CN-122020497-A - Small sample fault diagnosis method for on-load tap-changer based on wavelet elasticity measurement network

CN122020497ACN 122020497 ACN122020497 ACN 122020497ACN-122020497-A

Abstract

The utility model provides a small sample fault diagnosis method of on-load tap-changer based on wavelet elasticity measurement network, relates to on-load tap-changer fault diagnosis technical field, is used for solving on-load tap-changer fault diagnosis's recognition accuracy problem under the small sample fault condition that on-load tap-changer fault sample acquisition is difficult, the fault type is various and the difference is complicated between the category. The invention takes the original on-load tap-changer mechanical vibration signal as input, sequentially completes the learning wavelet enhancement, the multi-scale feature extraction, the elasticity measurement modeling and the fault type discrimination, and realizes the end-to-end modeling from the original vibration signal to the diagnosis result. The modules cooperate with each other in the whole flow, the wavelet enhancement improves the feature separability, the multi-scale feature extraction effectively characterizes the local and the whole change of the vibration signal, and the elasticity measurement mechanism carries out self-adaptive adjustment on the feature distance under the condition of a small sample, thereby improving the classification discrimination capability of the model.

Inventors

ZHAO TONG
CHEN ZHIXIN
WANG XIAOLONG
ZHANG YUANTAO
SUN YING
LIU YADI
QI RUNZE
DUAN TIANYU

Assignees

山东大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. The on-load tap-changer small sample fault diagnosis method based on the wavelet elasticity measurement network is characterized by comprising the following steps: s1, data acquisition The method comprises the steps of collecting a mechanical vibration signal x i of an on-load tap-changer, dividing the mechanical vibration signal into a training set and a testing set, and dividing the training set and the testing set into a supporting set and a query set respectively; s2, constructing wavelet enhancement multi-scale feature encoder S2.1, performing leachable wavelet transformation on the vibration signal x i , and then mapping the vibration signal x i to a multi-scale time-frequency representation space; S2.2, constructing a lightweight multi-scale convolution to perform feature extraction on the wavelet reinforced vibration signal so as to obtain the features of the vibration signal under different time scales, and splicing the features to obtain a fusion feature F ms ; S2.3 introduction of channel attention mechanism Providing fusion features Performing feature compression on each channel through global average pooling; s3, elasticity measurement modeling Introducing an elastic measuring device EMM, and carrying out self-adaptive adjustment on measuring scales of different categories; S4, discriminating fault types By an objective function (1) And judging the fault type, wherein, An elastic classification loss function of an elastic measuring device EMM, wherein alpha is a regularization coefficient; Is a regularization term for the elastic factor.
2. The method for diagnosing a small sample fault of an on-load tap changer based on a wavelet elasticity metric network as claimed in claim 1, wherein the specific contents of step S2.1 are that an input vibration signal is set L is the signal length, and the output of the jth wavelet neuron is expressed as: (2) Wherein, the method comprises the steps of, For the wavelet basis function, morlet wavelet is used: (3) W ji is the weight matrix of the wavelet neurons, b j is the translation coefficient of the wavelet, a j is the scale coefficient of the wavelet, and M is the number of the wavelet neurons.
3. The method for diagnosing a small sample fault of an on-load tap changer based on a wavelet elasticity metric network as claimed in claim 2, wherein in step S2.2, the multi-scale convolution content is S2.2.1 input features processed by a learnable wavelet (4) The multi-scale convolution module comprises a plurality of parallel one-dimensional convolution branches, and for the S-th convolution branch, s=1, 2..S, the one-dimensional convolution operation is as follows: (5) Wherein, the method comprises the steps of, The method is characterized by comprising the steps of (a) calculating a convolution kernel parameter of an s-th branch, wherein k s is the convolution kernel length corresponding to the branch, C s is the output channel number of the branch, and b (s) is a bias term; a characteristic representation representing the output of the s-th convolution branch.
4. The method for diagnosing a small sample fault of an on-load tap changer based on a wavelet elasticity metric network according to claim 3, wherein in step S2.2, a S2.2.2 batches of normalization is introduced to normalize the features after the convolution operation: (6); (7) Wherein, the method comprises the steps of, Is a standardized feature; representing characteristics Is the average value of (2); representing characteristics Is a variance of (2); a smoothing constant for preventing numerical instability; Is the final feature after translation and scaling; Is a leachable scale; Is a learnable bias parameter.
5. The method for diagnosing a small sample fault in an on-load tap changer based on a wavelet elastic metric network as claimed in claim 4, wherein in step S2.2, step S2.2.2 is followed by step S2.2.3 of stitching the output features of all branches in channel dimension to obtain a fused feature (8) Wherein Concat (,) represents a splice operation along the channel dimension.
6. The method for diagnosing a small sample fault of an on-load tap changer based on a wavelet elasticity metric network as claimed in claim 5, wherein in step S2.3, the characteristic compression of the channel is as follows: (9) Wherein, the method comprises the steps of, Representing the eigenvalue of the c-th channel at position i, m being the characteristic dimension, m=1, 2..d: (10) Wherein, the method comprises the steps of, Attention weight for each channel; Representing a Sigmoid function; And Is a learnable parameter of the full connection layer; representing a ReLU activation function; scaling the weight of each channel and the original characteristic channel by channel to obtain the characteristic representation after attention enhancement (11) Wherein β c represents the attention weight corresponding to the c-th channel; Representing the feature vector of the c-th channel.
7. The method for diagnosing a small sample fault in an on-load tap changer based on a wavelet elasticity measure network as claimed in claim 6, wherein in step S4, the elasticity classification loss function (12) Wherein Q k is a k-th query set, and phi is all the learnable parameters of the feature encoder; Sample for query set Elastic distance between the model and the class k prototype, N is the number of the current task class, exp is the natural exponential function e x ;d ij * and the elastic distance between the query set sample z i and the j-th class prototype p j .
8. The method for diagnosing a small sample fault in an on-load tap changer based on a wavelet elasticity metric network of claim 7, The calculation formula of (2) is as follows: (13) Wherein d k * is the elastic distance of the kth prototype, E φ (. Cndot.) represents the mapping function of the feature encoder, and P k is the query set sample Probability belonging to the k-th class.
9. The method for diagnosing a small sample fault in an on-load tap changer based on a wavelet elasticity metric network of claim 8, The calculation formula of (2) is as follows: (14) Wherein λ cm is an elasticity factor, and λ cm >0; As the elastic weight vector of the model is, 。
10. The method for diagnosing a small sample fault in an on-load tap changer based on a wavelet elastic metric network as recited in claim 9, wherein P k is calculated by Softmax function, P k = (15) Wherein d ij * is calculated as d ij * = (16) Wherein z im represents the value of the i-th query set sample embedding vector in the m-th dimension, P cm is the component of the c-th prototype vector in the m-th dimension, P c is the c-th prototype vector, S c is the class c support set, |S c | is the number of samples, and x j is the j-th original input sample in S c .

Description

Small sample fault diagnosis method for on-load tap-changer based on wavelet elasticity measurement network Technical Field The invention relates to the technical field of on-load tap-changer fault diagnosis, in particular to a small sample fault diagnosis method of an on-load tap-changer based on a wavelet elasticity measurement network. Background Around the problem of fault identification of an on-load tap-changer OLTC vibration signal, the existing research mostly adopts a method of combining manual experience with traditional signal processing, such as extraction of time domain statistical features, frequency domain features and time frequency analysis features, and combines a support vector machine, k nearest neighbor or random forest classification model to realize fault discrimination. Although the above method verifies the feasibility of vibration signals for OLTC fault diagnosis, its performance is highly dependent on feature engineering design and has limited generalization capability under complex operating conditions and multiple fault scenarios. Along with the development of deep learning technology, convolutional neural network and cyclic neural network models are gradually introduced into the field of OLTC fault diagnosis, distinguishing features are automatically extracted through end-to-end learning, and good diagnosis effects are achieved under the conditions of sufficient data scale and relatively stable working conditions. Although deep learning methods exhibit strong feature learning capabilities in vibration signal fault diagnosis, their effectiveness is generally based on large-scale and well-labeled data. However, in an actual engineering scenario, OLTC fault samples are high in acquisition cost and limited in field test conditions, it is difficult to construct a sufficient fault sample data set in an active test mode, and fault sample acquisition relies on long-term operation accumulation or manual labeling, so that the size of fault vibration data available for modeling is limited and classification distribution is uneven. Under the condition, the traditional deep learning model is directly adopted, so that over fitting is easy to occur, and the generalization capability of the model is limited. Modeling ideas which rely only on data scale expansion are difficult to meet engineering requirements of OLTC fault diagnosis, and it is highly desirable to introduce fault diagnosis strategies which can still have good adaptability under sample limited conditions. Unlike the traditional supervised learning which relies on large-scale annotation data, the small sample learning enables the model to have stronger generalization capability under the condition of data limitation by mining the common structure and potential rules among tasks. In many small sample learning methods, meta-learning enables models to quickly complete adaptation to unknown tasks relying on only a small number of new samples by learning migratable prior knowledge among multiple related tasks, thus exhibiting unique advantages in sample-limited engineering scenarios. The meta-learning framework achieves good effects in the task of diagnosing faults of small samples in the fields of rotary machinery and wind power generation equipment. In general, meta-learning has a strong potential in small sample diagnosis of mechanical equipment, but the performance of the meta-learning depends on capturing distinguishing characteristics of signals, and when a diagnosis object turns to a switching device with a more complex working mechanism from a rotating machine with a stable periodic characteristic, the prior method still has limitation in effect. Although the prior art introduces a small sample learning method into the field of fault diagnosis of mechanical or electrical equipment, the prior art focuses on the structural design of a general model, and has insufficient consideration on the characteristics of strong transient, strong noise and working condition difference contained in an OLTC vibration signal, and the applicability of the method in the fault diagnosis of an actual OLTC small sample is still to be studied intensively. Disclosure of Invention The invention aims to provide a small sample fault diagnosis method of an on-load tap-changer based on a wavelet elasticity measurement network, which is used for solving the problem of identification accuracy of fault diagnosis of the on-load tap-changer under the condition of small sample faults of difficult acquisition of the fault sample of the on-load tap-changer, various fault types and complex difference between categories. The technical scheme adopted by the invention for solving the technical problems is that the on-load tap-changer small sample fault diagnosis method based on the wavelet elasticity measurement network comprises the following steps. S1, data acquisition. The method comprises the steps of collecting mechanical vibration signals of an on-load tap change