CN-122020330-A - Substation GIS equipment partial discharge fault type identification method considering environmental interference compensation

CN122020330ACN 122020330 ACN122020330 ACN 122020330ACN-122020330-A

Abstract

The invention relates to the technical field of fault diagnosis of power system equipment, in particular to a substation GIS equipment partial discharge fault type identification method considering environmental interference compensation, which comprises multi-source multi-mode signal acquisition, mixed interference decoupling and preprocessing, physical information fusion dynamic interference compensation, fusion feature extraction, interpretability feature optimization and fault identification, full-flow optimization of cloud-side collaborative closed-loop updating, multi-mode signal synchronous acquisition through an FPGA+edge node, preprocessing such as improved FastICA decoupling, db4 wavelet denoising and the like, fusion feature vector construction, dynamic compensation by a random forest model, 16-dimensional traditional feature extraction and 8-dimensional deep learning feature fusion, and input migration learning-PSO-ELM model identification optimization through SHAP value analysis and genetic algorithm optimization, and model updating through cloud-side collaborative realization.

Inventors

GONG ZHENZHOU
XIONG HAIQIANG
ZHU JIANWU
ZHENG CHENQUAN
YU ZHONGSHU
WANG JIAXIN
WANG WEIZHANG
PENG MINGCONG

Assignees

国网江西省电力有限公司南昌供电分公司

Dates

Publication Date: 20260512
Application Date: 20251215

Claims (10)

1. The method for identifying the partial discharge fault type of the GIS equipment of the transformer substation by considering the environmental interference compensation is characterized by comprising the following steps of: S1, multi-source multi-mode signal acquisition, namely arranging a discharge sensor array and an interference-state sensor array at key monitoring points of GIS equipment, and acquiring multi-mode mixed signals comprising partial discharge, environmental interference and equipment states through an FPGA and edge node synchronous control architecture, so as to ensure that the error of a data time stamp is less than or equal to 1 mu s; S2, mixed interference decoupling and preprocessing, namely decoupling a pure discharge signal candidate set and an environment interference signal set from the multi-mode mixed signal in the step S1 by adopting an improved blind source separation algorithm, and respectively performing wavelet threshold denoising, polynomial baseline correction and standardization processing on the two types of signals to obtain a denoised discharge signal and a preprocessed environment interference signal; s3, dynamic interference compensation of physical information fusion, namely, based on the two types of signals in the step S2, extracting data characteristics and physical constraint characteristics of interference signals to construct a fusion characteristic vector, training a physical information enhancement type interference compensation model containing an incremental learning module, and correcting the de-noised discharge signal through a dynamic weight compensation function to obtain a compensated discharge signal; S4, fusion feature extraction, namely synchronously extracting traditional features of a time domain, a frequency domain, a time frequency domain and deep learning time-frequency diagram features of the compensated discharge signals in the step S3, and constructing a fusion feature set; S5, performing explanatory feature optimization and fault identification, namely performing explanatory analysis on the fusion feature set in the step S4, performing optimization by adopting a genetic algorithm to obtain an optimal feature subset, inputting an intelligent identification model with improved transfer learning to realize fault classification, and outputting an identification result; S6, cloud edge collaborative closed-loop updating, namely based on the identification result in the step S5, adopting a standard data set and a field new sample to verify, and updating a compensation model and an identification model through a cloud edge collaborative framework when the accuracy rate is less than 95% for 3 times continuously, wherein the updated model is used for subsequent fault identification.
2. The method of claim 1, wherein in step S1, the discharge sensor array is a combination of an ultrahigh frequency sensor, an ultrasonic sensor and a high frequency current sensor, the interference-state sensor array comprises an electromagnetic radiation sensor, a corona sensor, a pulse interference sensor, an infrared temperature sensor and an SF 6 air pressure sensor, the synchronization control adopts a clock synchronization algorithm, a core generates fixed frequency clock trigger sampling for an FPGA, and a clock frequency Sampling frequency The sampling duration t=10s, and the timestamp error verification formula is: , Wherein, the The time stamp differences of the same time signal are acquired for different sensors, 、 Respectively the first Sampling time stamps for individual sensors.
3. The method of claim 1, wherein in step S2, the modified blind source separation algorithm is a modified fastca algorithm, the negative entropy is maximized as an objective function, and the objective function formula is , Wherein, the For the separated and output independent signals, the larger the negative entropy value is, the stronger the signal independence is, the iteration number K=100, the convergence threshold epsilon=1e-6 is, and the decoupling core formula is as follows WX, wherein W is M×M dimension unmixed matrix, M is sensor channel number, updated by Newton iteration method, X is N×M dimension multi-mode mixed signal matrix, N is sampling point number, the wavelet threshold denoising adopts db4 wavelet base, and the number of decomposition layers The threshold value calculation formula is , Wherein sigma is a noise estimation value, and the calculation formula is W is a detail coefficient after wavelet decomposition, mean (·) is a median operation, N is the number of signal sampling points, and a coefficient reconstruction formula after denoising is Wherein sign (·) is a sign function, w' is a denoised wavelet coefficient, the polynomial baseline correction uses a 5 th order polynomial fit, and the formula is , Wherein x (t) is the original signal, In order to correct the post-signal, For 5 th order polynomial fitting coefficients, i=0, 1,..5, fitting errors are minimized by least squares solution, normalization employs maximum and minimum normalization, and the formula is , Wherein, the 、 Respectively corrected signals R minimum and maximum values, and mapping the normalized signal value range to the [0,1] interval to eliminate dimension difference.
4. The method according to claim 1, wherein in step S3, the fusion feature vector is The definition and calculation modes of each characteristic are as follows, In order to interfere with the signal peaks, For the pulse-width of the interference signal, The center of gravity of the frequency spectrum is calculated as , Wherein, the Is the frequency value of the kth frequency bin, K is the total number of frequency points and reflects the frequency position in which the signal energy is concentrated for the power spectral density of the corresponding frequency points, The energy entropy is calculated as the formula , Wherein, the , For the energy of the i-th frequency band, Is the total energy of the signal, I is the number of frequency bands, reflects the uniformity of energy distribution, Is a physical constraint characteristic set and comprises an electromagnetic wave propagation loss coefficient alpha and an acoustic wave attenuation coefficient beta, wherein alpha is calculated by the following formula , Wherein d is the distance between the sensor and the interference source, f is the center frequency of the interference signal, 、 The gain of the signal transmitting end and the gain of the signal receiving end are respectively obtained, the compensation model is a physical information enhanced random forest, and the number of decision trees is determined Tree depth The loss function formula is , Wherein N is the number of training samples, The true distortion amount of the discharge signal for the ith sample, For model predictive distortion, γ=0.3 is a physical constraint weight, In order to actually measure the physical characteristic value, Values are calculated for physical feature theory.
5. The method of claim 4, wherein in step 3, the incremental learning module uses a sliding window mechanism with a window sample size of 500 sets, the new interference identification uses a cosine similarity algorithm, and the identification formula is , Wherein, the As a feature vector for the interference to be identified, The method is characterized in that feature vectors in a known interference feature library, "·" is a vector dot product, |·| is a vector modular length, when Sim <0.7 is judged to be a new type of interference, on-line updating of model parameters is triggered, a random gradient descent method is adopted, full-scale retraining is avoided, and the dynamic weight compensation function is as follows , Wherein y is the de-noised discharge signal, y' is the compensated discharge signal, For the amount of discharge signal distortion predicted by the model, For the self-adaptive weight of the interference type, θ=0.9 during corona interference and θ=1.1 during pulse interference, the influence intensity of different interferences on the discharge signal is adapted.
6. The method according to claim 1, wherein in step S4, the conventional feature is 16 dimensions, specifically including: (1) Time domain 4-dimensional, peak factor Wherein As a result of the signal peaks, Kurtosis as signal effective value Wherein E is a mathematical expectation, a form factor Wherein Is the signal average value; (2) Frequency domain 3-dimensional, spectral peak I.e. power spectral density maximum, band energy I.e. 50MHz-1GHz frequency band power integration, spectral entropy Wherein The power ratio of each frequency point; (3) Time-frequency domain 9-dimensional, db6 wavelet packet 3-layer decomposed 8-frequency band energy duty ratio Entropy of 1 wavelet packet ; (4) The deep learning features are 8-dimensional, the extraction flow is that the compensated signal is converted into a 256×256 time-frequency diagram through short-time Fourier transform, light CNN extraction is input, the CNN structure is 3 layers 3×3 convolution kernel+2 layers 2×2 maximum pooling, and 8-dimensional feature vectors are output; (5) The fusion feature set is 24-dimensional, and the splicing formula is , wherein, For a 16-dimensional conventional feature vector, Feature vectors are deep learned for 8 dimensions.
7. The method of claim 1, wherein in step S5, the interpretability analysis uses a SHAP value analysis method, the calculation uses TreeExplainer algorithm, the random forest model is adapted, and the SHAP value calculation formula is , Wherein, the Reflecting the contribution degree of the features to the model output for the SHAP value of the ith feature, wherein positive is promotion, negative is inhibition, S is a feature subset, F is a full feature set, n is the total number of features, F (·) is a model output function, and eliminating The absolute value average value is less than 5% of redundant characteristics, the genetic algorithm is used for characteristic optimization, and the parameters are population scale Iteration number g=100, crossover probability Probability of variation The fitness function is , Wherein: In order to support the recognition accuracy of the vector machine, And the weight distribution gives consideration to both precision and differentiation for the variance of the characteristic SHAP value.
8. The method of claim 7, wherein in step S5, the smart recognition model is a transfer learning-PSO-ELM model with a structure of "input layer 8 node-hidden layer 20 node-output layer 4 node", and the ELM core formula is Wherein Y is the output vector class 4 fault state, H is the hidden layer output matrix, For outputting weight, domain self-adaption adopts maximum mean difference alignment distribution by Moore-Penrose generalized inverse solution, and the formula is , Wherein X is a laboratory sample set, n samples, Y is a field sample set, m samples, For a Gaussian kernel mapping function, the smaller the MMD value is, the smaller the domain distribution difference is, the domain adaptive factor eta=0.15, the input layer weight W and the hidden layer bias b of the PSO optimization ELM are, the parameters are that the population scale N=30, the learning factor c 1 =c 2 =2, and the inertia weight updating formula is that , Wherein G is the current iteration number, G is the total iteration number, inertia is dynamically adjusted along with iteration advancing and domain difference, and a particle position updating formula is as follows , Wherein, the In order to be able to achieve a particle velocity, For the optimal location of the individual, As a global optimum position for the device, 、 Is a [0,1] random number.
9. The method of claim 1, wherein in step S6, the standard data set is an IEEEP2810 standard data set, and the 20-class interference +4-class fault samples are included, and the verification index comprises accuracy rate , Wherein TP is true positive, TN is true negative, FP is false positive, FN is false negative, and F1 value is , And 4×4 confusion matrix, wherein the horizontal axis of SHAP thermodynamic diagram is characteristic, the vertical axis is sample, the color depth corresponds to SHAP value, the cloud edge collaborative update adopts incremental learning, and the parameter update formula is that , Wherein, theta is a model parameter, In order to update the pre-parameters, For updated parameters, η=0.01 is learning rate, L is loss gradient, and updating takes less than or equal to 10min.
10. The method according to claim 1, wherein in step S6, the field verification samples are collected into 100 groups each month, covering the working condition of 30% -100% of load rate, the training set and the test set are sampled in a 7:3 layered mode, the proportion of each type of sample is kept consistent, and the update triggering condition is the average value of the accuracy of 3 continuous verifications , Acc 1 、Acc 2 、Acc 3 is the accuracy rate of 3 times of verification, the updated model is issued in an incremental package mode, and the deployment time is less than or equal to 5 minutes.

Description

Substation GIS equipment partial discharge fault type identification method considering environmental interference compensation Technical Field The invention relates to the technical field of fault diagnosis of power system equipment, in particular to a substation GIS equipment partial discharge fault type identification method considering environmental interference compensation. Background The Gas Insulated Switchgear (GIS) is core distribution equipment of a transformer substation, the accurate identification of partial discharge faults is a key for predicting insulation degradation and preventing power failure accidents, although the current GIS partial discharge fault identification technology has evolved towards intellectualization, a plurality of technology short boards still exist, the existing scheme adopts a single sensor to collect signals, complex interference such as corona and electromagnetic radiation of the transformer substation is difficult to resist, the signal to noise ratio is only about 22dB, interference is easy to mask discharge real characteristics, interference compensation is static design, dynamic time-varying interference cannot be adapted, excessive or insufficient compensation of signals is easy to be caused, distortion is high, an identification model is mainly a pure data driving 'black box', physical mechanism support is lacked, interpretability is insufficient, small sample generalization capability is weak, meanwhile, the model is multi-static deployment and has no long-acting update mechanism, and identification precision can be gradually attenuated along with on-site environment change. Disclosure of Invention At present, the existing GIS partial discharge fault recognition technology has a plurality of defects, wherein the GIS partial discharge fault recognition technology adopts a single sensor to collect signals, is easily influenced by environmental factors such as transformer substation corona, electromagnetic radiation, pulse interference and the like, has low signal to noise ratio, simultaneously has interference compensation for static known interference design, has insufficient self-adaptive capacity to dynamic time-varying interference, and has the intelligent recognition model which is a black box model driven by pure data, lacks physical mechanism support, has poor interpretability and weak generalization capability of small samples, lacks a long-term updating mechanism after model training, and is difficult to adapt to on-site environment change, so that the fault recognition method with anti-interference capability, recognition precision and long-term adaptability is considered. Aiming at the defects of the prior art, the invention aims to provide a substation GIS equipment partial discharge fault type identification method considering environmental interference compensation, which aims to solve the problems that a single sensor collects signals, is easily influenced by environmental factors, has low signal-to-noise ratio, has insufficient self-adaptive capacity of dynamic time-varying interference, lacks physical mechanism support, has poor interpretability and weak generalization capability of small samples, lacks a long-acting update mechanism after model training, and is difficult to adapt to field environmental changes. A substation GIS equipment partial discharge fault type identification method considering environmental interference compensation comprises the following steps: S1, multi-source multi-mode signal acquisition, namely arranging a discharge sensor array and an interference-state sensor array at key monitoring points of GIS equipment, and acquiring multi-mode mixed signals comprising partial discharge, environmental interference and equipment states through an FPGA and edge node synchronous control architecture, so as to ensure that the error of a data time stamp is less than or equal to 1 mu s; S2, mixed interference decoupling and preprocessing, namely decoupling a pure discharge signal candidate set and an environment interference signal set from the multi-mode mixed signal in the step S1 by adopting an improved blind source separation algorithm, and respectively performing wavelet threshold denoising, polynomial baseline correction and standardization processing on the two types of signals to obtain a denoised discharge signal and a preprocessed environment interference signal; s3, dynamic interference compensation of physical information fusion, namely, based on the two types of signals in the step S2, extracting data characteristics and physical constraint characteristics of interference signals to construct a fusion characteristic vector, training a physical information enhancement type interference compensation model containing an incremental learning module, and correcting the de-noised discharge signal through a dynamic weight compensation function to obtain a compensated discharge signal; S4, fusion feature extraction, namely syn