Data self-correction fault diagnosis method and system

Through the self-correction fault diagnosis method of data, self-learning feature extraction and generation adversarial networks are used to reconstruct and label filling, solving the problems of incomplete data and lack of labels in special equipment fault diagnosis, improving diagnostic accuracy and efficiency. Enhanced information security.

CN118916753AInactive Publication Date: 2024-11-08YANGTZE DEITA GRADUATE SCHOOI OF BEIJING INST OF TECH (JIAXING) +1
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411406585.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-10
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the diagnosis of special equipment faults, the lack of label data and the insufficient generalization ability of traditional methods, the diagnosis accuracy and efficiency are low, and information security issues are involved.

Method used

Data self-correction fault diagnosis method is adopted, and data reconstruction is carried out through self-learning feature extraction network and generative adversarial network. Combined with semi-supervised label filling technology, a generative adversarial network model with integrated identity map jump connection blocks is constructed, and Euclidean distance loss is used. , sample frequency domain feature loss and sample feature frequency loss are optimized for parameter optimization to improve data reconstruction quality and label filling accuracy.

Benefits of technology

It significantly improves the accuracy and operating efficiency of special equipment fault diagnosis, reduces maintenance costs, improves information security, overcomes the problems of incomplete data and lack of labels, and enhances the generalization ability and diagnostic performance of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916753A_ABST
    Figure CN118916753A_ABST
Patent Text Reader

Abstract

The invention discloses a data self-correction fault diagnosis method and system, and relates to the technical field of fault diagnosis, and the method specifically comprises the steps: 1, obtaining motor fault data, carrying out the data integrity judgment of the motor fault data, and executing the step 2 if the motor fault data is not complete; if so, executing the step 3; step 2, inputting the motor fault data into a data reconstruction model to obtain reconstruction data, and executing step 3; and step 3, inputting the motor fault data or the reconstruction data into a label completion and classification model for feature extraction and fault prediction, and outputting a predicted fault. According to the invention, the performance and accuracy of the special equipment fault diagnosis system under the condition of imperfect information are obviously improved; and the operation efficiency of the special equipment and the safety of the equipment under the condition of information loss are greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and more particularly to a data self-correction fault diagnosis method and system. Background Art

[0002] Special equipment plays a vital role in key areas such as military, aviation and special industrial applications. Its stability and safety are essential to ensure the smooth progress of operations and tasks. The purpose of special equipment fault diagnosis technology is to detect and identify potential faults in a timely manner to prevent major safety accidents. In recent years, fault diagnosis methods that integrate artificial intelligence and big data technologies have shown great potential. However, fault diagnosis of special equipment faces many challenges in practical applications, especially data incompleteness. Due to factors such as sensor failure, data transmission interruption or multi-rate sampling, the obtained operating data is often missing or incomplete, which poses a great challenge to data-driven fault diagnosis methods. In addition, high-quality fault diagnosis model training requires a large amount of labeled data, and the labeling process relies on expert experience, which is costly and inefficient, and involves sensitive information security issues, which limits the application of information technology in this field.

[0003] Traditional fault diagnosis methods based on expert experience and simple models often have limited generalization capabilities when dealing with complex working conditions and incomplete data, and are difficult to meet actual needs. How to overcome the problems of incomplete data, lack of labeled data, and insufficient generalization capabilities of traditional methods is an issue that technicians in this field need to solve urgently. Summary of the invention

[0004] In view of this, the present invention provides a data self-correction fault diagnosis method and system to overcome the above-mentioned defects.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A data self-correction fault diagnosis method, the specific steps are:

[0007] Step 1: Obtain motor fault data, and perform data integrity judgment on the motor fault data. If the data is incomplete, execute step 2; if the data is complete, execute step 3;

[0008] Step 2: input the motor fault data into a data reconstruction model to obtain reconstructed data, and execute step 3;

[0009] Step 3: Input the motor fault data or the reconstructed data into a label completion and classification model for feature extraction and fault prediction, and output the predicted fault.

[0010] Optionally, the steps of constructing the data reconstruction model are:

[0011] Extracting fault features from a complete data set based on multiple parallel self-learning feature extraction networks, and generating a fault feature encoder using the fault features;

[0012] Constructing a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator;

[0013] The fault feature encoder is introduced into the generator; and identity mapping jump connection blocks are added to both the generator and the discriminator to obtain the data reconstruction model.

[0014] Optionally, during the training process of the data reconstruction model, the fault feature encoder and the generator are collaboratively trained together.

[0015] Optionally, during the training process, the data reconstruction model performs parameter adjustment and optimization based on an adversarial loss function, a Euclidean distance loss function, a sample feature frequency loss function, and a sample frequency domain feature loss function.

[0016] Optionally, the steps of constructing the label completion and classification model are:

[0017] Acquire source domain data and target domain data, and identify and link labels of the source domain data and the target domain data to obtain labeled samples and unlabeled samples;

[0018] Pre-training the self-learning feature extraction network using the label samples to obtain a feature extraction model;

[0019] Extracting target features from the target domain data using the feature extraction model;

[0020] Inputting the extracted target features into a multi-layer perceptron classifier for category prediction, and selecting pseudo-label samples according to the prediction confidence;

[0021] The self-learning feature extraction network and the multi-layer perceptron classifier are trained using the label samples and the pseudo-label samples in the source domain data to obtain the label completion and classification model.

[0022] Optionally, in the training of the label completion and classification model, parameter adjustment and optimization are performed based on feature differentiation loss function, cross loss function, consistency regularization loss function and LMMD loss function.

[0023] Optionally, the self-learning feature extraction network includes an input layer, a preprocessing layer, a deep feature extraction layer, a multi-scale pooling layer, a feature fusion layer, a fully connected layer and an output layer which are connected in sequence.

[0024] A data self-correction fault diagnosis system, comprising:

[0025] Data acquisition module, used to collect motor fault data;

[0026] An integrity judgment module, used for judging the data integrity of the motor fault data;

[0027] A data reconstruction module, used to reconstruct the missing motor fault data to generate reconstructed data;

[0028] The fault prediction module is used to extract features and predict faults of the complete motor fault data or the reconstructed data, and output the predicted faults.

[0029] It can be seen from the above technical solutions that the present invention discloses a data self-correction fault diagnosis method and system, which has the following beneficial effects compared with the prior art:

[0030] 1. Based on the self-learning feature extraction network, it can accurately extract various fault data features including frequency band energy, spectrum peak, power spectrum density, root mean square amplitude, waveform index, pulse index, peak index, margin index, kurtosis index, etc. This unique feature extraction capability avoids the common problem of lack of key fault feature frequency in the data when using generative adversarial network to generate data.

[0031] 2. Based on the constructed generative adversarial network model with integrated identity mapping jump connection blocks, the generative adversarial network training collapse caused by gradient vanishing and the network degradation problem caused by increased network depth are effectively prevented; at the same time, in order to reconstruct the original data more accurately, Euclidean distance loss, sample frequency domain feature loss and sample feature frequency loss are introduced, so that the semi-supervised data reconstruction model can more effectively restore high-quality data, thereby significantly improving the accuracy of diagnosis.

[0032] 3. Advanced semi-supervised label completion technology is used to effectively complete missing labels. The distribution between subdomains is optimized by introducing the local maximum average difference, and the cross entropy loss and feature differentiation loss are combined to further enhance the learning efficiency and generalization ability of label completion and classification models. In addition, the introduced consistency regularization loss ensures the high quality of label completion and effectively avoids the interference of noise labels.

[0033] 4. The present invention significantly improves the performance and accuracy of the special equipment fault diagnosis system under imperfect information conditions; it greatly improves the operating efficiency and equipment safety of special equipment under information loss conditions, and has important practical significance for improving production efficiency and reducing maintenance costs in industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0035] Figure 1 A schematic diagram of the method provided by the present invention;

[0036] Figure 2 A schematic diagram of the integrated identity mapping jump connection block structure provided by the present invention;

[0037] Figure 3 A network training flow chart provided by the present invention;

[0038] Figure 4 A schematic diagram of the Euclidean distance loss provided by the present invention;

[0039] Figure 5 A schematic diagram of sample characteristic frequency loss provided by the present invention;

[0040] Figure 6 A schematic diagram of fault classification provided by the present invention;

[0041] Figure 7 Schematic diagram showing the comparison of learning performance between the method of the present invention and the existing incomplete special equipment fault information reconstruction method. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] This embodiment discloses a data self-correction fault diagnosis method. Figure 1 As shown, the specific steps are:

[0044] Step 1: Obtain the motor fault data and determine the data integrity of the motor fault data. If the data is incomplete, proceed to step 2; if the data is complete, proceed to step 3.

[0045] Step 2: Input the motor fault data into the data reconstruction model to obtain the reconstructed data, and then execute step 3;

[0046] Step 3: Input the motor fault data or reconstructed data into the label completion and classification model for feature extraction and fault prediction, and output the predicted fault.

[0047] In one embodiment, the steps of constructing the data reconstruction model are:

[0048] Step 21: extracting fault features from the complete data set based on multiple parallel self-learning feature extraction networks, and generating a fault feature encoder using the fault features;

[0049] Step 22: construct a generative adversarial network, which includes a generator and a discriminator;

[0050] Step 23: Import the fault feature encoder into the generator; and add an identity mapping jump connection block to both the generator and the discriminator to obtain a data reconstruction model.

[0051] Furthermore, this embodiment embeds the fault feature encoder into the generator of the generative adversarial network to reconstruct data using unlabeled data, specifically:

[0052] The fault feature encoder, whose structure is similar to the deep feature extraction layer in the self-learning feature extraction network, compresses the raw sensor data obtained from the complete data set into a low-dimensional feature vector to capture the key information of the operating status of the special equipment. Subsequently, the real sample feature encoder is imported into the residual block generator with the identity mapping together with the low-dimensional feature vector (instead of random noise), so that the generator generates samples according to the extracted fault feature information.

[0053] In one embodiment, during the training process of the data reconstruction model, the fault feature encoder and the generator are collaboratively trained together.

[0054] Furthermore, during the training process, the fault feature encoder is trained collaboratively with the generator of the generative adversarial network. The fault feature encoder focuses on extracting richer and more representative fault features, so that the generator can generate more realistic samples that are closer to the real data distribution. Through this method, the present invention can significantly improve the quality of generated samples, enhance the generalization ability of the model, and reduce the dependence on labeled data, thereby more effectively reconstructing data, improving the model's ability to process incomplete special equipment fault information, and ultimately improving the accuracy and reliability of fault diagnosis.

[0055] Furthermore, in order to prevent the problem of the generative adversarial network crashing due to gradient disappearance during training and the network degradation caused by the increase of network depth, an identity mapping jump connection block is added to both the generator and the discriminator. The structure is as follows: Figure 2 shown.

[0056] In one embodiment, during the training process, the data reconstruction model performs parameter adjustment and optimization based on the adversarial loss function, the Euclidean distance loss function, the sample feature frequency loss function, and the sample frequency domain feature loss function.

[0057] Furthermore, the Euclidean distance loss is used to minimize the direct geometric distance between the generated samples and the real samples, which helps the generated samples to be closer to the real samples. Its loss function can be expressed as:

[0058] ;

[0059] In the formula, Represents the output of the generator (i.e. generated samples); represents the real sample; express norm; In the real sample and a random noise vector The expected value calculated above.

[0060] The sample frequency domain feature loss is mainly used to enhance the generator's ability to reproduce the real sample features in the frequency domain. This loss helps to better preserve the frequency domain features of the real sample. Its loss function can be expressed as:

[0061] ;

[0062] In the formula, represents the operation of converting samples to the frequency domain (using Fourier transform), represents the output of the generator (i.e., generated samples), represents the real sample; express Norm, which is the square of the Euclidean distance.

[0063] The sample feature frequency loss aims to minimize the difference between the generated samples and the real samples in specific frequency features. This is achieved by comparing their spectral responses. Its loss function can be expressed as:

[0064] ;

[0065] In the formula, represents a conversion function, which extracts the key frequency characteristics of the real sample. When processing the fault signal in this embodiment, the envelope spectrum is used to extract the key frequency.

[0066] In one embodiment, the steps of constructing the label completion and classification model are:

[0067] Step 31: Obtain source domain data and target domain data, and identify and link the labels of the source domain data and the target domain data to obtain labeled samples and unlabeled samples;

[0068] Step 32: pre-train the self-learning feature extraction network using the label samples to obtain a feature extraction model;

[0069] Step 33: extract target features from the target domain data using the feature extraction model;

[0070] Step 34: Input the extracted target features into the multi-layer perceptron classifier for category prediction, and select pseudo-label samples according to the prediction confidence;

[0071] Step 35: Use the labeled samples and pseudo-label samples in the source domain data to train the self-learning feature extraction network and the multi-layer perceptron classifier to obtain a label completion and classification model.

[0072] Furthermore, the labels of all data (including source domain data and target domain data) are first identified, and the labeled data are linked. Then, the self-learning feature extraction network is pre-trained on the labeled samples of the source domain and the target domain to obtain the feature extraction model. The pre-trained feature extraction model is used to extract features from all samples in the target domain. The extracted features are input into the multi-layer perceptron classifier for category prediction. Pseudo-label samples are selected according to the prediction confidence. The entire network is fine-tuned together with the labeled samples in the source domain. The parameters of the multi-layer perceptron classifier and the feature extraction model are updated. The fine-tuned network is used to extract features and classify the test samples in the target domain.

[0073] In one embodiment, in order to solve the class alignment problem, a multi-layer local maximum average difference module placed behind the MLP classifier is designed to measure the structural differences between domains, and a linear combination of Gaussian kernels is used to effectively determine the distribution differences between related subdomains. In the loss function design, in addition to the feature differentiation loss and the local maximum average difference loss, the prediction loss (cross entropy loss) and consistency regularization loss of the multi-layer perceptron classifier are also added to ensure that the multi-layer perceptron classifier can adapt well to the target domain data.

[0074] Furthermore, a local maximum mean difference module is introduced to adjust the distribution of relevant subdomains. The local maximum mean difference module calculates the Hilbert-Schmidt norm between the kernel mean embeddings of the empirical distribution of relevant subdomains in the source and target domains by considering the weight of each sample.

[0075] To adjust the distribution of relevant subdomains, the local maximum average difference can be defined as follows:

[0076] ;

[0077] In the formula, Indicates the target domain Subset of or The expected value of the feature map of the sample in; Indicates the source domain The expected value of the feature map of the sample in; is to transform the original data into the reproducing kernel Hilbert space The feature mapping function of and Represents the category label;

[0078] Unlike adaptive methods such as MMD that calculate global distribution differences, local maximum average difference techniques can calculate the expected value of local distribution differences. Assume that each sample is assigned a weighted category In this case, the following is the definition of the unbiased estimator of the equation:

[0079] ;

[0080] In the formula, and Respectively and The number of samples in ; Indicates that the The weight of each sample in is assigned to The weight of each sample in .

[0081] Furthermore, in addition to introducing the local maximum average difference module to adjust the distribution of related subdomains, the cross entropy loss and feature differentiation loss are also introduced to further improve the performance of label completion and classification models.

[0082] Among them, the cross entropy loss is defined as:

[0083] ;

[0084] in, represents the true label, represents the predicted label, Indicates the number of categories.

[0085] The feature differentiation loss is defined as:

[0086] ;

[0087] in, represents the features extracted by the self-learning extraction network, represents the discriminator, and Represent the positive and negative classes respectively; Represents the probability distribution of features when processing positive sample data; Represents the probability distribution of features when processing negative sample data.

[0088] Furthermore, we introduce consistency regularization loss to ensure the quality of label completion and avoid introducing too many noisy labels:

[0089] For each unlabeled sample, SLCN performs data augmentation to obtain different variants. SLCN requires the model to give consistent prediction results for these variants. The consistency regularization loss is defined as:

[0090] ;

[0091] In the formula, represents the number of unlabeled samples, and Represent the predicted probability distribution of the original sample and the enhanced sample respectively, Represents divergence.

[0092] In one embodiment, the self-learning feature extraction network includes an input layer, a preprocessing layer, a deep feature extraction layer, a multi-scale pooling layer, a feature fusion layer, a fully connected layer and an output layer connected in sequence.

[0093] Furthermore, the self-learning feature extraction network uses a multi-layer structure to extract fault features from multiple angles. The training process includes:

[0094] First, the input layer receives raw signal data from multiple sensors and segments them. The preprocessing layer normalizes each data segment and extracts statistical features in the time, waveform, and frequency domains of short, medium, and long time domains through sliding windows of different sizes. The deep feature extraction layer contains three sets of parallel multi-scale convolutional layers, which process feature data of different scales respectively. Each set of convolutional layers uses increasing convolution kernel size and ReLU activation function to enhance nonlinear expression capabilities.

[0095] Subsequently, the multi-scale pooling layer reduces the dimension of the features, and the feature fusion layer splices the feature maps of different scales to obtain a more comprehensive feature expression. Finally, the fully connected layer performs nonlinear transformation and feature abstraction on the fused feature vector, and the number of neurons in the output layer depends on the specific task. During the training process, each self-learning feature extraction network is trained using labeled data from a single sensor, enabling it to accurately identify different types of special equipment failures.

[0096] This embodiment also discloses a data self-correction fault diagnosis system, including:

[0097] Data acquisition module, used to collect motor fault data;

[0098] An integrity judgment module, used to judge the data integrity of motor fault data;

[0099] A data reconstruction module is used to reconstruct missing motor fault data and generate reconstructed data;

[0100] The fault prediction module is used to extract features and predict faults of the complete motor fault data or reconstructed data, and output the predicted faults.

[0101] This embodiment also discloses a specific application step, such as Figure 3 As shown, specifically:

[0102] 1. Unsupervised data reconstruction model:

[0103] Self-learning feature extraction: Learn fault feature representation from a complete special equipment fault data set to provide high-quality input for subsequent generative adversarial networks. Extract the time series and frequency domain features of the original data. Parallel self-learning feature extraction network , respectively This corresponds to a sensor source. The self-learning feature extraction networks have the same structure and do not share parameters with each other. Indicates The self-learning feature mapping network corresponding to each sensor source, Represents the parameters to be learned.

[0104] In order to consider the completeness of the extracted features, from the complete data set The statistical features involved in feature extraction include mean value, root mean square value, standard deviation, variance, third-order moment, fourth-order moment, waveform index, pulse index, peak index, margin index, etc. Each self-learning feature extraction network consists of three layers of convolutional layers and three layers of pooling layers stacked crosswise, which is used to extract the fault features of each sensor data.

[0105] In each convolutional layer, the input data is first convolved using a one-dimensional convolution kernel. The kernel size is set to 128, the number of kernels decreases step by step according to the number of layers, and the number of filters increases step by step.

[0106] After the convolutional layers, a batch normalization (BN) layer is employed to reduce the shift of the internal covariance.

[0107] Subsequently, the output of the convolution operation is activated nonlinearly to enhance the convergence of the model.

[0108] After nonlinear activation, maximum pooling is performed to reduce the size of fault feature maps and improve network computational efficiency. The convolutional layer has an output for:

[0109] ;

[0110] ;

[0111] In the formula, o j Represents the input of the convolution operation; Represents the input of the convolutional layer; represents the nonlinear correction unit activation function; Represents the output of the convolution operation; represents a learnable convolution kernel, ,in, represents the number of convolution kernels, Indicates the number of channels of the convolution kernel, Represents the dimension of the convolution kernel; Represents the bias vector of the convolutional layer. The output of the convolutional layer It will be used as the input of the subsequent pooling layer, and the first The output of the layer pooling layer is:

[0112] ;

[0113] In the formula, represents the maximum pooling function; Indicates the pooling size; Indicates the pooling sliding step size. In this embodiment .

[0114] For Self-learning feature mapping network , whose input is each sample A single sensor data in , and the fault features automatically extracted by the convolution layer are first adjusted to a one-dimensional vector, and then mapped to the common information metric space to obtain a single sensor fault feature ,Right now:

[0115] ;

[0116] The multi-sensor fault features of the input sample can be obtained from the self-learning feature extraction network corresponding to the data input .

[0117] According to the fault features extracted by the feature encoder, in order to solve the problem of discriminator gradient disappearance during training, this paper uses Wasserstein divergence to construct a generative adversarial network and defines the loss functions of the generator and discriminator as:

[0118] ;

[0119] ;

[0120] Where: represents random interpolation between true and false samples, , , is a real sample; express The sample distribution of express The last item is the gradient penalty item, is the coefficient, is the power of the norm. In this embodiment, set At the same time, in order to better reconstruct the original data, Euclidean distance loss, sample frequency domain feature loss and sample feature frequency loss are also set up.

[0121] The Euclidean distance loss is used to minimize the direct geometric distance between the generated samples and the real samples, which helps the generated samples to be closer to the real samples, such as Figure 4 As shown, its loss function can be expressed as:

[0122] ;

[0123] In the formula, represents the output of the generator, represents the real sample, express norm;

[0124] The sample frequency domain feature loss is mainly used to enhance the generator’s ability to reproduce the real sample features in the frequency domain. This loss helps to better preserve the frequency domain features of the real data. Its loss function can be expressed as:

[0125] ;

[0126] In the formula, represents the operation of converting samples to the frequency domain (using Fourier transform), represents the output of the generator, represents the real sample, express Norm, which is the square of the Euclidean distance.

[0127] The sample feature frequency loss aims to minimize the difference between the generated samples and the real samples in specific frequency features. This is achieved by comparing their spectral responses, such as Figure 5 As shown, the specific loss function can be expressed as:

[0128] ;

[0129] In the formula, represents a transformation function that extracts key frequency features of input samples; is the generator output (fake sample in this formula); represents the real sample; represents the squared Euclidean distance ( norm).

[0130] A more comprehensive loss function is formed to better optimize the performance of the generated model. The comprehensive loss function is as follows:

[0131] ;

[0132] In the formula, , and They are the weight coefficients of different loss terms, which are used to balance their impact on the overall training process.

[0133] 2. Semi-supervised label completion and classification model:

[0134] In label completion and classification models, the source domain is defined as ,in, is the number of labeled samples, is the source domain, yes The category label of the target domain Include the marked target domain and the unlabeled target domain ,in, Indicates the goal, and They represent the number of labeled samples in the target domain and the number of samples in the target domain, respectively.

[0135] The probability distribution in each domain is different. The label completion and classification model used in this embodiment adjusts the distribution of related subdomains by introducing LMMD. The LMMD method calculates the Hilbert-Schmidt norm between the kernel mean embeddings of the empirical distribution of the related subdomains in the source domain and the target domain by considering the weight of each sample.

[0136] To adjust the distribution of relevant subdomains, LMMD can be defined as follows:

[0137] ;

[0138] In the formula, Indicates the target domain Subset of or The expected value of the feature map of the sample in; Indicates the source domain The expected value of the feature map of the sample in; is to transform the original data into the reproducing kernel Hilbert space The feature mapping function.

[0139] Unlike domain adaptation methods such as MMD that calculate global distribution differences, LMMD techniques can calculate the expected value of local distribution differences. Assume that each sample is assigned a weighted category In this case, the following is the definition of the unbiased estimator of the equation:

[0140] ;

[0141] In the formula, and They are and The number of samples in . is assigned to The weight of each sample in is assigned to The weight of each sample in ;

[0142] And the following equation is used to calculate each sample Weight , specifically:

[0143] ;

[0144] In the formula, yes Medium Sample The characteristic value reflects the importance, frequency or other relevant measures of the sample, and the denominator is This ensures that the sum of the weights of all samples is 1.

[0145] In addition to introducing the LMMD method to adjust the distribution of related subdomains, cross entropy loss, feature differentiation loss and consistency regularization loss are also introduced to further improve the performance of SLCN.

[0146] Among them, the cross entropy loss is defined as:

[0147] ;

[0148] Among them, y is the true label, which is Predicted label, C is the number of categories.

[0149] The feature differentiation loss is defined as:

[0150] ;

[0151] in, represents the features extracted by the self-learning extraction network, represents the discriminator, and represent positive and negative classes respectively.

[0152] The consistency regularization loss is introduced to ensure the quality of label completion and avoid introducing too many noisy labels. For each unlabeled sample, SLCN performs data augmentation to obtain different variants. Then, SLCN requires the model to give consistent prediction results for these variants. The consistency regularization loss is defined as:

[0153] ;

[0154] in, is the number of unlabeled samples, and are the predicted probability distributions of the original sample and the enhanced sample, respectively. Represents divergence.

[0155] Finally, we get the loss function of the complete SLCN model:

[0156] ;

[0157] In the application, a special equipment fault diagnosis is used as an example. Multi-channel signal data in various health states are collected from the test bench of the special equipment. Four different data and label missing rates are set to verify the practicality of the method of the present invention. The classification results are as follows: Figure 6 shown.

[0158] The experiment considered four health states, including inner race fault, outer race fault, rolling element fault and cage fault. Each health state includes 1000 samples, and each sample includes 2048 sampling points. On this basis, the data is trained and tested in a 1:1 ratio.

[0159] The experiment compares the method proposed in the present invention with three existing fault diagnosis methods for incomplete information. The specific three methods are based on deep convolution, interpolation and direct discarding. The basic structure of the self-learning feature extraction network is shown in Table 1, and the training related parameters are shown in Table 2. The experimental results are shown in Table 2. Figure 7 shown.

[0160] Table 1 Summary of basic network structure

[0161]

[0162] Table 2 Summary of training related parameters

[0163]

[0164] The experimental results show that under different conditions of data and label missing rates, the present invention shows significant performance advantages compared with the other three methods, and always maintains the highest diagnostic accuracy. This excellent performance is due to the data self-correction strategy adopted by the present invention. This strategy enables the network to reconstruct data and complete labels more flexibly, effectively handles the information fusion problem from multiple sensor sources, and achieves the best balance between accuracy and stability in the application of special equipment fault diagnosis. In addition, this strategy significantly improves the performance upper limit of the model in data reconstruction. The above experimental results fully verify the superiority and practical application value of the method of the present invention in the field of special equipment fault diagnosis. These results not only demonstrate the high efficiency of the present invention in processing complex and changeable fault data, but also prove its significant advantages in improving diagnostic accuracy and system stability.

[0165] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0166] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data self-correction fault diagnosis method, characterized in that: The specific steps are: Step 1: Obtain motor fault data, and perform data integrity check on the motor fault data. If the data is incomplete, execute step 2. If complete, proceed to step 3; Step 2: input the motor fault data into a data reconstruction model to obtain reconstructed data, and execute step 3; Step 3: Input the motor fault data or the reconstructed data into a label completion and classification model for feature extraction and fault prediction, and output the predicted fault.

2. A data self-correction fault diagnosis method according to claim 1, characterized in that: The steps for constructing the data reconstruction model are: Extracting fault features from a complete data set based on multiple parallel self-learning feature extraction networks, and generating a fault feature encoder using the fault features; Constructing a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator; The fault feature encoder is introduced into the generator; and identity mapping jump connection blocks are added to both the generator and the discriminator to obtain the data reconstruction model.

3. A data self-correction fault diagnosis method according to claim 2, characterized in that: During the training process of the data reconstruction model, the fault feature encoder and the generator are collaboratively trained together.

4. A data self-correction fault diagnosis method according to claim 2, characterized in that: During the training process, the data reconstruction model performs parameter adjustment and optimization based on the adversarial loss function, the Euclidean distance loss function, the sample feature frequency loss function and the sample frequency domain feature loss function.

5. A data self-correction fault diagnosis method according to claim 2, characterized in that: The steps for constructing the label completion and classification model are as follows: Acquire source domain data and target domain data, and identify and link labels of the source domain data and the target domain data to obtain labeled samples and unlabeled samples; Pre-training the self-learning feature extraction network using the label samples to obtain a feature extraction model; Extracting target features from the target domain data using the feature extraction model; Inputting the extracted target features into a multi-layer perceptron classifier for category prediction, and selecting pseudo-label samples according to the prediction confidence; The self-learning feature extraction network and the multi-layer perceptron classifier are trained using the label samples and the pseudo-label samples in the source domain data to obtain the label completion and classification model.

6. A data self-correction fault diagnosis method according to claim 5, characterized in that: In the training of the label completion and classification model, parameter adjustment and optimization are performed based on feature differentiation loss function, cross loss function, consistency regularization loss function and LMMD loss function.

7. A data self-correction fault diagnosis method according to claim 5, characterized in that: The self-learning feature extraction network includes an input layer, a preprocessing layer, a deep feature extraction layer, a multi-scale pooling layer, a feature fusion layer, a fully connected layer and an output layer which are connected in sequence.

8. A data self-correction fault diagnosis system, characterized in that: include: Data acquisition module, used to collect motor fault data; An integrity judgment module, used for judging the data integrity of the motor fault data; A data reconstruction module, used to reconstruct the missing motor fault data to generate reconstructed data; The fault prediction module is used to extract features and predict faults of the complete motor fault data or the reconstructed data, and output the predicted faults.

Citation Information

Patent Citations

  • Multi-source-domain variable-working-condition mechanical fault diagnosis method for incomplete category data

    CN115600137A

  • Adaptive bearing fault diagnosis method based on depth discrimination confrontation domain

    CN116451022A

  • Mechanical fault diagnosis method, device and equipment under unavailable target working condition data

    CN116756483A

  • Motor fault diagnosis stepped learning method based on full-autonomous mover offset

    CN117150377A

  • Rotary machinery intelligent fault diagnosis method and system oriented to sensor data abnormity

    CN117407784A