CN-122020291-A - Rotor structure fault diagnosis method and system based on lightweight zero sample correlation discrimination space adaptation

CN122020291ACN 122020291 ACN122020291 ACN 122020291ACN-122020291-A

Abstract

The invention discloses a rotor structure fault diagnosis method and system based on lightweight zero sample correlation discrimination space adaptation, and belongs to the technical field of mechanical equipment state monitoring and fault diagnosis. According to the method, a zero sample correlation discrimination space adaptation mechanism and a lightweight model design are introduced, domain alignment is realized through task irrelevant data (normal samples), discrimination is kept for task relevant data (source domain fault samples), a feature extraction process is guided to generate feature vectors meeting multi-objective optimization, a model can mine a common fault mode hidden behind distribution differences through a game mechanism of correlation enhancement and discrimination constraint, zero sample cross-domain migration is realized, the model parameter quantity and the calculation complexity are greatly reduced, the method is suitable for real-time diagnosis of industrial edge equipment, lightweight high-efficiency deployment can be realized, only source domain labels are needed, target domain non-label migration is realized through task irrelevant/correlation separation, and real zero sample adaptation can be realized.

Inventors

QU YU
LIU CONGRUI
WU SHIYU
MENG LINGYONG
SU MENG
LI YUHENG
YANG YUJIE
LIU YANG
JIANG QINGBIN

Assignees

西安西热电站信息技术有限公司
西安交通大学

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (10)

1. A rotor structure fault diagnosis method based on lightweight zero sample correlation discrimination space adaptation is characterized by comprising the following steps: S1, collecting vibration signals of a rotor structure of a source domain and a rotor structure of a target domain, preprocessing the vibration signals to obtain processed signals, adopting short-time Fourier transform to convert the processed signals into fixed-size images, dividing the fixed-size images into a source domain sample and a target domain sample, marking the source domain sample with a fault type, and separating the target domain sample into task irrelevant data and task relevant data; S2, constructing a lightweight zero sample model, wherein the lightweight zero sample model comprises a feature extractor, a classifier and a loss component, the feature extractor is lightweight ViT, an obtained fixed-size image is input, the fixed-size image is mapped into a low-dimensional feature vector through patch embedding and lightweight attention mechanism and a few-layer encoder, the classifier is a full-connection layer, the input is the low-dimensional feature vector, and probability prediction distribution of fault types is output; S3, designing a joint loss function, training a lightweight zero sample model, dynamically dividing source domain samples into task irrelevant data and task relevant data in each batch in the training process, and carrying out pairing calculation loss on the task irrelevant data of a source domain and the task irrelevant data of a target domain by adopting the joint loss function to obtain a trained model; S4, inputting the rotor structure vibration signals acquired in real time into a trained model through the same preprocessing of S1, outputting fault types and confidence coefficient by the model, and simultaneously generating a training loss curve, a principal component analysis characteristic distribution diagram and a confusion matrix to realize the interpretability analysis of the diagnosis result.
2. The rotor structure fault diagnosis method based on lightweight zero sample correlation discrimination space adaptation according to claim 1, wherein in S1, the preprocessing includes denoising and normalization processing performed sequentially; The processed signal is recorded as: ; Wherein, the As the original signal is meant to be a signal, Is the minimum value in the original signal, Is the maximum value in the original signal, Is a very small constant value, and is a very small constant, Is the processed signal; the fixed-size image is noted as: ; Wherein, the As a function of the window(s), For the time-frequency representation, For the time frame index to be used, For the frequency index to be used, For the time series index, For the length of the signal, Is the first The signal values of the individual sampling points are, In imaginary units.
3. The rotor structure fault diagnosis method based on lightweight zero sample correlation discrimination space adaptation according to claim 2, wherein in S1, the task independent data is normal data, the task dependent data is fault data, and no label is provided.
4. A rotor structure fault diagnosis method based on lightweight zero-sample correlation discrimination space adaptation as claimed in claim 3, wherein in S2, said low-dimensional feature vector is noted as And (2) and Wherein, the method comprises the steps of, Representing an input image of a fixed size, The feature mapping function is represented as a function of the feature, Is a trainable parameter; the light-weight attention mechanism is Vision Transformer light-weight attention mechanism, which is recorded as: ; Wherein the method comprises the steps of , wherein, In order to query the matrix, In the form of a matrix of keys, In the form of a matrix of values, In order to pay attention to the number of heads, A learnable linear transformation matrix representing an ith attention head, The output linear transformation matrix is represented as such, The operation of the splice is indicated and, Representing a self-attention computing function.
5. The rotor structure fault diagnosis method based on lightweight zero-sample correlation discrimination space adaptation as set forth in claim 4, wherein in S2, the probability prediction distribution of the barrier type is noted as , ; Wherein the method comprises the steps of Representing the classification function, its parameters By minimizing classification loss on source domain samples Optimization is performed to ensure discrimination of the learned features for fault categories.
6. The rotor structure fault diagnosis method based on the lightweight zero-sample correlation discrimination space adaptation according to claim 5, wherein in S2, the simplified typical correlation analysis loss in the loss component is calculated by a projection matrix, the mean square error is minimized, the source domain and the target domain irrelevant feature difference MSE are calculated directly by the alignment loss, the discrimination loss is calculated based on the intra-class distance to inter-class distance ratio calculated by the class center so as to improve the discrimination, and the component processes the problem of inconsistent batch size and ensures stable calculation.
7. The rotor structure fault diagnosis method based on lightweight zero sample correlation discrimination space adaptation according to claim 6, wherein the simplified typical correlation analysis loss corresponds to a loss function as follows: ; Wherein, the , In order to project the matrix of the light, , For the source domain and the target domain features, For the number of samples to be taken, Represents an L2 norm; the inter-domain feature alignment loss function is: ; Wherein, the , For the source domain and the target domain features, For the number of samples to be taken, Represents an L2 norm; The discriminative feature learning loss function is: ; Wherein, the As a total number of fault categories, For the index of the current fault category, Is different from Is used for the indexing of other categories of (a), Is of the category Is used for the measurement of the characteristic of the sample, Is of the category Is characterized in that in the characteristic center of the (c), Is of the category Is characterized in that in the characteristic center of the (c), Representing the L2 norm.
8. The rotor structure fault diagnosis method based on lightweight zero-sample correlation discrimination space adaptation according to claim 7, wherein in S3, the joint loss function is: Wherein, the , In order to combine the total loss of the loss function, For the cross entropy loss of the classifier, For gradient decoupling contrast learning loss, invalid tags are ignored, In order to simplify the CCA loss, In order to achieve a loss of alignment, In order to discriminate the loss of the object, ; In order to classify the lost weight, Weights for gradient decoupling contrast learning loss; the weights of the loss functions corresponding to the typical correlation analysis loss, the weights of the inter-domain feature alignment loss functions and the weights of the discriminant feature learning loss functions are respectively calculated.
9. The rotor structure fault diagnosis method based on the lightweight zero sample correlation discrimination space adaptation according to claim 7, wherein in S3, training a lightweight zero sample model by adopting a batch random gradient descent method, dynamically dividing source domain samples into task irrelevant data and task relevant data in each batch in training engineering, pairing and calculating losses with target domain irrelevant data, synchronously updating parameters of a feature extractor and a classifier, and enabling the feature extractor to learn domain-invariant fault feature representation through a zero sample domain adaptation learning paradigm; the zero sample domain adaptation learning paradigm is: Source domain: Target domain: ; Wherein, the For the i-th source domain sample, For the corresponding fault class label, As a total number of source domain samples, For the i-th sample of the target field, For the corresponding fault class label, For the number of tagged target domain samples, The number of the target domain samples without labels; The training process includes gradient clipping and learning rate scheduling to promote stability.
10. A rotor structure fault diagnosis system based on lightweight zero sample correlation discrimination space adaptation, comprising: The data acquisition and preprocessing module is used for acquiring vibration signals of a rotor structure of a source domain and a rotor structure of a target domain, preprocessing the vibration signals to obtain processed signals, converting the processed signals into fixed-size images by adopting short-time Fourier transform, dividing the fixed-size images into a source domain sample and a target domain sample, marking the source domain sample with a fault type, and separating the target domain sample into task irrelevant data and task relevant data; The model building module is used for building a lightweight zero sample model, wherein the lightweight zero sample model comprises a feature extractor, a classifier and a loss component, the feature extractor is lightweight ViT, an obtained fixed-size image is input, the fixed-size image is mapped into a low-dimensional feature vector through patch embedding and lightweight attention mechanism and a few-layer encoder, the classifier is a full-connection layer, the input is the low-dimensional feature vector, and the probability prediction distribution of fault types is output; training the lightweight zero sample model, dynamically dividing source domain samples into task independent data and task related data in each batch in the training process, and carrying out pairing calculation loss on the task independent data of the source domain and the task independent data of the target domain by adopting the joint loss function to obtain a trained model; The fault diagnosis module is used for inputting the rotor structure vibration signals acquired in real time into a trained model through the same preprocessing of the S1, outputting fault types and confidence coefficient by the model, and generating a training loss curve, a principal component analysis characteristic distribution diagram and a confusion matrix at the same time to realize the interpretability analysis of the diagnosis result.

Description

Rotor structure fault diagnosis method and system based on lightweight zero sample correlation discrimination space adaptation Technical Field The invention belongs to the technical field of mechanical equipment state monitoring and fault diagnosis, and particularly relates to a rotor structure fault diagnosis method and system based on lightweight zero sample correlation discrimination space adaptation. Background With the rapid development of high-end equipment manufacturing industry, the rotor structure is used as the most core and most vulnerable part in the rotary machine, and the running reliability of the rotor structure directly determines the stability and the service life of the whole mechanical system. The rolling rotor structure bears complex alternating load, impact and friction in the actual service process, and various fault modes such as inner ring faults, outer ring faults, ball faults, cage damage and the like are very easy to occur. Once the rotor structure breaks down, the equipment vibration is aggravated and the efficiency is reduced due to light weight, and the interlocking failure, the shutdown accident and even the serious potential safety hazard are caused due to heavy weight, so that huge economic loss and production interruption are caused. Therefore, early warning, accurate diagnosis and residual life prediction of rotor structure faults are realized, and the method becomes a core requirement and research hot spot in the intelligent operation and maintenance field of mechanical equipment. At present, intelligent fault diagnosis technology based on data driving has become a mainstream direction in the field. The technology mainly collects vibration signals through an acceleration sensor arranged on a bearing seat or a shell, extracts fault sensitive characteristics through signal processing and a machine learning algorithm, and realizes automatic identification and classification of fault types. Early methods mostly employ time domain statistics (e.g., root mean square value, kurtosis), frequency domain analysis (e.g., envelope demodulation), and time-frequency analysis (e.g., wavelet transformation) in combination with shallow classifiers (e.g., support vector machine, random forest) for diagnosis. With the rise of deep learning, models such as a Convolutional Neural Network (CNN), a long and short term memory network (LSTM), vision Transformer (ViT) and the like are widely introduced, and hierarchical feature representations can be learned from original vibration signals end to end, so that the diagnosis precision is remarkably improved. However, in industrial practical applications, the existing deep learning diagnostic method still faces several significant technical bottlenecks, firstly, the running conditions (such as rotational speed, load, ambient temperature and lubrication state) of the rotor structure are highly dynamic and changeable, so that the vibration signal statistical distribution of the same rotor structure between different time periods or different rotor structure individuals has significant differences, namely, domain offset phenomenon occurs between a source domain (laboratory or historical data) and a target domain (actual running data). For example, models trained under constant load, when applied to variable load conditions, shift in failure characteristic frequency can occur, resulting in a significant drop in diagnostic accuracy. More seriously, due to different manufacturing tolerances, installation errors, service years and wear degrees, the vibration response characteristics of different rotor structures are larger, so that a model trained on one batch of rotor structures is difficult to directly migrate to another batch of rotor structures, the cross-equipment, cross-working condition generalization capability and engineering popularization value of a diagnosis model are seriously restricted, and secondly, in an actual industrial site, the rotor structures are in a normal health state most of the time, the occurrence probability of fault events is low (usually less than 1%), and fault samples with accurate labels, which can be used for supervision training, are extremely deficient. Meanwhile, the acquisition of the fault labels requires the professional to mark through modes of shutdown disassembly, expert experience or laboratory reproduction and the like, and has high cost, long period and great difficulty. For new production facilities or production lines not equipped with advanced diagnostic systems, the labeled fault data are hardly obtained in sufficient quantity and diversity, so that the traditional supervised learning model is difficult to train reliable diagnostic capability and has the problem of 'data starvation', and the traditional migration learning method, such as a distribution matching technology based on maximum mean value difference (MMD), correlation alignment (CORAL) or contrast Domain Adaptation (DANN), alleviates the do