Rolling bearing fault diagnosis method and system based on F&K domain incremental learning

Abstract

The present invention discloses a rolling bearing fault diagnosis method and system based on F&K domain incremental learning, the method comprising: collecting motor rolling bearing data samples under different working conditions that continuously change with time; constructing an expandable framework of a screening column and a knowledge base for the motor rolling bearing data samples under each working condition, and filtering and obtaining the corresponding representative samples of the new working condition by mixing the data samples of each new working condition with the representative samples in the old working condition knowledge base and storing them in the corresponding knowledge base; using the EWC algorithm, using the representative samples of the old working condition knowledge base to train the model and using the representative samples of the new working condition knowledge base to update the model parameters; obtaining the predicted value of the fault category of the new working condition according to the updated model parameters, and obtaining the fault diagnosis result. The present invention allows the model to process data of different working conditions that continuously change with time, and effectively retains the knowledge of the old working condition while learning the knowledge of the new working condition, so as to better perform real-time fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis, and in particular to a rolling bearing fault diagnosis method and system based on F&K domain incremental learning. Background Art

[0002] There are generally three types of fault diagnosis methods for rolling bearings in motors: model-based methods, signal processing methods, and data-driven methods. As a typical data-driven method, deep learning methods have strong feature mining capabilities and are widely used in industrial fault diagnosis scenarios where the system operating state is complex and mechanism modeling is difficult. Industrial machinery and equipment will continue to generate dynamic data information with real-time and variability under time-varying conditions (changing speed or load). Changes in data distribution will lead to differences in the underlying structure of the data space. It is difficult for new and old data to meet the requirements of deep learning for independent and identically distributed data, so it is difficult to use existing models to achieve high-precision fault diagnosis of new data. Retraining new data will lead to catastrophic forgetting of old data, that is, the model cannot adapt to scenarios with continuous dynamic changes.

[0003] To address the problem of generating dynamic data under time-varying working conditions, existing technologies mostly use transfer learning methods, the core of which is to apply knowledge learned in a certain field to another new field. Its main focus is on the performance of knowledge in the new field, rather than the performance of knowledge in the source field. In other words, transfer learning is more of a one-time learning under changing working conditions, and does not have the ability to continuously learn.

[0004] Domain incremental learning methods can be divided into three categories: network structure-based, replay-based, and regularization-based.

[0005] The network structure-based method adjusts the neural network structure so that the model can retain old knowledge to the greatest extent when learning new knowledge. The model usually contains multiple feature extraction layers and classification layers. In order to reduce the redundancy of the model, it has become a common practice to use shared feature extraction layers or classification layers. However, the shared feature extraction layers are usually trained on the old domain, and these features may not be fully adapted to the data distribution of the new domain.

[0006] Replay-based methods usually require a lot of space to store previous data samples for replay during the incremental learning process. Existing methods generally use the method of saving some sample sets of old domains or generating samples to reduce the storage volume, but it is easy to ignore some key and informative samples, resulting in the omission of important knowledge in the old domain. In addition, the representative samples selected during the model training stage cannot fully represent the distribution of new samples, and the failure to continuously update representative samples also reduces the performance of incremental learning.

[0007] Regularization-based methods protect old knowledge from being overwritten by new knowledge by imposing constraints on the loss function of the new domain, thereby alleviating catastrophic forgetting in incremental learning. Regularization methods can be divided into parameter regularization and knowledge distillation. However, most of the data used for knowledge distillation belongs to the new domain, which will cause the model to easily bias towards the new domain and fail to preserve the knowledge of the old domain well. Summary of the invention

[0008] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a method for rolling bearing fault diagnosis based on F&K domain incremental learning, which allows the model to process data of different working conditions that arrive continuously over time, and effectively retains the knowledge of old working conditions while learning new working condition knowledge, so as to better perform real-time fault diagnosis. This is a method that can better overcome the stability-plasticity dilemma.

[0009] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0010] A rolling bearing fault diagnosis method based on F&K domain incremental learning, comprising:

[0011] S101) collecting data samples of motor rolling bearings under different working conditions that change continuously over time;

[0012] S102) constructing a scalable framework of a screening column and a knowledge base for the motor rolling bearing data samples under each working condition, and filtering the data samples of each new working condition with the representative samples in the old working condition knowledge base to obtain the corresponding representative samples of the new working condition and save them in the corresponding knowledge base;

[0013] S103) using the EWC algorithm, using representative samples of the old working condition knowledge base to train the model and using representative samples of the new working condition knowledge base to update the model parameters;

[0014] S104) Obtaining a predicted value of the fault category of the new working condition according to the updated model parameters to obtain a fault diagnosis result.

[0015] Further, the motor rolling bearing data samples in step S101 include fault data samples and normal data samples of the motor rolling bearing vibration acceleration.

[0016] Furthermore, in step S102, when constructing a filter column and an expandable framework of a knowledge base for the motor rolling bearing data sample under each working condition, it includes:

[0017] If the new operating condition is the first operating condition, the motor rolling bearing data samples of the new operating condition are sorted by importance, and a specified first sparsity rate is used to filter the sorted motor rolling bearing data samples to obtain representative samples corresponding to the new operating condition;

[0018] If the new operating condition is not the first operating condition, the representative samples in the knowledge base of the old operating condition and the motor rolling bearing data samples of the new operating condition are mixed and sorted by importance, and the specified second sparsity rate is used to filter from the sorted motor rolling bearing data mixed samples to obtain representative samples corresponding to the new operating condition.

[0019] Furthermore, the importance sorting is specifically performed according to the L1 regularization standard.

[0020] Furthermore, when ranking the importance according to the L1 regularization standard, it includes:

[0021] On the trained network, each data in the motor rolling bearing data sample is traversed to obtain the predicted value of each data, the absolute value of the difference between the predicted value and the label value of each data is calculated, and the absolute value corresponding to each data is added to obtain the standard value of the motor rolling bearing data sample;

[0022] Sort each motor rolling bearing data sample from small to large according to the corresponding standard value.

[0023] Furthermore, the second thinning rate corresponding to each operating condition is different.

[0024] Furthermore, step S103 includes:

[0025] Use representative samples of the old working condition knowledge base to train the model and obtain the parameter vector of the old working condition Where i represents the parameter vector The parameter number in n-1 represents the number of the old working condition, and the cross entropy classification loss function L of the old working condition is calculated. (n-1) (θ); Parameter vector for evaluating the old condition using the Fisher information matrix The importance of each parameter in

[0026] Use the trained model to learn representative samples of the new working condition knowledge base and obtain the parameter vector θ of the new working condition i , calculate the cross entropy classification loss function L for the new working condition (n) (θ) and EWC importance loss function The cross entropy classification loss function L for the new working condition (n) (θ) and EWC importance loss function Add together to get the minimized total loss function L n (θ);

[0027] According to the EWC importance loss function The parameter vector The parameters with high importance in are retained, and the parameter vector The less important parameters in are replaced by the parameter vector θ i The corresponding parameters in the new working condition are obtained, and the parameter vector after the new working condition is updated, so that the loss function of the model is L n (θ).

[0028] Further, step S104 includes: the model after parameter update is classified according to the cross entropy loss function L (n) (θ), classify the representative samples of the new working condition knowledge base respectively, so as to determine whether the type of each sample is a fault data sample or a normal data sample.

[0029] Furthermore, the Fisher information matrix is ​​used to evaluate the parameter vector of the old condition. The importance of each parameter in specifically refers to the calculation of the parameter vector The value of the Fisher information matrix corresponding to each parameter in ; the parameters with high importance are the parameters whose values ​​of the Fisher information matrix are greater than the preset threshold, and the parameters with low importance are the parameters whose values ​​of the Fisher information matrix are less than the preset threshold. The value expression of the Fisher information matrix is ​​as follows:

[0030]

[0031] Among them, x i represents the i-th data in the representative sample of the old working condition, m is the number of data in the representative sample of the old working condition, p(x i |θ) is the probability of the i-th data in the representative sample of the old working condition.

[0032] The present invention also proposes a rolling bearing fault diagnosis system based on F&K domain incremental learning, comprising a microprocessor and a computer-readable storage medium connected to each other, wherein the microprocessor is programmed or configured to execute any one of the rolling bearing fault diagnosis methods based on F&K domain incremental learning.

[0033] Compared with the prior art, the advantages of the present invention are:

[0034] The present invention consists of two parts: an extensible framework (F&K) composed of a screening column and a knowledge base, and an elastic weight consolidation (EWC) parameter update module. In the F&K part, the screening column uses two L1 regularizations (L1-norm) to automatically screen the sampled vibration acceleration data, select representative samples containing important knowledge in different fields, and protect the previously learned knowledge in different fields while learning new knowledge; the knowledge base realizes knowledge sharing between different fields by dynamically integrating knowledge from different domains, so that it can be applied to application scenarios with continuously changing motor working conditions; in addition, in order to improve the scalability of the knowledge base capacity and save memory space, a sparsity rate parameter is introduced to flexibly adjust the number of representative samples stored. In the EWC part, based on the representative samples containing important features in the knowledge base, the EWC algorithm is used, and the Fisher information matrix is ​​used to evaluate the importance of parameters to old working conditions, and the parameter update process is performed based on this evaluation, effectively realizing domain incremental learning. The present invention makes full use of the advantages of the domain incremental learning method to improve the fault diagnosis accuracy of new working conditions while maintaining excellent diagnostic performance in old working conditions, thus overcoming the stability-plasticity dilemma. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a technical solution block diagram of an embodiment of the present invention.

[0036] Figure 2 is a flow chart of an embodiment of the present invention.

[0037] Figure 3 Schematic diagram of two L1-norm screening processes in an embodiment of the present invention.

[0038] Figure 4 Schematic diagram of the EWC in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0040] Before introducing this embodiment, relevant terms are explained first.

[0041] Fault diagnosis: The process of using certain inspection and testing methods to find out whether there is a fault in the system or equipment.

[0042] Domain incremental learning: allows the model to process continuously arriving data with different data distributions and effectively retain the knowledge of the old domain when learning the new domain. The domain refers to the data distribution. In motor fault diagnosis, differences in operating conditions such as speed, load or environmental noise will lead to differences in data distribution, that is, domain differences.

[0043] F&K: refers to the method of filtering and storing data using an extensible framework of filter columns and knowledge base (Filter Columns & Knowledge Base).

[0044] Rolling bearing failure: Typical rolling bearing failures include inner ring failure, outer ring failure, rolling element failure, cage failure, etc.

[0045] Example

[0046] In order to solve the current problems of domain increment based on network structure, we consider designing an extensible framework consisting of filter columns and knowledge bases, and using dynamically adjustable shared feature extraction layers to flexibly adjust the knowledge interaction between different domains. We can effectively carry out domain increment learning and overcome the stability-plasticity dilemma.

[0047] In order to solve the current problems of the replay-based method of domain increment, we consider selectively retaining the key features of different working condition data through a filter column module with fault feature extraction functions of different scales, which helps the model to maintain the protection of old working condition knowledge while learning new working condition knowledge. The sparse rate parameter is introduced to filter the current data and subsequent change data by autonomously setting the size of the sparse rate, which can selectively store important data samples according to storage requirements, effectively reducing storage overhead.

[0048] In order to solve the current problems of regularization-based methods for domain increment, we consider sorting the importance of different working condition data, screening and storing important features, and retaining high-importance data as representative samples stored in the knowledge base, which is conducive to achieving a balance between new and old domain knowledge. Based on the representative samples in the knowledge base, using EWC for parameter update can effectively alleviate catastrophic forgetting.

[0049] Based on the above concept, this embodiment proposes a rolling bearing fault diagnosis method based on domain incremental learning of filter columns and knowledge base (F&K). This fault diagnosis method can use the domain incremental learning method to perform high-precision fault diagnosis on continuously changing different working conditions while retaining the previously learned working condition knowledge. It is a method that can better overcome the dilemma of stability (ability to retain old domain knowledge)-plasticity (ability to integrate new domain knowledge). Figure 1 and Figure 2 As shown, the following steps are included:

[0050] S101) Collecting sample data under different working conditions: Collecting data samples of motor rolling bearings under different working conditions that change continuously over time;

[0051] S102) constructing a screening column and developing a knowledge base: constructing a screening column and an extensible framework of a knowledge base for the motor rolling bearing data samples under each working condition, and filtering and obtaining a corresponding representative sample of the new working condition by mixing the data sample of each new working condition with the representative sample in the knowledge base of the old working condition and saving it in the corresponding knowledge base;

[0052] S103) EWC parameter update: using the EWC algorithm, using representative samples of the old working condition knowledge base to train the model and using representative samples of the new working condition knowledge base to update the model parameters;

[0053] S104) Fault diagnosis: Obtain the predicted value of the fault category of the new working condition according to the updated model parameters to obtain the fault diagnosis result.

[0054] Through the above steps, the method of this embodiment designs an extensible framework consisting of a filter column and a knowledge base, which can better utilize the shared feature extraction layer to flexibly adjust the knowledge interaction of different domains. Figure 1 As shown in the figure, the filter column uses the L1-norm method to sort and filter the vibration acceleration data collected under different working conditions, realize the extraction of fault features of different scales, and selectively retain key features. The knowledge base realizes knowledge sharing between different working conditions by effectively integrating knowledge from different domains. By setting different sparsity rates, important data samples can be selectively stored according to storage requirements. The knowledge of new working conditions can be dynamically learned and the knowledge base can be updated while retaining the knowledge of old working conditions. The selection and storage of representative samples under different working conditions can be completed, and the capacity of the knowledge base can be dynamically adjusted. Memory space can be saved on the basis of overcoming the stability-plasticity dilemma. Combining the scalable framework of the filter column and the knowledge base with the EWC parameter update module can improve computational efficiency and alleviate catastrophic forgetting.

[0055] Each step is described in detail below.

[0056] In step S101 of the present embodiment, the motor rolling bearing data sample includes a fault data sample and a normal data sample of the motor rolling bearing vibration acceleration. Specifically, when collecting the motor rolling bearing data samples under different working conditions, the vibration acceleration sensor is fixed on the motor housing to measure its vibration acceleration signal and set the sampling frequency. Assume that the data of the vibration acceleration signal of n different working conditions (variable speed or variable load) that come in sequence with time are collected and form a database. The data of each working condition contains N samples, and the length of each sample is C. That is, the total amount of data is n*N*C. For each sample in the database, it is necessary to mark it to distinguish between the fault data sample and the normal data sample, and the corresponding text mark can be added, or the corresponding characteristic value can be added. For example, if the characteristic value is 1, it means that the sample is a fault data sample, and if the characteristic value is 0, it means that the sample is a normal data sample.

[0057] In step S102 of this embodiment, the screening column includes two processes: importance sorting and screening. Among them, the standard for importance sorting is L1 regularization (L1-norm), that is, traversing all data on the trained network to obtain its corresponding prediction value, calculating the absolute value of the difference between the prediction value and the label value for each data item in each field, and adding these absolute values ​​to calculate a normative value for each learned sample. The larger the calculated normative value, the greater the difference between the network prediction value and the label value, and the worse the representativeness of the data. Therefore, in the present invention, all field samples are sorted from small to large according to the obtained normative values. The screening process extracts representative samples according to the sparsity rate set by twice L1-norm, and can selectively store important samples according to storage requirements to achieve dynamic adjustment of the knowledge base capacity. The L1-norm calculation formula is as follows:

[0058]

[0059]

[0060] Among them, formula (2) is based on the data x to be learned k and its corresponding label y k The label is a one-hot encoding vector containing the actual category. A domain (i.e., the working condition in this embodiment) contains N samples, and the length of each sample is C. In formula (1), when the sigmoid function is applied in the activation stage, the predicted value f(x k ) is a real-valued vector between 0 and 1.

[0061] Therefore, in step S102 of this embodiment, when constructing a filter column and an expandable framework of a knowledge base for the motor rolling bearing data sample under each working condition, it includes:

[0062] If the new operating condition is the first operating condition, the motor rolling bearing data samples of the new operating condition are sorted by importance, and a specified first sparsity rate is used to filter the sorted motor rolling bearing data samples to obtain representative samples corresponding to the new operating condition;

[0063] If the new operating condition is not the first operating condition, the representative samples in the knowledge base of the old operating condition and the motor rolling bearing data samples of the new operating condition are mixed and sorted by importance, and the specified second sparsity rate is used to filter from the sorted motor rolling bearing data mixed samples to obtain representative samples corresponding to the new operating condition.

[0064] Correspondingly, the importance sorting in step S102 of this embodiment is specifically performed according to the L1 regularization standard. When the importance sorting is performed according to the L1 regularization standard, it includes:

[0065] On the trained network, each data in the motor rolling bearing data sample is traversed to obtain the predicted value of each data, the absolute value of the difference between the predicted value and the label value of each data is calculated, and the absolute value corresponding to each data is added to obtain the standard value of the motor rolling bearing data sample;

[0066] Sort each motor rolling bearing data sample from small to large according to the corresponding standard value.

[0067] like Figure 3 As shown, data from different fields are represented as D 1 , …, D n After the first screening, the representative sample of working condition 1 is D 1 ′, the data after the second screening are D 2 ′,…,D n As a specific embodiment, the screening process of D-L1-norm consists of n different parallel branches, where:

[0068] When L1-norm selects representative samples of the data of working condition 1 for the first time, the sparsity rate is set to α (α∈(0,1]). The total amount of data M stored in the knowledge base after screening is 1 =N*C*α, the total amount of data reduced compared to full storage ΔM 1 =(1-α)*N*C.

[0069] The second L1-norm selects representative samples from the data of working conditions 2 to working conditions n, and the sparsity rate is set to β (β∈(0,1]). The sparsity rate β for data of different working conditions can be set to different values. The sparsity rate β corresponding to working condition 2 1 , and so on, the sparseness rate corresponding to working condition n is β n-1 , the total amount of data stored in the knowledge base after screening is M n =(N*C+M n-1 )*β n-1 , (n=2,3,4,...), the total amount of data reduced compared to full storage is ΔM n =n*N*CM n ,(n=2,3,4,...).

[0070] In step S102 of this embodiment, the knowledge base is used to store representative samples that have been screened and sorted by importance. These samples contain rich information and can effectively represent the data distribution and characteristics under different working conditions. After L1-norm importance sorting, samples with high importance will be retained in the knowledge base first, while samples with low importance can be selectively discarded to reduce storage overhead. The sparsity rate determines the number of samples stored in the knowledge base. The sparsity rate of different working condition data can be set independently according to storage requirements, so that the knowledge base capacity is scalable. The knowledge base realizes knowledge sharing between different working conditions by effectively integrating knowledge of different working conditions. This allows the model to make full use of old working condition knowledge when learning new working condition knowledge, and realize balanced knowledge transfer and migration, which helps to alleviate catastrophic forgetting. Since the representative samples of the old working conditions previously stored in the knowledge base are mixed with the new working condition data and sorted each time the screening is performed, the representative samples finally obtained after two L1-norm screenings not only contain the knowledge learned under the new working condition, but also contain important knowledge of the old working condition. Therefore, the stability-plasticity dilemma can be overcome through step S102. At the same time, by adjusting the sparsity rates α and β, the capacity of the knowledge base can be dynamically adjusted to save memory while maintaining high-precision diagnostic performance.

[0071] In step S103 of this embodiment, EWC is a technology used to avoid catastrophic forgetting of neural networks in incremental learning. It is applied to rolling bearing fault diagnosis in the process of domain incremental learning. Its goal is to retain the knowledge of old working conditions when learning new working condition knowledge to reduce the performance impact of the old working conditions.

[0072] Using Bayes’ rule, we can calculate the conditional probability p(θ|D) using the prior probability p(θ) of the parameter and the probability p(D|θ) of the data. Assume that the data D is divided into two independent parts, one of which is defined as the old condition (D n-1 ), and the other is defined as the new working condition (D n ), we can get the following formula:

[0073] logp(θ|D)=logp(D n |θ)+logp(θ|D n-1 )-logp(D n ) (3)

[0074] In formula (3), the left side describes the posterior probability given the entire data set, while the right side logp(D n |θ) represents the loss function of the new working condition, logp(θ|D n-1 ) represents the posterior probability of containing important parameter information for the old working condition, logp(D n ) is a constant, so the right side depends only on the loss function logp(Dn |θ).

[0075] When learning new conditions, EWC constrains the parameters to The low error area of ​​the old working condition as the center is used to maintain the performance of the old working condition, such as Figure 3 As shown in the figure, if the gradient is updated only based on the knowledge of the new working condition (blue arrow), the minimum loss of the new working condition will be obtained while the knowledge of the old working condition will be forgotten; if the same coefficient (green arrow) is used to constrain each parameter, the model will remember the knowledge of the old working condition while failing to learn the knowledge of the new working condition. In order to overcome the stability-plasticity dilemma, EWC finds the solution to the new working condition by explicitly calculating the importance of the parameters of the old working condition, and the gradient update in the direction of the red arrow will not cause significant losses to the old working condition.

[0076] In this embodiment, step S103 specifically includes:

[0077] (1) Using a representative sample D of the old working condition knowledge base n-1 ′Train the model and obtain the parameter vector of the old working condition Where i represents the parameter vector The parameter number in n-1 represents the number of the old working condition, and the cross entropy classification loss function L of the old working condition is calculated. (n-1) (θ), if the old working condition is the first working condition, that is, working condition 1, the total loss function L of working condition 1 1 (θ) is its cross entropy classification loss function L (1) (θ), it should be noted that the parameter vector after calculating the model training and the cross entropy classification loss function are conventional technical means known to those skilled in the art. The present invention does not involve the improvement of the specific process, and the specific implementation process will not be repeated here;

[0078] (2) Use the Fisher information matrix to evaluate the parameter vector of the old working condition The importance of each parameter in the calculation is as follows: i To evaluate the parameter vector The importance of each parameter to the old working condition, the F corresponding to the most important parameter i The value of is large, and the original value will be kept as much as possible when the parameter is updated; the F corresponding to the less important parameter i The value of is small and will be updated when the parameters are updated. As shown in formula (4), the old working condition is D n-1 ′={x 1 ,x 2 ,...,x m}, m is the number of data.

[0079]

[0080] In the above formula, p(x i |θ) is the i-th data x in the representative sample of the old working condition i The probability of

[0081] (3) As time goes by, new working condition data arrives. The trained model is used to learn representative samples of the new working condition knowledge base to obtain the parameter vector θ of the new working condition. i , calculate the cross entropy classification loss function L for the new working condition (n) (θ) (used to guide the correct classification of samples in new working conditions) and EWC importance loss function (used to guide parameter updates), the calculation formula is as follows:

[0082] L (n) (θ) = logp(D n |θ) (5)

[0083]

[0084] Among them, p(D n |θ) represents the loss function of the new working condition, λ is the regularization coefficient, the purpose is to make the parameter vector θ of the new working condition as much as possible. i Parameter vector close to the old condition To balance the trade-off between learning new working condition knowledge and retaining old working condition knowledge;

[0085] (4) The posterior probability p(θ|D n-1 ) is the key to realize EWC. In this embodiment, we approximate the posterior probability to a Gaussian distribution, thereby converting the cross entropy classification loss function L of the new working condition into (n) (θ) and EWC importance loss function Add together to get the minimized total loss function L n (θ), the expression is as follows:

[0086]

[0087] (5) According to the EWC importance loss function The parameter update process is: the parameter vector The parameters with high importance in are retained, and the parameter vector The less important parameters in are replaced by the parameter vector θ i The corresponding parameters in the new working condition are obtained, and the parameter vector after the new working condition is updated, so that the loss function of the model is L n (θ), where the most important parameter is the Fisher information matrix F i The parameters whose values ​​are greater than the preset threshold and the parameters with low importance are the Fisher information matrix Fi Parameters whose values ​​are less than the preset threshold.

[0088] In step S104 of this embodiment, the model after parameter update is based on the cross entropy classification loss function L (n) (θ), classify the representative samples of the new working condition knowledge base respectively, so as to determine whether the type of each sample is a fault data sample or a normal data sample. It should be noted that the trained model is based on the EWC importance loss function The model after parameter update is classified according to the cross entropy loss function L (n) (θ) Sample classification is a conventional technical means known to those skilled in the art. The present invention does not involve improvements to the specific process, and the specific implementation process will not be repeated here.

[0089] Through the above steps, the trained model builds an extensible framework of filtering columns and knowledge base for the motor rolling bearing data samples under the new working conditions and saves it in the corresponding knowledge base, then classifies the representative samples of the mixed samples of motor rolling bearing data under the new working conditions, and learns the representative samples of the mixed samples of motor rolling bearing data under the new working conditions to update the model parameters.

[0090] Embodiment 2

[0091] This embodiment proposes a rolling bearing fault diagnosis system based on F&K domain incremental learning, including a microprocessor and a computer-readable storage medium connected to each other, and the microprocessor is programmed or configured to execute the rolling bearing fault diagnosis method based on F&K domain incremental learning described in Example 1.

[0092] The system of this embodiment includes four parts: data acquisition module, filter column module, knowledge base module and EWC parameter update module.

[0093] The data acquisition module is used to execute step S101 of the first embodiment to collect data samples of the motor rolling bearing under different working conditions that continuously arrive over time;

[0094] The filter column module is used to execute the step of constructing the filter column in step S102 of the first embodiment, that is, constructing a filter column for the motor rolling bearing data sample under each working condition to obtain a representative sample corresponding to each working condition;

[0095] The knowledge base module is used to execute the step of developing the knowledge base in step S102 of the first embodiment, that is, to store the representative samples corresponding to each working condition in the knowledge base corresponding to each working condition;

[0096] The EWC parameter updating module is used to execute step S103 of the first embodiment, using the EWC algorithm, using representative samples of the old working condition knowledge base to train the model and using representative samples of the new working condition knowledge base to update the model parameters.

[0097] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of 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 rolling bearing fault diagnosis method based on F&K domain incremental learning, characterized in that: include: S101) collecting data samples of motor rolling bearings under different working conditions that change continuously over time; S102) constructing a scalable framework of a screening column and a knowledge base for the motor rolling bearing data samples under each working condition, and filtering the data samples of each new working condition with the representative samples in the old working condition knowledge base to obtain the corresponding representative samples of the new working condition and save them in the corresponding knowledge base; S103) using the EWC algorithm, using representative samples of the old working condition knowledge base to train the model and using representative samples of the new working condition knowledge base to update the model parameters; S104) Obtaining a predicted value of the fault category of the new working condition according to the updated model parameters to obtain a fault diagnosis result.

2. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 1 is characterized in that: The motor rolling bearing data samples in step S101 include fault data samples and normal data samples of the motor rolling bearing vibration acceleration.

3. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 1 is characterized in that: In step S102, when constructing a filter column and an expandable framework of a knowledge base for the motor rolling bearing data sample under each working condition, it includes: If the new operating condition is the first operating condition, the motor rolling bearing data samples of the new operating condition are sorted by importance, and a specified first sparsity rate is used to filter the sorted motor rolling bearing data samples to obtain representative samples corresponding to the new operating condition; If the new operating condition is not the first operating condition, the representative samples in the knowledge base of the old operating condition and the motor rolling bearing data samples of the new operating condition are mixed and sorted by importance, and the specified second sparsity rate is used to filter from the sorted motor rolling bearing data mixed samples to obtain representative samples corresponding to the new operating condition.

4. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 3 is characterized in that: The importance sorting is specifically performed according to the L1 regularization standard.

5. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 4 is characterized in that: When ranking by importance according to the L1 regularization criterion, it includes: On the trained network, each data in the motor rolling bearing data sample is traversed to obtain the predicted value of each data, the absolute value of the difference between the predicted value and the label value of each data is calculated, and the absolute value corresponding to each data is added to obtain the standard value of the motor rolling bearing data sample; Sort each motor rolling bearing data sample from small to large according to the corresponding standard value.

6. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 3 is characterized in that: The second thinning rate corresponding to each operating condition is different.

7. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 1 is characterized in that: Step S103 includes: Use representative samples of the old working condition knowledge base to train the model and obtain the parameter vector of the old working condition Where i represents the parameter vector The parameter number in n-1 represents the number of the old working condition, and the cross entropy classification loss function L of the old working condition is calculated. (n-1) (θ); Parameter vector for evaluating the old condition using the Fisher information matrix The importance of each parameter in Use the trained model to learn representative samples of the new working condition knowledge base and obtain the parameter vector θ of the new working condition i , calculate the cross entropy classification loss function L for the new working condition (n) (θ) and EWC importance loss function The cross entropy classification loss function L for the new working condition (n) (θ) and EWC importance loss function Add together to get the minimized total loss function L n (θ); According to the EWC importance loss function The parameter vector The parameters with high importance in are retained, and the parameter vector The less important parameters in are replaced by the parameter vector θ i The corresponding parameters in the new working condition are obtained, and the parameter vector after the new working condition is updated, so that the loss function of the model is L n (θ).

8. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 7 is characterized in that: Step S104 includes: the model after parameter update is classified according to the cross entropy loss function L (n) (θ), classify the representative samples of the new working condition knowledge base respectively, so as to determine whether the type of each sample is a fault data sample or a normal data sample.

9. The rolling bearing fault diagnosis method based on F&K domain incremental learning according to claim 7 is characterized in that: Using the Fisher Information Matrix to Evaluate the Parameter Vector of the Old Case The importance of each parameter in specifically refers to the calculation of the parameter vector The value of the Fisher information matrix corresponding to each parameter in ; the parameters with high importance are the parameters whose values ​​of the Fisher information matrix are greater than the preset threshold, and the parameters with low importance are the parameters whose values ​​of the Fisher information matrix are less than the preset threshold. The value expression of the Fisher information matrix is ​​as follows: Among them, x i represents the i-th data in the representative sample of the old working condition, m is the number of data in the representative sample of the old working condition, p(x i |θ) is the probability of the i-th data in the representative sample of the old working condition.

10. A rolling bearing fault diagnosis system based on F&K domain incremental learning, characterized in that: It comprises a microprocessor and a computer-readable storage medium connected to each other, wherein the microprocessor is programmed or configured to execute the rolling bearing fault diagnosis method based on F&K domain incremental learning as claimed in any one of claims 1 to 9.

Images