Method for classifying multi-source partial discharge in gas insulated switch cabinet based on fluorescent optical fiber

Abstract

The invention provides a multi-source partial discharge classification method in a gas insulated switch cabinet based on fluorescent optical fibers. The method comprises the following steps: constructing four typical partial discharges in the gas insulated switch cabinet; fluorescent optical fibers are adopted to surround the discharge defects, distributed partial discharge detection is carried out, and various multi-source partial discharge optical signals are obtained. And performing feature extraction processing on the partial discharge optical signal to obtain a data set. A data set is divided according to a certain proportion, one part serves as a training set, and the other part serves as a test set period. And the XGBoost and the LightGBM are optimized and fused, so that pre-classification is realized. And dynamic weight adjustment is applied to the training process of each type of partial discharge according to a pre-classification result, and final classification is carried out. The method has remarkable technical advantages in the aspects of improving the multi-source partial discharge identification accuracy, enhancing the anti-interference capability, reducing the calculation and deployment complexity, improving the engineering application feasibility and the like.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment fault diagnosis, specifically to a method for classifying multi-source partial discharges in a fluorescent fiber optic gas-insulated switchgear. Background Technology

[0002] High-voltage switchgear is an indispensable core device in power systems, undertaking critical functions such as power transmission, distribution, control, and protection. During its manufacturing, processing, transportation, operation, and long-term service, internal insulation defects such as burrs and particles inevitably occur. These defects can cause electric field distortion, leading to excessively high local electric field strength, which in turn induces partial discharge. Detecting partial discharge is an important part of the routine maintenance work for power workers.

[0003] Partial discharges in gas-insulated switchgear can be mainly classified into four categories: corona discharge caused by metal burrs, surface discharge caused by contamination of the insulator surface, air gap discharge caused by gaps formed due to loose metal components, and suspension discharge caused by charged particles suspended between the high-voltage end and the grounding end. Different types of discharges have their own unique characteristics and degrees of harm; therefore, corresponding prevention and control measures need to be developed based on their causes and characteristics to ensure the safe and stable operation of the equipment. Therefore, in-depth research on the detection and classification of partial discharges not only helps in the early identification and prevention of potential insulation defects, effectively extending equipment service life and reducing operation and maintenance costs, but also provides solid theoretical support for the design, manufacture, and operation of gas-insulated switchgear.

[0004] Currently, the classification methods for partial discharge in gas-insulated switchgear include direct and indirect methods. Direct methods include: ultrasonic detection of partial discharge, UHF signal detection of partial discharge, optical signal detection of partial discharge, pulse current signal detection of partial discharge, and transient ground wave (TEV) detection of partial discharge. Indirect methods include: gas decomposition component detection. Traditional ultrasonic detection methods have strong resistance to electromagnetic interference, but significant on-site interference problems exist. UHF detection methods have high sensitivity and good resistance to low-frequency interference, and can identify discharge types, but are costly and suffer from significant signal attenuation in complex multi-conductor environments. Pulse current detection methods have high sensitivity and accurate quantification, but are susceptible to background electromagnetic interference. Transient ground wave (TEV) detection methods can quickly screen for discharge phenomena, but are susceptible to electromagnetic noise interference. Compared to the above detection methods, optical signal detection technology has received increasing attention from the academic community due to its unique advantages. This method utilizes fluorescent optical fibers that can be embedded inside the switchgear, enabling flexible deployment and wide-area coverage even in environments with complex equipment structures and limited space. Fluorescent materials exhibit high sensitivity to partial discharge radiation signals, enabling high-frequency, real-time monitoring. They also effectively avoid the problem of traditional electrical sensors being susceptible to interference in strong electromagnetic environments, significantly improving the stability and reliability of detection.

[0005] In actual operating conditions, switchgear often exhibits aliasing of multi-source and multi-type partial discharge signals. Considering the significantly different hazards posed by different types of partial discharge to equipment—for example, discharges caused by loose or detached connectors are usually difficult to self-eliminate, while discharges caused by burrs may gradually ablate and passivate over time, eventually disappearing—accurate identification of various partial discharges within multi-source and multi-type aliased signals is crucial. Therefore, researchers have focused on classifying multi-source and multi-type aliased discharge signals. Currently, research datasets in this field mainly originate from experiments using discharge simulation cavities and discharge models, rather than from actual equipment. Existing research has confirmed the feasibility of using mathematical models and algorithms for multi-source aliased partial discharge pattern recognition. However, existing datasets and algorithms do not fully cover the four typical discharge types and their combinations; more importantly, the aforementioned datasets were not obtained from actual power equipment. This presents challenges and risks of failure and misjudgment in engineering applications. Summary of the Invention

[0006] 1. The technical problem to be solved:

[0007] Accurately identify various partial discharges in multi-source, multi-type aliased signals.

[0008] 2. Technical Solution:

[0009] To address the above problems, this invention provides a multi-source partial discharge classification method for gas-insulated switchgear based on fluorescent optical fiber. The method involves preprocessing the collected partial discharge signals using windowing and feature extraction to generate a dataset. This dataset is then trained on an XGBoost-LightGBM model optimized using Bayesian methods and fused using soft voting. This accurately identifies single-source and multi-source partial discharge types, ensuring the long-term reliable operation of the gas-insulated switchgear. The specific method includes the following steps:

[0010] Step S01: Four typical partial discharges are constructed in a real gas-insulated switchgear for detection. The four typical partial discharges are: air gap discharge, surface discharge, needle tip discharge, and floating discharge.

[0011] Step S02: Use fluorescent optical fiber to surround these discharge defects to perform distributed partial discharge detection and obtain various multi-source partial discharge optical signals.

[0012] Step S03: Preprocess the partial discharge signal to obtain the total number of discharges and the average discharge value for each type of discharge in each cycle. Perform feature extraction processing on the partial discharge optical signal for each cycle to obtain the dataset.

[0013] Step S04: Divide the dataset according to a certain ratio, with one part as the training set and the other part as the test set.

[0014] Step S05: Optimize and merge XGBoost and LightGBM to achieve pre-classification.

[0015] Step S06: Apply dynamic weight adjustments to the training process of each type of partial discharge based on the pre-classification results and perform final classification.

[0016] 3. Beneficial effects:

[0017] This invention has significant technical advantages in improving the accuracy of multi-source partial discharge identification, enhancing anti-interference capabilities, reducing computational and deployment complexity, and improving the feasibility of engineering applications. It can provide reliable technical support for power equipment condition monitoring, intelligent early warning, and operation and maintenance decision-making, and has good engineering application prospects and promotion value. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the multi-source partial discharge mode recognition process according to an embodiment of this application.

[0019] Figure 2 This is a waveform diagram of the air gap partial discharge optical signal according to an embodiment of this application.

[0020] Figure 3 The above is a comparison chart of the prediction results and effects of the embodiments of this application.

[0021] Figure 4 This is a single-source discharge confusion matrix diagram showing the prediction results of an embodiment of this application.

[0022] Figure 5 This is a dual-source discharge confusion matrix diagram representing the prediction results of an embodiment of this application.

[0023] Figure 6 The diagram shows the three-source discharge confusion matrix for the prediction results of an embodiment of this application.

[0024] Figure 7 This is a multi-source aliasing discharge confusion matrix diagram representing the prediction results of an embodiment of this application.

[0025] Figure 8 The image shows a T-SNE scatter plot of the predicted multi-source aliasing discharge for an embodiment of this application. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] like Figure 1As shown, a method for classifying multi-source partial discharges in a fluorescent fiber optic gas-insulated switchgear includes the following steps:

[0028] Step S01: Four typical partial discharges are constructed in a real gas-insulated switchgear for detection. The four typical partial discharges are: air gap discharge, surface discharge, needle tip discharge, and floating discharge.

[0029] This invention addresses the problem of difficulty in distinguishing and susceptibility to interference of multi-source partial discharge signals in gas-insulated switchgear. By deploying a fluorescent fiber optic sensing system within a real switchgear, it achieves distributed acquisition of partial discharge optical signals, effectively overcoming the shortcomings of traditional electrical detection methods, such as weak electromagnetic interference resistance, difficulty in localization, and severe superposition of multi-source signals, thus improving the stability and reliability of the detection process. Based on this, a measured dataset covering four typical partial discharge types and their various combinations is constructed, making model training more closely resemble the actual operating environment and significantly enhancing the engineering applicability of the method.

[0030] Step S02: Fluorescent optical fibers are used to surround these discharge defects to perform distributed partial discharge detection, acquiring various multi-source partial discharge optical signals. The electro-optic signals are as follows: Figure 2 As shown.

[0031] In one embodiment, in step S02: the signal detected by the fluorescent optical fiber obtains single-source discharges of four typical partial discharges in the switch cabinet, and multi-source partial discharge optical signals of dual-source, triple-source and quadruple-source discharges, resulting in air gap discharge, surface discharge, suspension discharge, corona discharge, air gap + surface discharge, air gap + suspension discharge, air gap + corona discharge, surface + suspension discharge, surface + corona discharge, suspension + corona discharge, air gap + surface + suspension discharge, air gap + surface + corona discharge, surface + suspension + corona discharge, air gap + suspension + corona discharge, and air gap + surface + suspension + corona discharge, totaling 15 types of discharges.

[0032] Step S03: Preprocess the partial discharge signal to obtain the total number of discharges and the average discharge value for each type of discharge in each cycle. Perform feature extraction processing on the partial discharge optical signal for each cycle to obtain the dataset.

[0033] In step S03, the effective discharge of each cycle is extracted using the peak detection method. The feature extraction method utilizes the following formula:

[0034] (1)

[0035] (2)

[0036] Among them, S eff The effective discharge amplitude is represented by SA, and the effective discharge threshold is represented by z. i Represents the value of the data point. Let Q be the arithmetic mean, and Q be the total number of data collection points.

[0037] The peak interval of the peak detection function is set to 10 data points, where the highest value within this interval is considered the discharge amplitude.

[0038] (3)

[0039] (4)

[0040] Where, N eff AverageS represents the number of effective discharges per cycle. eff The molecule ΣS represents the average effective discharge amplitude per cycle. eff It is the sum of the total discharge amplitudes of the effective discharge within a given period.

[0041] In one embodiment, to more accurately analyze phase features and introduce additional features to improve diagnostic accuracy, a cycle is divided into 15 regions. In equation (5), the time-domain signal in each time region is represented as C. t , where t ranges from 1 to 15. In equation (6), N eff (t) represents the discharge count for each region, which will generate a dataset of 15,000 cycles, each with the above characteristics. Then, the discharge count matrix can be represented as follows.

[0042] (5)

[0043] (6).

[0044] In one embodiment, each cycle lasts 20 milliseconds and contains 100,000 data points. After collecting discharge signals for 15,000 cycles, feature extraction is performed.

[0045] This invention effectively extracts features from partial discharge optical signals and introduces a fusion classification model composed of Extreme Gradient Boosting Tree (XGB) and Lightweight Gradient Boosting Machine (LGBM). It fully leverages the complementary advantages of the two models in nonlinear feature modeling and complex pattern discrimination. Compared with a single classification model, it exhibits higher diagnostic accuracy and generalization ability in multi-class and multi-source mixed discharge identification tasks. At the same time, it combines a Bayesian optimization strategy to adaptively optimize the model hyperparameters, which reduces the cost of manual parameter tuning while ensuring recognition accuracy and improving model construction efficiency.

[0046] Step S04: Divide the dataset according to a certain ratio, with one part as the training set and the other part as the test set.

[0047] In one embodiment, the ratio is 8:2, with the former serving as the training set and the latter as the test set.

[0048] Step S05: Optimize and merge XGBoost and LightGBM to achieve pre-classification.

[0049] The specific details of the optimization and integration of XGBoost and LightGBM are as follows:

[0050] (7)

[0051] Given the similarity between the principles of LGBM and XGB, this invention proposes a formula (7) to represent the objective function of both, where l(y i ,y i ') is the loss function for the training data, used to measure the predicted value y. i 'and truth value y i The error between them, Ω(f) k f(x) is a regularization term used to prevent overfitting. It penalizes the complexity of the tree, typically involving the number of leaf nodes and their weights. Here, N represents the total number of samples, K represents the total number of trees in the model, and f(x) = f(x). k This represents the output of model K rounds. The regularization term is shown in formula (8).

[0052] (8)

[0053] In the formula, T is the number of leaf nodes in the tree, and ω j is the weight of leaf node j, and γ and λ are the regularization parameters that control the number of leaf nodes and the weight of leaf nodes, respectively.

[0054] Step S06: Apply dynamic weight adjustments to the training process of each type of partial discharge based on the pre-classification results and perform final classification.

[0055] The weights are applied as follows:

[0056] (9)

[0057] Where N represents the number of samples, ω i The weights for sample i (obtained by manually setting class weights in the code). l(y i ,f(x i )) is the loss function, which measures the true label y. i and predicted value f(x) i The error between Ω(f) k ) is a regularization term for model complexity, used to prevent overfitting, f k This represents the output of the model in k rounds. K is the total number of trees in the model.

[0058] The specific method for step 6 includes the following steps:

[0059] Step S61: After determining the sample weights, use Bayesian optimization to optimize the hyperparameters of the XGBoost and LightGBM models.

[0060] (10)

[0061] Here, f(x) represents the objective function, m(x) represents the mean function, and k(x, x') represents the covariance function or kernel function, which characterizes the correlation between input points x and x'. The higher the correlation, the larger the value of k.

[0062] Step S62: For the hyperparameter tuning of XGB and LGBM, select the radial basis function (RBF) kernel as the kernel function of GP.

[0063] (11)

[0064] Where, σ 2 The value represents the amplitude of the kernel function, and 'l' represents the length scale. The length scale determines the range by which changes in hyperparameters affect the objective function.

[0065] Step S63: Obtain the predicted distribution of each hyperparameter combination through a GP, so that uncertainty can be quantified and a choice can be made, as described by formula (12):

[0066] (12)

[0067] Where μ(x) represents the prediction mean, reflecting the overall performance, and Σ(x, x) represents the prediction variance at x.

[0068] Step S64: Use AF to select the optimal combination of hyperparameters. This study uses EI as the AF. EI calculates the expected improvement relative to the current best solution, especially the expected value of the objective function f(x) near the current maximum value. The formula for EI is as follows: (13)

[0069] Among them, f best Let f(x) represent the current best solution, and f(x) be the predicted value of the objective function at a given point x.

[0070] Step S65: After optimizing the parameters of the two models and obtaining their respective optimal solutions, the optimal solutions of the two models are screened and merged using a soft voting method.

[0071] (14)

[0072] (15)

[0073] Where score(n) represents the weighted sum of the predicted probabilities for each class n from the classifier. Meanwhile, p 1,n and p 2,n Y represents the predicted probabilities of the two classifiers for class n, and Y is the final prediction result of the fusion model.

[0074] In one embodiment, the weights are applied as follows: air gap discharge mixed suspension discharge is 2.0, air gap discharge is 0.8, the four types of discharge mixed is 1.4, surface discharge mixed suspension discharge is 1.4, and the remaining categories are all 1.0.

[0075] This invention introduces a dynamic category weight adjustment mechanism, which effectively alleviates the identification bias problem caused by the uneven distribution of multi-source discharge samples. This significantly improves the model's recall rate for models with few samples and easily confused discharge types, enhancing its ability to identify key high-risk discharge patterns. Experimental results show that this method can achieve high-precision and stable pattern recognition under real-world multi-source partial discharge conditions in switchgear, and exhibits good adaptability to both single-source and multi-source mixed discharges.

[0076] Let A be the air gap discharge label, B be the surface discharge label, C be the suspension discharge label, and D be the corona discharge label.

[0077] like Figure 3 As shown, the existing models have a classification prediction accuracy of at least 79% of the sample count, a maximum of 91%, and generally around 85%.

[0078] like Figure 4 The diagram shown is a single-source discharge confusion matrix of the prediction results from an embodiment of this application. The predicted samples and actual structures have almost no error, ranging from a minimum of 97% to a maximum of 100%.

[0079] like Figure 5 The figure shows the dual-source discharge confusion matrix of the prediction results in the embodiment. As can be seen from the figure, the number of predicted classifications ranges from 94% to 100%.

[0080] like Figure 6 As shown in the figure, the three-source discharge confusion matrix of the prediction results in the embodiment shows that the number of predicted classifications ranges from 94% to 100%.

[0081] like Figure 7 As shown in the example, the multi-source aliasing discharge confusion matrix of the prediction results shows that, in the most complex case, the number of predicted classifications reaches up to 98%, and is generally above 90%.

[0082] like Figure 8The figure shows a T-SNE scatter plot of the prediction results of multi-source aliasing discharge according to an embodiment of this application. This is a visualization of the prediction results. As can be seen from the figure, regional clusters clearly appeared for the 15 types of discharge, with not many adjacent clusters and doping. This indicates that the model of the present invention effectively distinguishes these discharge types, making the classification effect better and the classification more accurate.

Claims

1. A method for classifying multi-source partial discharges in a fluorescent fiber optic gas-insulated switchgear, characterized in that: Includes the following steps: Step S01: Four typical partial discharges are constructed in a real gas-insulated switchgear for detection. The four typical partial discharges are: air gap discharge, surface discharge, needle tip discharge, and floating discharge. Step S02: Use fluorescent optical fiber to surround these discharge defects to perform distributed partial discharge detection and obtain various multi-source partial discharge optical signals; Step S03: Preprocess the partial discharge signal to obtain the total number of discharges and the average discharge value for each type of discharge in each cycle. Perform feature extraction processing on the partial discharge optical signal for each cycle to obtain the dataset. Step S04: Divide the dataset according to a certain ratio, with one part as the training set and the other part as the test set. Step S05: Optimize and merge XGBoost and LightGBM to achieve pre-classification; Step S06: Apply dynamic weight adjustments to the training process of each type of partial discharge based on the pre-classification results and perform final classification.

2. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 1, characterized in that: In step S02: The signal detected by the fluorescent optical fiber obtains single-source discharges of four typical partial discharges in the switch cabinet, as well as multi-source partial discharge optical signals of dual-source, triple-source, and quadruple-source discharges, resulting in 15 types of discharges: air gap discharge, surface discharge, suspension discharge, corona discharge, air gap + surface discharge, air gap + suspension discharge, air gap + corona discharge, surface + suspension discharge, surface + corona discharge, suspension + corona discharge, air gap + surface + suspension discharge, air gap + surface + corona discharge, surface + suspension + corona discharge, air gap + suspension + corona discharge, and air gap + surface + suspension + corona discharge.

3. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 1, characterized in that: In step S03, the effective discharge of each cycle is extracted using the peak detection method. The feature extraction method utilizes the following formula: (1) (2) Among them, S eff The effective discharge amplitude is represented by SA, and the effective discharge threshold is represented by z. i Represents the value of the data point. Let Q be the arithmetic mean, and Q be the total number of data collection points. The peak interval of the peak detection function is set to 10 data points, where the highest value within this interval is considered the discharge amplitude. (3) (4) Where, N eff AverageS represents the number of effective discharges per cycle. eff The molecule ΣS represents the average effective discharge amplitude per cycle. eff It is the sum of the total discharge amplitudes of the effective discharge within a given period.

4. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 3, characterized in that: One cycle is divided into 15 regions, generating a dataset of 15,000 cycles. Each cycle has the aforementioned characteristics, and the discharge counting matrix is ​​represented as follows. (5) (6) Where Ct represents the time-domain signal in each time region, where t ranges from 1 to 15, and Neff(t) represents the discharge count in each region.

5. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 4, characterized in that: Each cycle lasts 20 milliseconds and contains 100,000 data points. After collecting discharge signals for 15,000 cycles, the data is extracted.

6. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 1, characterized in that: In step S04, a certain ratio of 8:2 is used, with the former serving as the training set and the latter as the test set.

7. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 1, characterized in that: In step S05: the objective functions of XGBoost and LightGBM are expressed by formula (7). (7) Where N represents the total number of samples, K represents the total number of trees in the model, and f k This represents the output of model K rounds, l(y) i ,y i ') is the loss function for the training data, used to measure the predicted value y. i 'and truth value y i The error between them, Ω(f) k ) is a regularization term used to prevent overfitting, as shown in formula (8). （8） Where T is the number of leaf nodes in the tree, ω j is the weight of leaf node j, and γ and λ are the regularization parameters that control the number of leaf nodes and the weight of leaf nodes, respectively.

8. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 7, characterized in that: In step S06, the weights are as follows: （9） Where N represents the number of samples, ω i The weights for sample i are obtained by manually setting the class weights in the code; l(y i ,f(x i )) is the loss function used to measure the true label y. i and predicted value f(x) i Error between ) ; Ω(f k ) is a regularization term for model complexity, used to prevent overfitting, f k This represents the output of the model in k rounds, where K is the total number of trees in the model.

9. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 8, characterized in that: The specific method for step S06 is as follows: Step S61: After determining the sample weights, use Bayesian optimization to optimize the hyperparameters of both the XGBoost and LightGBM models. （10） Where f(x) represents the objective function, m(x) represents the mean function, and k(x, x') represents the covariance function (or kernel function), which characterizes the correlation between input points x and x'. The higher the correlation, the larger the value of k. Step S62: For the hyperparameter tuning of XGB and LGBM, the radial basis function (RBF) kernel is selected as the kernel function of GP, as shown in the following formula: （11） Where, σ 2 The amplitude of the kernel function is represented by l, and the length scale is represented by l. The length scale determines the range of how changes in hyperparameters affect the objective function. Step S63: Obtain the predicted distribution of each hyperparameter combination through a GP, so that uncertainty can be quantified and a choice can be made, as described by formula (12): （12） Where μ(x) represents the prediction mean, reflecting the overall performance, and Σ(x, x) represents the prediction variance at x, describing the uncertainty in the prediction of the objective function f(x); Step S64: After the above processing, use AF to select the optimal hyperparameter combination, and use EI as AF. The formula for EI is as follows: （13） Among them, f best Let f(x) represent the current best solution, and f(x) be the predicted value of the objective function at a given point x. Step S65: After optimizing the parameters of the two models and obtaining their respective optimal solutions, a soft voting method is used to filter and merge the optimal solutions of the two models, as shown in the following formula: （14） （15） Where score(n) represents the weighted sum of the predicted probabilities for each class n from the classifier, and p 1,n and p 2,n Y represents the predicted probabilities of the two classifiers for class n, and Y is the final prediction result of the fusion model.

10. The multi-source partial discharge classification method in a fluorescent fiber optic gas-insulated switchgear as described in claim 9, characterized in that: The sample weights are as follows: air gap discharge mixed suspension discharge is 2.0, air gap discharge is 0.8, the four types of discharge mixed is 1.4, surface discharge mixed suspension discharge is 1.4, and the remaining categories are all 1.0.

Images