelope neural network ANN, support vector machine SVM and k-nearest neighbor classification method, vibration signal data is input into the model for supervised training, and the trained model is applied to test data to obtain the classification of faults. In addition to shallow models, deep learning includes convolutional neural network, recurrent neural network and ResNet network, which can effectively process high-dimensional nonlinear data, and can effectively process original time series data and image video data.

[0006] In actual production process, the environment of equipment is complex and changeable. With the increase of the number of nursing equipment, the accumulation of production data, and the explosive growth of online data, the effective fault diagnosis model is brought great challenges.

[0007] For mechanism-based fault diagnosis, most of them rely on artificial experts to analyze and diagnose faults. Although mechanism-based can well analyze the cause and effect relationship of the diagnosis result and provide reasonable analysis support for the diagnosis result, the whole process is time-consuming and laborious, and in the face of a large number of equipment, it is basically impossible to complete the accurate mechanism diagnosis of a large number of rotating equipment under the interference of complex and changeable environmental factors. For the actual needs of enterprises, it is more limited.

[0008] Data-driven intelligent fault diagnosis can effectively process high-dimensional nonlinear big data and extract features from complex environments for effective classification. However, the intelligent method model is limited in actual equipment fault diagnosis. The construction of data-driven fault diagnosis method needs to meet the following two conditions: 1) have sufficient labeled data containing fault information; 2) training data and test data need to come from the same probability distribution. However, it is difficult to meet such conditions in actual industrial production. In real production, unexpected equipment downtime can cause significant losses, so equipment failure is not allowed, and fault is a long cracking process, so it is difficult for the equipment itself to record the accurate time of fault occurrence, and relying on expert fault time marking is time-consuming and laborious. On the other hand, the intelligent algorithm model constructed based on the effective data set collected from a certain device often needs to be applied to other devices or the same device under different working conditions. In such cases, the training data and application data often have distribution offset problems, which can cause the performance of intelligent diagnosis method to decrease rapidly, the scope of application is limited, and the utilization rate of a small amount of valuable labeled case information is also very limited.

[0009] Although the proposed domain adaptation algorithms for data domain shift are developing rapidly, the research applied to the field of fault diagnosis is still very limited and does not combine the domain knowledge well. Domain knowledge, including machine speed, fault bearing geometric parameters, fault degree, work load and other factors, will affect the data shift. Most algorithms directly use the original data as input, extract features through intelligent algorithm "black box" learning, and do not use domain knowledge to guide the model correctly. SUMMARY

[0010] The present application is aimed at the situation of label information missing and data distribution shift in actual industry, combines domain knowledge into the model, proposes a bearing fault classification system and method based on multi-scale domain adaptation network, can make full use of existing label data resources, can still migrate multi-device and multi-condition label data knowledge to the equipment to be diagnosed when the equipment data distribution is shifted due to the change of working environment, greatly makes up for the problem of label missing in real industrial production, greatly improves the use value of the label, and also improves the efficiency of bearing fault diagnosis.

[0011] The present application is realized by the following technical solutions:

[0012] The bearing fault classification system based on multi-scale domain adaptation network comprises a data signal processing module, a shared network module, a fault diagnosis classification module, a domain discriminator module and a distribution difference measurement module.

[0013] The data signal processing module pre-processes the original vibration data in the database to obtain an output; the database is a database for storing source domain and target domain.

[0014] The shared network module uses the output of the data signal processing module as input, uses different size convolution kernels to process the input in different scales, and combines the features in different scales on the full connection layer to output.

[0015] The fault diagnosis classification module uses the output of the shared network module as input, uses the data of the source domain to supervise the training of the fault diagnosis classifier, outputs the classification result, and optimizes the output loss L of the fault diagnosis classifier module c ; the trained fault diagnosis classifier is used for data classification of the target domain.

[0016] The domain discriminator module uses the output of the shared network module as input, maximizes the loss function to pull in the distribution difference between the source domain and the target domain, and obtains an output; the output loss L of the domain discriminator module is optimized d .

[0017] The distribution difference measurement module takes the output of the shared network module as input, obtains the maximum mean difference between the source domain and the target domain by minimizing the loss function, and outputs; the output loss D of the distribution difference measurement module is optimized 2 .

[0018] Preferably, the preprocessing is sequentially performed by normalization processing, feature data extraction, angular domain resampling, and envelope spectrum calculation.

[0019] The bearing fault classification method based on the multi-scale domain adaptive network is based on the system, including step 1, the training phase:

[0020] Step 1.1, preprocessing the original vibration data in the source domain and the target domain in the database;

[0021] Step 1.2, using different size convolution kernels to perform different scale convolution processing on the output of step 1.1 to extract features, and then full connecting the features obtained in different scales;

[0022] Step 1.3, using the output of step 1.2 from the source domain data, the fault diagnosis classifier is supervised trained;

[0023] Step 1.4, using the output of step 1.2 from the source domain and the target domain data, the distribution difference between different domains is pulled in by maximizing the loss function;

[0024] Step 1.5, using the output of step 1.2 from the source domain and the target domain data, the maximum mean difference between the source domain and the target domain is obtained by minimizing the loss function;

[0025] Step 1.6, using the output of step 1.3, step 1.4 and step 1.5, calculating and optimizing the output loss L of the fault diagnosis classifier module c and the output loss L of the domain discriminator module d and the output loss D of the distribution difference measurement module 2 to complete the fault diagnosis classifier training;

[0026] Step 2, classification phase

[0027] The trained fault diagnosis classifier is used for data classification of the target domain.

[0028] Preferably, in step 1.1, the preprocessing is sequentially performed by normalization processing, feature data extraction, angular domain resampling, and envelope spectrum calculation.

[0029] Further, step 1.1 includes:

[0030] Step 1.11, the original vibration data is normalized according to the following formula

[0031]

[0032] wherein μ is the mean value of the original vibration data, and σ is the standard deviation of the original vibration data;

[0033] Step 1.12, based on the vibration data after standardization, extracting the rotating speed table information or constant rotating speed information, and obtaining the rotating frequency f r according to the rotating speed table information or constant rotating speed information, and estimating the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI and the rolling element fault characteristic frequency f BFF respectively according to formula (2), (3) and (4);

[0034]

[0035]

[0036]

[0037] Step 1.13, based on the obtained rotating frequency f r , the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI and the rolling element fault characteristic frequency f BFF , performing oversampling operation on the rotating speed and the pulse to obtain oversampling data;

[0038] Step 1.14, for the vibration data after standardization, performing anti-aliasing filtering processing according to formula (5) for each fault characteristic frequency to obtain vibration data after anti-aliasing;

[0039]

[0040] wherein O s is the number of samples per rotation relative to the fault characteristic frequency f, f inst (i), i = 1, 2,..., N is the instantaneous frequency of the fault characteristic frequency f; the fault characteristic frequency f is the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI or the rolling element fault characteristic frequency f BFF .

[0041] Step 1.15, using the oversampling data and the vibration data after anti-aliasing to perform equiangular interpolation resampling to obtain four angular domain resampling signals;

[0042] Step 1.16, using Hilbert transform to calculate the envelope signal of each angular domain resampling signal, which is respectively represented as normal rotating speed frequency Outer ring fault characteristic frequency Inner ring fault characteristic frequency And rolling body fault characteristic frequency Each segment length N re The envelope signal is merged into a matrix vector with a size of N re ×4, denoted as

[0043] Further, in step 1.2, in the convolution processing, the output feature of the given l-1 layer The output feature of the l-th layer can be calculated by the following formula:

[0044]

[0045] Where D={s,t} represents the index value of the source domain or the target domain, represents the transfer feature learned from the feature representation of the l-1 layer, b l refers to the bias of the l-th layer, f r (g) represents the commonly used activation function ReLU to solve the vanishing gradient problem;

[0046] The feature representation between the full connection is as follows

[0047]

[0048] Where is the connected feature representation after global maximum pooling, and respectively represent the weight and bias of the full connection F2 layer.

[0049] Further, in step 1.3, the data from the source domain in the output of step 1.2 is calculated by formula (8) and (9) to obtain the output:

[0050]

[0051]

[0052] Where C represents the number of bearing fault categories.

[0053] Further, in step 1.4, it is obtained by processing formula (10) and (11);

[0054]

[0055]

[0056] Further, in step 1.5, the source domain F33s characteristic representation of the target domain F 33t The maximum mean difference is represented by formula (13), (14), (15) respectively:

[0057]

[0058]

[0059]

[0060] Wherein, k(·,·) represents a characteristic kernel function, n1 and n2 represent the sample number of the source domain and the target domain respectively.

[0061] Further, L c and L d are calculated by formula (16) and (17) respectively:

[0062]

[0063]

[0064] Wherein, I(·) is an indicator function, d m ={0,1} represents a domain label;

[0065] The optimization target is represented as minimizing L c , maximizing L d and minimizing D 2 , represented by the following formula:

[0066] minL c -λL d +βD 2 (18)

[0067] Wherein, λ and β are hyperparameters.

[0068] Compared with the prior art, the present application has the following beneficial effects:

[0069] The system of the multi-scale deep field self-adaptive network based on prior knowledge comprises a data signal processing module, a shared network module, a fault diagnosis classification module, a field discriminator module and a distribution difference measurement module, which are clear in module and function, realize end-to-end algorithm flow design, reduce manual participation, save labor and improve work efficiency. The system can make full use of the existing valuable label database information to effectively make up for the lack of label information in the real industrial scene, thereby effectively solving the problem of modeling difficulty caused by label loss; the system integrates a deep learning intelligent algorithm network structure, sets different scale convolution kernels, uses multiple different size convolution kernels to extract features from input feature data, enriches feature extraction, better mines effective information contained in the data, uses the excellent performance of deep algorithms in high-dimensional nonlinearity to quickly and accurately extract effective features, and comprehensively supports fault classification and diagnosis of rotary bearings; the field discriminator is based on the idea of adversarial learning and is integrated into the training process of the model, so that the network cannot distinguish whether the data comes from the source domain or the target domain, thereby reducing the distribution difference between the data; the design of the distribution difference measurement model constructs a difference measurement method of the feature vector and integrates it into the deep framework, reduces the distribution difference between the source domain and the target domain, quickly adapts to the data distribution difference caused by different domain offsets, further improves the classification accuracy, and also widens the application range of intelligent algorithms.

[0070] In order to solve the problems of data distribution offset and label information loss in real industrial applications, promote the successful application of intelligent diagnosis in real industrial scenes, overcome the limitation that the deep learning model cannot cope with the different distribution differences between training data and test data, realize field adaptive fault feature extraction and classification identification, the application combines application scenes, takes prior diagnosis field knowledge as a guide, combines field prior knowledge of fault diagnosis, uses the limited fault diagnosis case label information of enterprise case database, extracts rich features of data by means of multi-scale, simultaneously migrates the data knowledge learned by the training set to the equipment data to be diagnosed by means of field self-adaptive algorithm, and the intelligent model automatically extracts field invariance features to solve the problem of distribution offset caused by individual or running condition change, so that the fault diagnosis migration task can be better realized.

[0071] Further, the application preprocesses data by means of industry field knowledge, so that the model can better learn specific industrial fault diagnosis problems, better mine data features, and also enable the field knowledge to better guide the model to learn and train.

[0072] Further, the application considers normal data of the bearing, inner ring failure of the bearing, outer ring failure of the bearing and rolling body failure, in actual use, more failure types can be added to build higher-dimensional data input, so as to improve the type of model failure diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 A multi-scale deep domain adaptation network framework based on prior knowledge of the application is shown in the figure.

[0074] Figure 2 The application is a rotating bearing fault diagnosis application process. DETAILED DESCRIPTION

[0075] In order to further understand the application, the application will be described below in conjunction with examples, which are only used to further explain the features and advantages of the application, and are not used to limit the claims of the application.

[0076] The application designs a multi-scale deep domain adaptive intelligent algorithm model combined with prior knowledge for bearing fault diagnosis classification method, and the model structure is as shown in the figure. Figure 1 It should be noted that the application analyzes the fault diagnosis of the bearing, the main common faults are inner ring failure of the bearing, outer ring failure of the bearing and rolling body failure, plus normal data, the model faces a four-classification problem; in the domain adaptive algorithm, the source domain refers to the data set with labeled fault classification information, and the target domain refers to the data set without labeled fault information which needs to be classified. The application is designed to include five components of data signal processing module, shared network module, fault diagnosis classification module, domain discriminator module and distribution difference measurement module, as shown in the figure. Figure 1 The technical scheme is realized as follows:

[0077] (1) Data signal processing module

[0078] The original vibration data is standardized, in order to avoid frequency ambiguity phenomenon, equal-angle speed synchronous sampling algorithm is used for resampling to obtain angular domain resampling signal, and Hilbert transform is used to calculate envelope spectrum of the angular domain resampling signal as the input of the model. Given the original vibration data x={x(1),...,x(N)} with length N, the relevant basic information includes bearing pitch diameter D p , rolling body diameter D b , rolling body contact angle θ and rolling body number n e . The specific processing steps are as follows:

[0079] 1) The original vibration data is standardized according to the following formula

[0080]

[0081] wherein μ is the mean of the original vibration data, and σ is the standard deviation of the original vibration data.

[0082] The original vibration data includes data of the source domain and the target domain.

[0083] 2) Based on the vibration data after the standardization processing, the rotating speed table information or the constant rotating speed information is extracted, and the rotating frequency f is obtained according to the rotating speed table information or the constant rotating speed information. r And the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI and the rolling element fault characteristic frequency f BFF are respectively estimated according to the formulas (2), (3) and (4).

[0084]

[0085]

[0086]

[0087] Based on the obtained rotating frequency f r , the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI and the rolling element fault characteristic frequency f BFF , the oversampling operation is performed on the rotating speed and the pulse to obtain the oversampling data.

[0088] Meanwhile, for the vibration data after the standardization processing, the anti-aliasing filtering (low-pass) processing according to the formula (5) is implemented for each fault characteristic frequency to obtain the vibration data after the anti-aliasing. The Nyquist frequency (half of the discrete signal sampling rate) corresponding to each fault characteristic frequency is calculated by the following equation:

[0089]

[0090] wherein O s is the number of samples per rotation relative to the fault characteristic frequency f, f inst (i), i = 1, 2,..., N is the instantaneous frequency of the fault characteristic frequency f. The fault characteristic frequency f is the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI or the rolling element fault characteristic frequency f BFF .

[0091] The oversampling data and the vibration data after the anti-aliasing are used for the equiangular interpolation resampling to obtain four angular domain resampling signals. The previous anti-aliasing processing can avoid the aliasing of the angular domain resampling signals after the equiangular interpolation resampling.

[0092] 3) The envelope signal of each angle domain resampling signal is calculated using the Hilbert transform, denoted as normal speed frequency Outer ring fault characteristic frequency Inner ring fault characteristic frequency And the rolling body fault characteristic frequency Each segment of length N re The envelope signal is merged into a matrix vector of size N re ×4, denoted as The resampling sequence is obtained as the input of the next stage.

[0093] (2) Shared network module

[0094] The shared network module constructs a multi-scale network structure, uses different size convolution kernels to process the input information in different scales, aims to extract more rich features, and finally combines the features in different scales on the full connection layer. It is composed of a plurality of convolution layers, pooling layers and full connection layers.

[0095] In the present application, when the constructed resampling sequence is subjected to convolution processing, the square matrix size convolution kernel is no longer used for operation due to the particularity of the sequence. Because the meaning of transverse convolution is difficult to explain, therefore, the size of the convolution kernel used in the present application is k l ∈R h·1·d , wherein h represents the height of the convolution kernel, 1 and d respectively represent the length and depth of the convolution kernel, that is, the convolution operation is performed on each resampling sequence without operating on two different resampling sequences. Given the output feature of the l-1 layer The output feature of the l layer can be calculated by the following formula:

[0096]

[0097] Where D={s,t} represents the index value of the source domain or the target domain, represents the transfer feature learned from the feature representation of the l-1 layer, b l refers to the bias of the l layer, f r (g) represents the commonly used activation function ReLU to solve the vanishing gradient problem. On this basis, the down-sampling process of the pooling layer is used to reduce the number of training parameters, thereby effectively overcoming the overfitting problem. In this invention, multiple size convolution kernels are used to process data at the same time, so three different multi-size branches are formed. For each depth of the convolution layer, the number of convolution kernels is the same, that is, the number of convolution kernels of C 11 , C 21 and C 31 is the same.

[0098] The present application adopts a down-sampling form of max-pooling, divides the transferable features into several non-overlapping parts, and returns the maximum value of each part. Similarly, when performing the pooling operation, only the longitudinal matrix is subjected to the pooling operation, so that the size of the pooled operation is h p ×1, and the features in the source domain and the target domain After max-pooling, it can be represented as After several convolution and pooling operations in different branches, a global max-pooling operation is performed, that is, the maximum value in each dimension is extracted as the result after pooling. Since the number of convolution kernels is set to be the same, each branch will generate a group of one-dimensional vectors, and the number of elements N zl of the vector is the same as the number of convolution kernels of the last convolution layer. The three groups of one-dimensional vectors are connected to form a one-dimensional vector group with an element number of 3*N zl , which constitutes the full connection layer F1. Further, the feature representation between the full connections can be represented as the following formula

[0099]

[0100] wherein is the connected feature representation after global max-pooling, and represent the weight and bias of the F2 layer, respectively.

[0101] The result of formula (7) is input as the output of the shared network module into the fault diagnosis classifier module, the domain discriminator module and the distribution difference measurement module.

[0102] (3) Fault diagnosis classifier module

[0103] The fault diagnosis classifier module uses the source domain data with labeled data to perform supervised training on the classifier, and the trained classifier is also used as the output of the fault classification result of the test sample. Therefore, only the labeled data from the source domain can be input into this module during the supervised training stage of the classifier.

[0104] The output of the shared network module with respect to the source domain data is input into the fault diagnosis classifier module, and the output formula of the full connection layer F 31 is

[0105]

[0106] Meanwhile, the result of formula (8) can be estimated by Softmax regression through the output of the classifier M1, which can be represented as

[0107]

[0108] Wherein, C represents the rotating equipment fault category.

[0109] That is, the result of formula (9) is the output of the fault diagnosis classifier module.

[0110] (4) Field discriminator module

[0111] The field discriminator module is constructed based on the idea of adversarial learning, and the network is confused by maximizing the loss function, so that the network cannot distinguish whether the data comes from the source domain or the target domain, i.e. the training set and the test set, for extracting domain-invariant feature representation and compressing domain-specific features, and narrowing the distribution difference between different domains.

[0112] The input of the field discriminator module is the result of formula (7) Including the source domain and the target domain. The field discriminator module is composed of a fully connected layer and a binary classifier M2 of Logistics regression, which are respectively represented as:

[0113]

[0114]

[0115] The result of formula (7) is processed by formula (10) and (11) in turn to obtain the output of the field discriminator module.

[0116] (5) Distribution difference measurement module

[0117] In the distribution difference measurement module, the maximum mean difference (MMD) is used as the measurement criterion of the distribution difference, and the difference between the distributions is reduced by minimizing the loss function. The empirical estimation of the MMD method based on the kernel method can be expressed as:

[0118]

[0119] Wherein, k(·,·) represents a feature kernel function, and n1 and n2 represent the number of samples of the source domain and the target domain respectively. In the present application, a Gaussian kernel is used for calculation, that is, Wherein γ is the kernel width. In the present application, the feature representation of the source domain F 33s and the target domain F 33t and the MMD distance can be represented by formulas (13), (14), (15) respectively:

[0120]

[0121]

[0122]

[0123] (6) Construction of optimization target

[0124] The optimization objective of the multi-scale domain adaptation network model based on prior knowledge designed by the application comprises a fault diagnosis classifier module output loss L c and a domain discriminator module output loss L d and a distribution difference measurement module output loss D 2 wherein L c and L d may be calculated by the following formula respectively:

[0125]

[0126]

[0127] wherein I(·) is an indicator function, C represents the number of classes of faults, d m ={0,1} represents a domain label. Therefore, the optimization objective can be represented as minimizing L c , maximizing L d and minimizing D 2 , which is represented by the following formula:

[0128] minL c -λL d +βD 2 (18)

[0129] wherein λ and β are hyperparameters, which determine the strength of domain adaptation, and their values are selected by cross-validation or dynamic parameters.

[0130] The optimization objective is constructed, and a batch stochastic gradient descent algorithm is used to estimate and optimize the parameters of the model. After the model training is completed, the fault diagnosis classifier can classify and diagnose the actual faults of the to-be-tested sample.

[0131] The application adopts a multi-scale structure, uses three branches to extract features, and in actual industry, more branch designs can be used to extract features in depth under the condition that the computing resources are allowed. Meanwhile, only two convolution pooling operations are used on the branch in the application, and under the condition that the conditions are allowed, convolution pooling layers can be increased or deeper network structures such as ResNet series networks and the like can be tried to obtain more superior performance.

[0132] The distribution difference measurement method adopted by the application is the maximum mean difference (MMD), and the measurement method can also use the KL divergence or the joint MMD and the like, and the data object for measurement can also be multi-layer or only the difference of the fully connected layer.

[0133] The application is directed to a rotating bearing fault diagnosis practical application case, taking normal data, bearing inner ring fault, bearing outer ring fault and rolling body fault as examples, based on the above-mentioned prior knowledge multi-scale deep domain adaptive network model, a specific application process is given, as shown in Figure 2 In actual industrial scenarios, the database contains case records of multiple devices, and the common types of bearing faults are screened, including inner ring fault, outer ring fault and rolling body fault (including but not limited to this) and the like. The data with effective labels are taken as the source domain, the high-frequency vibration acceleration vibration waveform of the corresponding data is extracted, the original data is processed by the angular domain synchronous resampling method for feature extraction and integration as the training data; at the same time, the high-frequency vibration acceleration waveform of the to-be-tested sample is also processed by the same method as the target domain data for training and testing; the source domain and the target domain data after data preprocessing are input into the multi-scale deep domain adaptive network model for training, the model parameters are updated and optimized according to the corresponding objective function and optimization algorithm, and finally the trained fault diagnosis classifier is obtained; the data of the to-be-tested sample after processing are input into the trained network, and the fault diagnosis type of the to-be-tested bearing can be output by the classifier for analysis.

[0134] The multi-scale deep domain adaptive network based on prior knowledge can make full use of existing label data resources, and can still migrate multi-device multi-condition label data knowledge to the device to be diagnosed when the device data distribution deviates due to the working environment, greatly compensating for the label missing problem in real industrial production, greatly improving the use value of the label, and also improving the efficiency of bearing fault diagnosis.

Claims

1. A bearing fault classification method based on multi-scale domain adaptive networks, characterized in that, Based on the bearing fault classification system, the bearing fault classification system includes a data signal processing module, a shared network module, a fault diagnosis and classification module, a domain discriminator module, and a distribution difference measurement module; The data signal processing module preprocesses the raw vibration data in the database to obtain the output; the database is a database that stores the source domain and the target domain; the source domain is data with valid labels. The shared network module takes the output of the data signal processing module as input, uses convolutional kernels of different sizes to process the input at different scales, and merges the features at different scales into the fully connected layer for output. The fault diagnosis classification module takes the output of the shared network module as input, adopts the data of the source domain to supervise the training of the fault diagnosis classifier, outputs a classification result, and optimizes the output loss L of the fault diagnosis classifier module c The trained fault diagnosis classifier is used for data classification of the target domain. The field discriminator module takes the output of the shared network module as input, and maximizes the loss function to pull in the distribution difference between the source domain and the target domain to obtain an output; and the output loss L of the field discriminator module is optimized d ; The distribution difference measurement module takes the output of the shared network module as input, and obtains and outputs the maximum mean difference between the source domain and the target domain by minimizing the loss function. The optimization distribution difference measure module outputs a loss D 2 ; The methods include: Step 1, Training Phase: Step 1.1: Preprocess the raw vibration data from the source and target domains in the database, where the source domain consists of data with valid labels; including: Step 1.11: Standardize the raw vibration data according to the following formula. Where μ is the mean of the original vibration data and σ is the standard deviation of the original vibration data; Step 1.12, based on the standardized vibration data, extract the rotating speed table information or constant rotating speed information, and obtain the rotating frequency f according to the rotating speed table information or constant rotating speed information r , and estimate the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI and the rolling element fault characteristic frequency f BFF according to the formulas (2), (3) and (4) respectively; Step 1.13, based on the obtained rotation frequency f r , outer ring fault characteristic frequency f BPFO , inner ring fault characteristic frequency f BPFI , and rolling element fault characteristic frequency f BFF , oversampling operation is performed on the rotation speed and the pulse to obtain oversampling data; Step 1.14: For the standardized vibration data, anti-aliasing filtering is performed for each fault characteristic frequency according to formula (5) to obtain the anti-aliasing vibration data; wherein O s is the number of samples per revolution with respect to the fault characteristic frequency f inst (i), i = 1, 2,..., N is the instantaneous frequency of the fault characteristic frequency f; the fault characteristic frequency f is the outer ring fault characteristic frequency f BPFO , the inner ring fault characteristic frequency f BPFI or the rolling element fault characteristic frequency f BFF ; Step 1.15: Use the oversampled data and the vibration data after anti-aliasing to perform equal-angle interpolation resampling to obtain the four corner domain resampled signals; Step 1.16: Use Hilbert transform to calculate the envelope signal of each angular domain resampled signal, which is represented by the normal rotational speed frequency. Outer ring fault characteristic frequency Inner ring fault characteristic frequency and rolling element failure characteristic frequency Each segment has a length of N. re The envelope signals are merged into a group of size N. re A matrix-vector of size ×4 is represented as Step 1.2: Perform convolution processing on the output of Step 1.1 at different scales using convolution kernels of different sizes to extract features, and then fully connect the features obtained at different scales. Step 1.3: Using the data from the source domain and the output of Step 1.2, perform supervised training on the fault diagnosis classifier; Step 1.4: Using the data from the source and target domains and the output of Step 1.2, the distribution differences between the different domains are narrowed down by maximizing the loss function; Step 1.5: Using the data from the source domain and the target domain, and the output of Step 1.2, the maximum mean difference between the source domain and the target domain is obtained by minimizing the loss function; Step 1.6, calculate and optimize the fault diagnosis classifier module output loss L using the outputs of Step 1.3, Step 1.4 and Step 1.5 c and the domain discriminator module output loss L d and the distribution difference measure module output loss D 2 to complete the fault diagnosis classifier training; Step 2, Classification Stage A trained fault diagnosis classifier is used for data classification in the target domain.

2. The bearing fault classification method based on multi-scale domain adaptive networks according to claim 1, characterized in that, In step 1.2, during the convolution process, given the output features of the (l-1)th layer... The output features of layer l can be calculated using the following formula: Where D = {s, t} represents the index value of the source or target domain. This represents the feature representation from the (l-1)th layer. The transfer features learned in the middle, b l This refers to the deviation of the l-th layer, f r (·) represents ReLU, a commonly used activation function for solving the vanishing gradient problem; The features between fully connected layers are represented by the following formula: in This represents the feature representation concatenated after global max pooling. as well as These represent the weights and biases of the fully connected F2 layer, respectively.

3. The bearing fault classification method based on multi-scale domain adaptive networks according to claim 2, characterized in that, In step 1.3, the data from the source domain in the output of step 1.

2. The output is obtained by calculation using formulas (8) and (9): Where C represents the number of bearing failure categories.

4. The bearing fault classification method based on multi-scale domain adaptive networks according to claim 3, characterized in that, In step 1.4, the specific results are obtained through formulas (10) and (11); 5. The bearing fault classification method based on multi-scale domain adaptive networks according to claim 4, characterized in that, In step 1.5, the source domain F is obtained through a fully connected layer. 33s With the target domain F 33t The characteristic representation and the maximum mean difference are expressed by formulas (13), (14), and (15), respectively: Where k(·,·) represents the feature kernel function, and n1 and n2 represent the number of samples in the source domain and the target domain, respectively.

6. The bearing fault classification method based on multi-scale domain adaptive networks according to claim 5, characterized in that, L c and L d are calculated from equations (16) and (17), respectively: Where I(·) is the indicator function, d m ={0,1} represents the domain label; The optimization objective is expressed as minimizing L c Maximize L d and minimizing D 2 It can be expressed by the following formula. min L c −λL d +βD 2 (18) Where λ and β are hyperparameters.

Images