Sample adaptive weighted training method for bearing fault diagnosis model under unbalanced data

By adopting adaptive weighting training method in the bearing fault diagnosis model, the sample weight is dynamically adjusted, and the fault diagnosis model training problem under unbalanced data is solved, improving the model performance and rapid optimization capabilities.

CN113505817BActive Publication Date: 2024-11-26ZHEJIANG UTE BEARING
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110672571.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-06-17
Publication Date
2024-11-26
Estimated Expiration
2041-06-17

AI Technical Summary

Technical Problem

In the training of bearing fault diagnosis models under uneven data, data-level and model-level methods are suitable for situations where the degree of imbalance is small. The algorithm-level methods have sample weighting weights that require manual settings, making it difficult to quickly optimize and apply.

Method used

The adaptive weighted training method of bearing fault diagnosis model samples under unbalanced data is adopted. By obtaining training samples and verification samples, feature extraction, prediction label calculation, error calculation and backpropagation update are performed, and the sample weight is dynamically adjusted to adapt to the training process.

Benefits of technology

The sample adaptive weighting of the intelligent fault diagnosis model training process under unbalanced data is realized, which weakens the dominance of most classes on training, and improves the performance and rapid optimization capabilities of the diagnostic model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113505817B_ABST
    Figure CN113505817B_ABST
Patent Text Reader

Abstract

The present invention provides a method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, comprising the following steps: S1. Obtain training samples; S2. Construct a bearing fault diagnosis model; S3. Input the training samples into the bearing fault diagnosis model and extract features from each training sample; S4. Predict the label of each training sample based on the features of each training sample; S5. Calculate the error between the predicted label and the true label, as well as the weight of each training sample; S6. Weight the error based on the weight of each training sample and then back propagate to update the model; S7. Repeat steps S3-S6 until the bearing fault diagnosis model converges or reaches a specified number of iterations. This solution implements adaptive weighting of samples in the training process of the bearing intelligent fault diagnosis model under unbalanced data, which can focus the training process on non-converged samples, weaken the dominance of the majority class in the unbalanced data on the training, and improve the performance of the intelligent diagnosis model under unbalanced data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a sample adaptive weighted training method for a bearing intelligent fault diagnosis model under unbalanced data. Background Art

[0002] Bearings are the core components of mechanical equipment and the "heart" of the slewing support system. Their reliability is crucial to ensuring the performance, function and efficiency of the main machine. Once a bearing in a mechanical equipment fails, it will lead to reduced safety and performance degradation of the equipment. In severe cases, it will cause the loss of equipment function, leading to unexpected downtime, huge production capacity loss and cost loss. Therefore, it is necessary to develop a fault diagnosis method that can effectively identify the health status of bearings.

[0003] In the field of fault diagnosis, the identification of bearing health status is mainly divided into signal processing-based methods and data-driven methods. The signal processing-based method uses various signal processing techniques, such as fast Fourier transform, wavelet decomposition, empirical mode decomposition, etc., to analyze the collected bearing signals and determine whether there are key sensitive features corresponding to different bearing faults in the signal. Since these key sensitive features are only for typical bearing faults under a single condition, without considering factors such as fault distribution attributes and size attributes, and the bearing signal transmission path collected from mechanical equipment is complex and the signal-to-noise ratio is low, the generalization ability of the fault diagnosis method based on signal processing is seriously restricted in practical applications. With the development of sensor technology and communication technology, the monitoring of bearings by modern mechanical equipment has entered the era of big data. These data contain rich bearing operation information and status information. If the key sensitive features of different faults under different conditions (different speeds, different loads, different fault distributions, and different fault sizes) are directly mined from big data, the most similar features under different conditions of the same fault and the most distinguishing features between different faults can be extracted, which can solve the problem of insufficient feature generalization ability of signal processing methods. Therefore, data-driven fault diagnosis methods have become the most popular and cutting-edge technology in the current field of fault diagnosis.

[0004] However, most of the current data-driven methods are based on the assumption that the data of different health states of bearings are balanced, which is unreasonable in practice: on the one hand, bearings run in a healthy state for a long time, and the time of running with faults is short, so most of the collected bearing operation data is bearing health state data, and there is less fault data; on the other hand, affected by the properties of various bearing components, the probability of failure of each component during operation is also different, so the amount of data between different faults in the bearing operation data is also inconsistent. In general, there is a serious imbalance between the operation data of mechanical equipment bearings, which directly affects the application effect of data-driven methods in practice.

[0005] In order to eliminate the impact of using unbalanced data training on the fault diagnosis model, three types of methods have been proposed, namely: data level, model level, and algorithm level. The data level method restores the balance of the bearing monitoring data by performing undersampling on the majority class and oversampling on the minority class, and then trains the fault diagnosis model. The patents published for this type of method include: CN111639461A, CN112257767A, CN111240279A, and CN112395558A. The model level method weakens the negative impact of unbalanced data on the fault diagnosis model by constructing a model that is insensitive to class imbalance. Some patents published for this type of method include: CN112364706A. The algorithm-level method weights samples of different categories to suppress the dominance of categories with more data on the training of the fault diagnosis model, thereby eliminating the impact of unbalanced data on the model. Patents published for this type of method include: CN111881159A, CN111860658A, and CN111340107A.

[0006] However, the above three methods still have various problems:

[0007] (1) Since the data-level method only relies on existing data, when there is less fault data, the information gain that the oversampling method can bring is limited; when the data imbalance is large, there is still a certain imbalance problem after undersampling;

[0008] (2) Since the model-level method only relies on the model’s own optimization properties to adjust the imbalanced data training process, when the data imbalance is large, it is easy for the dominant effect of the majority class on training to be greater than the model’s own adjustment ability.

[0009] In summary, data-level and model-level methods are only applicable to situations with a small degree of imbalance, and their ability to cope with actual big data scenarios is limited. Algorithm-level methods are a fundamental solution to the problem of imbalanced learning because they directly adjust the model's training process. However, the current algorithm-level methods have the problem that sample weights need to be manually set and weight selection is difficult, which is not conducive to the rapid optimization and rapid application of fault diagnosis models. Therefore, there is an urgent need for a new method that can adaptively adjust weights during the training process to facilitate the training of fault diagnosis models under imbalanced data. Summary of the invention

[0010] The purpose of the present invention is to provide a method for adaptively weighting training of bearing fault diagnosis model samples under unbalanced data in order to solve the above problems.

[0011] In order to achieve the above object, the present invention adopts the following technical solutions:

[0012] A method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data includes the following steps:

[0013] S1. Obtain training samples;

[0014] S2. Construct bearing fault diagnosis model;

[0015] S3. Inputting the training samples into the bearing fault diagnosis model, and extracting features from each training sample;

[0016] S4. predicting the label of each training sample according to the characteristics of each training sample;

[0017] S5. Calculate the error between the predicted label and the true label, as well as the weight of each training sample;

[0018] S6. weighting the error according to the weight of each training sample and then back-propagating to update the model;

[0019] S7. Repeat steps S3-S6 until the bearing fault diagnosis model converges or reaches the specified number of iterations.

[0020] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, in step S1, training samples and verification samples are obtained simultaneously, and after step S7, the following steps are further included:

[0021] S7. Input the verification sample into the bearing fault diagnosis model to verify the effectiveness of the model.

[0022] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, in step S1, the original signal during the operation of the bearing is collected by a data acquisition system and / or a sensor acquisition system, and the original signal is divided into the training sample and the verification sample.

[0023] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, in step S1, the original signal is intercepted with a determined length and divided into the training sample and the verification sample.

[0024] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, before step S2, data preprocessing including fast Fourier transform and normalization is performed on the training samples and the verification samples.

[0025] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, the bearing fault diagnosis model includes a feature extraction module and a classification module; in step S3, the feature extraction module performs feature extraction, and the calculation process of feature extraction is as follows:

[0026] All training samples are input into the fault diagnosis model, and the features of the training samples are extracted through the nonlinear transformation of the neural network:

[0027] F=h(X)

[0028] Where: X = [x1; x2; ...; x N ] represents a set of training samples, and X∈R N×d ; h represents the nonlinear mapping from input to feature of the fault diagnosis model; F = [f1; f2; ...; f N ] represents the set of training sample features, and F∈R N ×k .

[0029] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, in step S4, the classification module classifies the features, and step S4 specifically includes:

[0030] Input all features into the classification module to obtain predictions for the training sample labels

[0031]

[0032] Where: Ψ represents the activation function, represents the classification module's prediction of the training sample label, and W represents the classification weight of the classification module, and W∈R k×M , ||W|| represents the vector composed of the modulus of each classification weight, ||W||∈R 1×M .

[0033] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, in step S5, the weight of each training sample is calculated based on directional similarity.

[0034] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, in step S5, the weight of each training sample is calculated in the following manner:

[0035] Calculate the directional similarity between training sample features:

[0036] A=(F·F T ) / .(||F||·||F|| T )

[0037] Among them: ||F|| represents the vector composed of the modulus of each sample feature, and ||F||∈R N×1 ;||F|| T represents the transpose of the corresponding vector; A represents the directional similarity matrix between the features of each sample, and A∈RN×N ; / . Indicates the division of the corresponding bits of the matrix;

[0038] Calculate the directional similarity between each sample feature and the classification weight:

[0039] B=(F·W) / .(||F||·||W||)

[0040] Where: B represents the directional similarity matrix between each sample feature and each classification weight, and B∈R N×M ;

[0041] Convert the sample true label into "one-hot encoding" form

[0042] y i =onehot(y i )

[0043] Where: y i represents the one-hot encoding of the sample label, and y i ∈R 1×M ; onehot represents the one-hot encoding conversion function;

[0044] The one-hot encodings of all samples are integrated to form the one-hot encoding matrix of the real labels of the training samples: Y = [y1; y2; ...; y N ];

[0045] Calculate the mean directional similarity between each sample feature and other similar sample features:

[0046] A mean =mean_row(A⊙(Y·Y T -eye(N)))

[0047] Among them: A mean represents the mean matrix of the directional similarity between each sample feature and other similar sample features, and A mean ∈R N×1 ; mean_row means finding the mean of non-zero elements in a row of a matrix; eye(N) means the unit matrix of size N×N; ⊙ means multiplying corresponding elements;

[0048] Calculate the distribution mean of the directional similarity between each sample feature of each category and other sample features of the same category:

[0049] A mean_mean =(Y·Y T ·A mean ) / .sum_row(Y·Y T )

[0050] Among them: A mean_meanThe distribution mean of the directional similarity between the features of each sample in each category and the features of other samples of the same category, and A mean_mea n∈R N×1 ;

[0051] Calculate the directional similarity between each sample feature and the corresponding category classification weight:

[0052] B sum =sum_row(B⊙Y)

[0053] Among them: B sum ∈R N×1 , sum_row means to find the sum of non-zero elements in a row of the matrix;

[0054] The error weight of each training sample is calculated based on the calculation results of the above directional similarity:

[0055] λ=sigmoid(A mean +A mean_mean ×(sigmoid(B sum -a)-0.5)-b)

[0056] Among them: a and b are hyperparameters.

[0057] In the above-mentioned method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, in step S5, the error between the predicted label and the true label is calculated by the following formula:

[0058]

[0059] l represents the error function, y i represents the true label of sample i, represents the predicted label of sample i;

[0060] In step S6, each sample error is weighted by the following formula:

[0061]

[0062] Where: i represents the error weight of sample i, L i represents the error of sample i, N represents the number of training samples, and M represents the number of categories of bearing health status;

[0063] And in step S6, the weighted result is input into the back propagation algorithm to update the model.

[0064] The advantages of the present invention are: it realizes adaptive weighting of samples in the training process of the bearing intelligent fault diagnosis model under unbalanced data, which can focus the training process on non-converged samples, weaken the dominance of the majority class in the unbalanced data on the training, and improve the performance of the intelligent diagnosis model under unbalanced data. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 The present invention is a flow chart of the method for adaptively weighting training of bearing fault diagnosis model samples under unbalanced data. DETAILED DESCRIPTION

[0066] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0067] like Figure 1 As shown, this embodiment discloses a method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, comprising the following steps:

[0068] 1) Use the data acquisition system and sensor acquisition system to collect the bearing operation process data, that is, the original signal during the bearing operation process. The sensor acquisition system includes various sensors.

[0069] 2) The original signal without any processing is intercepted with a certain length d and divided into training samples D train ={x i ,y i}(i∈{1, 2, ..., N}) and validation samples D test ={x j ,y j}(j∈{1,2,...,L}), x i represents the training sample, y i ∈{1, 2, ..., M} represents the label corresponding to the training sample, x j represents the validation sample, y j ∈{1, 2, ..., M} represents the label corresponding to the verification sample, N represents the number of training samples, M represents the number of bearing health status categories, and L represents the number of verification samples.

[0070] 3) Perform data preprocessing including fast Fourier transform and normalization on all training samples and validation samples. The preprocessing step can be performed based on the difficulty of the diagnostic task.

[0071] 4) Construct a bearing fault diagnosis model including a feature extraction module and a classification module based on deep learning theory. Available deep learning theories include but are not limited to: convolutional neural network, autoencoder, recurrent neural network, etc. Figure 1 ,The feature extraction module includes input layer, convolution layer, pooling layer, fully connected layer, etc. and the classification module includes feature layer, etc.

[0072] 4.1) The calculation process of feature extraction by the feature extraction module is as follows:

[0073] All training samples D train Input the fault diagnosis model, and after the nonlinear transformation of the neural network, the characteristics of the training samples can be obtained as follows:

[0074] F=h(X)

[0075] Where: X = [x1; x2; ...; x N ], represents the set of training samples, and h represents the nonlinear mapping of the fault diagnosis model from input to features; F = [f1; f2; ...; f N ], represents the set of identification features proposed by the bearing fault diagnosis model for the training samples, and .

[0076] 4.2) The calculation process of the classification module is as follows:

[0077] First, normalize the weights of the classification module

[0078]

[0079] Where: W is the classification weight matrix of the classification module, and ; ||W|| represents the vector composed of the modulus of each weight matrix, and .

[0080] Input all features into the classification module to obtain the fault diagnosis model's prediction of the training sample label

[0081]

[0082] Where: Ψ represents the activation function, including but not limited to sigmoid, softmax, etc.; represents the classification module's prediction of the training sample label, expressed in the form of "one-hot encoding", and Represents the prediction of the sample i label, expressed in the form of "one-hot encoding", f i represents the identification features proposed by the bearing fault diagnosis model for the training samples, .

[0083] Calculate the directional similarity between training sample features

[0084] A=(F·F T ) / .(||F||·||F|| T )

[0085] Among them: F T represents the transpose of the feature matrix, ||F|| represents the vector composed of the modulus of each sample feature, and ;||F|| T represents the transpose of the corresponding vector; A represents the directional similarity matrix between the features of each sample, and ; / .Indicates the division of the corresponding bits of the matrix.

[0086] Calculate the directional similarity between each sample feature and the classification weight

[0087] B=(F·W) / .(||F||·||W||)

[0088] Where: B represents the directional similarity matrix between each sample feature and each classification weight, and .

[0089] Convert the sample true label into "one-hot encoding" form

[0090] y i =onehot(y i )

[0091] Where: y i represents the one-hot encoding of the sample label, and ; onehot represents the one-hot encoding conversion function.

[0092] The one-hot encodings of all training samples are integrated to form the one-hot encoding matrix of the true labels of the training samples: Y = [y1; y2; ...; y N ].

[0093] Calculate the mean directional similarity between each sample feature and other similar sample features

[0094] A mean =mean_row(A⊙(Y·Y T -eye(N)))

[0095] Among them: A mean represents the mean matrix of the directional similarity between each sample feature and other similar sample features, and ; mean_row means finding the mean of non-zero elements in a row of the matrix; eye(N) means the unit matrix of size N×N; ⊙ means multiplying corresponding elements.

[0096] Calculate the distribution mean of the directional similarity between the features of each sample in each category and the features of other samples of the same category

[0097] A mean_mean =(Y·Y T ·A mean) / .sum_row(Y·Y T )

[0098] Among them: A mean_mean represents the distribution mean of the directional similarity between the features of each sample in each category and the features of other samples of the same category, and .

[0099] Calculate the directional similarity between each sample feature and the corresponding category classification weight

[0100] B sum =sum_row(B⊙Y)

[0101] Among them: B sum Indicates the directional similarity between each sample feature and the corresponding category classification weight, ; sum_row means to find the sum of non-zero elements in a row of the matrix.

[0102] Calculate the error weight of each sample:

[0103] λ=sigmoid(A mean +A mean_mean ×(sigmoid(B sum -a)-0.5)-b)

[0104] Among them: a and b are hyperparameters. According to the weight calculation process, it can be seen that: when the directional similarity between the sample and the same type of samples is smaller, the sample error weight is larger; when the directional similarity between the sample and the corresponding weight is smaller, the sample weight is larger.

[0105] Calculate the error between the predicted label and the true label of the classification module

[0106]

[0107] l represents the error function, including but not limited to mean square error, root mean square error, cross entropy, etc.; y i represents the true label of sample i, represents the predicted label of sample i.

[0108] Taking the root mean square as an example, the error between the predicted label of the classification module and the actual label is:

[0109]

[0110] Weight the error of each sample and find the total error of the group of training samples:

[0111]

[0112] Among them: mean means to find the mean of all elements in the matrix.

[0113] 4.3) Input the weighted results into the back propagation algorithm to update the model.

[0114] 5) Repeat step 4) until the model converges or the specified number of cycles is reached.

[0115] 6) Verify sample D test Input into the model to verify the validity of the model.

[0116] Through the above training process, it can be found that the sample training weights are calculated based on the convergence characteristics of the samples themselves. The classification characteristics of the samples and the classification characteristics of the categories are taken into account in the calculation. Compared with the traditional method of using fixed weights, the adaptive ability is stronger, and the training of the fault diagnosis model focuses more on non-convergent samples, so the convergence speed is faster, especially weakening the dominance of the majority class in the imbalanced data over the training.

[0117] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0118] Although this article uses more terms such as training samples, verification samples, bearing fault diagnosis models, predicted labels, and true labels, it does not exclude the possibility of using other terms. These terms are used only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional restrictions is contrary to the spirit of the present invention.

Claims

1. A method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data, characterized in that: The following steps are involved: S1. Obtain training samples; S2. Construct bearing fault diagnosis model; S3. Inputting the training samples into the bearing fault diagnosis model, and extracting features from each training sample; S4. predicting the label of each training sample according to the characteristics of each training sample; S5. Calculate the error between the predicted label and the true label, as well as the weight of each training sample; S6. weighting the error according to the weight of each training sample and then back-propagating to update the model; S7. Repeat steps S3-S6 until the bearing fault diagnosis model converges or reaches a specified number of iterations; The bearing fault diagnosis model includes a feature extraction module and a classification module. In step S3, the feature extraction module performs feature extraction, and the calculation process of feature extraction is as follows: All training samples are input into the fault diagnosis model, and the features of the training samples are extracted through the nonlinear transformation of the neural network: F=h(X) Where: X = [x1; x2; ...; x N ] represents the set of training samples, and h represents the nonlinear mapping of the fault diagnosis model from input to features; F = [f1; f2; ...; f N ] represents the set of training sample features, and In step S4, the classification module classifies the features, and step S4 specifically includes: Input all features into the classification module to obtain the prediction of the training sample label Where: Ψ represents the activation function, represents the classification module's prediction of the training sample label, and W represents the classification weight of the classification module, and ||W|| represents the vector composed of the modulus of each classification weight, In step S5, the weight of each training sample is calculated based on the directional similarity; In step S5, the weight of each training sample is calculated in the following way: Calculate the directional similarity between training sample features: A=(F·F T ) / .(||F||·||F|| T ) Among them: ||F|| represents the vector composed of the modulus of each sample feature, and ||F|| T represents the transpose of the corresponding vector; A represents the directional similarity matrix between the features of each sample, and / . Indicates the division of the corresponding bits of the matrix; Calculate the directional similarity between each sample feature and the classification weight: B=(F·W) / .(||F||·||W||) Where: B represents the directional similarity matrix between each sample feature and each classification weight, and Convert the sample true label into "one-hot encoding" form y i =onehot(y i ) Where: y i represents the one-hot encoding of the sample label, and Onehot represents the one-hot encoding conversion function; The one-hot encodings of all samples are integrated to form the one-hot encoding matrix of the real labels of the training samples: Y = [y1; y1; ...; y N ]; Calculate the mean directional similarity between each sample feature and other similar sample features: A mean =mean_row(A⊙(Y·Y T -eye(N))) Among them: A mean represents the mean matrix of the directional similarity between each sample feature and other similar sample features, and mean_row means finding the mean of non-zero elements in a row of a matrix; eye(N) means the unit matrix of size N×N; ⊙ means multiplying corresponding elements; Calculate the distribution mean of the directional similarity between each sample feature of each category and other sample features of the same category: TO mean_mean =(Y·Y T ·TO mean ) / .sum_row(Y·Y T ) Among them: A mean_mean The mean of the distribution of the directional similarity between the features of each sample in each category and the features of other samples of the same category, and Calculate the directional similarity between each sample feature and the corresponding category classification weight: b sum =sum_row(B⊙Y) in: sum_row means to find the sum of non-zero elements in a row of the matrix; The error weight of each training sample is calculated based on the calculation results of the above directional similarity: λ=sigmoid(A mean +A mean_mean ×(sigmoid(B sum -a)-0.5)-b) Among them: a and b are hyperparameters.

2. The method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data according to claim 1 is characterized in that: In step S1, a training sample and a verification sample are obtained simultaneously, and after step S7, the following steps are further included: S7. Input the verification sample into the bearing fault diagnosis model to verify the effectiveness of the model.

3. The method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data according to claim 2 is characterized in that: In step S1, the original signal during the operation of the bearing is collected by a data collection system and / or a sensor collection system, and the original signal is divided into the training sample and the verification sample.

4. The method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data according to claim 3 is characterized in that: In step S1, the original signal is cut into a determined length and divided into the training sample and the verification sample.

5. The method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data according to claim 4 is characterized in that: Before step S2, data preprocessing including fast Fourier transform and normalization is performed on the training samples and the validation samples.

6. The method for adaptive weighted training of bearing fault diagnosis model samples under unbalanced data according to claim 5, characterized in that: In step S5, the error between the predicted label and the true label is calculated by the following formula: l represents the error function, y i represents the true label of sample i, represents the predicted label of sample i; In step S6, each sample error is weighted by the following formula: Where: i represents the error weight of sample i, L i represents the error of sample i, N represents the number of training samples, and M represents the number of categories of bearing health status; And in step S6, the weighted result is input into the back propagation algorithm to update the model.