Search

CN-121980252-A - Hydraulic system fault diagnosis method based on small sample learning

CN121980252ACN 121980252 ACN121980252 ACN 121980252ACN-121980252-A

Abstract

The invention discloses a hydraulic system fault diagnosis method based on small sample learning, which is characterized in that data preprocessing is carried out on multisource sensor signals obtained in advance in the operation process of a hydraulic system, the multisource sensor signals are converted into a two-dimensional time-frequency characteristic diagram through short-time Fourier transform, gram angle and field transform, then a data set is divided according to small sample learning characteristics to form a task set, an improved relation network and Marsdh metric learning double-branch fault diagnosis model is constructed, a relation network branch generates an enhanced prototype through self-attention, global semantic enhancement is realized by combining a global prototype modulation mechanism, the Marsdh metric learning branch processes matrix singular problems under the small sample through an improved covariance matrix, and classification output scores are calculated by the two branches respectively and are subjected to self-adaptive weighting fusion to form a fault diagnosis result. The invention improves the fault classification accuracy under the condition of small samples, and is suitable for fault identification of the hydraulic system with limited samples.

Inventors

  • CHEN NINGNING
  • LIU JIANWEI
  • XUE QING
  • ZENG MENGJIE
  • ZHANG CHENGRUI

Assignees

  • 南京航空航天大学

Dates

Publication Date
20260505
Application Date
20260107

Claims (10)

  1. 1. The hydraulic system fault diagnosis method based on small sample learning is characterized by comprising the following steps of: (1) Carrying out data preprocessing on the multisource sensor signals obtained in advance in the running process of the hydraulic system; (2) Converting the preprocessed one-dimensional time sequence signal into a two-dimensional time-frequency complementary characteristic diagram through short-time Fourier transform, gram angle and field transform; (3) Dividing a data set according to the learning characteristics of the small sample, and forming a task set; (4) The method comprises the steps of constructing a hydraulic system fault diagnosis model, wherein the hydraulic system fault diagnosis model comprises a feature extraction and fusion module, a Markov measurement learning branch, an improved relation network branch and a self-adaptive weighting fusion and classification module, wherein the Markov measurement learning branch calculates the Markov distance between a query sample and a class prototype by using an improved covariance matrix to realize measurement classification; (5) The training set task and the verification set task are input to respectively train and verify, model parameters are updated, and a model with the best diagnosis effect in verification is stored as an optimal diagnosis model; (6) And 5, performing model performance evaluation on the test set task by using the optimal diagnosis model selected in the verification stage in the step 5, and finally realizing hydraulic system fault diagnosis based on small sample learning.
  2. 2. The hydraulic system fault diagnosis method based on small sample learning according to claim 1, wherein the multi-source sensor signal in step (1) comprises a pressure, a flow rate, a temperature, a motor power, and a vibration one-dimensional time sequence signal.
  3. 3. The hydraulic system fault diagnosis method based on small sample learning according to claim 1, wherein the implementation process of the step (2) is as follows: STFT is essentially a transformation of a basis function whose mathematical expression is as follows: ; Wherein, the Representing the input signal and the output signal, In order to analyze the window function, In order to be a frequency of the light, Namely, is A spectrogram of the moment is taken as output; GASF is time sequence code transformation based on the angle sum of the polar coordinate system, and the mathematical expression is as follows: ; Wherein, the And (3) with Representing the time of day of the original signal And The polar angle value after normalization mapping, The normalized signal maps the one-dimensional time sequence into the two-dimensional time domain feature map through STFT and GASF.
  4. 4. The hydraulic system fault diagnosis method based on small sample learning according to claim 1, wherein the data set in step (3) comprises a training set, a verification set and a test set, and the three sets are divided into a task set in the form of N-way K-shot, and the task set comprises a support set and a query set.
  5. 5. The hydraulic system fault diagnosis method based on small sample learning is characterized in that in the feature extraction and fusion module in the step (4), a feature extraction part sequentially comprises convolution, pooling, normalization and multi-scale attention enhancement modules, wherein the multi-scale attention enhancement modules are composed of a plurality of parallel large-core attention branches, each branch adopts depth separable convolution and cavity convolution with different convolution core sizes to generate a space attention map and is fused with an input residual, features extracted from all channels are spliced in a channel dimension, the channel attention network is used for self-adaptive weighting, and finally fusion features are output through convolution compression channel numbers, so that the efficient extraction and fusion of multi-scale features of a cross sensor are realized.
  6. 6. The hydraulic system fault diagnosis method according to claim 1, wherein the improved relation network branch in the step (4) comprises a feature encoder module and a relation matching module, wherein the feature encoder module optimizes a prototype generation process through a self-attention-enhancing prototype mechanism, and generates support set each type of enhanced prototype vector and a modulated query set feature vector through a global prototype modulation mechanism.
  7. 7. The hydraulic system fault diagnosis method based on small sample learning according to claim 6, wherein the self-attention-enhancing prototype mechanism calculation formula is as follows: ; Wherein, the The class fault number to which the sample belongs, For the number of each type of sample in the support set samples in the constructed N-way K-shot task, For the enhanced class prototype vector, To support set Class K sample features Is a set of (a) and (b), Representative of the use of self-attention mechanism calculations thereon; Representative support set number Class 1 The individual samples correspond to weights.
  8. 8. The hydraulic system fault diagnosis method based on small sample learning according to claim 6, wherein the global prototype modulation scheme is calculated as follows: ; Wherein, the Representing a support set enhanced class prototype after a global prototype modulation mechanism, And In order for the parameters to be able to be learned, The representative GELU is used to activate a function, Is a layer normalization method.
  9. 9. The hydraulic system fault diagnosis method based on small sample learning as claimed in claim 1, wherein the mahalanobis metric learning branch of step (4) is formed by learning a diagonal positive lower triangular matrix Thereby constructing a metric matrix for the calculation of the mahalanobis distance, the mathematical expression of which is as follows: ; Wherein, the In order to support a certain category in the set, To learn the resulting initial lower triangular matrix, As the diagonal elements of the lower triangular matrix, For other elements than the diagonal line, In order to finally learn the obtained lower triangular matrix after complete processing, To learn the resulting prototype vectors using support sets of classes, Then it is the feature vector of a sample of the query set, I.e., the score of the query set sample in each category in the classification.
  10. 10. The hydraulic system fault diagnosis method based on small sample learning according to claim 1, wherein in the adaptive weighted fusion and classification module in step (4), weights are ensured to be between [0,1] through a Sigmoid function based on the learnable adaptive fusion weights: ; Wherein, the For the mahalanobis branch score, For the score of a branch of a relationship, The final classification output is carried out by the method And taking the maximum value to obtain a prediction category, and realizing comprehensive judgment.

Description

Hydraulic system fault diagnosis method based on small sample learning Technical Field The invention belongs to the field of hydraulic system fault diagnosis, and particularly relates to a hydraulic system fault diagnosis method based on small sample learning, which is suitable for multiple fault recognition scenes with limited sample numbers. Background The hydraulic system is used as a power transmission and control device widely applied in modern industry, engineering machinery, aerospace and automatic production, and the performance of the hydraulic system directly influences the safety and reliability of equipment operation. The hydraulic system consists of a hydraulic pump, a hydraulic valve, an energy accumulator, a cooler, a pipeline and the like, wherein the components have close fluid coupling and energy transfer relations, and the operation environment is often accompanied with high temperature, high pressure, vibration and external interference factors. In the long-term operation process, the hydraulic system is easy to have faults such as leakage, insufficient pressure, abnormal flow, overhigh temperature, switching delay and the like, and if the problems can not be diagnosed and processed in time, the whole machine can be stopped, and even serious safety accidents can be caused. Traditional hydraulic system fault diagnosis relies on manual experience or statistical and machine learning methods based on a large number of samples, such as neural networks, support vector machines and the like. However, in practical engineering application, because of low occurrence frequency and high acquisition cost of certain fault types and complex condition of on-site operation environment, only a very small amount of characteristic samples can be obtained, which causes serious limitation on the performance of traditional machine learning. Under the condition of rare samples, the model is easy to generate overfitting and reduced generalization capability, and the requirements of quick and accurate on-site diagnosis are difficult to meet. In the field of image recognition, voice recognition, fault diagnosis and the like, the technology of small sample Learning (Few-Shot Learning, FSL) is paid attention to in recent years, and the core idea is to quickly extract the discrimination characteristics related to tasks from a small number of support set samples by utilizing strategies such as metric Learning, relational modeling, knowledge migration and the like under the condition of limited samples, so as to realize classification recognition of query samples. In the hydraulic system fault diagnosis scene, small sample learning has potential to break through the bottleneck of insufficient data. However, the signals output by the multisource sensors of the hydraulic system are various in types, including pressure, flow, temperature, vibration and the like, and the data distribution difference is obvious and the correlation is strong. The existing small sample fault diagnosis research is concentrated on bearing vibration data, the information of a multi-source sensor of a hydraulic system cannot be fully utilized, and the limitation exists in the aspects of coping with hydraulic multi-element and multi-mode compound faults. Therefore, the realization of efficient and accurate fault diagnosis of the hydraulic system under the condition of a small sample is an important problem to be solved urgently. Disclosure of Invention The invention aims to provide a hydraulic system fault diagnosis method based on small sample learning, which integrates a Marsdian metric learning branch with an improved relational network branch, and improves the accuracy and reliability of hydraulic system fault diagnosis by utilizing two-dimensional time-frequency feature extraction, multi-mode fusion, self-attention-enhancing prototype, global prototype modulation mechanism, self-adaptive weight fusion and other technologies. The hydraulic system fault diagnosis method based on small sample learning comprises the following steps: (1) Carrying out data preprocessing on the multisource sensor signals obtained in advance in the running process of the hydraulic system; (2) Converting the preprocessed one-dimensional time sequence signal into a two-dimensional time-frequency complementary characteristic diagram through short-time Fourier transform, gram angle and field transform; (3) Dividing a data set according to the learning characteristics of the small sample, and forming a task set; (4) The method comprises the steps of constructing a hydraulic system fault diagnosis model, wherein the hydraulic system fault diagnosis model comprises a feature extraction and fusion module, a Markov measurement learning branch, an improved relation network branch and a self-adaptive weighting fusion and classification module, wherein the Markov measurement learning branch calculates the Markov distance between a query sample and a class prototype by using an i