Multi-source partial discharge mode identification method and system

Abstract

The invention relates to the technical field of partial discharge, in particular to a multi-source partial discharge mode recognition method and system, and the method comprises the steps: firstly constructing a target discharge mode through a defect simulation module, then starting a voltage regulating table, gradually increasing the output voltage of a test transformer to a discharge critical value, and triggering partial discharge; an optical signal generated by discharging is captured by the fluorescent optical fiber in the corresponding area and is guided into the PMT through the transmission optical fiber; the PMT converts an optical signal into an electric signal under the power supply of the switching power supply, and transmits the electric signal to the oscilloscope in real time; the oscilloscope stores the electric signal data periodically and screens effective discharge signals through a peak detection algorithm to complete data cleaning; and inputting the training set into an optimized two-stage model of decision tree coarse classification and XGBoost fine classification, completing training, and realizing efficient and accurate identification of partial discharge types.

Description

Technical Field

[0001] This invention relates to the field of partial discharge technology, specifically to a method and system for multi-source partial discharge pattern recognition. Background Technology

[0002] Partial discharge is a key indicator of insulation degradation in high-voltage switchgear. Essentially, it's a phenomenon of concentrated electric field discharge caused by internal insulation defects. As critical equipment in the power transmission and distribution chain, high-voltage switchgear directly connects the grid to end users. If internal partial discharge is not detected and identified in a timely manner, it will lead to a continuous decline in insulation strength, ultimately causing equipment burnout, explosions, or even large-scale power outages, resulting in huge economic losses and disruptions to residential electricity supply. Therefore, accurate detection, signal acquisition, and pattern recognition of partial discharge in high-voltage switchgear are core technical requirements for ensuring the safe and stable operation of the distribution network.

[0003] Existing partial discharge detection and identification technologies can be categorized into three main types based on their principles: electrical signal detection technology, non-electrical signal detection technology, and pattern recognition technology. The development history and inherent bottlenecks of each type of technology are as follows: (1) Development and bottlenecks of electrical signal detection technology Electrical signal detection technology was the earliest technology applied to the diagnosis of partial discharge. Its core principle is to detect discharge by capturing signals such as electrical pulses and electromagnetic waves generated during the discharge process. Its development and limitations are detailed below: Pulse current method: This method captures the pulse current signal generated by the discharge by connecting a series detection impedance. It has the advantages of high quantitative analysis accuracy and high technical maturity, and is a commonly used method for partial discharge laboratory calibration. However, this method relies on the equipment grounding loop, is easily affected by background electromagnetic interference from the power grid, which leads to signal distortion, and cannot directly distinguish the discharge type, only achieving a qualitative judgment of "whether there is a discharge".

[0004] Transient ground voltage (TEV) method: This method uses the transient voltage wave induced on the metal casing of the switchgear by the electromagnetic waves generated by partial discharge for detection. The equipment is highly portable and suitable for rapid on-site screening. However, the TEV signal attenuates quickly and can only reflect the induced voltage on the surface of the switchgear, which cannot accurately quantify the discharge intensity. It is also susceptible to interference from the cabinet structure and external electromagnetic noise. The positioning accuracy can only cover the "cabinet area" and cannot pinpoint the specific defect location.

[0005] Ultra-high frequency (UHF) method: This method captures 300MHz~3GHz ultra-high frequency electromagnetic waves generated by partial discharge. It has better resistance to low-frequency electromagnetic interference than the pulse current method and TEV method, and can achieve discharge location through a multi-sensor array. However, this method suffers from significant signal attenuation in the complex conductive environment inside the switchgear, and has extremely high requirements for sensor placement. At the same time, UHF sensors are expensive and cannot effectively detect "weak discharges," and are prone to missed detections due to insufficient signal amplitude.

[0006] (2) Breakthroughs and limitations of non-electrical signal detection technology Non-electrical signal detection techniques avoid electromagnetic interference by capturing non-electrical physical quantities such as sound and light that accompany discharges. Among these, ultrasonic detection and optical detection methods are currently research hotspots, and their technical characteristics and bottlenecks are as follows: Ultrasonic testing: This method detects defects by receiving ultrasonic signals (20kHz~200kHz) generated by partial discharge, completely avoiding electromagnetic interference and allowing for location based on sound path difference. However, ultrasonic waves attenuate rapidly in air, making it impossible to cover the entire switch cabinet area. Furthermore, the cabinet structure easily leads to multi-path reflections of sound waves, creating false signals and resulting in location deviations of 10~20cm, making it impossible to accurately pinpoint minute defects.

[0007] Optical detection technology utilizes the ultraviolet-visible radiation accompanying partial discharge to achieve detection, with fluorescent fiber optic sensing technology being a core development direction. Fluorescent fibers possess advantages such as strong resistance to electromagnetic interference, good flexibility, and the ability to be embedded (e.g., surrounding insulators and busbars), enabling full coverage within switchgear. However, early optical detection technologies had significant limitations: ① Photomultiplier tubes were large and required external high-voltage power supplies, making distributed deployment impossible; ② Early research focused only on the signal characteristics of single discharge types, neglecting the detection of multi-source aliasing discharges; ③ Existing fluorescent fiber optic detection largely relied on a single correspondence between "light intensity and discharge intensity," failing to consider the periodicity and regional distribution characteristics of the signal, thus unable to provide comprehensive feature support for pattern recognition.

[0008] (3) Development and bottlenecks of partial discharge pattern recognition technology Partial discharge pattern recognition is the core component for achieving accurate discharge type determination. Its development can be divided into two stages: traditional manual processing and machine learning applications. The bottlenecks are as follows: Traditional manual processing methods rely on manually extracting statistical features from discharge signals and determining the discharge type through empirical thresholds. This method is highly subjective, with different operators having 30% to 50% differences in feature selection criteria. Furthermore, it cannot handle massive amounts of data, resulting in low recognition efficiency and an accuracy rate of less than 70%.

[0009] Machine learning identification methods: In recent years, algorithms such as neural networks, support vector machines, and decision trees have been applied to partial discharge identification. Among them, neural networks (such as CNN and LSTM) have strong nonlinear fitting capabilities, but they suffer from the black box problem, with opaque processes and poor interpretability, failing to provide maintenance personnel with a basis for "why a certain type of discharge is identified"; support vector machines are suitable for small sample data, but their generalization ability is insufficient in complex scenarios of multi-source mixed discharges; while decision tree algorithms have the advantages of process transparency and visualization, existing research has not optimized model parameters for "partial discharge signals collected by fluorescent optical fibers" and has not designed input features in conjunction with the characteristic differences of multi-source discharges, resulting in an identification rate of less than 80% for multi-source mixed discharges. Summary of the Invention

[0010] Purpose of the invention: In view of the above-mentioned problems, the present invention proposes a multi-source partial discharge pattern recognition method, which makes creative improvements to address the shortcomings of low recognition accuracy and high false judgment rate of multi-source mixed discharge. The present invention also provides a multi-source partial discharge pattern recognition system.

[0011] Summary of the Invention: This invention provides a method for multi-source partial discharge pattern recognition, the method comprising: Single-source discharge defects are constructed within the switchgear, with each type of discharge defect arranged in an independent defect area; dual-source discharge defects are constructed based on the single-source discharge defects, which are obtained by superimposing single-source discharge defects in pairs; simultaneously, a high-aliasing discharge defect is constructed by superimposing multiple single-source discharge defects. Discharge optical signals in each defect area are collected using an independent optical fiber path and converted into electrical signals. After preprocessing the collected raw electrical signal data, features that characterize the differences between different discharge types are extracted from the effective discharge signals to obtain the input feature vector. The training set feature vector obtained from the input feature vector is input into the optimized decision tree model. The decision tree structure is generated through iterative splitting, that is, the following three categories are generated: single-source discharge, dual-source discharge and high aliasing discharge. The optimized decision tree model sets different misclassification cost matrices for discharge types with different numbers of sources during the splitting process of constructing decision tree nodes, and selects the optimal splitting feature by calculating the cost-sensitive information gain. For single-source discharge, dual-source discharge, and high aliasing discharge, an ensemble learning model based on a tree model is used for accurate classification. The features representing the differences between different discharge types are used as the input of the model, and the output is the corresponding single-source discharge defect or the superposition of single-source discharge defects. For the superposition of high-risk single-source discharge defects, a misclassification cost matrix is ​​added during accurate classification.

[0012] Furthermore, including: The construction of single-source discharge defects within the switchgear, with each type of discharge defect arranged within an independent defect area, includes: Four independent defect areas were delineated within the 10kV high-voltage switchgear, each corresponding to one of the four typical partial discharge types, and each area was physically isolated. Defect construction methods: Single-source discharge defects, including: Area A corresponding to surface discharge: copper foil is attached to the surface of the insulator near the high-voltage end to simulate the concentrated discharge of the surface electric field caused by surface contamination of the insulator; Area B corresponding to air gap discharge: the fixing screws between the high-voltage end and the insulator are removed to leave an air gap to simulate the air gap breakdown discharge caused by loose connection; Area C corresponding to needle tip discharge: a thin copper needle is inserted between the high-voltage end and the switch cabinet wall to simulate the concentrated discharge of the tip electric field caused by metal burrs; Area D corresponding to suspension discharge: several small copper pieces are attached to cotton thread and suspended between the high-voltage end and the grounding end to simulate the gap discharge caused by the suspension of charged metal particles; Dual-source discharge defects include: constructing a dual-source discharge mode by combining the activation of two regions in regions A, B, C, and D; High aliasing discharge defects include: constructing a high aliasing discharge mode by combining activation of four regions in regions A, B, C, and D.

[0013] Furthermore, including: The method of acquiring discharge optical signals in each defect area using independent optical fiber paths includes: Fluorescent optical fibers are wrapped around the four defect regions A, B, C, and D, with the fiber length for each region set to ensure coverage of the light radiation range generated by the defects; transmission optical fibers are connected to both ends of the fluorescent optical fibers to form a signal transmission path.

[0014] Furthermore, including: In the process of splitting decision tree nodes, different misclassification cost matrices are set for discharge types with different source numbers, and the optimal splitting feature is selected by calculating cost-sensitive information gain, including: Let the set of categories be: ; The cost of misclassification is constructed, which is defined by the cost matrix and expressed as: , ;in, Indicates the true category is However, it was incorrectly classified as The cost of time; Therefore, for a certain node sample set D Its cost-sensitive entropy is defined as: ; in, For the sample set belonging to category proportion, For the sample set belonging to category The proportion; The feature A that characterizes the differences between different discharge types is used to divide the threshold. t After splitting the sample set, a subset is obtained. , Its cost-sensitive posterior entropy is: ; Therefore, the cost-sensitive information gain is expressed as: .

[0015] Furthermore, including: In the process of splitting decision tree nodes, different misclassification cost matrices are set for discharge types with different numbers of sources, and the optimal splitting feature is selected by calculating cost-sensitive information gain. This also includes: The features that characterize the differences between different discharge types A and its splitting threshold t In the middle, select to make The largest feature and the threshold are used as the optimal splitting condition for the current node, that is: .

[0016] On the other hand, the present invention also provides a multi-source partial discharge pattern recognition system, the system comprising: The discharge defect construction module is used to construct single-source discharge defects in the switch cabinet, and each discharge defect is arranged in an independent defect area; dual-source discharge defects are constructed on the basis of single-source discharge defects, which are obtained by superimposing single-source discharge defects in pairs; at the same time, a high-aliasing discharge defect is constructed by superimposing multiple single-source discharge defects. The feature extraction module is used to collect discharge optical signals in each defect area using independent optical fiber paths and convert the collected optical signals into electrical signals. After preprocessing the collected raw electrical signal data, features that can characterize the differences between different discharge types are extracted from the effective discharge signals to obtain the input feature vector. The coarse classification module is used to input the training set feature vector obtained from the input feature vector into the optimized decision tree model, and generate the decision tree structure through iterative splitting, that is, to generate the following three classifications: single-source discharge, dual-source discharge and high aliasing discharge. The optimized decision tree model sets different misclassification cost matrices for discharge types with different numbers of sources during the splitting process of constructing decision tree nodes, and selects the optimal splitting feature by calculating the cost-sensitive information gain. The fine classification module is used to accurately classify single-source discharge, dual-source discharge, and high-aliasing discharge using an ensemble learning model based on a tree model. The model takes the features that characterize the differences between different discharge types as input and outputs the corresponding single-source discharge defect or the superposition of single-source discharge defects. For the superposition of high-risk single-source discharge defects, a misclassification cost matrix is ​​added during the accurate classification.

[0017] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the multi-source partial discharge pattern recognition method described above.

[0018] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the multi-source partial discharge pattern recognition method described above.

[0019] Beneficial effects: Compared with the prior art, the present invention has the following advantages: Breaking through the electromagnetic interference bottleneck of traditional electrical signal detection, this method uses fluorescent optical fiber to collect partial discharge optical signals in a distributed manner, combined with multi-dimensional feature extraction, to accurately capture weak discharge signals, thereby improving anti-interference capability and signal capture rate. By eliminating reliance on manual labor, an integrated automated process of "signal acquisition - feature extraction - model recognition" is constructed. The process transparency of the decision tree algorithm is utilized to achieve automated identification of discharge types, thereby improving identification efficiency and result repeatability. To further improve the accuracy and stability of partial discharge pattern recognition in multi-source aliasing scenarios, this invention introduces a discharge pattern recognition algorithm based on a two-stage cascade approach, building upon existing feature extraction and decision tree recognition methods. This algorithm fully leverages the advantages of decision tree models—clear structure, high computational efficiency, and strong interpretability—to perform a preliminary coarse classification of partial discharge samples, categorizing them into different types based on complexity, such as single-source discharge, multi-source aliasing discharge, and high-aliasing discharge. Based on this, a dedicated fine classification model, XGBoost, is constructed for each coarse classification result, enabling differentiated recognition of various discharge patterns. This "coarse classification followed by fine classification" algorithm structure effectively reduces feature overlap and classification boundary ambiguity caused by multi-source aliasing discharge in a unified model. This ensures that the recognition process maintains interpretability while maintaining accuracy in complex scenarios, significantly improving the overall performance of partial discharge pattern recognition without increasing hardware complexity, ultimately achieving a recognition accuracy of approximately 88%.

[0020] Finally, to enhance the interpretability and operational adaptability of the model, feature importance analysis and confusion matrix output are used to make the identification logic traceable, providing maintenance personnel with the basis for "why it is judged to be a certain type of discharge". At the same time, it adapts to the distributed detection needs of switchgear, provides technical support for early diagnosis of equipment faults and operation and maintenance decisions, and ultimately ensures the safe and stable operation of the distribution network. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the multi-source partial discharge mode recognition method according to an embodiment of the present invention; Figure 2 This is a flowchart of the anchor frame marking method according to an embodiment of the present invention; Figure 3 This is a flowchart of the key region extraction method described in an embodiment of the present invention; Figure 4 This is a flowchart of the non-critical region extraction method described in an embodiment of the present invention; Figure 5 This is a flowchart of the iterative reconstruction process method described in an embodiment of the present invention; Figure 6 This is a schematic diagram of the recognition model structure described in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating image classification as described in an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1: Refer to Figure 1 An embodiment of the present invention provides a method for multi-source partial discharge pattern recognition, the method comprising the following steps: Step 1: Construct single-source discharge defects within the switchgear, with each type of discharge defect arranged within an independent defect area; construct dual-source discharge defects based on the single-source discharge defects, which are obtained by superimposing single-source discharge defects in pairs; simultaneously, construct a highly aliased discharge defect by superimposing multiple single-source discharge defects.

[0025] The construction of single-source discharge defects within the switchgear, with each type of discharge defect arranged within an independent defect area, includes: Four independent defect areas were delineated within the 10kV high-voltage switchgear, each corresponding to one of the four typical partial discharge types, and each area was physically isolated. Defect construction methods: Single-source discharge defects, including: Area A corresponding to surface discharge: copper foil is attached to the surface of the insulator near the high-voltage end to simulate the concentrated discharge of the surface electric field caused by surface contamination of the insulator; Area B corresponding to air gap discharge: the fixing screws between the high-voltage end and the insulator are removed to leave an air gap to simulate the air gap breakdown discharge caused by loose connection; Area C corresponding to needle tip discharge: a thin copper needle is inserted between the high-voltage end and the switch cabinet wall to simulate the concentrated discharge of the tip electric field caused by metal burrs; Area D corresponding to suspension discharge: several small copper pieces are attached to cotton thread and suspended between the high-voltage end and the grounding end to simulate the gap discharge caused by the suspension of charged metal particles; Dual-source discharge defects include: constructing a dual-source discharge mode by combining the activation of two regions in regions A, B, C, and D; High aliasing discharge defects include: constructing a high aliasing discharge mode by combining activation of four regions in regions A, B, C, and D.

[0026] In a preferred embodiment of this invention, such as Figure 2 As shown, firstly, a typical partial discharge defect in the switchgear is constructed to simulate single-source and multi-source mixed discharge scenarios that may occur in actual operation. The specific structural design is as follows: Defect location: Within the 10kV high-voltage switchgear provided by Nanjing Switchgear Co., Ltd., four independent defect areas (marked as A, B, C, and D) are designated, each corresponding to one of the four typical partial discharge types. Each area is physically isolated to avoid cross-interference of discharge signals. Defect construction method: Surface discharge (Area A): Copper foil is attached to the surface of the insulator near the high-voltage end to simulate the concentrated discharge of the surface electric field caused by contamination on the insulator surface; Air gap discharge (B area): Remove the fixing screws between the high voltage end and the insulator to leave an air gap and simulate air gap breakdown discharge caused by loose connection; Needle tip discharge (C area): A thin copper needle is inserted between the high voltage end and the switch cabinet wall to simulate the concentrated discharge of the electric field at the tip caused by metal burrs. Suspension discharge (D region): Several small copper sheets are attached to cotton thread and suspended between the high voltage end and the grounding end to simulate the gap discharge caused by the suspension of charged metal particles. Multi-source cascading is achieved by combining and activating defects in regions A, B, C, and D (e.g., activating region A alone results in single-source surface discharge, activating regions B and D simultaneously results in dual-source air gap + suspension discharge, and activating regions A, B, C, and D results in four-source cascading discharge). Finally, eight typical discharge modes are constructed: four single-source, three dual-source, and one four-source, covering the main discharge scenarios in actual operation and maintenance.

[0027] This embodiment also includes providing a stable high voltage to generate partial discharge, ensuring the repeatability and consistency of the discharge signal. The composition and connection relationship are as follows: Core components: Test transformer (YDW-10KVA / 50KV), voltage regulating platform (HZTC-104D), protective resistor, isolation transformer (HZLB-10KVA); The voltage regulating platform is connected to 220V / 50Hz mains power, and controls the input of the test transformer by adjusting the output voltage; The test transformer outputs high voltage (up to 50KV), which, after being current-limited by the series protective resistor, is connected to the C-phase bus of the switchgear to provide the electric field required for discharge in the defect area; The isolation transformer is connected in parallel to the power supply circuit to achieve electrical isolation, avoid interference from high voltage to subsequent signal conversion modules, and ensure the safe operation of the system.

[0028] Step 2: Collect the discharge optical signals in each defect area using an independent optical fiber path, and convert the collected optical signals into electrical signals; after preprocessing the collected raw electrical signal data, extract features that can characterize the differences between different discharge types from the effective discharge signals, thereby obtaining the input feature vector.

[0029] In this embodiment, this step is responsible for the distributed capture of partial discharge optical signals in each defect region, and the structural design is as follows: Fiber optic selection: Fluorescent fiber with high responsivity in the ultraviolet-visible band is adopted, which takes into account the characteristics of low signal transmission loss and good flexibility, and is suitable for complex spatial layout in switch cabinets. Arrangement: Fluorescent optical fibers are wrapped around the four defect areas A, B, C, and D, with each area having a fiber length of ≥2m to ensure coverage of the light radiation range generated by the defects; both ends of the fluorescent optical fibers are connected to transmission optical fibers to form a signal path of "fluorescent optical fiber acquisition → transmission optical fiber light guiding". Distributed advantages: By collecting discharge optical signals from each region through independent optical fiber paths, it is possible to distinguish between spatially displaced multi-source aliasing discharges. For example, the optical signals of air gap discharge in region B and suspension discharge in region D are collected by corresponding optical fibers, avoiding identification interference caused by signal aliasing.

[0030] In addition to this, the method is also responsible for converting the optical signals collected by the fluorescent optical fiber into analyzable electrical signals and completing data storage, as follows: Core components: photomultiplier tube (PMT), switching power supply (SMPS), oscilloscope (RIGOL DS1102E); Conversion and storage logic: The switching power supply outputs +4V and ±12V DC voltages to provide a stable operating power supply for the photomultiplier tube; the transmission optical fiber guides the optical signal into the photomultiplier tube, and the PMT amplifies the weak optical signal and converts it into an electrical signal; the oscilloscope is connected to the electrical signal output terminal of the PMT through a signal cable, the sampling frequency is set to ≥1MHz, and the discharge signal data is collected and stored. Each discharge mode collects 500 cycles, and finally forms a raw dataset of 4000 cycles.

[0031] In a preferred embodiment, after preprocessing the acquired raw electrical signal data, features that characterize the differences between different discharge types are extracted from the effective discharge signals to obtain an input feature vector. Effective discharge signals are then filtered from the raw data to remove noise interference. The specific process can be implemented as follows: Data reading: The raw voltage-time data exported from the oscilloscope is read according to the "discharge mode-cycle" classification to form a structured dataset; Effective discharge screening: A peak detection algorithm is used, with three screening criteria set: If the data point value zi > the previous data point value zi-1 and zi > the next data point value zi+1, then zi is a local peak value; zi > the effective discharge threshold SA, which is experimentally calibrated to 0.05V to avoid misjudging low-amplitude noise. The interval between adjacent effective peak values ​​is ≥75 data points to avoid the same discharge pulse being counted repeatedly; Dataset partitioning: The preprocessed 4000 cycles of data were divided into a training set (3200 cycles) and a test set (800 cycles) in an 8:2 ratio for model training and performance evaluation. The original discharge signals for the eight categories are as follows: Figure 3 As shown. Figure 3 (a) in the diagram is a schematic diagram showing the relationship between the amplitude and time corresponding to the air gap discharge. Figure 3 (b) in the diagram shows the relationship between the amplitude and time corresponding to surface discharge. Figure 3 (c) in the diagram is a schematic diagram showing the relationship between the amplitude and time corresponding to the levitation discharge. Figure 3 (d) in the diagram is a schematic diagram showing the relationship between the amplitude and time corresponding to the needle tip discharge. Figure 3 (e) in the diagram is a schematic diagram showing the relationship between amplitude and time for air gap discharge and hybrid suspension discharge. Figure 3(f) in the diagram illustrates the relationship between amplitude and time for mixed air gap discharge and surface discharge. Figure 3 (g) in the diagram represents the relationship between amplitude and time for air gap discharge mixed with tip discharge. Figure 3 (h) in the figure is a schematic diagram showing the relationship between amplitude and time for the mixed discharge of four typical partial discharges.

[0032] In this embodiment, the physical meaning and statistical characteristics of different features are fully considered during feature selection and model structure design, and a reasonable division of labor is carried out based on the fit between features and classification models. Features with clear physical meaning, obvious threshold characteristics, and strong monotonicity, such as the number of effective discharges per cycle, average effective discharge intensity, and number of regional discharges, are preferentially used in the first-stage decision tree model to fully leverage the advantages of decision trees in rule division and interpretability, achieving rapid coarse classification of partial discharge samples. This module extracts core features that characterize the differences between different discharge types from the effective discharge signal, constructing the model input vector. Specific features and extraction methods are as follows: Feature 1: Number of effective discharges per cycle (Neff) The number of effective discharge peaks that meet the screening criteria in each cycle is counted to reflect the density of discharge. For example, surface discharge has a higher Neff, while floating discharge has a lower Neff. Feature 2: Average effective discharge per cycle (AVerageSeff) Calculate the sum of the amplitudes of all effective discharge peaks in each cycle (ΣSeff), and then divide it by the Neff of that cycle. The formula is: AVerageSeff = ΣSeff / Neff, which reflects the overall intensity of the discharge. For example, for surface discharge, AVerageSeff ≥ 0.4V, and for air gap discharge, AVerageSeff ≤ 0.3V. Feature 3: Number of effective discharges in the region (Neff(t)) Each cycle (20ms) is evenly divided into 15 time regions (each region is about 1.33ms). The number of effective discharges in each region is counted to form a 15-dimensional region feature vector, which reflects the time distribution characteristics of the discharge. For example, the needle tip discharge is relatively evenly distributed in each region, while the air gap discharge is irregularly distributed. Since time-domain features alone cannot accurately identify some easily confused discharges, this module performs additional time-frequency conversion on the signal. For features with high nonlinearity, complex coupling relationships between features, and difficulty in being distinguished by a single threshold, such as wavelet packet energy distribution and FFT main frequency amplitude, these features are introduced into the second-stage XGBoost-based fine classification model. By iteratively superimposing multiple weak classifiers, higher-order combination relationships between features are learned, thereby effectively improving the recognition accuracy in multi-source mixed discharge scenarios.

[0033] Feature 4: Wavelet packet energy distribution It represents the energy distribution characteristics of the signal in different frequency bands and is used to characterize the detailed differences in the frequency components of the discharge signal over time.

[0034] Feature 5: FFT main frequency amplitude It represents the main frequency of the signal in the frequency domain and its energy proportion, and is used to reflect the dominant frequency characteristics and spectral structure differences of the discharge signal.

[0035] Meanwhile, the correspondence between the above-mentioned features and the model is not a simple superposition, but a targeted design based on the mechanism of partial discharge signals and the characteristics of the model, which reflects the professionalism and innovation of this invention in terms of algorithm structure and feature engineering.

[0036] Feature vector construction: The feature vector for each period is [Neff, AVerageSeff, Neff(t)1, Neff(t)2, ..., Neff(t)15, wavelet packet energy distribution, dominant frequency amplitude], totaling 19 dimensions, ensuring comprehensive characterization of discharge properties. Feature importance analysis is as follows: Figure 4 As shown.

[0037] Step 3: Input the training set feature vector obtained from the input feature vector into the optimized decision tree model, and generate the decision tree structure through iterative splitting, that is, generate the following three categories: single-source discharge, dual-source discharge and high aliasing discharge. The optimized decision tree model sets different misclassification cost matrices for discharge types with different numbers of sources during the splitting process of constructing decision tree nodes, and selects the optimal splitting feature by calculating the cost-sensitive information gain.

[0038] In this embodiment, during the splitting process of constructing decision tree nodes, different misclassification cost matrices are set for discharge types with different source numbers. The optimal splitting feature is selected by calculating the cost-sensitive information gain, including: Let the set of categories be: ; The cost of misclassification is constructed, which is defined by the cost matrix and expressed as: , ;in, Indicates the true category is However, it was incorrectly classified as The cost of time; Therefore, for a certain node sample set D Its cost-sensitive entropy is defined as: ; in, For the sample set belonging to category proportion, For the sample set belonging to category The proportion; The feature A that characterizes the differences between different discharge types is used to divide the threshold. t After splitting the sample set, a subset is obtained. , Its cost-sensitive posterior entropy is: ; Therefore, the cost-sensitive information gain is expressed as: .

[0039] Furthermore, including: In the process of splitting decision tree nodes, different misclassification cost matrices are set for discharge types with different numbers of sources, and the optimal splitting feature is selected by calculating cost-sensitive information gain. This also includes: The features that characterize the differences between different discharge types A and its splitting threshold t In the middle, select to make The largest feature and the threshold are used as the optimal splitting condition for the current node, that is: .

[0040] In a preferred embodiment of this invention, the basic decision tree first includes: The CART (Classification and Regression Tree) algorithm is used. The root node is the entire training set, the internal nodes are the feature judgment conditions, such as "AVerageSeff>0.35V", and the leaf nodes are the discharge type labels, such as "single-source discharge" or "dual-source discharge". Feature splitting criterion: "Entropy" is selected as the splitting criterion. The optimal splitting feature is chosen by calculating the information gain (the difference in entropy between the dataset before and after splitting). The improved information gain formula is as follows: Information gain = Pre-split entropy - Σ (Subset sample proportion × Subset entropy) The entropy is calculated as follows: Entropy = -Σ(pk × log2pk) (pk is the proportion of a certain discharge type in the subset).

[0041] This embodiment improves information gain by introducing a cost-sensitive entropy design into the decision tree classification model, taking into account the differences in importance and misclassification risk of different partial discharge types in engineering scenarios. Specifically, during the splitting process of constructing decision tree nodes, different misclassification cost matrices are set for discharge types with different numbers of sources. Discharge types with higher misclassification probability or greater impact on equipment safety (such as dual-source discharge) are given higher classification cost weights. This allows the model to consider both "feature separability" and "misclassification risk" when calculating information gain or Gini index, thereby prompting the decision tree to prioritize more refined node division in key categories.

[0042] Let the set of categories be: ; The cost of misclassification is constructed, which is defined by the cost matrix and expressed as: , ;in, Indicates the true category is However, it was incorrectly classified as The cost of time; Therefore, for a certain node sample set D Its cost-sensitive entropy is defined as: ; in, For the sample set belonging to category proportion, For the sample set belonging to category The proportion; The feature A that characterizes the differences between different discharge types is used to divide the threshold. t After splitting the sample set, a subset is obtained. , Its cost-sensitive posterior entropy is: ; Therefore, the cost-sensitive information gain is expressed as: .

[0043] The features that characterize the differences between different discharge types A and its splitting threshold t In the middle, select to make The largest feature and the threshold are used as the optimal splitting condition for the current node, that is: .

[0044] Among them, the cost-sensitive design essentially represents the model's measurement of the "risk sensitivity difference" of different discharge signals, so that the identification process not only classifies based on feature differences, but also takes into account the importance of discharge to the safety of equipment operation.

[0045] Hyperparameter optimization: The Bayesian optimization algorithm is used to traverse the value space of key hyperparameters of the decision tree and finally determine the optimal combination of hyperparameters to avoid overfitting or underfitting of the model. Model Training and Evaluation: The feature vectors from the training set are input into the optimized decision tree model, and the decision tree structure is generated through iterative splitting. The model performance is evaluated using a test set, and the accuracy, recall, and overall recognition rate for each discharge type are calculated to verify the model's effectiveness. The decision tree diagram is shown below. Figure 5 As shown, it is ultimately divided into three types: "single-source discharge", "dual-source discharge", and "high aliasing discharge".

[0046] Step 4: For single-source discharge, dual-source discharge, and high aliasing discharge, an ensemble learning model based on a tree model is used for accurate classification. The features representing the differences between different discharge types are used as the input of the model, and the corresponding single-source discharge defect or the superposition of single-source discharge defects is output. For the superposition of high-risk single-source discharge defects, a misclassification cost matrix is ​​added during accurate classification.

[0047] In this embodiment, after completing the coarse classification of the first-stage decision tree, the present invention further constructs multiple sets of fine classifiers for different coarse classification results to achieve higher resolution partial discharge type identification. This step aims to perform deep discrimination on dual-source aliasing signals and highly aliased signals, which are more difficult to classify, thereby significantly improving the overall identification accuracy.

[0048] The second-stage fine classifiers are divided into three categories based on the coarse classification categories: single-source discharge fine classifier; dual-source aliasing fine classifier; and high-aliasing fine classifier. Each type of fine classifier is trained using XGBoost, and is constructed by combining strategies such as feature selection, sample resampling, and misclassification cost allocation.

[0049] The second-stage fine classifier uses a 19-dimensional feature vector generated by the feature extraction module as input, including: Time-domain statistical characteristics: number of effective discharges per period (Neff), average amplitude of effective discharges per period (AverageSeff); regional temporal distribution characteristics (15 dimensions): Neff(t1) to Neff(t15); frequency domain and time-frequency domain characteristics: wavelet packet energy distribution, FFT dominant frequency amplitude. A single-source discharge fine classifier is used to accurately distinguish samples coarsely classified as "Single" in the first stage, outputting the following four types of single-source discharges: tip discharge, surface discharge, air gap discharge, and surface discharge. The model's hyperparameters are: n_estimators = 260, max_depth = 6, learning_rate = 0.05, subsample = 0.8, colsample_bytree = 0.6, min_child_weight = 2, and scale_pos_weight = 1, using 5-fold cross-validation.

[0050] The single-source discharge classifier achieves an accuracy rate of over 95% in single-source discharge scenarios. A dual-source aliasing classifier is used to identify the following dual-source aliasing types: air gap discharge + levitation discharge, air gap discharge + surface discharge, and air gap discharge + tip discharge. The model's hyperparameters are: n_estimators = 250, max_depth = 6, eta = 0.05, gamma = 0.3, lambda = 2, and scale_pos_weight = 2.

[0051] Because "air gap + suspension" is considered dangerous, this invention incorporates a misclassification cost matrix, assigning it a higher cost weight to improve the recognition priority of high-risk types. The values ​​for air gap discharge + suspension discharge, air gap discharge + surface discharge, and air gap discharge + needle tip discharge are 5, 3, and 3, respectively. The weights for this classification can be adjusted using `scale_pos_weight` or `sample_weight`.

[0052] A high-aliasing fine classifier is used to identify complex aliasing, specifically four-source aliasing discharges (A+B+C+D). The model's hyperparameters are: n_estimators=300, max_depth=8, learning_rate=0.03, subsample=0.6, colsample_bytree=0.6, and scale_pos_weight=4. Given the high overlap of high-aliasing information, this invention uses SHAP feature importance ranking to filter feature dimensions, reducing them from 19 to 12, thus improving the generalization performance of the fine classifier while reducing the risk of overfitting.

[0053] Finally, the recognition results are visualized to intuitively demonstrate the model's recognition effectiveness and decision-making logic, assisting maintenance personnel in understanding and verification. The output includes: Confusion matrix: Displays the correspondence between the true type and the predicted type of the test set samples in the form of a heat map, intuitively reflecting the misjudgment of various types of discharge (such as the number of cross-misjudgments between air gap discharge and suspension discharge). t-SNE scatter plot: The 19-dimensional features are reduced to 2-dimensionality using the t-SNE algorithm, and clustering diagrams of samples with different discharge types are plotted to verify the distinguishability of the features (e.g., surface discharge samples are clustered compactly, while air gap + suspension discharge samples partially overlap). The confusion matrix is ​​as follows: Figure 6 As shown, Figure 6 In the matrix (a), the single-source confusion matrix is ​​shown. Figure 6 (b) in the matrix represents the multi-source confusion matrix.

[0054] Therefore, this invention, based on the synergy of "the specificity of partial discharge optical signals" and "the feature classification ability of decision trees," is specifically divided into two parts: (1) Principle of partial discharge optical signal acquisition During partial discharge, the breakdown of the insulating medium releases optical signals, and the intensity and pulse frequency of these signals are strongly correlated with the type of discharge. This invention uses fluorescent optical fibers (anti-electromagnetic interference) to collect optical signals from various defect areas in a distributed manner. These signals are then converted into electrical signals by photomultiplier tubes (PMTs) and stored in an oscilloscope. This method avoids the electromagnetic interference of traditional electrical detection and accurately captures weak discharge signals. (2) Principle of Multi-Source Partial Discharge Pattern Recognition in a Two-Stage Model of Decision Tree Coarse Classification and XGBoost Fine Classification The decision tree uses a "divide and conquer" approach for identification: the root node is initialized with the training set and its entropy (reflecting sample disorder) is calculated; feature splitting selects the feature with the highest information gain and the split point, splitting the root node into child nodes; iterative splitting of child nodes continues until the sample category is singular or the gain is insufficient, generating leaf nodes (discharge types); during prediction, the node output types are traversed, and single-source / multi-source aliasing discharges are distinguished based on multi-dimensional features (intensity, frequency, distribution). After completing the coarse classification of the first-stage decision tree, this invention further introduces a second-stage fine classifier to address the problem of high overlap and blurred boundaries in the feature space of multi-source aliasing discharges, performing fine classification on the coarse types.

[0055] On the other hand, the present invention also provides a multi-source partial discharge pattern recognition system, the system comprising: The discharge defect construction module is used to construct single-source discharge defects in the switch cabinet, and each discharge defect is arranged in an independent defect area; dual-source discharge defects are constructed on the basis of single-source discharge defects, which are obtained by superimposing single-source discharge defects in pairs; at the same time, a high-aliasing discharge defect is constructed by superimposing multiple single-source discharge defects. The feature extraction module is used to collect discharge optical signals in each defect area using independent optical fiber paths and convert the collected optical signals into electrical signals. After preprocessing the collected raw electrical signal data, features that can characterize the differences between different discharge types are extracted from the effective discharge signals to obtain the input feature vector. The coarse classification module is used to input the training set feature vector obtained from the input feature vector into the optimized decision tree model, and generate the decision tree structure through iterative splitting, that is, to generate the following three classifications: single-source discharge, dual-source discharge and high aliasing discharge. The optimized decision tree model sets different misclassification cost matrices for discharge types with different numbers of sources during the splitting process of constructing decision tree nodes, and selects the optimal splitting feature by calculating the cost-sensitive information gain. The fine classification module is used to accurately classify single-source discharge, dual-source discharge, and high-aliasing discharge using an ensemble learning model based on a tree model. The model takes the features that characterize the differences between different discharge types as input and outputs the corresponding single-source discharge defect or the superposition of single-source discharge defects. For the superposition of high-risk single-source discharge defects, a misclassification cost matrix is ​​added during the accurate classification.

[0056] Other technical features of the multi-source partial discharge pattern recognition system described in this embodiment are similar to those of the multi-source partial discharge pattern recognition method, and will not be repeated here.

[0057] Thirdly, the present invention also provides a multi-source partial discharge mode recognition device, such as... Figure 7 As shown, the device includes: a power supply unit (PMT), a high-voltage switchgear, coupling capacitors, a test transformer, an isolation transformer, a voltage regulating platform, an oscilloscope, a switching power supply, and a control module. These components are interconnected to form an identification device. The control modules within this device operate in a coordinated manner according to a time sequence, with the specific operational relationships as follows: Defect activation and high voltage application: First, construct the target discharge mode through the defect simulation module, then start the voltage regulating table and gradually increase the output voltage of the test transformer to the discharge critical value to trigger partial discharge; Optical signal acquisition and conversion: The optical signal generated by the discharge is captured by the fluorescent optical fiber in the corresponding area and introduced into the PMT through the transmission optical fiber; the PMT converts the optical signal into an electrical signal under the power supply of the switching power supply and transmits it to the oscilloscope in real time. Data storage and preprocessing: The oscilloscope stores electrical signal data in "cycles". After 500 cycles are collected, the software system reads the data and uses a peak detection algorithm to filter out valid discharge signals, thus completing the data cleaning. Feature extraction and model input: The software system extracts 19-dimensional features such as Neff, AvalageSeff, Neff(t), and wavelet packet energy distribution from the preprocessed data to form feature vectors, and divides them into training and test sets in an 8:2 ratio; Model training and recognition: The training set is input into the optimized decision tree coarse classification and XGBoost fine classification two-stage model to complete the training; the test set is input into the model to output the predicted discharge type of each sample; Results Output and Verification: The software system generates a confusion matrix and a t-SNE scatter plot, calculates the overall recognition rate, and completes a discharge pattern recognition process.

[0058] If a new discharge signal needs to be identified, steps 1-6 only need to be repeated. There is no need to adjust the hardware structure or software algorithm parameters, which has good versatility and repeatability.

[0059] This device includes: a multi-source discharge defect simulation device containing four types of isolation units: surface, air gap, tip, and suspension, which can be activated independently or in combination. A fluorescent fiber distributed acquisition system: fluorescent fibers are arranged according to the defects, combined with a PMT (photoelectric converter) and an oscilloscope (for storage) to achieve anti-interference optical signal acquisition. A multi-dimensional feature extraction method: effective signals are filtered by peak detection, and three types of features in the time domain ("frequency-intensity-regional distribution") and frequency domain features ("wavelet packet energy distribution, FFT main frequency amplitude") are extracted to construct a 19-dimensional vector. A second-order multi-source partial discharge pattern recognition method integrating decision tree and XGBoost: Bayesian optimization of decision tree parameters, 8:2 dataset partitioning, and training to identify three coarse-grained discharge types. Subsequently, XGBoost further classifies eight fine-grained types based on the output coarse-grained types, achieving an overall recognition rate of ≥88%. An integrated "acquisition-recognition" system: integrating the above devices and methods to achieve a closed loop of "defect activation → acquisition → feature extraction → recognition → visualization", automatically identifying multi-source discharges with traceable logic.

[0060] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. "A plurality of" means two or more, unless otherwise explicitly specified.

[0061] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0062] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0063] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0064] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0065] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0066] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0067] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0068] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0069] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for recognizing multi-source partial discharge patterns, characterized in that, The method includes: Single-source discharge defects are constructed within the switchgear, with each type of discharge defect arranged in an independent defect area; dual-source discharge defects are constructed based on the single-source discharge defects, which are obtained by superimposing single-source discharge defects in pairs; simultaneously, a high-aliasing discharge defect is constructed by superimposing multiple single-source discharge defects. Discharge optical signals in each defect area are collected using an independent optical fiber path and converted into electrical signals. After preprocessing the collected raw electrical signal data, features that characterize the differences between different discharge types are extracted from the effective discharge signals to obtain the input feature vector. The training set feature vector obtained from the input feature vector is input into the optimized decision tree model. The decision tree structure is generated through iterative splitting, that is, the following three categories are generated: single-source discharge, dual-source discharge and high aliasing discharge. The optimized decision tree model sets different misclassification cost matrices for discharge types with different numbers of sources during the splitting process of constructing decision tree nodes, and selects the optimal splitting feature by calculating the cost-sensitive information gain. For single-source discharge, dual-source discharge, and high aliasing discharge, an ensemble learning model based on a tree model is used for accurate classification. The features representing the differences between different discharge types are used as the input of the model, and the output is the corresponding single-source discharge defect or the superposition of single-source discharge defects. For the superposition of high-risk single-source discharge defects, a misclassification cost matrix is ​​added during accurate classification.

2. The multi-source partial discharge pattern recognition method according to claim 1, characterized in that, The construction of single-source discharge defects within the switchgear, with each type of discharge defect arranged within an independent defect area, includes: Four independent defect areas were delineated within the 10kV high-voltage switchgear, each corresponding to one of the four typical partial discharge types, and each area was physically isolated. Defect construction methods: Single-source discharge defects, including: Area A corresponding to surface discharge: copper foil is attached to the surface of the insulator near the high-voltage end to simulate the concentrated discharge of the surface electric field caused by surface contamination of the insulator; Area B corresponding to air gap discharge: the fixing screws between the high-voltage end and the insulator are removed to leave an air gap to simulate the air gap breakdown discharge caused by loose connection; Area C corresponding to needle tip discharge: a thin copper needle is inserted between the high-voltage end and the switch cabinet wall to simulate the concentrated discharge of the tip electric field caused by metal burrs; Area D corresponding to suspension discharge: several small copper pieces are attached to cotton thread and suspended between the high-voltage end and the grounding end to simulate the gap discharge caused by the suspension of charged metal particles; Dual-source discharge defects include: constructing a dual-source discharge mode by combining the activation of two regions in regions A, B, C, and D; High aliasing discharge defects include: constructing a high aliasing discharge mode by combining activation of four regions in regions A, B, C, and D.

3. The multi-source partial discharge pattern recognition method according to claim 2, characterized in that, The method of acquiring discharge optical signals in each defect area using independent optical fiber paths includes: Fluorescent optical fibers are wrapped around the four defect regions A, B, C, and D, with the fiber length for each region set to ensure coverage of the light radiation range generated by the defects; transmission optical fibers are connected to both ends of the fluorescent optical fibers to form a signal transmission path.

4. The multi-source partial discharge pattern recognition method according to claim 1, characterized in that, In the process of splitting decision tree nodes, different misclassification cost matrices are set for discharge types with different source numbers, and the optimal splitting feature is selected by calculating cost-sensitive information gain, including: Let the set of categories be: The misclassification cost is constructed, which is defined by the cost matrix and expressed as: , ;in, Indicates the true category is However, it was incorrectly classified as The cost of time; Therefore, for a certain node sample set D Its cost-sensitive entropy is defined as: ; in, For the sample set belonging to category proportion, For the sample set belonging to category The proportion; The feature A that characterizes the differences between different discharge types is used to divide the threshold. t After splitting the sample set, a subset is obtained. , Its cost-sensitive posterior entropy is: ; Therefore, the cost-sensitive information gain is expressed as: .

5. The multi-source partial discharge pattern recognition method according to claim 4, characterized in that, In the process of splitting decision tree nodes, different misclassification cost matrices are set for discharge types with different numbers of sources, and the optimal splitting feature is selected by calculating cost-sensitive information gain. This also includes: The features that characterize the differences between different discharge types A and its splitting threshold t In the middle, select to make The largest feature and the threshold are used as the optimal splitting condition for the current node, that is: .

6. A multi-source partial discharge pattern recognition system, characterized in that, The system includes: The discharge defect construction module is used to construct single-source discharge defects in the switch cabinet, and each discharge defect is arranged in an independent defect area; dual-source discharge defects are constructed on the basis of single-source discharge defects, which are obtained by superimposing single-source discharge defects in pairs; at the same time, a high-aliasing discharge defect is constructed by superimposing multiple single-source discharge defects. The feature extraction module is used to collect discharge optical signals in each defect area using independent optical fiber paths and convert the collected optical signals into electrical signals. After preprocessing the collected raw electrical signal data, features that can characterize the differences between different discharge types are extracted from the effective discharge signals to obtain the input feature vector. The coarse classification module is used to input the training set feature vector obtained from the input feature vector into the optimized decision tree model, and generate the decision tree structure through iterative splitting, that is, to generate the following three classifications: single-source discharge, dual-source discharge and high aliasing discharge. The optimized decision tree model sets different misclassification cost matrices for discharge types with different numbers of sources during the splitting process of constructing decision tree nodes, and selects the optimal splitting feature by calculating the cost-sensitive information gain. The fine classification module is used to accurately classify single-source discharge, dual-source discharge, and high-aliasing discharge using an ensemble learning model based on a tree model. The model takes the features that characterize the differences between different discharge types as input and outputs the corresponding single-source discharge defect or the superposition of single-source discharge defects. For the superposition of high-risk single-source discharge defects, a misclassification cost matrix is ​​added during the accurate classification.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the multi-source partial discharge pattern recognition method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-source partial discharge pattern recognition method according to any one of claims 1 to 5.

Images