CN-121980453-A - Transformer fault recognition model training method, fault recognition method and device

CN121980453ACN 121980453 ACN121980453 ACN 121980453ACN-121980453-A

Abstract

The application discloses a training method of a transformer fault recognition model, a fault recognition method and a device, and relates to the technical field of power system fault diagnosis; each data sample is marked in coarse granularity based on various running states, each data sample marked in coarse granularity is marked in fine granularity to obtain original sample data, the original sample data is preprocessed to obtain standard sample data, and the identification model is trained based on the standard sample data to obtain a transformer fault identification model. According to the application, through marking the coarse granularity and the fine granularity of the three-phase current signals in various running states, the trained model can learn richer and finer fault characteristics, so that the accuracy and the robustness of transformer fault identification are improved.

Inventors

YANG KEWEI
LI XIANGYU
LI YONGJIAN
YANG YI

Assignees

浙江江山变压器股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260115

Claims (10)

1. The transformer fault identification model training method is characterized by comprising the following steps of: continuously sampling three-phase current signals of various running states of the transformer to obtain a plurality of data samples, wherein the various running states comprise a normal state and various abnormal states, and the fault types corresponding to the various abnormal states at least comprise open circuit, short circuit and magnetic abnormality; Performing coarse granularity labeling on each data sample based on a plurality of running states, wherein the coarse granularity labeling is that normal labels are labeled on the data samples corresponding to normal states, and abnormal labels are labeled on the data samples corresponding to abnormal states according to corresponding fault categories; carrying out fine granularity labeling on each data sample subjected to coarse granularity labeling to obtain original sample data, wherein the fine granularity labeling is to label sub-abnormal labels on each data sample labeled with abnormal labels, and the sub-abnormal labels are fine granularity sub-categories corresponding to the fault categories; Preprocessing the original sample data to obtain standard sample data; and training the identification model based on the standard sample data to obtain a transformer fault identification model.
2. The transformer fault recognition model training method according to claim 1, wherein the preprocessing the raw sample data to obtain standard sample data specifically comprises: Randomly sampling each three-phase current signal in the original sample data to obtain a plurality of standard three-phase current signals, wherein the number of sampling points of the standard three-phase current signals is the same, and the sampling points are continuous; and adding Gaussian white noise to each standard three-phase current signal to simulate a noise interference environment, so as to obtain the standard sample data.
3. The transformer fault recognition model training method according to claim 2, wherein the training of the recognition model based on the standard sample data specifically comprises: carrying out weighted summation on any one of the standard three-phase current signals in the standard sample data through a learnable decoupling weight matrix to obtain a linear combination result; constraining the linear combination result to be non-negative through a ReLU activation function, and executing Hadamard product operation of element-by-element multiplication with a complex exponential phase rotation factor to realize decoupling of the standard three-phase current signals and obtain complex signals of a plurality of decoupling phase channels; And splicing a plurality of complex signals according to rows to form a phase channel complex signal matrix.
4. The transformer fault recognition model training method according to claim 3, wherein the training of the recognition model based on the standard sample data specifically further comprises: extracting the characteristics of the complex signals of each channel in the phase channel complex signal matrix by adopting a plurality of characteristic extraction units with different sizes, and carrying out characteristic enhancement processing on the characteristics of each complex signal by using a Gaussian error linear unit to obtain a plurality of scale characteristics corresponding to each channel; Quantifying significance of each scale feature based on an L1 norm of the scale feature, and calculating attention weight of each scale feature by combining a scale normalization factor and a first Softmax function, wherein the attention weight is used for representing importance degree of the corresponding scale feature; And carrying out weighted fusion on the multiple scale features of each channel according to the attention weight to obtain the multi-scale fusion features corresponding to each channel.
5. The transformer fault recognition model training method according to claim 4, wherein the training of the recognition model based on the standard sample data specifically further comprises: flattening the multi-scale fusion features, inputting the flattened multi-scale fusion features into a long-short-term memory network, and extracting time domain feature vectors capable of reflecting time sequence dependency relations; Performing fast Fourier transform on the multi-scale fusion features to obtain frequency domain features, extracting amplitude spectrums of the frequency domain features, and extracting the amplitude spectrums by a feature extraction unit with a fixed size to obtain frequency domain feature vectors; Splicing the time domain feature vector and the frequency domain feature vector in a feature dimension, and generating a gating signal vector through linear transformation of a learnable weight matrix and a Sigmoid activation function; And carrying out weighted fusion on the time domain feature vector and the frequency domain feature vector based on the gating signal vector to obtain a fusion feature vector.
6. The transformer fault recognition model training method according to claim 5, wherein the training of the recognition model based on the standard sample data specifically further comprises: Calculating the average value of the amplitude values of all elements of the fusion feature vector, and mapping by a hyperbolic tangent function to obtain amplitude significance weights, wherein the amplitude significance weights are used for amplifying the fine fault features with low amplitude values; converting the fusion feature vector into probability distribution through a second Softmax function, calculating information entropy of the probability distribution, and carrying out normalization processing to obtain a normalized entropy value; Calculating to obtain entropy uncertainty weight based on the normalized entropy value, wherein the entropy uncertainty weight is used for strengthening fine fault characteristics with disordered feature distribution; and combining the amplitude significance weight with the entropy uncertainty weight according to a preset coefficient, and multiplying the combined feature vector by element to obtain a weighted enhancement feature vector.
7. The transformer fault recognition model training method according to claim 6, wherein the training of the recognition model based on the standard sample data specifically further comprises: Mapping the weighted enhancement feature vector to a coarse-granularity class space through a learnable coarse-granularity weight matrix to obtain a first mapping value, and inputting the first mapping value to a third Softmax function to obtain a coarse-granularity classification probability distribution vector; splicing the weighted enhancement feature vector and the coarse-granularity classification probability distribution vector in a feature dimension to obtain a splicing result; and inputting the splicing result into a fine classification weight matrix corresponding to the coarse category so as to map the splicing result into a fine-granularity category space to obtain a second mapping value, and inputting the second mapping value into a fourth Softmax function to obtain a fine-granularity classification probability distribution vector.
8. The transformer fault recognition model training method according to claim 7, wherein the training of the recognition model based on the standard sample data specifically further comprises: calculating the balance weight of each standard three-phase current signal corresponding category based on the ratio of the total number of samples of each fault category to the total number of samples of the training data; and constructing a gradient penalty term for the L2 norm square of the weighted enhancement feature vector by combining the balance weight and the fine granularity classification probability distribution vector to form a synchronous gradient weighting loss.
9. A transformer fault identification method, characterized in that it is implemented based on the transformer fault identification model obtained by the training method according to any one of claims 1-8, said transformer fault identification method comprising: Acquiring a historical three-phase current signal of a target transformer; inputting the historical three-phase current signals into the transformer fault recognition model to obtain a target recognition result comprising coarse-granularity classification probability distribution vectors and fine-granularity classification probability distribution vectors; And outputting a visual report based on the target identification result.
10. Training device for a transformer fault identification model, characterized in that it is adapted to implement a method for training a transformer fault identification model according to any of claims 1-8, said training device for a transformer fault identification model comprising: The system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module is used for continuously sampling three-phase current signals of a plurality of running states of the transformer to acquire a plurality of data samples, the plurality of running states comprise a normal state and a plurality of abnormal states, and fault types corresponding to the plurality of abnormal states at least comprise open circuits, short circuits and magnetic anomalies; The first labeling module is used for carrying out coarse granularity labeling on each data sample based on a plurality of running states, wherein the coarse granularity labeling is used for labeling normal labels on the data samples corresponding to normal states, and labeling abnormal labels on the data samples corresponding to abnormal states according to corresponding fault categories; The second labeling module is used for carrying out fine granularity labeling on each data sample subjected to coarse granularity labeling to obtain original sample data, wherein the fine granularity labeling is used for labeling sub-abnormal labels on each data sample labeled with abnormal labels, and the sub-abnormal labels are fine granularity sub-categories corresponding to the fault categories; The data processing module is used for preprocessing the original sample data to obtain standard sample data; and the model training module is used for training the identification model based on the standard sample data to obtain a transformer fault identification model.

Description

Transformer fault recognition model training method, fault recognition method and device Technical Field The application relates to the technical field of fault diagnosis of power systems, in particular to a transformer fault identification model training method, a fault identification method and a device. Background The power electronic transformer is used as core conversion equipment of a modern power system and plays an irreplaceable role in new energy grid connection, regional power grid interconnection and intelligent power distribution network. The operation environment has obvious complexity characteristics, and the equipment is easy to generate multiple fault modes due to long-term exposure to multiple stress effects such as high voltage, high current, electromagnetic interference and the like. The open circuit fault is usually caused by loosening of connecting parts or breakage of wires, so that a current path is interrupted, the short circuit fault is mainly caused by ageing or manufacturing defects of insulating materials, abnormal large current impact is generated, and the magnetic abnormality relates to magnetic circuit problems such as iron core saturation, winding deformation and the like, so that the energy conversion efficiency is influenced. The faults can accelerate the performance degradation of equipment, are more likely to cause chain reactions such as power grid voltage fluctuation, frequency instability and the like, and seriously threaten the safe and stable operation of a power system. However, the conventional fault identification method mainly relies on threshold comparison or simple feature extraction, is difficult to effectively distinguish the fine differences of multiple types of faults, and cannot meet the requirements of real-time performance, accuracy and environmental adaptability of fault identification. Disclosure of Invention The application aims to provide a transformer fault identification model training method, a fault identification method and a device, which can improve the accuracy of transformer fault identification and have the advantage of effectively distinguishing the fine differences of different fault categories. In order to achieve the above object, the present application provides the following solutions: The application provides a training method of a transformer fault identification model, which comprises the steps of continuously sampling three-phase current signals of various running states of a transformer to obtain a plurality of data samples, wherein the various running states comprise a normal state and various abnormal states, fault categories corresponding to the various abnormal states at least comprise open circuits, short circuits and magnetic anomalies, rough granularity labeling is carried out on each data sample based on the various running states, the rough granularity labeling is carried out on the data samples corresponding to the normal state and marks the data samples corresponding to the abnormal states according to the corresponding fault categories, fine granularity labeling is carried out on each data sample after rough granularity labeling to obtain original sample data, the fine granularity labeling is carried out on each data sample marked with an abnormal label, the sub-abnormal labels are fine granularity sub-categories corresponding to the fault categories, preprocessing is carried out on the original sample data to obtain standard sample data, and the transformer fault identification model is obtained based on the standard sample data. Optionally, the preprocessing the raw sample data to obtain standard sample data specifically includes: Randomly sampling each three-phase current signal in the original sample data to obtain a plurality of standard three-phase current signals, wherein the number of sampling points of the standard three-phase current signals is the same, and the sampling points are continuous; and adding Gaussian white noise to each standard three-phase current signal to simulate a noise interference environment, so as to obtain the standard sample data. Optionally, the training the recognition model based on the standard sample data specifically includes: carrying out weighted summation on any one of the standard three-phase current signals in the standard sample data through a learnable decoupling weight matrix to obtain a linear combination result; constraining the linear combination result to be non-negative through a ReLU activation function, and executing Hadamard product operation of element-by-element multiplication with a complex exponential phase rotation factor to realize decoupling of the standard three-phase current signals and obtain complex signals of a plurality of decoupling phase channels; And splicing a plurality of complex signals according to rows to form a phase channel complex signal matrix. Optionally, the training the recognition model based on the standard sample data specifically f