Power equipment transient fault probability estimation method
By constructing a recurrent Bayesian neural network model, the problems of high-dimensional nonlinear feature representation and reliable quantification in the transient fault probability estimation of power equipment are solved, realizing real-time fault prediction and online adaptive updating of power equipment, and improving the safety monitoring capability of the power system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-03-31
AI Technical Summary
Existing fault probability estimation methods are difficult to accurately handle transient faults such as partial discharge and corona discharge in power equipment, especially due to their limited ability to express high-dimensional nonlinearity and time-series characteristics, and the lack of reliable quantification of prediction results.
A recurrent Bayesian neural network model integrating GRU-attention mechanism and Bayesian inference is constructed. By processing normalized current-time series waveforms, time-series features are extracted, and Hamiltonian Monte Carlo sampling is used for posterior inference to achieve uncertainty quantification and online adaptive updating of fault probability.
It improves the reliability and stability of fault prediction, adapts to changes in equipment status, and achieves real-time risk warning and high interpretability. It is applicable to transient fault estimation of power equipment such as gas-insulated switchgear and transformers.
Smart Images
Figure CN121765249A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and in particular to a method for estimating the probability of transient faults in power equipment. Background Technology
[0002] In modern power systems, the operating status of key equipment such as gas-insulated switchgear, transformers, and surge arresters is directly related to the safety and stability of the power grid. In order to achieve predictive maintenance, it is urgent to accurately predict and assess the early faults of the equipment. Among them, transient faults such as partial discharge and corona discharge are characterized by strong randomness, short duration and complex dynamic characteristics. The accurate estimation of their occurrence probability has become one of the core challenges in power equipment condition monitoring.
[0003] Existing methods for estimating fault probability mainly include analytical models based on physical mechanisms and statistical inference based on historical data. The former relies on a large number of environmental and structural parameters that are difficult to obtain accurately, making it difficult to adapt to complex and ever-changing actual operating conditions. The latter, although it can consider random factors, has limited expressive power and is difficult to extract high-dimensional nonlinear and temporal features from waveforms. Although some classification models based on machine learning have been applied to discharge detection, most of them are black-box deterministic models, lacking the ability to quantify the reliability of prediction results, and are difficult to meet the power system's requirements for high interpretability and high reliability. Summary of the Invention
[0004] This invention provides a method for estimating the probability of transient faults in power equipment. It can directly process raw transient time-series signals, has the ability to quantify endogenous uncertainties, and supports online adaptive updates of power equipment fault probability estimates, thereby improving the reliability, stability, and practicality of prediction results and meeting the requirements of intelligent power systems for real-time risk warning.
[0005] A method for estimating the probability of transient faults in power equipment includes the following steps: S1. Construct a needle-plate electrode discharge experimental platform to simulate internal defects in power equipment, collect current-time series waveforms, including discharge waveforms labeled 1 and background signals labeled 0, normalize the current-time series waveforms and perform sliding window segmentation to form a standardized training dataset. S2, Construct a recurrent Bayesian neural network model. The recurrent Bayesian neural network model includes a GRU-attention feature extraction module and a Bayesian inference module. The GRU-attention feature extraction module is used to extract temporal features from the standardized training dataset and generate a context vector. The Bayesian inference module is used to receive the context vector and output the fault probability value in the interval [0,1]. Posterior inference is performed by Hamiltonian Monte Carlo sampling method to obtain the posterior distribution sample set of network parameters. S3 combines the newly acquired current-time series waveforms, uses the posterior distribution sample set of the current network parameters as the prior, performs recursive Bayesian update to obtain the updated posterior distribution sample set, and performs online estimation based on the updated posterior distribution sample set to calculate the transient failure probability of power equipment, while outputting a quantitative index of prediction uncertainty.
[0006] Optionally, the parameters of the needle-plate electrode discharge experimental platform include: The needle electrode uses a brass needle with a radius of curvature of 0.5 mm and a cone tip angle of 60°; The plate electrode is a grounding aluminum disc with a diameter of 30cm and a thickness of 0.5cm; The experiment was conducted under standard atmospheric conditions, with a continuously adjustable DC negative high voltage of 0 to -25 kV applied between the electrodes via a high-voltage power supply. The current-time sequence waveform flowing through a 1MΩ impedance was acquired using a digital storage oscilloscope with a bandwidth of 500MHz.
[0007] Optionally, the normalization and sliding window segmentation of the current-time series waveform includes: The Min-Max method is used to normalize the current-time series waveform, mapping the data to the [0,1] interval; A sliding window with a step size of 150 sampling points and a window length of 2000 sampling points is used to segment the current-time series waveform.
[0008] Optionally, the GRU-attention feature extraction module includes a stacked 2-layer GRU layer, a hidden state with a dimension of 100, a dropout rate of 0.2, and an attention mechanism introduced after the last GRU layer. The attention weights at each time step are calculated through a trainable fully connected layer and then weighted and summed to obtain the context vector.
[0009] Optionally, the Bayesian inference module is a fully connected layer with random weights, all of which are assigned a standard Gaussian prior distribution, i.e., θ ~ N(0, I). The posterior inference is implemented using a Hamiltonian Monte Carlo sampler, wherein the simulation step size Step_size = 0.001, the number of frog-jump integration steps for each sampling point L = 30, the total number of samples is 120, and the first 20 samples are discarded as the burning period.
[0010] Optionally, the recursive Bayesian update includes: Treat the distribution P(θ|D) represented by the posterior distribution sample set of the currently stored network parameters as the new prior distribution, and treat the newly accumulated batch data D... new The initial position of the HMC sampling chain is determined by using the state of the last batch of samples in the posterior distribution sample set of the current network parameters, with P(θ|D) as the prior and D as the initial position.new Calculate the likelihood, restart the HMC sampling process, keep the sampling parameters unchanged, and replace the sample set with the updated posterior distribution sample set after completing a new round of sampling.
[0011] Optionally, the online estimation includes: When a new current-time series waveform arrives, it is normalized and segmented by a sliding window. 100 different parameter samples are randomly drawn from the updated posterior distribution sample set. The network instantiated for each sample is used for forward propagation to obtain 100 probability predictions. The average of the 100 probability predictions is calculated as the final fault probability point estimate. At the same time, the standard deviation σ and the 95% confidence interval are calculated as the quantitative output of the prediction uncertainty.
[0012] Optionally, the power equipment includes gas-insulated switchgear, transformers, or surge arresters, and the transient fault includes partial discharge and corona discharge.
[0013] Optionally, the recurrent Bayesian neural network model is built using the Python programming language and the PyTorch deep learning framework.
[0014] The beneficial effects of this invention are: This invention constructs a recurrent Bayesian neural network model that integrates a GRU-attention mechanism and a Bayesian inference structure. This model can accurately extract time-series features from normalized current-time series waveforms and quantify the uncertainty of fault probability in a high-dimensional parameter space. Based on a Hamiltonian Monte Carlo sampler, it performs posterior inference on model parameters to generate a complete distribution of fault probability rather than a single predicted value. This overcomes the overconfidence of traditional deep models and introduces credibility assessment capabilities into the prediction results, significantly improving the interpretability and robustness of the model in power system safety monitoring.
[0015] This invention, by introducing a recursive Bayesian update mechanism, can dynamically use the current posterior as the prior distribution to start the next round of sampling when new data arrives. This allows the model parameters to continuously evolve and adapt to changes in equipment state, forming a closed-loop inference process of prior-sampling-update-estimation. Combined with an online estimation strategy, it uses the posterior sample set for probability prediction and confidence interval output, realizing the transformation from static training to real-time online learning. This effectively improves the adaptability under actual working conditions, the generalization ability of small samples, and the real-time early warning capability. It is applicable to transient fault probability estimation and intelligent monitoring of various types of power equipment such as GIS, transformers, and surge arresters. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the estimation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the probability distribution of the latest current waveform diagnosis category probability in an embodiment of the present invention. Detailed Implementation
[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.
[0019] like Figures 1-2 As shown, a method for estimating the probability of transient faults in power equipment includes the following steps: Step 1: Construct a needle-plate electrode discharge experimental platform to simulate the metal burr defects inside a gas-insulated switchgear (GIS). Collect raw current signals at each discharge stage under different known spacings, labeled as Category 1; and background signals under normal operating conditions, including environmental noise, equipment operation interference and other non-discharge transient phenomena, labeled as Category 0.
[0020] The specific parameters of the needle-plate electrode discharge experimental platform simulating internal metal burr defects in gas-insulated switchgear (GIS) are as follows: the needle electrode is a brass needle with a radius of curvature of 0.5 mm and a cone tip angle of 60° to simulate the tip of a metal burr; the plate electrode is a grounded aluminum plate with a diameter of 30 cm and a thickness of 0.5 cm. The experiment was conducted under standard atmospheric conditions. A continuously adjustable DC negative high voltage of 0 to -25 kV was applied between the electrodes through a high-voltage power supply, and the transient voltage waveform flowing through a 1 MΩ impedance was acquired using a digital storage oscilloscope with a bandwidth of 500 MHz.
[0021] In this example, environmental factors such as temperature, air pressure, and humidity significantly affect the corona current. Under the premise of keeping these environmental factors relatively stable, the needle-plate electrode discharge behavior simulating internal metal burr defects in a gas-insulated switchgear (GIS) exhibits specific regularities with variations in the external electric field strength, thus forming a stable evolution trend. To ensure the reliability and consistency of the experimental results, we strive to maintain the stability of the experimental environment conditions, keeping variations within a small range of 5%.
[0022] After data acquisition, the raw dataset is preprocessed. First, the waveform signals are normalized. Specifically, the discharge transient signal and the non-discharge transient signal are normalized using the Min-Max method. The calculation formula is as follows: ; After processing, all data are mapped to the [0,1] interval to form standardized model input samples. This transforms continuous waveforms into a sample sequence that the model can process, eliminating dimensions and accelerating subsequent model convergence. A sliding window with a step size of 150 sampling points is used to segment the waveform, with each window having a fixed length of 2000 sampling points, corresponding to a time length of 330.78 μs at the oscilloscope sampling rate. This length is optimized to cover the complete occurrence and development process of a single typical corona discharge pulse.
[0023] Step 2: Construct a recurrent Bayesian neural network and perform initial training. In this embodiment, the Bayesian neural network model is built using the Python programming language and the PyTorch deep learning framework, and mainly includes the following two parts: The GRU attention feature extraction module sets the hidden state dimension of the GRU layers to 100, stacking two layers to fully learn the temporal dependency patterns in the waveform. The dropout rate is set to 0.2 to prevent overfitting. The attention layer applies an attention mechanism to the output of the last GRU layer at all time steps. The attention weights for each time step are calculated through a trainable fully connected layer, and the weighted output is used to obtain the context vector c, which is then fed into a Bayesian layer. The calculation formulas involved are as follows: ; ; ; Among them, h t v is the hidden state at time step t. a W a b a Here are the learnable parameters for the attention mechanism, where T is the sequence length and α is the value of α. t This indicates the degree of importance of time step t in affecting the final result.
[0024] The Bayesian inference module is a fully connected layer with random weights. Its input is the context vector *c* from the previous output, and its output is a scalar mapped to a fault probability value between [0,1] using a sigmoid activation function. All weight parameters in this module... Assigned a standard Gaussian prior distribution, i.e. Posterior inference was implemented using a Hamiltonian Monte Carlo (HMC) sampler. Specific parameters were set as follows: simulation step size Step_size = 0.001, number of frog-jump integration steps per sampling point L = 30, and total number of samples set to 120, with the first 20 samples discarded as a "burning period" to ensure chain convergence.
[0025] The model initially uses the first batch of 60 sets of valid waveform data (including discharge and non-discharge samples) that have been collected and preprocessed as the initial dataset D0. The HMC sampler is run to obtain the first set of posterior distribution samples for the network parameters. The initial training was completed using the first 60 sets of pre-processed waveform data with explicit labels (0 or 1), which yielded the first set of posterior distribution samples for the network parameters, indicating that the model possessed basic waveform recognition and probability prediction capabilities.
[0026] Step 3: After the model completes offline initialization, it is deployed to the online monitoring system. When the system starts, it loads a representative sample set of the posterior distribution of the current parameters. The sample set initially consists of the aforementioned 100 samples and will be replaced in subsequent updates. During real-time estimation, when a new transient waveform segment... Upon arrival, the system performs the same sliding window segmentation and normalization preprocessing on it, and randomly selects 100 different parameter samples from the current parameter posterior sample set. Use each Instantiated network pairs Forward propagation is performed to obtain 100 probability predictions, and the average of these 100 predictions is calculated as the final fault probability point estimate. The calculation formula is as follows: ; Simultaneously, calculate the standard deviation σ of these 100 values and provide the 95% confidence interval. This serves as a quantitative output of the prediction uncertainty. During the recursive Bayesian update process, the system periodically or based on triggering conditions initiates model updates, updating the currently stored posterior sample set of parameters. The distribution it represents Treating the newly accumulated batch data as a new prior distribution Add. Then repeat the HMC continuous sampling process in step 2, using the state of the last batch of samples in the current sample set as the initial position of the HMC sampling chain, and so on. As a priori, with Calculate the likelihood and restart the HMC sampling process. Sampling parameters such as step size and number of steps remain unchanged. After completing the new round of sampling, use the resulting new sample set. Replace the original This completes a full recursive Bayes update, enabling the model knowledge to dynamically adapt to the latest state of the device. Figure 2 This represents the probability distribution of the latest waveform at a given moment having a diagnostic category of 1.
[0027] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.
[0028] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method of estimating a transient fault probability of a power equipment, characterized by, The method comprises the following steps: S1, a needle-plate electrode discharge experiment platform simulating internal defects of power equipment is constructed, current-time sequence waveforms are collected, including discharge waveforms labeled as 1 and background signals labeled as 0, the current-time sequence waveforms are normalized and segmented by a sliding window to form a standardized training data set; S2, a recursive Bayesian neural network model is constructed, the recursive Bayesian neural network model comprises a GRU-attention feature extraction module and a Bayesian inference module, wherein the GRU-attention feature extraction module is used for extracting time sequence features from the standardized training data set and generating a context vector, the Bayesian inference module is used for receiving the context vector and outputting a fault probability value in the interval [0, 1], posterior inference is performed by a Hamilton Monte Carlo sampling method to obtain a posterior distribution sample set of network parameters; S3, the posterior distribution sample set of the current network parameters is taken as a priori in combination with newly collected current-time sequence waveforms, recursive Bayesian updating is performed to obtain an updated posterior distribution sample set, online estimation is performed according to the updated posterior distribution sample set, a transient fault probability of the power equipment is calculated, and a quantitative index of prediction uncertainty is output.
2. The method of claim 1, wherein, The parameters of the needle-plate electrode discharge experiment platform comprise: The needle electrode adopts a brass needle with a curvature radius of 0.5 mm and a conical tip angle of 60°; The plate electrode is a grounded aluminum plate with a diameter of 30 cm and a thickness of 0.5 cm; The experiment is performed under standard atmospheric conditions, a continuous adjustable DC negative high voltage of 0 to -25 kV is applied between the electrodes by a high-voltage power supply; A digital storage oscilloscope with a bandwidth of 500 MHz is used to collect current-time sequence waveforms flowing through a detection impedance of 1 MΩ.
3. The method of claim 2, wherein, The normalization processing and sliding window segmentation of the current-time sequence waveforms comprise: The Min-Max method is used for normalization processing of the current-time sequence waveforms, and the data is mapped to the interval [0, 1]; A sliding window with a step length of 150 sampling points and a window length of 2000 sampling points is used for segmentation of the current-time sequence waveforms.
4. The method of claim 3, wherein, The GRU-attention feature extraction module comprises 2 stacked GRU layers, a hidden state with a dimension of 100, a dropout rate of 0.2, and an attention mechanism introduced after the last GRU layer, attention weights of each time step are calculated by a trainable fully connected layer, and a context vector is obtained by weighted summation.
5. The method of claim 4, wherein, The Bayesian inference module is a fully connected layer with random weights, all weight parameters θ of the fully connected layer are assigned a standard Gaussian prior distribution, that is, θ ~ N(0, I), and posterior inference is realized by a Hamilton Monte Carlo sampler, wherein a simulation step size Step_size=0.001, a frog hopping integral step number L=30 for each sampling point, and a total sampling sample number is 120, and the first 20 samples are discarded as a burning period.
6. The method of claim 5, wherein, The recursive Bayesian updating comprises: The distribution P(θ|D) represented by the current stored set of posterior distribution samples of the network parameters is taken as the new prior distribution, and the newly accumulated batch of data D new is added. The last sample in the current set of posterior distribution samples of the network parameters is taken as the initial position of the HMC sampling chain, and P(θ|D) is taken as the prior. The HMC sampling process is restarted, and the parameters remain unchanged. After a new round of sampling is completed, the updated set of posterior distribution samples obtained is used for replacement. new The likelihood is calculated, the HMC sampling process is restarted, the parameters remain unchanged, and after a new round of sampling is completed, the updated set of posterior distribution samples obtained is used for replacement.
7. The method of claim 6, wherein, The online estimation comprises: When a new current-time sequence waveform arrives, it is normalized and divided into a sliding window, 100 different parameter samples are randomly selected from the updated posterior distribution sample set, the network is instantiated using each sample, and forward propagation is performed to obtain 100 probability prediction values, the average of the 100 probability prediction values is calculated as the final fault probability point estimate, and the standard deviation sigma and the 95% confidence interval are calculated as the quantitative output of the prediction uncertainty.
8. The method of claim 7, wherein, The power equipment includes a gas insulated combined electric appliance, a transformer or a lightning arrester, and the transient fault includes partial discharge or corona discharge.
9. The method of claim 8, wherein, The recurrent Bayesian neural network model is built by using a Python programming language and a PyTorch deep learning framework.