CN-121980348-A - Small sample fault diagnosis method and system based on physical prior element optimization

CN121980348ACN 121980348 ACN121980348 ACN 121980348ACN-121980348-A

Abstract

The application discloses a small sample fault diagnosis method and system based on physical prior element optimization, and belongs to the technical field of electric data processing. The method comprises the steps of receiving signal data samples of industrial equipment, extracting statistics and physical characteristics, carrying out physical priori enhancement on fault signal characteristics to form a signal characteristic set, training a fault diagnosis model aiming at each signal characteristic to generate a data sequence set containing signal characteristics, model parameters, diagnosis results and probabilities, utilizing the set to train a parameter optimization model, adjusting learning rate according to a loss curve and outputting an optimized model parameter predicted value, taking the predicted value as a new initial parameter to update the data sequence set until the model parameters are converged, and finally utilizing the trained model to carry out real-time fault diagnosis.

Inventors

TAO LAIFA
LIU HAIFEI
LIAN ZHIXUAN
HUANG QIXUAN
FENG CHANGCHUN
SUN JUNRU

Assignees

杭州市北京航空航天大学国际创新研究院(北京航空航天大学国际创新学院)
北京航空航天大学

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (10)

1. A method for small sample fault diagnosis based on physical prior element optimization, the method comprising: Receiving signal data samples of industrial equipment, and carrying out statistical feature extraction and physical feature extraction to obtain corresponding signal features; performing feature enhancement on signal features of fault signal samples in the signal data samples to generate signal features with physical interpretation, wherein all the signal features form a signal feature set; Training a fault diagnosis model aiming at each signal characteristic of the signal characteristic set, wherein the output of the fault diagnosis model is the diagnosis result and probability of industrial equipment corresponding to the signal characteristic, and obtaining different diagnosis results and probabilities by adjusting the model parameters of the fault diagnosis model from the initial model parameters of the fault diagnosis model so as to obtain a data sequence set of the signal characteristic, the corresponding fault diagnosis model and model parameters, the diagnosis results and probabilities; Training a parameter optimization model by using the data sequence set, wherein the parameter optimization model adjusts the learning rate according to the loss curve and outputs an optimized model parameter predicted value; Inputting the optimized model parameter predicted value into the fault diagnosis model as an initial model parameter to update the data sequence set until the variation amplitude of the model parameter is smaller than a preset threshold value; and receiving signal data of industrial equipment, and performing fault diagnosis by using the trained fault diagnosis model.
2. The small sample fault diagnosis method based on physical prior element optimization according to claim 1, wherein the steps of receiving signal data samples of industrial equipment, and performing statistical feature extraction and physical feature extraction to obtain corresponding signal features include: receiving signal data samples of industrial equipment, and carrying out multi-mode signal acquisition and preprocessing; Carrying out statistical feature extraction on the preprocessed signal data to obtain a feature vector of a statistical encoder; extracting physical characteristics of the preprocessed signal data to obtain a multi-scale physical characteristic matrix; And carrying out weighted fusion on the feature vector of the statistical encoder and the multi-scale physical feature matrix to obtain corresponding signal features.
3. The small sample fault diagnosis method based on physical prior element optimization according to claim 2, wherein the feature enhancement is performed on the signal features of the fault signal samples in the signal data samples, and generating the signal features with physical interpretation includes: performing physical priori embedding processing on signal features of fault signal samples in the signal data samples to generate feature representations with physical constraints; Performing multi-scale feature enhancement on the feature representation with physical constraint, gradually adding noise through a forward diffusion process of a diffusion model, and recovering signal features from the noise by utilizing a reverse denoising process; and fusing the recovered signal characteristics with the signal characteristics of the fault signal samples to obtain corresponding signal characteristics.
4. A small sample fault diagnosis method based on physical prior element optimization according to claim 3, wherein the obtaining the different diagnosis results and probabilities by adjusting model parameters of the fault diagnosis model from initial model parameters of the fault diagnosis model comprises: For each signal feature, forward propagation is carried out on the input signal feature based on initial model parameters of the fault diagnosis model, so that diagnosis results and probability are obtained; calculating a loss between the diagnostic result and the real label and back-propagating a gradient to update the model parameters; Recording signal characteristics, model parameters, diagnosis results and probability in the iteration of the round to form a single round data sequence; repeating the steps according to the updated model parameters until convergence conditions are reached.
5. The method of claim 4, wherein said calculating the loss between the diagnostic result and the true label and back-propagating gradients to update the model parameters comprises: Calculating cross entropy loss between the diagnosis result and the real label, and combining regularization items based on physical prior to form a loss function; Weighting the updating direction of the model parameters according to the inverse propagation gradient of the loss function and the physical enhancement gradient of the signal characteristics; and adjusting the updating step length of the model parameters according to the gradient until the model parameters are converged.
6. The small sample fault diagnosis method based on physical prior element optimization according to any one of claims 1 to 4, wherein the training a parameter optimization model by using the data sequence set, the parameter optimization model adjusting a learning rate according to a loss curve, and outputting an optimized model parameter predicted value comprises: constructing a space-time feature matrix by using the data sequence set, and performing dimension reduction on the space-time feature matrix by using principal component analysis to form an input-output pair for training the parameter optimization model; A deep neural network comprising a multi-head attention mechanism and residual connection is adopted as a parameter optimization model, the network weight is updated by calculating the loss between the output of the parameter optimization model and a target value, and the change trend of a loss curve is recorded; and adjusting the learning rate and the updating step length according to the derivative of the loss curve until the parameter optimization model converges to obtain the model parameter predicted value after the output optimization.
7. The method for diagnosing a small sample fault based on physical prior meta optimization according to claim 6, wherein the constructing a space-time feature matrix by using the data sequence set and reducing the dimension of the space-time feature matrix by using principal component analysis, forming an input-output pair for training the parameter optimization model, comprises: arranging signal characteristics, model parameters, diagnosis results and probability in the data sequence set in time sequence to construct a space-time characteristic matrix; performing principal component analysis on the space-time feature matrix, and reserving principal components with contribution rate larger than a preset threshold value to realize dimension reduction; Taking the space-time characteristic matrix after dimension reduction as input, taking corresponding model parameters as output, and forming an input-output pair for training the parameter optimization model.
8. The method for diagnosing a small sample fault based on physical prior meta optimization according to claim 6, wherein the steps of calculating the loss between the output of the parameter optimization model and the target value, updating the network weight, and recording the change trend of the loss curve include: Calculating the mean square error loss between the model parameter predicted value and the target value output by the parameter optimization model, and forming a comprehensive loss function by using a regularization term; Updating the network weight of the parameter optimization model based on the comprehensive loss function; recording the current loss value, iteration times and learning rate in each iteration, and constructing a loss curve data sequence; And carrying out smoothing treatment on the loss curve data sequence, and calculating the loss change rate.
9. The method for small sample fault diagnosis based on physical a priori meta optimization of claim 8 wherein the adjusting the learning rate and update step size according to the derivative of the loss curve comprises: calculating a first derivative and a second derivative of the loss curve, and respectively judging a loss descending trend and a convergence speed; adjusting the learning rate according to the change of the first derivative, and reducing the learning rate when the first derivative approaches zero; And adjusting the updating step according to the change of the second derivative, and reducing the step when the second derivative is increased.
10. A system comprising a storage medium and one or more processors, the storage medium storing a computer program, characterized in that the computer program is invoked by the one or more processors to implement a small sample fault diagnosis method based on physical a priori optimization as claimed in any one of claims 1 to 9.

Description

Small sample fault diagnosis method and system based on physical prior element optimization Technical Field The application relates to the technical field of fault diagnosis, in particular to a small sample fault diagnosis method and system based on physical prior element optimization. Background With the rapid development of industrial equipment, a rapid diagnostic capability is required. The most common in the prior art is to train a learning model in advance and then use the model to execute classification tasks of industrial equipment signal data to perform fault diagnosis. However, the scheme of the fixed parameter model has obvious problems that the model parameters cannot be adaptively adjusted according to the specific data distribution of the novel equipment, so that the diagnosis accuracy is obviously reduced in a rare sample scene, and secondly, the fixed network structure and parameters are difficult to effectively fuse physical characteristics in industrial data, so that the requirements of the actual industrial application on the adaptability and the universality of a diagnosis system cannot be met. Disclosure of Invention The embodiment of the application provides a small sample fault diagnosis method and system based on physical prior element optimization. The technical scheme is as follows: In a first aspect, an embodiment of the present application provides a method, including receiving a signal data sample of an industrial device, and performing statistical feature extraction and physical feature extraction to obtain a corresponding signal feature; performing feature enhancement on signal features of fault signal samples in the signal data samples to generate signal features with physical interpretation, wherein all the signal features form a signal feature set; Training a fault diagnosis model aiming at each signal characteristic of the signal characteristic set, wherein the output of the fault diagnosis model is the diagnosis result and probability of industrial equipment corresponding to the signal characteristic, and obtaining different diagnosis results and probabilities by adjusting the model parameters of the fault diagnosis model from the initial model parameters of the fault diagnosis model so as to obtain a data sequence set of the signal characteristic, the corresponding fault diagnosis model and model parameters, the diagnosis result and probability; Training a parameter optimization model by using the data sequence set, wherein the parameter optimization model adjusts the learning rate according to the loss curve and outputs an optimized model parameter predicted value; Inputting the optimized model parameter predicted value into the fault diagnosis model as an initial model parameter to update the data sequence set until the variation amplitude of the model parameter is smaller than a preset threshold value; and receiving signal data of industrial equipment, and performing fault diagnosis by using the trained fault diagnosis model. In one possible implementation manner, the receiving the signal data sample of the industrial device, and performing statistical feature extraction and physical feature extraction to obtain a corresponding signal feature includes: receiving signal data samples of industrial equipment, and carrying out multi-mode signal acquisition and preprocessing; Carrying out statistical feature extraction on the preprocessed signal data to obtain a feature vector of a statistical encoder; extracting physical characteristics of the preprocessed signal data to obtain a multi-scale physical characteristic matrix; And carrying out weighted fusion on the feature vector of the statistical encoder and the multi-scale physical feature matrix to obtain corresponding signal features. In one possible implementation manner, the feature enhancement is performed on the signal features of the fault signal samples in the signal data samples, so as to generate signal features with physical interpretation, which includes: performing physical priori embedding processing on signal features of fault signal samples in the signal data samples to generate feature representations with physical constraints; Performing multi-scale feature enhancement on the feature representation with physical constraint, gradually adding noise through a forward diffusion process of a diffusion model, and recovering signal features from the noise by utilizing a reverse denoising process; and fusing the recovered signal characteristics with the signal characteristics of the fault signal samples to obtain corresponding signal characteristics. In one possible implementation, starting from the initial model parameters of the fault diagnosis model, obtaining the different diagnosis results and probabilities by adjusting the model parameters of the fault diagnosis model includes: For each signal feature, forward propagation is carried out on the input signal feature based on initial model parameters of the fa