Rolling bearing intelligent fault diagnosis method for generalization in single field

Abstract

A rolling bearing intelligent fault diagnosis method oriented to generalization in a single field is characterized in that feature extractor branches of different scales are designed, branch features which are different in distribution but contain the same fault information are mixed through adversarial training, and quantized entropy indexes are designed to weight the branch features; a feature extractor can extract robust diagnosis features with cross-branch consistency; and meanwhile, the compactness of the classification result of each branch is improved by using the minimum classification entropy, and the consistency of the classification boundary of each branch is further constrained through a norm, so that the learning effect of the classifier is maximized. According to the method, the problem of effectively generalizing to an unknown target domain for fault diagnosis under the limitation of single source domain data can be solved in a simpler way.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing intelligent fault diagnosis, and in particular to a rolling bearing intelligent fault diagnosis method generalized to a single field. Background Art

[0002] Rolling bearings are one of the most critical components of rotating machinery, and their operating conditions directly affect the overall performance of the equipment. During long-term operation, due to repeated changes in load, speed and other working conditions, rolling bearings are prone to varying degrees of wear or failure, which can easily lead to a decrease in the overall performance of the equipment or even system failure. Therefore, the study of rolling bearing state identification and fault diagnosis has important theoretical value and practical significance.

[0003] In recent years, with the powerful automatic feature extraction capability of deep learning, rolling bearing intelligent fault diagnosis methods have been widely used and achieved excellent performance. An important prerequisite for the success of deep learning is that the training data and test data must have the same data distribution. However, in actual industrial scenarios, bearings operate in complex and time-varying environments. There are significant differences in the distribution of sampled data under different working conditions. This leads to a serious decline in the performance of intelligent fault diagnosis methods.

[0004] In order to solve the problem of cross-domain bearing fault diagnosis caused by changes in working conditions, transfer learning methods based on domain adaptation and domain generalization have developed rapidly in the field of intelligent fault diagnosis. The success of domain adaptation methods relies on the prerequisite that the prior distribution of target domain data can be accessed during the training phase. However, the target domain test data is often invisible in actual transfer diagnosis tasks, which seriously weakens its usability. Intelligent fault diagnosis methods based on domain generalization learn cross-domain invariant general diagnostic knowledge from multiple source domains. Its significant advantage is that no target domain data is required for training. However, limitations are still inevitable, which is caused by the scarcity of available bearing fault data. Considering factors such as data privacy and equipment heterogeneity, it is very difficult for the industry to obtain comprehensive and available fault data from multiple source domains.

[0005] Currently, in order to solve the more challenging situation that the target domain distribution is unknown and only supported by a single source domain data, which is more common in practical applications, some recent single-domain generalized fault diagnosis methods have been proposed. The latest methods rely on generating artificial extended domains from the source domain to obtain domain-invariant features. However, reliable domain-invariant features need to be obtained from real domains with different distributions but containing the same fault information. This means that in order to simulate the real domain, the artificial extended domain must be significantly different from the source domain distribution while inheriting the complete fault information of the source domain, which makes it difficult to guarantee the quality of the artificial extended domain in actual operation. At the same time, the high requirements for the artificial extended domain make its generation process very cumbersome and complicated, which limits the engineering application of intelligent fault diagnosis methods. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a rolling bearing intelligent fault diagnosis method generalized to a single domain, which solves the problem of effectively generalizing to an unknown target domain for fault diagnosis under the limitation of a single source domain data in a simpler way.

[0007] The rolling bearing intelligent fault diagnosis method provided by the present disclosure for a single field generalization includes the following steps:

[0008] S1: The original time domain signal of rolling bearing vibration collected in single source domain is divided into N equal lengths. S samples, and then perform maximum and minimum normalization to obtain the preprocessed source domain dataset

[0009] in:

[0010] Represents the i-th sample in the source domain dataset;

[0011] represents the label information of the i-th sample in the source domain dataset, and K represents the total number of categories of bearing health status;

[0012] S2: Establish an intelligent fault diagnosis model generalized for a single domain, which includes a multi-branch one-dimensional convolutional feature extractor module with different convolution kernel sizes, a discriminator module and a classifier module;

[0013] S3: A feature extractor is used to extract features under different receptive fields from the source domain signal. The feature distributions from different branches are aligned through adversarial training between the feature extractor and the discriminator to extract generalization features that are invariant across branches.

[0014] S4: Introduce entropy method to evaluate and adaptively weight generalization features;

[0015] S5: Send the weighted generalization features of each branch to the classifier, introduce the constraints of minimizing the L1 norm and minimizing the classification entropy in the classifier to align the classification boundaries of the generalization features of each branch and maximize the optimization effect of the classifier;

[0016] S6: Construct the overall optimization goal. Under the guidance of the overall optimization goal, the source domain dataset Input the generalized intelligent fault diagnosis model for a single field for training, and fix the model parameters after the training is completed;

[0017] S7: Collect the original time domain signal of bearing vibration in the unknown target domain, and obtain the target domain sample after the same preprocessing as the source domain. After inputting the trained feature extractor module and classifier module in turn, the rolling bearing fault diagnosis result is obtained by adding the classification results of each classifier.

[0018] Compared with the prior art, the beneficial effects of the present disclosure are: (1) solving the problem of effectively generalizing to an unknown target domain for fault diagnosis under the limitation of single source domain data in a concise manner;

[0019] (2) Different from the mainstream multi-scale feature fusion idea, the present invention designs feature extractor branches of different scales, confuses branch features with different distributions but containing the same fault information through adversarial training, and designs a quantitative entropy index to weight the features of each branch, so that the feature extractor can extract robust diagnostic features with cross-branch consistency;

[0020] (3) The compactness of the classification results of each branch is improved by minimizing the classification entropy, and the consistency of the classification boundaries of each branch is further constrained by the norm, thereby maximizing the learning effect of the classifier;

[0021] (4) There is no need to manually expand the domain and perform tedious and complex design processes. Effective generalization can be performed directly based on a single real source domain data, and the overall architecture is more simplified.

[0022] (5) More suitable for practical industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and other objects, features and advantages of the present disclosure will become more apparent through a more detailed description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present disclosure.

[0024] Figure 1 A flowchart of the training process of the rolling bearing intelligent fault diagnosis method generalized to a single field according to the present disclosure;

[0025] Figure 2 This is a schematic diagram of the fault diagnosis network model training process;

[0026] Figure 3 This is a flow chart of the testing process for the intelligent fault diagnosis method for rolling bearings generalized to a single field. DETAILED DESCRIPTION

[0027] The preferred embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0028] The present disclosure provides a rolling bearing intelligent fault diagnosis method for a single field generalization. According to an exemplary embodiment of the present disclosure, as shown in the attached Figure 1-3 As shown, the specific steps include:

[0029] Step 1: Divide the original time domain signal of rolling bearing vibration collected in the single source domain into N equal lengths S samples, and then perform maximum and minimum normalization to obtain the preprocessed source domain dataset in:

[0030] Represents the i-th sample in the source domain dataset;

[0031] represents the label information of the i-th sample in the source domain dataset, and K represents the total number of categories of bearing health status.

[0032] Step 2: Establish an intelligent fault diagnosis model generalized for a single field, which includes a multi-branch one-dimensional convolution feature extractor module with different convolution kernel sizes, a discriminator module and a classifier module;

[0033] As a preferred embodiment, it specifically includes:

[0034] The feature extractor module F consists of three different one-dimensional convolutional neural network branches (defined as F1, F2 and F3), each with a convolution kernel of different sizes to obtain multi-scale features under different receptive fields that contain the same diagnostic knowledge. Batch normalization layers and activation layers (activation functions are all ReLU functions) are introduced between each convolution layer and pooling layer;

[0035] The discriminator module D consists of two spectral normalization layers (which help to achieve stable adversarial training) and one fully connected layer. The number of neurons in the fully connected layer is three, corresponding to three different branches. Activation layers (activation functions are all ReLU functions) and Dropout operations are introduced between layers.

[0036] The classifier module C contains three fully connected layers. The number of neurons in the last fully connected layer is consistent with the total number of bearing health states K. Batch normalization layers, activation layers (activation functions are all ReLU functions) and Dropout operations are introduced between the fully connected layers. The output of the last fully connected layer is converted into a probability distribution with a sum of 1 through the softmax function, that is, the output of the network model is directly converted into the probabilities of different fault categories, realizing bearing fault classification.

[0037] Step 3: Use a feature extractor to extract features under different receptive fields from the source domain signal. Through adversarial training between the feature extractor and the discriminator, align the feature distributions from different branches to extract generalization features that are invariant across branches.

[0038] Specifically include:

[0039] Step 3.1, embed the gradient reversal layer (GRL) between the feature extractor module F and the discriminator module D. The function of the GRL layer is to make the optimization directions of F1, F2 F3 and D opposite to each other, thus forming an adversarial game.

[0040] In step 3.2, the existence of the GRL layer ensures that during the continuous iterative adversarial training process, D is optimized to identify the source of the feature, while F is optimized to deceive D to achieve the confusion effect. By labeling the branch attributes and setting the optimization goal of D to minimize the cross entropy loss between the branch label and the output prediction value, when the training reaches the Nash equilibrium, F can extract features whose source D cannot identify. Cross entropy loss Defined as:

[0041]

[0042] in:

[0043] Represents the predicted probability of source domain sample i after the feature input D on branch j;

[0044] They represent the labels of the three different scale branches respectively;

[0045] L(·,·) represents the cross entropy loss function;

[0046] Step 3.3, introduce the concept of entropy to further confuse D. When the predicted probabilities of the features belonging to different branch labels are the same, the confusion effect reaches its peak, and the discriminator classification entropy reaches its maximum value. It should be noted that, unlike step 3.2, here D and F1, F2 and F3 have a common optimization goal, which is to maximize the discriminator classification entropy loss.

[0047]

[0048] Represents the discriminant entropy of the sample, that is, the entropy value of the probability distribution obtained after the sample features enter the discriminator.

[0049] Step 4: Introduce entropy method to evaluate and adaptively weight generalization features;

[0050] Specifically include:

[0051] Step 4.1: The maximum value of the discriminator classification entropy in step 3 is a fixed value log(3). The closer the value is to log(3), the higher the cross-branch invariance of the branch j feature of the source domain sample i is, and the larger the weight should be. The weight coefficient is expressed as:

[0052]

[0053] Step 4.2: Adaptively weight the branch features of each sample in the source domain according to the weight coefficient to obtain the weighted features, which are expressed as

[0054] Step 5: Send the weighted generalization features of each branch into the classifier, introduce the constraints of minimizing the L1 norm and minimizing the classification entropy in the classifier to align the classification boundaries of the generalization features of each branch and maximize the optimization effect of the classifier;

[0055] Specifically include:

[0056] Step 5.1: Input the weighted multi-branch features into the classifier. First, ensure that the decision boundary of each branch has a good classification discrimination effect, that is, minimize the cross entropy loss L between the classification prediction value of each branch and the sample label. C :

[0057]

[0058] Step 5.2: Ensure the certainty of the classification prediction results of each branch by minimizing the classification entropy. The classification entropy constraint is defined as:

[0059]

[0060] in:

[0061] Represents the predicted probability of source domain sample i being classified as the kth health state after the feature input C on branch j is output.

[0062] In step 5.3, the L1 norm is introduced and minimized to improve the consistency of decision boundaries between different branches. The consistency loss is defined as:

[0063]

[0064] in:

[0065] L1(·,·) represents the L1 norm.

[0066] Step 6: Construct the overall optimization goal. Under the guidance of the overall optimization goal, the source domain dataset Input the generalized intelligent fault diagnosis model for a single field for training, and fix the model parameters after training. Figure 2 As shown;

[0067] Among them, the overall optimization goal is constructed as:

[0068]

[0069] in:

[0070] θ D ,θ C represents the trainable parameters of the feature extractor (F1, F2 and F3), the discriminator D and the classifier C, and {α, β, ε, δ} are the trade-off parameters.

[0071] Step 7: Collect the original time domain signal of bearing vibration in the unknown target domain, and obtain the target domain sample after the same preprocessing as the source domain After inputting the trained feature extractor module and classifier module in turn, the rolling bearing fault diagnosis result is obtained by adding the classification results of each classifier, as shown in the attached figure. Figure 3 shown.

[0072] The above technical scheme is only an exemplary embodiment of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the application methods and principles disclosed in the present invention, and it is not limited to the method described in the above specific embodiment of the present invention. Therefore, the method described above is only preferred and does not have a restrictive meaning.

Claims

1. A rolling bearing intelligent fault diagnosis method for generalization in a single field, comprising the following steps: S1: Based on the original time domain signal of bearing vibration in a single source domain, obtain the source domain data set in: represents the i-th sample; represents the label information of the i-th sample, and K represents the total number of categories of bearing health status; S2: Establish an intelligent fault diagnosis model generalized for a single domain, which includes a multi-branch one-dimensional convolutional feature extractor module with different convolution kernel sizes, a discriminator module, and a classifier module; S3: A feature extractor is used to extract features in different receptive fields from the source domain signal. Through adversarial training between the feature extractor and the discriminator, generalization features that are invariant across branches are extracted. S4: Introduce entropy method to evaluate and adaptively weight generalization features; S5: Send the weighted generalization features of each branch to the classifier, introduce the constraints of minimizing the L1 norm and minimizing the classification entropy in the classifier, and maximize the optimization effect of the classifier; S6: Constructing the overall optimization goal and using the source domain dataset Training the fault diagnosis model; S7: Collect the original time domain signal of bearing vibration in the unknown target domain, obtain the target domain sample according to the same preprocessing as the source domain, input it into the trained feature extractor module and classifier module in turn, and obtain the rolling bearing fault diagnosis result by adding the classification results of each classifier.

2. The method according to claim 1, characterized in that The specific method of step S1 includes: The original time domain signal of rolling bearing vibration collected in single source domain is divided into N equal lengths. S samples, and then perform maximum and minimum normalization to obtain the preprocessed source domain dataset in: Represents the i-th sample in the source domain dataset; represents the label information of the i-th sample in the source domain dataset, and K represents the total number of categories of bearing health status.

3. The method according to claim 1 or 2, characterized in that: The fault diagnosis model in step S2 specifically includes: Feature extractor module F: It consists of three different one-dimensional convolutional neural network branches F1, F2 and F3. Each branch has a convolution kernel of different sizes to obtain multi-scale features under different receptive fields containing the same diagnostic knowledge. Batch normalization layers and activation layers are introduced between each convolution layer and pooling layer. Discriminator module D: It consists of two spectral normalization layers and one fully connected layer. The number of neurons in the fully connected layer is three, corresponding to three different branches. Activation layers and Dropout operations are introduced between layers. Classifier module C: It contains three fully connected layers. The number of neurons in the last fully connected layer is consistent with the total number of bearing health states K. Batch normalization layers, activation layers and Dropout operations are introduced between the fully connected layers. Through the softmax function, the output of the last fully connected layer is converted into a probability distribution with a sum of 1, that is, the output of the network model is directly converted into the probabilities of different fault categories, thereby realizing bearing fault classification.

4. The method according to claim 3, characterized in that In the intelligent fault diagnosis model in step S2: the activation functions of the activation layer are all ReLU functions.

5. The method according to claim 3, characterized in that: The specific method of step S3 includes: Step S3.1, embed the gradient reversal layer GRL between the feature extractor module F and the discriminator module D. The function of the GRL layer is to make the optimization directions of F1, F2 and F3 opposite to those of D, thus forming a confrontation game; Step S3.2, continuously iterate adversarial training so that D is optimized to identify the source of features, while F is optimized to deceive D to achieve confusion effect; in this process, by labeling branch attributes and setting the optimization goal of D to minimize the cross entropy loss between branch labels and output prediction values, when the training reaches Nash equilibrium, F can extract features whose sources cannot be identified by D; Among them, the cross entropy loss Defined as: in: Represents the predicted probability of source domain sample i after the feature input D on branch j; They represent the labels of the three different scale branches respectively; L(·,·) represents the cross entropy loss function; Step S3.3, introduce the concept of entropy to further confuse D: When the predicted probabilities of features belonging to different branch labels are the same, the confusion effect reaches its peak. At this time, the discriminator classification entropy reaches its maximum value, and D and F1, F2 and F3 have a common optimization goal, that is, maximizing the discriminator classification entropy loss 6. The method according to claim 5, characterized in that The specific method of step S4 includes: Step S4.1, classify entropy according to the discriminator The degree of proximity to its maximum value is given a weight coefficient to indicate the degree of cross-branch invariance of branch j features of source domain sample i. The weight coefficient is expressed as: Step S4.2, adaptively weight the branch features of each sample in the source domain according to the weight coefficient to obtain the weighted features, expressed as 7. The method according to claim 6, characterized in that The specific method of step S5 includes: Step S5.1: Input the weighted multi-branch features into the classifier. First, ensure that the decision boundary of each branch has a certain classification discrimination effect, that is, minimize the cross entropy loss L between the classification prediction value of each branch and the sample label. C : Step S5.2, the certainty of the classification prediction results of each branch is ensured by minimizing the classification entropy constraint. The classification entropy constraint is defined as: in, Represents the predicted probability of source domain sample i being classified as the kth health state after the feature input C on branch j; In step S5.3, the L1 norm is introduced and minimized to improve the consistency of decision boundaries between different branches. The consistency loss is defined as: Here, L1(·,·) represents the L1 norm.

8. The method according to claim 7, characterized in that The overall optimization goal in step S6 is constructed as: in: represents the trainable parameters of the feature extractor (F1, F2 and F3), the discriminator D and the classifier C, and {α, β, ε, δ} are the trade-off parameters.

Images