Device sensor diagnosis method and system based on unsupervised domain adaptation and entropy optimization

Abstract

Unsupervised domain adaptive and entropy optimized equipment sensor diagnosis method and system. The following steps are included: 1) Collect vibration sensor signals of equipment under different working conditions, and pre-process the data using Gram angular field technology; 2) Construct a fault diagnosis model based on convolutional neural network, including feature extraction module, feature consistency judgment module and feature recognition module; 3) Train the labeled samples in the source domain on the feature extraction module and feature recognition module of the fault diagnosis model; 4) Use data enhancement strategy to perform consistency training on the target domain unlabeled training samples on the fault diagnosis model; 5) Input the unlabeled samples in the target domain that are not involved in the training into the trained diagnosis model and obtain the health status of the equipment. By combining RandAugment technology and domain adaptive technology, the recognition accuracy of the model for the health status of the equipment can be improved under the conditions of data imbalance and data distribution shift.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical equipment health management and computer artificial intelligence technology, and in particular relates to an equipment sensing diagnosis method based on unsupervised domain adaptation and convolutional neural network entropy optimization under multiple working conditions and data imbalance. Background Art

[0002] Rolling bearings are one of the most widely used components in mechanical equipment (such as pharmaceutical equipment) but are prone to failure due to minor faults. Fault diagnosis based on equipment sensor information can support the maintenance of reliable operation of the equipment, extend the service life of the equipment, and prevent major accidents. Rolling bearings usually work continuously under random noise, impact loads and thermal stress. Signal processing-based methods can achieve a certain degree of mechanical diagnosis of faulty bearings - by revealing abnormal characteristics of bearings. However, this method has high requirements on the professional knowledge of practitioners. Due to the complexity and variability of the working environment, more effective diagnostic technology is needed to solve the problem. The development of artificial intelligence has promoted the advancement of fault diagnosis technology, and many methods with superior performance have emerged. With the increasing complexity of industrial sensor networks, deep learning methods that are good at processing complex, nonlinear, and variable data have gradually become the mainstream research method.

[0003] However, existing studies have a premise assumption: the distribution of training data and test data is consistent. Due to constraints such as working conditions and sensor performance, there are usually differences in the data distribution of target samples and source domain samples, and data distribution shift will occur. Therefore, the requirement of consistent distribution is too strict in the variable working condition production process of actual industry, and its applicability is limited. In addition, in practical applications, since the normal operation time of the equipment is greater than its failure time, the data collected from healthy samples and faulty samples are often unbalanced, showing a class imbalance phenomenon. However, there is a lack of application research on fault diagnosis methods for mechanical equipment (such as pharmaceutical equipment) under multiple working conditions and data imbalance. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a device sensor diagnosis method based on unsupervised domain adaptation and entropy optimization with good adaptability and high accuracy under multiple working conditions and data imbalance. This method uses a data enhancement strategy that combines RandAugment technology and domain adaptation technology to improve the accuracy of fault diagnosis under data imbalance and multiple working conditions, and to a certain extent alleviates the impact of data imbalance and inconsistent data distribution between the source domain and the target domain on performance degradation.

[0005] The objectives of the present invention can be achieved through the following technical solutions:

[0006] Step 1: Collect vibration sensor signals under multiple working conditions on the equipment. Each working condition contains multiple equipment health states. The data of one working condition is a data domain. The data distribution in different data domains is inconsistent, and the data domain is divided into samples for long-tail distribution.

[0007] Step 2: Preprocess the data, convert the collected one-dimensional data into two-dimensional data through the Gram angular field algorithm, and store it on the disk; build a fault diagnosis model based on convolutional neural network, including feature extraction module, feature consistency judgment module and feature recognition module;

[0008] Step 3: Randomly select a domain from the data repository as the source domain and randomly sample and divide the data set; train the labeled samples on the feature extraction module and feature recognition module of the fault diagnosis model;

[0009] Step 4: Randomly select a target domain from the remaining databases and randomly sample and divide the data set; Based on the model trained in the source domain in step 3, use the data enhancement strategy to train the target domain unlabeled training samples on the fault diagnosis model;

[0010] Step 5: Input the unlabeled samples in the target domain that have not participated in the training into the trained diagnosis model and obtain the health status of the equipment.

[0011] Furthermore, the step 2 specifically includes the following steps:

[0012] The one-dimensional time series vibration signal is collected on the equipment, and the one-dimensional data is preprocessed by the Gram angular field algorithm to transform it into two-dimensional data. The Gram angular field conversion formula is as follows:

[0013]

[0014] Where X = [x1, x2, …, x n ] is n time series data sets, is the regularized data set, α∈[0,π] is the angle value after the regularized data is encoded, t i is the timestamp, r is the polar radius, and N is a constant factor.

[0015] Furthermore, the step 3 specifically includes the following steps:

[0016] Modify the ReNet50 network structure and replace the last linear layer with a fully connected layer with Xavier initialized weights and no bias. Use the modified convolutional neural network to extract features. The process is as follows:

[0017] z(i)=g(w T x(i)+b)#(2)

[0018] where x(i)={x1,x2,…,x n} is n input data, g(*) is the linear unit ReLU activation function, w is the weight matrix, b is the bias term, and z(*) is the output.

[0019] Furthermore, the step 4 specifically includes the following steps:

[0020] Step 4.1: Randomly sample the target domain unlabeled image data to form a training set and a test set;

[0021] Step 4.2: The data enhancement strategy in the present invention consists of RandAugment technology and self-training algorithm;

[0022] Step 4.3: Transform the training set images using RandAugment technology to generate a data set P = {r1(x τ ),r2(x τ ),…r n (x τ )}, where x τ is the unlabeled training data from the target domain, r i is the function of transforming the i-th operation;

[0023] Step 4.4: Input the dataset P into the data augmentation strategy, select the most reliable pseudo sample from the output results, and decide the consistency between its label and the true value label;

[0024] Step 4.5: If the decision result is consistent, minimize the entropy loss value of the pseudo sample that is consistent with the true value label, otherwise maximize the entropy loss value of the inconsistent pseudo sample;

[0025] Step 4.6: Combine the conditional entropy loss of the target domain unlabeled training set and the conditional distribution distance with the classification loss to form the objective function and train the model parameters;

[0026] Furthermore, the consistency in step 4.4 is described as:

[0027]

[0028] Among them, prediction + The generated pseudo-label is consistent with the true value label, prediction - The generated pseudo labels are inconsistent with the true value labels.

[0029] Furthermore, the entropy loss description formula in step 4.5 is:

[0030]

[0031] Among them, L EO is the complete entropy minimization function, is the consistency entropy loss function, is the inconsistency entropy loss function, and the latter two functions are defined as follows:

[0032]

[0033]

[0034] in, is the moving average exponential of the sample label y, and p is the probability distribution function.

[0035] Furthermore, the step 4.6 specifically includes the following steps:

[0036] Step 4.6.1: Calculate the target domain conditional distribution distance. The process description formula is as follows:

[0037]

[0038] Where S is the source domain, is the target domain, X * is the source domain or target domain, k is the number of Gaussian kernels, MMD Hk Represents the kernel space embedding k Gaussian kernels, defined as follows:

[0039]

[0040] Where φ is the nonlinear mapping function of the reproducing kernel Hilbert space H, are samples from the source or target domain.

[0041] Step 4.6.2: Combine the classification cross entropy loss and the conditional distribution distance loss with the conditional entropy loss to form the objective function to be optimized. The formula of the objective function is as follows:

[0042] L final =minL c +αL M +βL EO #(9)

[0043] Among them, α and β are penalty factors.

[0044] The present invention also provides a system for an unsupervised domain adaptive and entropy optimized device sensor diagnosis method, characterized in that the system comprises:

[0045] A feature extraction module is used to extract high-dimensional feature representations from source domain samples of preprocessed images;

[0046] The feature consistency module is used to determine the consistency between the pseudo samples generated by the self-training algorithm and the real samples, and select the most reliable pseudo samples to train the model;

[0047] The feature recognition module is used to identify samples in the target domain that are unlabeled and not involved in training, so as to determine the health category corresponding to the test sample.

[0048] Furthermore, the network frameworks of the feature extraction module and the feature recognition module are both convolutional neural networks, and the basis of the feature consistency module is RandAugment technology and self-training algorithm.

[0049] Beneficial effects of the present invention:

[0050] 1. There is no need to manually extract features. Through deep learning convolutional neural networks, the deep features of the data can be automatically mined, saving manpower and time and improving efficiency.

[0051] 2. The feature consistency module in the present invention can improve the quality of pseudo samples generated by semi-supervised learning and select the most reliable pseudo samples for model training.

[0052] 3. In the framework of convolutional neural network, RandAugment technology and domain adaptation technology are combined to calculate the conditional entropy loss and conditional distribution distance loss and optimize them as part of the objective function. This can effectively break through the limitations of the diagnostic model under multiple working conditions and data imbalance, so that the diagnostic model can adapt well to diagnostic tasks under different working conditions and unbalanced sample sizes, improve the adaptability and generalization ability of the model, and can be widely used in equipment health monitoring tasks in complex systems such as machinery, aviation, and railways. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the fault preprocessing image of the outer ring, inner ring and cage corresponding to the label in the embodiment of the present invention.

[0054] Figure 2 It is a schematic diagram of data transformed by using the RandAugemnt technology in an embodiment of the present invention.

[0055] Figure 3 It is a schematic diagram of the model structure of the deep learning framework in the present invention.

[0056] Figure 4 It is a method flow chart corresponding to the present invention.

[0057] Figure 5 4 is a diagram showing the classification results of the test data in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The embodiments described below are part of the embodiments of the present invention and are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0059] Embodiment 1

[0060] The experimental data adopts the bearing data set of Xi'an Jiaotong University, and the LDK UER204 model AC motor is collected at a sampling frequency of 64 kHz. The fault types mainly occur in the inner ring, outer ring and cage.

[0061] like Figure 4 As shown, the present invention comprises the following steps:

[0062] Step 1: Collect vibration signals under multiple working conditions on the equipment. Each working condition contains multiple equipment health states. The data of one working condition is a data domain. The data distribution in different data domains is inconsistent, and the data domain is divided into samples for long-tail distribution;

[0063] Step 2: Preprocess the data, convert the collected one-dimensional data into two-dimensional data through the Gram angular field algorithm, and store it on the disk; build a fault diagnosis model based on convolutional neural network, including feature extraction module, feature consistency judgment module and feature recognition module;

[0064] Step 3: Randomly select a domain from the data repository as the source domain and randomly sample and divide the data set; train the labeled samples on the feature extraction module and feature recognition module of the fault diagnosis model;

[0065] Step 4: Randomly select a target domain from the remaining databases and randomly sample and divide the data set; Based on the model in step 3, use the data enhancement strategy to train the unlabeled training samples of the target domain on the fault diagnosis model;

[0066] Step 5: Input the unlabeled samples in the target domain that have not participated in the training into the trained diagnosis model and obtain the health status of the equipment.

[0067] Furthermore, the step 1 specifically includes the following steps:

[0068] Step 1.1: This embodiment includes three working conditions, namely 1hp, 2hp, and 3hp. The three working conditions form three data sets, each of which is named after the working condition and contains two fault states, as shown in Table 1;

[0069] Step 1.2: Because some fault categories are more likely to occur than others, a power-law distribution is used to re-partition the dataset sample distribution, including two data imbalance distribution settings;

[0070] Table 1: Dataset details

[0071]

[0072] In the embodiment, the diagnostic task is represented by symbol X AB_C Indicates that the source domain is the Ahp dataset, the target domain is the Bhp dataset, and C is the data imbalance distribution 1 or 2. Under two different data imbalance distributions, the labeled samples in the source domain and the unlabeled training samples in the target domain participate in the training phase, and the unlabeled test samples in the target domain are only used in the testing phase.

[0073] Furthermore, the step 2 specifically includes the following steps:

[0074] The one-dimensional time series vibration signal is collected on the equipment, and the one-dimensional data is preprocessed by the Gram angular field algorithm to transform it into two-dimensional data. The Gram angular field conversion formula is as follows:

[0075]

[0076] Where X = [x1, x2, …, x n ] is n time series data sets, is the regularized data set, α∈[0,π] is the angle value after the regularized data is encoded, t i is the timestamp, r is the polar radius, and N is a constant factor.

[0077] Furthermore, the step 3 specifically includes the following steps:

[0078] The fault diagnosis model established in this embodiment is as follows: Figure 2 As shown in the figure, the fault diagnosis process is as follows Figure 3 shown.

[0079] Modify the ReNet50 network structure and replace the last linear layer with a fully connected layer with Xavier initialized weights and no bias. Use the modified convolutional neural network to extract features. The process is as follows:

[0080] z(i)=g(w T x(i)+b)#(2)

[0081] where x(i)={x1,x2,…,x n} is n input data, g(*) is the linear unit ReLU activation function, w is the weight matrix, b is the bias term, and z(*) is the output.

[0082] Furthermore, the step 4 specifically includes the following steps:

[0083] Step 4.1: Randomly sample the target domain unlabeled image data to form a training set and a test set;

[0084] Step 4.2: The data enhancement strategy in the present invention consists of RandAugment technology and self-training algorithm;

[0085] Step 4.3: Transform the training set images using RandAugment technology to generate a data set P = {r1(x τ ),r2(x τ ),…r n (x τ )}, where x τ is the unlabeled training data from the target domain, r i is the function of transforming i times, and the transformation result of each sample is as follows Figure 4 As shown;

[0086] Step 4.4: Input the dataset P into the self-training algorithm, and select the most reliable pseudo sample from the output results to determine the consistency of its label with the true value label;

[0087] Step 4.5: If the decision result is consistent, minimize the entropy loss value of the pseudo sample that is consistent with the true value label, otherwise maximize the entropy loss value of the inconsistent pseudo sample;

[0088] Step 4.6: Combine the conditional entropy loss of the target domain unlabeled training set and the conditional distribution distance with the classification loss to form the objective function and train the model parameters;

[0089] Furthermore, the consistency in step 4.4 is described as:

[0090]

[0091] Among them, prediction + The generated pseudo-label is consistent with the true value label, prediction - The generated pseudo labels are inconsistent with the true value labels.

[0092] Furthermore, the entropy loss description formula in step 4.5 is:

[0093]

[0094] Among them, L EO is the complete entropy minimization function, is the consistency entropy loss function, is the inconsistency entropy loss function, and the latter two functions are defined as follows:

[0095]

[0096]

[0097] in, is the moving average exponential of sample label y, and p is the probability distribution function.

[0098] Furthermore, the step 4.6 specifically includes the following steps:

[0099] Step 4.6.1: Calculate the target domain conditional distribution distance. The process description formula is as follows:

[0100]

[0101] Where S is the source domain, is the target domain, X * is the source domain or target domain, k is the number of Gaussian kernels, MMD Hk Represents the kernel space embedding k Gaussian kernels, defined as follows:

[0102]

[0103] Where φ is the nonlinear mapping function of the reproducing kernel Hilbert space H, are samples from the source or target domain.

[0104] Step 4.6.2: Combine the classification cross entropy loss and the conditional distribution distance loss with the conditional entropy loss to form the objective function to be optimized. The formula of the objective function is as follows:

[0105] L final =minL c +αL M +βL EO #(9)

[0106] Among them, α and β are penalty factors. The test results under 8 working conditions are as follows Figure 5 As shown. Figure 5 From the classification results, it can be seen that the average diagnostic accuracy of the model under 8 tasks reaches 99.32%, which shows that the present invention can maintain generalization performance and transferability under multiple working conditions and data imbalance conditions, which is conducive to the model monitoring the health status of equipment under the above working conditions.

[0107] The above-described embodiments are part of the embodiments of the present invention, which are used to illustrate the present invention, but are not used to limit the scope of the present invention. For those skilled in the art, the present invention may have various changes and variations; any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. An unsupervised domain adaptive and entropy optimized device sensor diagnosis method, characterized in that: The method comprises the following steps: Step 1: Collect vibration sensor signals under multiple working conditions on the equipment. Each working condition contains multiple equipment health states. The data of one working condition is a data domain. The data distribution in different data domains is inconsistent, and the data domain is divided into samples for long-tail distribution. Step 2: Preprocess the data, convert the collected one-dimensional data into two-dimensional data through the Gram angular field algorithm, and store it on the disk; build a fault diagnosis model based on convolutional neural network, which at least includes a feature extraction module, a feature consistency judgment module and a feature recognition module; Step 3: Randomly select a domain from the data repository as the source domain and randomly sample and divide the data set; train the labeled samples on the feature extraction module and feature recognition module of the fault diagnosis model; Step 4: Randomly select one of the remaining domains in the data repository as the target domain, and randomly sample and divide the data set; Based on the fault diagnosis model trained in the source domain in step 3, the target domain unlabeled training samples are trained on the fault diagnosis model using the data enhancement strategy; The step 4 comprises: Step 4.1: Randomly sample the target domain unlabeled image data to form a training set and a test set; Step 4.2: The data augmentation strategy consists of RandAugment technology and self-training algorithm; Step 4.3: Transform the training set images using RandAugment technology to generate a data set P = {r1(x τ ),r2(x τ ),…r n (x τ )}, where x τ is the unlabeled training data from the target domain, r i is the function of transforming the i-th operation; Step 4.4: Input the dataset P into the data augmentation strategy, select the most reliable pseudo sample from the output results, and decide the consistency between its label and the true value label; Step 4.5: If the decision result is consistent, minimize the entropy loss value of the pseudo sample that is consistent with the true value label, otherwise maximize the entropy loss value of the inconsistent pseudo sample; Step 4.6: Combine the conditional entropy loss of the target domain unlabeled training set and the conditional distribution distance with the classification loss to form the objective function and train the model parameters; The consistency in step 4.4 is described as: Among them, prediction + The generated pseudo-label is consistent with the true value label, prediction - The generated pseudo-labels are inconsistent with the true value labels; The entropy loss description formula in step 4.5 is: Among them, L EO is the complete entropy minimization function, is the consistency entropy loss function, is the inconsistency entropy loss function, and the latter two functions are defined as follows: in, is the moving average index of sample label y, and p is the probability distribution function; Step 5: Input the unlabeled samples in the target domain that have not participated in the training into the trained diagnosis model and obtain the health status of the equipment.

2. The device sensor diagnosis method with unsupervised domain adaptation and entropy optimization according to claim 1 is characterized in that: The step 2 comprises: The one-dimensional time series vibration sensor signal is collected on the equipment, and the one-dimensional data is preprocessed by the Gram angular field algorithm to transform it into two-dimensional data. The Gram angular field conversion formula is as follows: Where X = [x1, x2, …, x n ] is n time series datasets, is the regularized data set, α∈[0,π] is the angle value after the regularized data is encoded, t i is the timestamp, r is the polar radius, and N is a constant factor.

3. The device sensor diagnosis method with unsupervised domain adaptation and entropy optimization according to claim 1 is characterized in that: The step 3 comprises: Modify the ReNet50 network structure and replace the last linear layer with a fully connected layer with Xavier initialized weights and no bias; use the modified convolutional neural network to extract features, the process is as follows: z(i)=g(w T x(i)+b) where x(i)={x1,x2,…,x n } is n input data, g(*) is the linear unit ReLU activation function, w is the weight matrix, b is the bias term, and z(*) is the output.

4. The device sensor diagnosis method with unsupervised domain adaptation and entropy optimization according to claim 1 is characterized in that: The step 4.6 specifically includes the following steps: Step 4.6.1: Calculate the target domain conditional distribution distance. The process description formula is as follows: Where S is the source domain, is the target domain, X * is the source domain or target domain, k is the number of Gaussian kernels, MMD Hk Represents the kernel space embedding k Gaussian kernels, defined as follows: Where φ is the nonlinear mapping function of the reproducing kernel Hilbert space H, are samples from the source or target domain; Step 4.6.2: Combine the classification cross entropy loss and the conditional distribution distance loss with the conditional entropy loss to form the objective function to be optimized. The formula of the objective function is as follows: L final =minL c +αL M +βL EO Among them, α and β are penalty factors.

5. A system using the unsupervised domain adaptation and entropy optimization device sensor diagnosis method according to any one of claims 1 to 4, characterized in that: The system includes: A feature extraction module is used to extract high-dimensional feature representations from source domain samples of preprocessed images; The feature consistency module is used to determine the consistency between the pseudo samples generated by the self-training algorithm and the real samples, and select the most reliable pseudo samples to train the model; The feature recognition module is used to identify samples in the target domain that are unlabeled and not involved in training, so as to determine the health category corresponding to the test sample.

6. The system using unsupervised domain adaptation and entropy optimization device sensor diagnosis method according to claim 5, characterized in that: The network frameworks of the feature extraction module and the feature recognition module are both convolutional neural networks, and the basis of the feature consistency module is the RandAugment technology and the self-training algorithm.

Images