CN-122020249-A - Transformer fault diagnosis method and system based on deep learning and sensor fusion

CN122020249ACN 122020249 ACN122020249 ACN 122020249ACN-122020249-A

Abstract

The invention discloses a transformer fault diagnosis method and system based on deep learning and sensor fusion, wherein the method comprises the steps of collecting original signals generated by a multi-source sensor arranged on a transformer; the method comprises the steps of calculating noise parameters of a transformer, preprocessing an original signal based on the noise parameters to generate a preprocessing feature tensor, inputting the preprocessing feature tensor into a deep learning model with multiple channels, modeling the preprocessing feature tensor in each channel to obtain corresponding domain features, adaptively adjusting fusion weights based on the noise parameters in a fusion layer of the deep learning model, fusing the domain features of the preprocessing feature tensor to generate a fusion feature tensor, and judging the fault type or health state of the transformer based on the fusion feature tensor and the noise parameters. The invention has high precision and real-time performance under low signal-to-noise ratio, has generalization capability of crossing equipment and working conditions, and remarkably improves the reliability and applicability of intelligent operation and maintenance and state evaluation of the transformer.

Inventors

HUANG ZHIGUO
PENG SIMIN
MAO LIUMING
ZHAO MIAO
ZHOU HENGYI
SUN LIPENG
YI MIN
ZHANG ZHIDAN
HE SILIN
DUAN XUJIN

Assignees

国网湖南省电力有限公司电力科学研究院
国网湖南省电力有限公司
国家电网有限公司

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. The transformer fault diagnosis method based on deep learning and sensor fusion is characterized by comprising the following steps: S1, collecting original signals generated by at least two types of multi-source sensors arranged on a transformer to be monitored; S2, calculating noise parameters of the transformer to be monitored, and preprocessing an original signal based on the noise parameters to generate a preprocessing characteristic tensor; S3, inputting the preprocessing feature tensor into a deep learning model with multiple channels, and respectively modeling the preprocessing feature tensor in each channel to obtain corresponding time domain features, frequency domain features and sensor domain features; S4, in a fusion layer of the deep learning model, self-adaptively adjusting fusion weights based on the noise parameters, and fusing time domain features, frequency domain features and sensor domain features of the preprocessing feature tensor to generate fusion feature tensors; and S5, judging the fault type or health state of the transformer to be monitored based on the fusion characteristic tensor and the noise parameter.
2. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, wherein in step S2, the noise parameter of the transformer to be monitored is calculated, specifically comprising: According to the local fluctuation degree and instantaneous anomaly degree of the time-frequency tensor of the original signal at each time point and frequency point, a pixel-level noise diagram is obtained by calculation, and the expression is as follows: ; In the above-mentioned method, the step of, Is a pixel level noise map; is the degree of fluctuation in the frequency dimension; is the degree of anomaly in the time dimension; is a compression function; 、 Is a weight parameter; Index for frequency and time; And calculating the channel-level noise index according to the average value of the pixel-level noise diagram at all time frequency points, wherein the expression is as follows: ; In the above-mentioned method, the step of, Is the noise figure at the channel level; 、 The total number of frequency points and the total number of time points are respectively; Representing summation over all time-frequency points.
3. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 2, wherein the fluctuation degree of the frequency dimension is determined by comparing the frequency points The power sequence at all time points was calculated variance obtained as follows: ; In the above-mentioned method, the step of, Is a time-frequency power value; Representing frequency points Average power at all time points; The degree of abnormality in the time dimension is characterized by comparing the deviation of the point power from the time window mean, expressed as follows: ; In the above-mentioned method, the step of, To be at the point of time The average power of the window is the center, To prevent the denominator from being a constant of zero.
4. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, wherein in step S2, preprocessing is performed on an original signal based on the noise parameter to generate a preprocessing feature tensor, which specifically includes: performing threshold adjustment on the original signal based on the noise parameter to adjust a filter coefficient, and performing nonlinear filtering on the original signal to obtain a filtered signal; performing variable resolution time-frequency conversion on the filtered signal, and generating time-frequency characteristics; And extracting signal characteristics at a plurality of time scales or frequency scales, and carrying out weighting processing on the signal characteristics by combining the noise parameters to generate the preprocessing characteristic tensor.
5. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 4, wherein when threshold adjustment is performed on an original signal based on the noise parameters, specifically, based on a pixel-level noise map and a channel-level noise index in the noise parameters, a contraction process is performed on the original signal to obtain a denoising signal, and the expression is as follows: ; In the above-mentioned method, the step of, Is a denoising signal; is the original signal; is a reference threshold; is a threshold adjustment factor; as a positive function; is a sign function; is the noise figure at the channel level; is a pixel level noise figure.
6. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, wherein in step S4, when the fusion weight is adaptively adjusted based on the noise parameter, specifically, the weighting coefficient of the multi-domain feature of the preprocessing feature tensor is adjusted based on the channel-level noise index in the noise parameter, and a fusion feature is generated, and the expression is as follows: ; In the above-mentioned method, the step of, Is a multi-domain feature, including time domain, frequency domain, and sensor domain features; for the feature domain index to be used, Representing accumulation over all feature fields; is a weighting coefficient; domain sensitivity coefficients determined in a model training stage; Is a fusion feature; Is the channel-level noise figure.
7. The transformer fault diagnosis method based on deep learning and sensor fusion according to claim 1, wherein step S5 comprises: Receiving a fusion characteristic tensor output by a fusion layer and a channel-level noise index in noise parameters through a noise perception classifier network; Adjusting a classification threshold or a parameter of a normalization function of a noise perception classifier network based on the channel-level noise figure; Inputting the fusion feature tensor into a noise perception classifier network and outputting a corresponding fault type or health state label; And when the channel-level noise index is lower, the temperature coefficient of the classification threshold or the normalization function is reduced.
8. The transformer fault diagnosis method based on deep learning and sensor fusion according to any one of claims 1-7, wherein the deep learning model comprises a time domain modeling channel, a frequency domain modeling channel and a sensor domain modeling channel, wherein the time domain modeling channel comprises a multi-layer self-attention network, the frequency domain modeling channel comprises a combined structure of convolution layers and self-attention layers, the sensor domain modeling channel comprises self-attention layers with residual connections, and the scaling coefficients of the residual connections of the multi-layer self-attention network, the convolution kernel weights or the attention heads of the combined structure, and the self-attention layers are respectively adjusted based on channel level noise figures in the noise parameters.
9. The deep learning and sensor fusion based transformer fault diagnosis method according to any one of claims 1 to 7, wherein the multi-source sensor comprises two or more of a current sensor, a voltage sensor, a temperature sensor, a vibration sensor, an acoustic emission sensor, or a gas sensor, each of which is used for synchronous acquisition of raw signals under a unified clock.
10. A transformer fault diagnosis system based on deep learning and sensor fusion, the system being applied to the method according to any one of claims 1 to 9, the system comprising: The sensor acquisition module is used for acquiring original signals generated by at least two types of multi-source sensors arranged on the transformer to be monitored; The preprocessing module is used for calculating noise parameters of the transformer to be monitored, and preprocessing the original signals based on the noise parameters to generate preprocessing characteristic tensors; the deep learning modeling module is used for inputting the preprocessing feature tensor into a deep learning model with multiple channels, and modeling the preprocessing feature tensor in each channel respectively to obtain corresponding time domain features, frequency domain features and sensor domain features; The feature fusion module is used for adaptively adjusting fusion weights based on the noise parameters in a fusion layer of the deep learning model, fusing time domain features, frequency domain features and sensor domain features of the preprocessing feature tensor, and generating fusion feature tensors; And the fault judging module is used for judging the fault type or the health state of the transformer to be monitored based on the fusion characteristic tensor and the noise parameter.

Description

Transformer fault diagnosis method and system based on deep learning and sensor fusion Technical Field The invention relates to the technical field of data acquisition and processing, in particular to a transformer fault diagnosis method and system based on deep learning and sensor fusion. Background The transformer is used as a key device of the power system, and the operation state of the transformer is directly related to the safety and stability of the power network. However, the existing transformer fault diagnosis method still has a plurality of defects. The traditional method often depends on a single sensor signal or manual experience rules, has limited diagnosis basis, and is difficult to comprehensively reflect the real running condition of the equipment. Secondly, in an actual operating environment, the sensor signal is susceptible to noise interference, and especially under the condition of low signal-to-noise ratio, the accuracy and stability of the existing diagnostic method are remarkably reduced. Because of the difference between the equipment type and the operation condition, the conventional diagnosis method has insufficient generalization capability, usually needs to be remodelled or trained for different equipment, and is difficult to meet the universal application requirements of the cross-field. In addition, the existing method has the defects in real-time performance and calculation efficiency, and the requirements of high-precision diagnosis and online application are difficult to be met. Therefore, a new technical scheme for diagnosing faults of a transformer is needed, noise interference can be effectively restrained under the support of multi-source sensor data, reliability of diagnosis results is improved, meanwhile, strong adaptability across equipment is achieved, instantaneity and calculation efficiency are achieved, and accordingly accurate judgment of fault types and health states of the transformer is achieved. Disclosure of Invention Aiming at the problems in the prior art, the invention provides a transformer fault diagnosis method and a system based on deep learning and sensor fusion, which are used for solving the problems of insufficient utilization of multi-source signals, poor diagnosis reliability in a noise environment, insufficient adaptability across equipment, limited instantaneity and calculation efficiency and the like. In order to solve the technical problems, the invention adopts the following technical scheme: a transformer fault diagnosis method based on deep learning and sensor fusion comprises the following steps: S1, collecting original signals generated by at least two types of multi-source sensors arranged on a transformer to be monitored; S2, calculating noise parameters of the transformer to be monitored, and preprocessing an original signal based on the noise parameters to generate a preprocessing characteristic tensor; S3, inputting the preprocessing feature tensor into a deep learning model with multiple channels, and respectively modeling the preprocessing feature tensor in each channel to obtain corresponding time domain features, frequency domain features and sensor domain features; S4, in a fusion layer of the deep learning model, self-adaptively adjusting fusion weights based on the noise parameters, and fusing time domain features, frequency domain features and sensor domain features of the preprocessing feature tensor to generate fusion feature tensors; and S5, judging the fault type or health state of the transformer to be monitored based on the fusion characteristic tensor and the noise parameter. Further, in step S2, the calculating the noise parameter of the transformer to be monitored specifically includes: According to the local fluctuation degree and instantaneous anomaly degree of the time-frequency tensor of the original signal at each time point and frequency point, a pixel-level noise diagram is obtained by calculation, and the expression is as follows: ; In the above-mentioned method, the step of, Is a pixel level noise map; is the degree of fluctuation in the frequency dimension; is the degree of anomaly in the time dimension; is a compression function; 、 Is a weight parameter; Index for frequency and time; And calculating the channel-level noise index according to the average value of the pixel-level noise diagram at all time frequency points, wherein the expression is as follows: ; In the above-mentioned method, the step of, Is the noise figure at the channel level;、 The total number of frequency points and the total number of time points are respectively; Representing summation over all time-frequency points. Further, the fluctuation degree of the frequency dimension is calculated by the frequency pointThe power sequence at all time points was calculated variance obtained as follows: ; In the above-mentioned method, the step of, Is a time-frequency power value; Representing frequency points Average power at all time poi