A Motor Bearing Fault Diagnosis Method and Device Based on Incremental Learning

Abstract

The present invention discloses a method and device for motor bearing fault diagnosis based on incremental learning. The steps of the method include: in the model training stage, collecting the operation data and domain data of different types of motor bearings in different states respectively to form a training sample set, training a fault diagnosis model, and screening out representative samples; in the incremental learning stage, when new samples are input, determining whether they belong to a new category or a new domain, correcting the weights of the new samples according to the judgment result, using the trained network parameters as the initial conditions, and inputting the representative samples, new samples and the corrected weights into the fault diagnosis model for model fine-tuning, wherein the corrected weights are used to calculate the loss function in the model fine-tuning process; in the fault diagnosis stage, inputting the data to be diagnosed into the finally obtained fault diagnosis model for diagnosis. The present invention has the advantages of simple implementation method, low cost, high diagnostic accuracy and efficiency, and wide application range, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and in particular to a motor bearing fault diagnosis method and device based on incremental learning. Background Art

[0002] Motor bearing faults include various fault types, such as inner ring faults, outer ring faults, rolling element faults, cage faults, etc., and different degrees of faults are involved according to the fault size. Motor bearing fault diagnosis is to automatically diagnose whether there is a fault in the motor bearing through detection and data processing. The motor bearing fault diagnosis method based on deep learning can obtain high-precision diagnosis results, but such methods mainly rely on sufficient training data and the assumption that the training set and test set data satisfy the same distribution. In actual applications, most motor bearings operate in a normal state, and it is not easy to obtain fault data. It is very likely to obtain data in stages and batches, and the differences in motor bearing models and working conditions will lead to differences in data distribution. The above factors will limit the final diagnosis effect. If the original data and new data are used for complete retraining, it will cause problems of consuming computing and storage resources.

[0003] The incremental learning-based model training method can solve the problem of progressive training of the model when data is gradually obtained. The incremental learning method is to obtain data in a streaming manner. When new data appears, while learning knowledge from the new data, the knowledge that has been learned is retained to avoid the waste of resources caused by complete retraining. However, when the incremental learning method is used to implement equipment fault diagnosis in the prior art, the new samples and representative samples are usually directly used as a whole for training, and only the fault categories of the equipment are concerned, that is, the class incremental problem (whether the new samples are new fault categories). When applied to motor bearing fault diagnosis, the following problems will occur:

[0004] 1. In motor bearing fault diagnosis, differences such as motor bearing models and operating conditions will cause changes in fault characteristics, that is, although the fault categories are the same, different domain data will show different characteristics. The traditional incremental learning method only focuses on the class incremental problem and cannot meet the requirements of domain increment in motor bearing fault diagnosis. Domain increment refers to whether the new samples involve a new domain, that is, data with different working conditions or different bearing types from the original training data. The existence of domain increment data will lead to the failure of the final diagnosis model and affect the accuracy of fault diagnosis.

[0005] 2. Since in incremental learning, the representative samples and the newly added samples are directly used as a whole for training, when the number of samples of the two is quite different, the trained model cannot well represent the characteristics of the side with fewer samples, resulting in the impact on the model accuracy. The commonly used methods in the existing technology, such as oversampling the minority class, undersampling the majority class, or a combination of both, are prone to problems such as overfitting or information loss. The traditional data generation method based on generative adversarial networks will have problems such as large training difficulty and difficult convergence guarantee.

[0006] In summary, in the existing technology solutions for realizing motor bearing fault diagnosis using incremental learning, it is impossible to balance diagnosis efficiency, accuracy, and cost. Summary of the Invention

[0007] The technical problem to be solved by the present invention lies in: aiming at the technical problems existing in the existing technology, the present invention provides a method and device for motor bearing fault diagnosis based on incremental learning with simple implementation method, low cost, high diagnosis accuracy and efficiency, and wide application range.

[0008] To solve the above technical problems, the technical solution proposed by the present invention is:

[0009] A method for motor bearing fault diagnosis based on incremental learning, the steps include:

[0010] In the model training stage, collect the operation data of different types of motor bearings in different states and the domain data of different types of motor bearings in different states respectively to form a training sample set, use the training sample set to train a fault diagnosis model based on a deep learning network, and screen out the representative samples under the current fault diagnosis model, and output the screened representative samples and the trained model network parameters;

[0011] In the incremental learning stage, when a newly added motor bearing data sample is input, judge whether the newly added sample belongs to a newly added category or a newly added domain, correct the weight of the sample according to the judgment result and the quantity relationship between the representative sample and the newly added sample to obtain the corrected weight, use the trained model network parameters as the initial condition, input the representative sample, the newly added sample and the corrected weight into the trained fault diagnosis model for model fine-tuning, wherein the corrected weight is used to calculate the loss function in the model fine-tuning process, and the representative sample and the newly added sample are used as fine-tuning data;

[0012] In the fault diagnosis stage, input the data to be diagnosed into the fault diagnosis model finally obtained after the incremental learning stage for diagnosis, and output the diagnosis result.

[0013] Further, the operating data includes vibration acceleration signals, the domain data includes operating conditions and bearing models, the operating conditions include any one or more of rotational speed, load, and load torque, the states include normal state and fault states with different fault degrees, the fault states include any combination of inner race fault, outer race fault, rolling element fault, and cage fault states, and the fault categories are formed by combining the fault states and fault degrees.

[0014] Further, when the operating data of the motor bearing and / or new samples are collected, it also includes a data preprocessing step, and the data preprocessing step includes: converting the operating data from the time domain signal to the frequency domain, normalizing the converted data points and domain data respectively, and setting class labels and domain auxiliary labels for each data point. The class labels are used to mark the categories of each data point, and the domain auxiliary labels are used to mark the domains to which each data point belongs.

[0015] Further, the fault diagnosis model is trained using a deep learning network based on a stacked autoencoder. The deep learning network includes an input layer x, multiple hidden layers, and an output layer L. Among them, the hidden layer obtained after the previous autoencoder is pre-trained is used as the input of the next layer. During the pre-training process, multiple hidden layers are obtained through layer-by-layer training, and then the parameters of each layer are fine-tuned using part of the labeled data through the backpropagation algorithm to complete the construction of the network structure of the deep learning network.

[0016] Further, the dimension of the input layer x is N + 3, where N is half of the number of sampling points, and the dimension of the output layer L is K, where K > C, and C is the number of fault categories included in the current data.

[0017] Further, the calculation expression of the loss function is:

[0018]

[0019] where, J c represents the loss function, N total represents the total number of representative samples and new samples, I{·} is an indicator function, y i represents the classification result of the i-th sample, k is the actual label, K is the dimension of the output layer L. Among them, if the classification result is the same as the actual label, that is, y i = k, then I{·} is 1, otherwise it is 0; x i is the i-th sample, p(x i ) represents the probability that the i-th sample belongs to each label type, λ is the corrected weight, M = (m1, m2,..., m K ) is the class flag bit, m1, m2,..., m KRespectively represent the numerical values of each flag bit. The numerical value m = 1 of each flag bit indicates that this category has appeared, and the numerical value m = 0 of each flag bit indicates that this category has not appeared yet.

[0020] Further, in the step S01, the representative samples are screened out by using the method based on the optimal sub-optimal distance. The steps include: calculating the highest value p of the probability estimation of each sample Best and the second highest value p SecondBest , and calculating the probability difference value x Best between the highest value p SecondBest and the second highest value p BvSB . Screen out the samples with the highest probability difference value x BvSB from the samples with the same category label and the same domain auxiliary label as the top several samples as the representative samples under the current fault diagnosis model.

[0021] Further, when modifying the weights of the newly added samples according to the judgment result, if only the number of samples is increased and no new categories and domains are added, that is, c z ≤C and d z ≤D, then the weights λ of the N z newly added samples are modified to N r / N z , and the weight λ of the representative sample is 1; if new categories are added and no new domains are added, that is, c z >C and d z ≤D, then the weights λ of these N z newly added samples are modified to αN r / N z , α>1, and the weight λ of the representative sample is 1. If new domains are added and no new categories are added, that is, c z ≤C and d z >D, then the weights λ of the N z newly added samples are modified to βN r / N z , β>1, and the weight λ of the representative sample is 1, where c z is the category label in the newly added samples, N z is the number of newly added samples with the domain auxiliary label d z , α and β are different preset coefficients, C represents the number of category labels included in the current representative samples, D represents the number of domain auxiliary labels included in the current representative samples, and N r is the number of samples of any category c and domain auxiliary label d in the representative samples.

[0022] A motor bearing fault diagnosis device based on incremental learning, comprising:

[0023] A model training module, which is used in the model training stage to collect the operation data of different types of motor bearings in different states and the domain data of different types of motor bearings in different states respectively to form a training sample set, train a fault diagnosis model using the training sample set based on a deep learning network, and screen out representative samples under the current fault diagnosis model, and output the screened representative samples and the trained model network parameters;

[0024] An incremental learning module, which is used in the incremental learning stage. When new motor bearing data samples are input, it determines whether the new samples belong to a new fault category or new domain data, modifies the weights of the samples according to the judgment result and the quantitative relationship between the representative samples and the new samples to obtain the modified weights. With the trained model network parameters as the initial conditions, the representative samples, new samples and the modified weights are input into the trained fault diagnosis model for model fine-tuning. The modified weights are used to calculate the loss function in the model fine-tuning process, and the representative samples and new samples are used as fine-tuning data;

[0025] A fault diagnosis module, which is used in the fault diagnosis stage to input the data to be diagnosed into the fault diagnosis model finally obtained after the incremental learning stage for diagnosis and output the diagnosis result.

[0026] A computer-readable storage medium storing a computer program, and when the computer program is executed, it implements the method as described above.

[0027] Compared with the prior art, the advantages of the present invention are as follows:

[0028] 1. In the model training stage of the present invention, a fault diagnosis model is trained based on the operation data and domain data of motor bearings and representative samples are screened out. In the incremental learning stage, when new samples are available, it determines whether the new samples belong to a new category or a new domain, modifies the sample weights accordingly, and then uses the modified weights to calculate the loss function during model fine-tuning, so that not only can the diagnosis and recognition of new categories be realized, but also the diagnosis and recognition of new domains can be realized, which can simultaneously meet the requirements of class increment and domain increment in motor bearing fault diagnosis, and thus can be applied to various fault diagnosis applications across motors, variable working conditions, and multiple categories, and better meet the actual application requirements of motor bearing fault diagnosis.

[0029] 2. In the incremental learning process of the present invention, by considering the quantity problem of representative samples and new samples, the weights are modified according to the quantitative relationship between representative samples and new samples, and then the loss function in the fine-tuning process is calculated, so that the characteristics of both representative samples and new samples can be fully represented, and the accuracy of the fault diagnosis model is further improved.

[0030] 3. The present invention forms a complete fault diagnosis model by adopting an incremental learning-based progressive training method, which conforms to the characteristic that the operation data of the motor bearing is obtained in stages and batches. It can not only take into account the diagnostic accuracy, reliability, precision, and cost, but also balance the stability and plasticity of the diagnostic network, and does not require complex calculations or a large amount of storage space, facilitating the accurate and rapid implementation of the fault diagnosis of the motor bearing.

[0031] 4. The present invention further adopts the selection of representative samples based on the optimal sub-optimal distance, which can quantitatively describe the relationship between the sample and the classification boundary, and can better reflect the classification effect of each sample in the current diagnosis model, thus ensuring the representativeness and accuracy of the sample selection, and further improving the diagnostic performance after incremental training. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a schematic diagram of the implementation process of the motor bearing fault diagnosis method based on incremental learning in this embodiment.

[0033] Figure 2 is a schematic diagram of the principle of realizing the fault diagnosis of the motor bearing in this embodiment.

[0034] Figure 3 is a schematic diagram of the detailed process of realizing the fault diagnosis of the motor bearing in this embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The present invention will be further described below in conjunction with the accompanying drawings of the specification and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0036] As Figure 1 shown, the steps of the motor bearing fault diagnosis method based on incremental learning in this embodiment specifically include:

[0037] S01. In the model training stage, the operation data of different types of motor bearings in different states and the domain data of different types of motor bearings in different states are respectively collected to form a training sample set. The fault diagnosis model is trained based on the deep learning network using the training sample set, and the representative samples under the current fault diagnosis model are screened out, and the screened representative samples and the model network parameters obtained by training are output;

[0038] S02. In the incremental learning stage, when new motor bearing data samples are input, it is determined whether the new samples belong to a new fault category or new domain data. According to the judgment result and the quantitative relationship between the representative samples and the new samples, the weights of the samples are corrected to obtain the corrected weights. Taking the model network parameters obtained by training as the initial conditions, the representative samples, the new samples, and the corrected weights are input into the trained fault diagnosis model for model fine-tuning, where the corrected weights are used to calculate the loss function during the model fine-tuning process, and the representative samples and the new samples serve as the fine-tuning data;

[0039] S03. In the fault diagnosis stage, the data to be diagnosed is input into the fault diagnosis model finally obtained after the incremental learning stage for diagnosis, and the diagnosis result is output.

[0040] Considering that the operation data such as the vibration data of the motor bearing is usually obtained in stages and batches, and there are different situations of new category addition (different categories) and new domain addition (different working conditions or different bearing types) in the data. In this embodiment, a fault diagnosis model is first trained based on the operation data and domain data of the motor bearing in the model training stage, and representative samples are selected. In the incremental learning stage, when new samples appear, it is determined whether the new samples belong to a new category or a new domain. According to the judgment result and the quantitative relationship between the representative samples and the new samples, the weights of the samples are corrected accordingly. Then, when the model is fine-tuned, the corrected weights are used to calculate the loss function, so that not only can the diagnosis and recognition of new categories be realized, but also the diagnosis and recognition of new domains can be realized, meeting the requirements of class increment and domain increment in the motor bearing fault diagnosis at the same time, and can be applied to various fault diagnosis applications across motors, variable working conditions, and multiple categories, which more conforms to the actual application requirements of the motor bearing fault diagnosis. At the same time, in the incremental learning process, the quantity problem of the representative samples and the new samples is considered. Since the weights are corrected according to the quantitative relationship between the representative samples and the new samples, and then the loss function in the fine-tuning process is calculated, it can fully represent the characteristics of both the representative samples and the new samples, further improving the accuracy of the fault diagnosis model.

[0041] In this embodiment, the operation data collected in step S01 can specifically be vibration acceleration signals, etc., and other operation signals can also be further introduced. The domain data includes information such as the operation working conditions and the motor bearing model, where the operation working conditions include rotational speed, load, load torque, etc. The state of the motor bearing specifically includes the normal state and the fault states with different fault degrees. The fault states include inner ring fault, outer ring fault, rolling element fault, cage fault state, etc., and the fault degrees can be divided into mild, moderate, and severe, etc. The fault categories are formed by combining the bearing fault states and the fault degrees, such as mild, moderate, and severe inner ring faults. The specific selection of the above various types of information can be determined according to actual needs.

[0042] In a specific application embodiment, during the model training stage, vibration acceleration signals of different types of motor bearings in normal states, different types of faults (inner ring faults, outer ring faults, rolling element faults, cage fault states, etc.), and fault states of different degrees (mild, moderate, severe, etc.) are collected to characterize the fault characteristics of different types of motor bearings in different fault states. At the same time, information such as the operating conditions of the motor bearings (speed, load, load torque, etc.) and the motor bearing models is collected as domain data. The bearing operation data and the domain data are jointly used for model training to obtain a preliminary fault diagnosis model. Since vibration data, motor operating conditions, bearing model information, etc. are introduced as features together, the fault characteristics of different types of motor bearings can be fully characterized, that is, the fault characteristics of motor bearings in different fault states can be characterized, and the fault characteristics of different types of motor bearings and during faults under different operating conditions can also be characterized, so that the model can simultaneously meet the diagnostic requirements of both class increment and domain increment.

[0043] In this embodiment, when the operation data of the motor bearing is collected in step S01 and new samples are input in step S02, it also includes a data preprocessing step. The data preprocessing step includes: converting the operation data from the time domain signal to the frequency domain, normalizing the converted data points and the domain data, and setting class labels and domain auxiliary labels for each data point. The class labels are used to mark the classes of each data point, and the domain auxiliary labels are used to mark the domains to which each data point belongs.

[0044] In a specific application embodiment, taking the preprocessing of the data collected in step S01 as an example, the detailed steps include:

[0045] 1) For each fault category, the collected vibration acceleration signals are divided into multiple sequences with a length of N0; a fast Fourier transform is performed on each sequence to convert the time domain signal to the frequency domain, where the number of sampling points is 2N.

[0046] Due to the symmetry of the Fourier spectrum, only the first half of the data after the Fourier transform is retained in this embodiment, that is, N data points are stored.

[0047] 2) The speed, load, and bearing model corresponding to each sequence are stored as the (N + 1)th, (N + 2)th, and (N + 3)th data points of the sequence as domain data respectively. That is, each sequence contains N + 3 data points after processing, forming a sample.

[0048] 3) The first N data points are normalized using formula (1) to ensure the unity of the new data and the original data.

[0049]

[0050] Among them, x and x' are the values before and after normalization respectively. In this step, x is specifically the amplitude after fast Fourier transform, and min(x) and max(x) are the minimum and maximum values of x respectively. The maximum and minimum values are not generated by the current sequence but are set in the initial training stage, and the same maximum and minimum values are used for subsequent newly added data.

[0051] 4) Use formula (1) to perform normalization processing on the N+1, N+2, and N+3 data points respectively. In this step, x is the rotational speed, load, and bearing model respectively.

[0052] 5) For each category of data, multiple normalized samples with a length of N+3 are obtained.

[0053] 6) Set category labels for each data point, and at the same time set domain auxiliary labels according to the N+1, N+2, and N+3 data points, that is, the same N+1, N+2, and N+3 data points are set to the same domain auxiliary label.

[0054] Train a fault diagnosis model for the preprocessed training sample set based on a deep learning network. In this embodiment, the fault diagnosis model is specifically trained using a deep learning network based on a stacked autoencoder. The deep learning network includes an input layer x, multiple hidden layers {h1, h2,..., h M}, and an output layer L. Among them, the hidden layer obtained after pre-training of the previous autoencoder is used as the input of the next layer, and a Softmax classifier is added as the output layer at the top layer. During the pre-training process, multiple hidden layers are obtained through layer-by-layer training, and then the parameters of each layer are fine-tuned using partially labeled data through the backpropagation algorithm to complete the construction of the network structure of the deep learning network. The dimension of the above input layer x is specifically N+3 (the length of the preprocessed data), N is half of the number of sampling points, and the dimension of the output layer L is K, where K > C (the ability to reserve class increments), and C is the number of fault categories included in the current data. Except for the first training, subsequent incremental training does not perform pre-training again and directly proceeds to the fine-tuning step. Save the trained diagnostic network parameters and representative samples, and do not save the remaining data participating in the training, which can effectively reduce the demand for storage space. When new samples are collected, preprocess the new samples according to the above preprocessing steps.

[0055] In the incremental learning stage of this embodiment, using the saved diagnostic network parameters as the initial condition, the representative samples and new samples as the fine-tuning data, and introducing the corrected weights into the loss function, repeating the model training set can complete the fine-tuning of the current model to obtain the latest bearing fault diagnosis network. The calculation expression of the loss function is specifically:

[0056]

[0057] Among them, J c represents the loss function, N total represents the total number of representative samples and new samples, I{·} is the indicator function, and y i represents the classification result of the i-th sample, k is the actual label, and K is the dimension of the output layer L. If the classification result is the same as the actual label, that is, y i = k, then I{·} is 1, otherwise it is 0; x i is the i-th sample, and p(x i ) represents the probability that the i-th sample belongs to each label type. λ is the corrected weight, and M = (m1, m2,..., m K ) is the class flag bit. m1, m2,..., m K represent the values of each flag bit respectively. If m = 1 in each flag bit, it means that this class has appeared; if m = 0, it means that this class has not appeared yet.

[0058] In this embodiment, by using the above weighted loss function, considering the number of representative samples and new samples, introducing the corrected sample weight, and combining the probability that the sample belongs to each label type, etc., the diagnostic performance can be further improved, the stability and plasticity of the diagnostic network can be balanced, and problems such as overfitting or information loss can be avoided, and the calculation is simple.

[0059] In step S01 of this embodiment, after the training of the diagnostic model is completed, the representative samples are specifically selected by using the method based on the optimal sub-optimal distance. The specific steps include: calculating the highest value p i of the probability estimate x Best (x i ) and the second highest value p SecondBest (x i ), and calculating the probability difference value x Best (x i ) between the highest value p SecondBest (x i ) and the second highest value p BvSB (x i ). Select the top several samples with the highest probability difference value x BvSB from the samples with the same class label and the same domain auxiliary label as the representative samples under the current fault diagnosis model. Usually, the selection of representative samples based on the optimal sub-optimal distance can quantitatively describe the relationship between the sample and the classification boundary, and can better reflect the classification effect of each sample in the current diagnostic model, so as to ensure the representativeness and accuracy of sample selection, and further improve the diagnostic performance after incremental training.

[0060] The specific expression for calculating the probability difference value x BvSB is as follows:

[0061] xBvSB = p Best (x i ) - p SecondBest (x i ) (3)

[0062] In step S02 of this embodiment, it is assumed that c z is the class label in the newly added samples, N z is the number of newly added samples with the domain auxiliary label d z , C represents the number of class labels included in the current representative samples, c = {1, 2,..., C} represents any class of the representative samples, D represents the number of domain auxiliary labels included in the current representative samples, d = {1, 2,..., D}, N r is the number of samples with any class c of the representative samples and the domain auxiliary label d. When correcting the weights of the newly added samples according to the judgment results in the incremental learning stage, the correction is performed in the following three cases:

[0063] (1) If only the number of samples is increased, and no new classes and domains are added, that is, c z ≤ C and d z ≤ D, then the weights λ of the N z newly added samples are corrected to N r / N z , and the weights λ of the representative samples are 1;

[0064] (2) If a new class is added and no new domain is added, that is, c z > C and d z ≤ D, then the weights λ of these N z newly added samples are corrected to αN r / N z , α > 1, and the weights λ of the representative samples are 1;

[0065] (3) If a new domain is added and no new class is added, that is, c z ≤ C and d z > D, then the weights λ of the N z newly added samples are corrected to βN r / N z , β > 1, and the weights λ of the representative samples are 1.

[0066] The above α and β are different preset coefficients, and β > α, which can be specifically configured according to actual needs.

[0067] Such as Figure 2 , 3As shown in the figure, in this embodiment, specifically during the first training, the collected bearing vibration acceleration signals, operating condition information, and bearing model information are preprocessed to form a sample set, and a deep learning algorithm based on a stacked autoencoder is used to train the diagnostic model; a method based on the optimal suboptimal distance is used to screen out representative samples, and the representative samples and the diagnostic model parameters in the current state are saved; when new samples are collected, the data is preprocessed to obtain new samples. According to the relationship between the new samples and the representative samples, the sample weights are corrected; the new samples and the representative samples are input into the trained diagnostic network based on deep learning, and on this basis, the network parameters are fine-tuned. Among them, the corrected weights are introduced into the loss function. The fine-tuned diagnostic network can retain the original knowledge and learn the knowledge of the new samples. The training results of the diagnostic network are used to update the representative samples again and save them for training when new samples are collected next time. After one training, only the representative samples and the network structure parameters need to be saved, and there is no need to save other training data. Through the above steps of this embodiment, a complete model can be formed through the progressive training of the diagnostic model, which can not only take into account diagnostic accuracy, reliability, precision, and cost, but also balance the stability and plasticity of the diagnostic network, and does not require complex calculations or a large amount of storage space, facilitating the accurate and rapid implementation of the fault diagnosis of motor bearings.

[0068] The motor bearing fault diagnosis device based on incremental learning in this embodiment includes:

[0069] A model training module, which is used in the model training stage to collect the operation data of different types of motor bearings in different states and the domain data of different types of motor bearings in different states respectively to form a training sample set, and use the training sample set to train the fault diagnosis model based on a deep learning network, and screen out the representative samples under the current fault diagnosis model, and output the screened representative samples and the model network parameters obtained by training;

[0070] An incremental learning module, which is used in the incremental learning stage. When new motor bearing data samples are input, it is judged whether the new samples belong to a new fault category or new domain data. According to the judgment result and the quantitative relationship between the representative samples and the new samples, the weights of the samples are corrected to obtain the corrected weights. Taking the model network parameters obtained by training as the initial conditions, the representative samples, the new samples, and the corrected weights are input into the trained fault diagnosis model for model fine-tuning. Among them, the corrected weights are used to calculate the loss function in the model fine-tuning process, and the representative samples and the new samples are used as fine-tuning data;

[0071] A fault diagnosis module, which is used in the fault diagnosis stage to input the data to be diagnosed into the fault diagnosis model finally obtained after the incremental learning stage for diagnosis, and output the diagnosis result.

[0072] The motor bearing fault diagnosis device based on incremental learning in this embodiment corresponds one-to-one with the above-mentioned motor bearing fault diagnosis method based on incremental learning, and will not be elaborated here one by one.

[0073] This embodiment also provides a computer-readable storage medium storing a computer program, and when the computer program is executed, it implements the above method.

[0074] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A motor bearing fault diagnosis method based on incremental learning, characterized in that the steps Including: In the model training stage, the operation data of different types of motor bearings in different states and the domain data of different types of motor bearings in different states are respectively collected to form a training sample set. Based on a deep learning network, the training sample set is used to train a fault diagnosis model, and representative samples under the current fault diagnosis model are screened out, and the screened representative samples and the trained model network parameters are output; In the incremental learning stage, when new motor bearing data samples are input, it is determined whether the new samples belong to new fault categories or new domain data. According to the judgment result and the quantitative relationship between the representative samples and the new samples, the weights of the samples are corrected to obtain the corrected weights. Taking the trained model network parameters as the initial conditions, the representative samples, the new samples and the corrected weights are input into the trained fault diagnosis model for model fine-tuning, where the corrected weights are used to calculate the loss function in the model fine-tuning process, and the representative samples and the new samples are used as fine-tuning data; In the fault diagnosis stage, the data to be diagnosed is input into the fault diagnosis model finally obtained after the incremental learning stage for diagnosis, and the diagnosis result is output; The representative samples are selected by using the method based on the optimal sub-optimal distance, and the steps include: calculating the highest value of the probability estimation of each sample p Best and the second highest value p SecondBest , and calculating the probability difference value p Best between the highest value p SecondBest and the second highest value x BvSB . From the samples with the same class label and the same domain auxiliary label, the top several samples with the highest probability difference value x BvSB are selected as the representative samples under the current fault diagnosis model; When correcting the weight of the sample according to the judgment result and the quantitative relationship between the representative sample and the newly added sample, if only the number of samples is increased and no new categories and domains are added, that is c z ≤ C and d z ≤ D , then the weight λ of the N z newly added samples is corrected to N r / N z , and the weight λ of the representative sample is 1; if a new category is added and no new domain is added, that is c z > C and d z ≤ D , then the weight λ of these N z newly added samples is corrected to αN r / N z , α > 1, and the weight λ of the representative sample is 1. If a new domain is added and no new category is added, that is c z ≤ C and d z > D , then the weight λ of the N z newly added samples is corrected to βN r / N z , β > 1, and the weight λ of the representative sample is 1, where c z is the class label in the newly added samples, N z is the domain auxiliary label and is d z the number of newly added samples, α、β are different preset coefficients, β>α , C represents the number of class labels included in the current representative sample, D represents the number of domain auxiliary labels included in the current representative sample, N r is any class of the representative sample c and the domain auxiliary label is d the number of samples.

2. The method for diagnosing motor bearing faults based on incremental learning according to claim 1, wherein The operation data includes vibration acceleration signals, the domain data includes operating conditions and bearing models, the operating conditions include any one or more of rotational speed, load, and load torque, the states include normal state and fault states with different fault degrees, and the fault states include any combination of inner race fault, outer race fault, rolling element fault, and cage fault states. The fault categories are formed by the combination of fault states and fault degrees.

3. The method for diagnosing motor bearing faults based on incremental learning according to claim 1, wherein When the operation data of the motor bearing and / or new samples are collected, it further includes a data preprocessing step. The data preprocessing step includes: converting the operation data from the time domain signal to the frequency domain, respectively normalizing the converted data points and domain data, and setting class labels and domain auxiliary labels for each data point. The class labels are used to mark the categories of each data point, and the domain auxiliary labels are used to mark the domains to which each data point belongs.

4. The method for diagnosing motor bearing faults based on incremental learning according to claim 1, wherein The fault diagnosis model is trained using a deep learning network based on a stacked autoencoder. The deep learning network includes an input layer x, multiple hidden layers, and an output layer L. The hidden layer obtained after the pre-training of the previous autoencoder is used as the input of the next layer. In the pre-training process, multiple hidden layers are obtained through layer-by-layer training, and then the network structure of the deep learning network is constructed by fine-tuning the parameters of each layer using some labeled data through the backpropagation algorithm.

5. The method for diagnosing motor bearing faults based on incremental learning according to claim 4, characterized in that, The dimension of the input layer x is N +3, N which is half of the number of sampling points ， The dimension of the output layer L is K, where K > C and C is the number of fault categories included in the current data.

6. The method for diagnosing motor bearing faults based on incremental learning according to any one of claims 1 to 5, characterized in that, The calculation expression of the loss function is: （2） Among them, represents the loss function, N total represents the total number of samples of the representative samples and the newly added samples, I {•} is the indicator function, y i represents the classification result of the i th sample, k is the actual label ，K is the dimension of the output layer L, where if the classification result is the same as the actual label, that is y i = k , then I {•} is 1, otherwise it is 0; x i is the i th sample, p ( x i ) represents the probability that the i th sample belongs to each label type, λ is the corrected weight, M =( m 1, m 2,…, m K ) is the category flag bit, m 1、 m 2…, m K respectively represent the values of each flag bit, and the values of each flag bit m =1 indicates that this category has appeared, and the value of each flag bit m =0 indicates that this category has not appeared yet.

7. A motor bearing fault diagnosis device based on incremental learning, characterized in that, Including: A model training module, which is used in the model training stage to respectively collect the operation data of different types of motor bearings in different states and the domain data of different types of motor bearings in different states to form a training sample set, use the training sample set to train a fault diagnosis model based on a deep learning network, screen out representative samples under the current fault diagnosis model, and output the screened representative samples and the trained model network parameters; An incremental learning module, which is used in the incremental learning stage. When new motor bearing data samples are input, it determines whether the new samples belong to new fault categories or new domain data. According to the judgment result and the quantitative relationship between the representative samples and the new samples, it corrects the weights of the samples to obtain the corrected weights. With the network parameters of the trained model as the initial conditions, it inputs the representative samples, new samples, and the corrected weights into the trained fault diagnosis model for model fine-tuning. The corrected weights are used to calculate the loss function in the model fine-tuning process, and the representative samples and new samples are used as fine-tuning data; A fault diagnosis module, which is used to control the fault diagnosis stage. It inputs the data to be diagnosed into the fault diagnosis model finally obtained after the incremental learning stage for diagnosis and outputs the diagnosis result; In the model training module, the representative samples are selected by using the method based on the optimal sub-optimal distance. The steps include: calculating the highest value of the probability estimation of each sample p Best and the second highest value p SecondBest , and calculating the probability difference value p Best between the highest value p SecondBest and the second highest value x BvSB . From the samples with the same class label and the same domain auxiliary label, the top multiple samples with the highest probability difference value x BvSB are selected as the representative samples under the current fault diagnosis model; When correcting the weights of samples according to the judgment results and the quantitative relationship between the representative samples and the newly added samples, if only the number of samples is increased, and no new categories and domains are added, that is c z ≤ C and d z ≤ D , then the weight λ of the N z newly added samples is corrected to N r / N z , and the weight λ of the representative samples is 1; if new categories are added and no new domains are added, that is c z > C and d z ≤ D , then the weight λ of these N z newly added samples is corrected to αN r / N z , α > 1, and the weight λ of the representative samples is 1. If new domains are added and no new categories are added, that is c z ≤ C and d z > D , then the weight λ of the N z newly added samples is corrected to βN r / N z , β > 1, and the weight λ of the representative samples is 1, where c z is the class label in the newly added samples, N z is the domain auxiliary label as d z of the number of newly added samples, α、β are different preset coefficients, β>α , C represents the number of class labels included in the current representative samples, D represents the number of domain auxiliary labels included in the current representative samples, N r is any class of the representative samples c and the domain auxiliary label is d of the number of samples.

8. A computer-readable storage medium storing a computer program, characterized in that, The computer program, when executed, implements the method according to any one of claims 1 to 6.

Images