Switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation and control

Abstract

The invention relates to the technical field of intelligent diagnosis of power equipment, and discloses a switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation, which comprises the following steps: firstly, collecting multi-source partial discharge signals and constructing a defect basic database, then carrying out multi-source feature generation, principal component dimension reduction and BPNN key feature screening on the signals, and carrying out multi-source feature extraction on the signals; then, a dynamic BPA value is constructed based on the key features; and then, the defect type is identified by fusing multi-source information through an improved D-S evidence theory, and the defect is positioned in combination with a TDOA and an optimization algorithm. According to the method, the accuracy and reliability of defect type judgment after multi-source information fusion are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for power equipment, specifically to a method for diagnosing partial discharge in switchgear based on multi-information fusion and dynamic intelligent control. It is applicable to the detection, identification, location, and trend prediction of insulation defects in 10kV-500kV air-insulated, gas-insulated, and solid-insulated switchgear, providing standardized and intelligent technical support for the operation and maintenance of switchgear throughout its entire life cycle. It can be widely applied in key scenarios of power systems such as substations and distribution rooms. Background Technology

[0002] As a core piece of equipment in the power transmission and distribution system, the insulation performance of switchgear directly determines the safety and stability of the power grid. During long-term operation, factors such as manufacturing defects (e.g., uneven conductor surfaces), installation errors (e.g., excessive gaps in insulation assembly), and environmental corrosion (e.g., temperature and humidity changes, dust accumulation) can cause insulation defects inside the switchgear, including metal protrusions, air gaps, suspended metal particles, and internal insulation damage. These defects can trigger partial discharge, and the electrical, thermal, and chemical effects of partial discharge accelerate the aging of insulation materials. If not addressed promptly, this can ultimately lead to insulation breakdown, equipment burnout, or even large-scale power outages, causing significant economic losses to the power system (e.g., power outage losses for industrial users can reach tens of thousands of yuan per hour) and adverse social impacts.

[0003] Existing partial discharge diagnostic technologies for switchgear suffer from three major bottlenecks, making it difficult to meet the needs of intelligent operation and maintenance of power grids:

[0004] 1. Passive anti-interference capability and lack of guaranteed data validity: Existing technologies use fixed sensor parameters (such as fixed TEV sampling frequency and ultrasonic filtering mode) to collect signals, which cannot adjust strategies according to the dynamic interference environment on site. For example, when the electromagnetic interference intensity of a substation increases by 40%-60% during peak hours compared to normal times, the signal-to-noise ratio of UHF signals collected with fixed parameters can drop below 1.5, resulting in severe signal distortion. At the same time, the lack of a unified data validity standard means that low-quality data collected is directly used for subsequent analysis, causing the defect identification accuracy to plummet to below 70%.

[0005] 2. Static reliability of information fusion, lacking dynamic adaptability: When fusing multiple technologies, the basic probability allocation (BPA) value is often fixed based on historical data, without considering the impact of real-time interference intensity and sensor operating status (such as poor TEV grounding or low ultrasonic coupling) on ​​detection reliability. For example, the reliability of UHF detection in high-interference environments has decreased from 90% to 60%, but a fixed BPA value (such as 0.9) is still used, causing the fusion result to be biased in the wrong direction, increasing the false judgment rate by more than 30%.

[0006] 3. Defect management is lagging behind, lacking proactive early warning capabilities: Existing technologies can only identify defects after they have occurred, and cannot predict the development trend of defects based on historical characteristic data. Some metal suspension defects may develop from "minor" to "serious" within 24 hours, but routine periodic maintenance (interval of 1-3 months) cannot detect them in time, easily missing the best time for treatment, and ultimately causing equipment to trip. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for diagnosing partial discharge in switchgear based on multi-information fusion and dynamic intelligent control. Through a closed-loop technical system of defect database construction, dynamic interference adaptive acquisition, multi-dimensional feature optimization, and improved evidence fusion identification, the invention achieves accurate, efficient, and forward-looking diagnosis of insulation defects in switchgear.

[0008] This invention is achieved through the following technical solution:

[0009] A method for diagnosing partial discharge in switchgear based on multi-information fusion and dynamic intelligent control includes:

[0010] Theoretical partial discharge signals and corresponding operating condition parameters for each stage of typical insulation defects are collected to construct a defect database. The theoretical partial discharge signals include transient ground voltage signals, ultrasonic signals and ultra-high frequency signals, as well as corresponding defect type labels and corresponding signal propagation speed calibration data.

[0011] Based on real-time interference monitoring results, and referring to the operating condition parameters and signal validity standards in the defect database, the sensor parameters and signal acquisition strategy are dynamically adjusted. Based on the adjusted sensor parameters and signal acquisition strategy, the actual partial discharge signal of the switchgear to be diagnosed is used to obtain a multi-source signal feature set. The actual partial discharge signal includes transient ground voltage signal, ultrasonic signal, and ultra-high frequency signal. At least M TEV and ultrasonic sensors are arranged on each side of the switchgear shell to form a distributed array. The arrival timestamp of the actual partial discharge signal of each sensor is collected by a clock synchronization module with an accuracy higher than a first threshold, and the arrival time difference TDOA of the actual partial discharge signal is calculated.

[0012] Based on a preset four-level processing flow, the multi-source signal feature set is optimized to obtain key feature vectors and dynamic BPA values.

[0013] Using the key feature vector and dynamic BPA value as evidence sources, and referring to the defect type labels in the defect database, the actual defect type of the actual partial discharge signal in the multi-source signal feature set is determined by integrating information from multiple detection technologies through improved DS evidence theory. At the same time, the coordinates of the distributed array formed by the sensors, the time difference of arrival (TDOA) of the signal, and the signal propagation speed calibration data are used as inputs to construct and solve a three-dimensional time difference positioning model. The specific coordinates of the defect inside the switchgear are determined by particle swarm optimization-least squares method.

[0014] As an optimization, the typical insulation defects include insulation defects based on metal protrusions, insulation defects based on air gaps, insulation defects based on metal suspensions, and insulation defects based on internal discharges within the insulation.

[0015] Each stage includes a mild stage, a moderate stage, and a severe stage. The mild stage is defined as a theoretical partial discharge signal with an average pulse amplitude of <100mV, the moderate stage is defined as a theoretical partial discharge signal with an average pulse amplitude of 100-300mV, and the severe stage is defined as a theoretical partial discharge signal with an average pulse amplitude of >300mV.

[0016] The operating parameters include ambient temperature, ambient humidity, switchgear operating load, and voltage level.

[0017] As an optimization, the defect type label adopts a structured labeling format, which is defect type-key parameter-severity level; the signal propagation speed calibration data is classified and stored according to signal type-propagation medium; the full-cycle time series data of defect development consists of the pulse average amplitude of the theoretical partial discharge signal continuously collected at preset time intervals, and the duration of a single full-cycle time series data of defect development is not less than 72 hours, and it contains at least 3 complete characteristic change cycles.

[0018] As an optimization, based on real-time interference monitoring results and referring to the operating condition parameters and signal validity standards in the defect database, sensor parameters and signal acquisition strategies are dynamically adjusted. Transient ground voltage, ultrasonic, and ultra-high frequency technologies are used to collaboratively acquire the actual partial discharge signals of the switchgear under diagnosis, thereby obtaining a multi-source signal feature set. The actual partial discharge signals include transient ground voltage signals, ultrasonic signals, and ultra-high frequency signals. Simultaneously, at least M TEV sensors and ultrasonic sensors are arranged in a distributed array on each side of the switchgear casing. The arrival timestamps of the signals from each sensor are acquired through a clock synchronization module with an accuracy higher than a first threshold. The specific process for calculating the signal arrival time difference (TDOA) is as follows:

[0019] Several interference monitoring sensors are installed within a preset distance range from the switch cabinet to be diagnosed. The interference detection sensors include electromagnetic interference sensors and mechanical vibration sensors.

[0020] Interference monitoring sensors collect interference data at preset time intervals, and after continuous collection of As, the average value of the interference data is taken. The interference intensity I is calculated according to a set formula, and the interference level of the interference intensity I is classified with reference to the defect database. The interference data includes electromagnetic interference amplitude and mechanical interference amplitude. The interference intensity I = electromagnetic interference amplitude E × 0.65 + mechanical vibration amplitude V × 0.35.

[0021] A mapping table based on interference level and optimal parameters is constructed using the theoretical partial discharge signal. The optimal parameters include sensor operating parameters and data acquisition parameters. Based on the mapping table, a combination scheme of sensor parameters and acquisition strategies adapted to the current interference level and operating conditions is determined by combining the current operating condition parameters.

[0022] Based on the aforementioned combination scheme, several TEV sensors and ultrasonic sensors are arranged on the switch cabinet to be diagnosed. The actual partial discharge signals are collected through the TEV sensors and ultrasonic sensors, and the original pulse waveforms of the actual partial discharge signals are stored in the original pulse waveform database. A timestamp is recorded when the TEV sensors and ultrasonic sensors collect the actual partial discharge signals.

[0023] For the same discharge event, the time difference between any two sensors of the same type is calculated based on the timestamp, and the time difference is bound to the corresponding sensor coordinates and stored.

[0024] The actual partial discharge signal is preprocessed and its features are extracted to obtain a multi-source signal feature set.

[0025] As an optimization, the basic features in the multi-source signal feature set include the pulse average amplitude and pulse maximum amplitude of the transient ground voltage signal, the pulse duration, pulse amplitude standard deviation, center frequency and bandwidth of the ultrasonic signal, and the UHF pulse average amplitude, UHF pulse amplitude variance, percentage of pulses with amplitude exceeding the threshold, percentage of positive half-cycle pulse count, percentage of negative half-cycle pulse count, standard deviation of positive half-cycle phase distribution, standard deviation of negative half-cycle phase distribution, phase interval corresponding to the maximum pulse amplitude, total pulse count, positive half-cycle pulse count, negative half-cycle pulse count and the rate of change of pulse count over time.

[0026] As an optimization, the preset four-level processing flow is as follows:

[0027] The first stage involves extracting multi-dimensional features, including using the multi-source signal feature set as a basis, calling the original pulse waveform data of transient ground voltage signals, ultrasonic signals, and ultra-high frequency signals in the same batch of the original pulse waveform database through the batch index, extracting the depth features of the original pulse waveform data of the transient ground voltage signals, ultrasonic signals, and ultra-high frequency signals, and merging the depth features into the multi-source signal feature set to obtain an initial extended feature set;

[0028] The second stage involves principal component dimensionality reduction. Using the theoretical partial discharge signals in the defect database as a reference, principal component analysis is performed on the initial extended feature set. The covariance matrix of the initial extended feature set is calculated, the eigenvalues ​​and eigenvectors are solved, and the principal components with a cumulative contribution rate of ≥90% are selected, thereby reducing the feature dimension of the initial extended feature set to obtain the principal component feature set.

[0029] The third stage involves optimizing the backpropagation neural network. A backpropagation neural network model is constructed, and 70% of the theoretical partial discharge signals from the defect database are used as the training set, 15% as the validation set, and 15% as the test set to train the model. Then, redundant features with sensitivity <0.05 are removed through sensitivity analysis, and 5-7 key feature vectors are output to obtain a key feature set. The backpropagation neural network model includes an input layer, two hidden layers, and an output layer. The node dimension of the input layer is the same as the dimension of the features in the dimensionality-reduced feature set. The output layer outputs defect type labels, and corresponding key feature vectors are selected based on these labels.

[0030] The fourth layer is to construct a dynamic BPA. An initial BPA value is set based on the defect base database, and the BPA value is assigned to each of the key features. A feature reliability coefficient is introduced to correct the initial BPA value. The corrected BPA value is equal to the initial BPA value multiplied by the reliability coefficient.

[0031] As an optimization, the depth features include the average peak factor and peak factor standard deviation of the transient ground voltage signal, the spectral energy ratio and pulse rise time of the ultrasonic signal, and the phase entropy H and amplitude variation coefficient of the ultra-high frequency signal.

[0032] As an optimization, using the key feature vector and dynamic BPA value as evidence sources, and referring to the defect type labels in the defect database, the specific process of determining the actual defect type of the actual partial discharge signal in the multi-source signal feature set by improving the DS evidence theory and fusing information from multiple detection technologies is as follows:

[0033] The key feature vectors are divided into three independent evidence sources according to signal type: TEV evidence source, ultrasound evidence source, and UHF evidence source. The dynamic BPA value corresponding to each independent evidence source is normalized.

[0034] Calculate the conflict coefficient k among the three independent sources of evidence and classify the conflict level according to the value of the conflict coefficient k, as follows:

[0035] ;

[0036] in, This indicates that the defect type A pointed to by the TEV evidence source, the defect type B pointed to by the ultrasonic evidence source, and the defect type C pointed to by the UHF evidence source have no overlap. , , These are the normalized dynamic BPA values ​​of the TEV evidence source for defect type A, reflecting the confidence level that the TEV signal features support defect A. The larger the value, the higher the support. This represents the normalized dynamic BPA value of the ultrasonic evidence source for defect type B, reflecting the confidence level that the ultrasonic signal characteristics support defect B. ∑ represents the normalized dynamic BPA value of the UHF evidence source for defect type C, reflecting the confidence level that the UHF signal features support defect C, and ∑ is the summation over all combinations that satisfy A∩B∩C=∅;

[0037] A differentiated fusion strategy is adopted for different conflict levels to fuse BPA values ​​and obtain a comprehensive BPA value;

[0038] The defect type with the largest comprehensive BPA value is selected as the candidate type. If the comprehensive BPA value of the candidate type is greater than or equal to the judgment threshold, the candidate type is determined to be the actual defect type. If the comprehensive BPA value of the candidate type is less than the judgment threshold, the backpropagation neural network is re-optimized until the judgment condition is met.

[0039] As an optimization, the differentiated fusion strategy is specifically as follows:

[0040] If the conflict level is low, the normalized dynamic BPA values ​​of the three independent evidence sources are directly synthesized using the traditional DS synthesis rules to obtain the comprehensive BPA value.

[0041] If the conflict level is medium conflict, the normalized dynamic BPA value of the conflicting independent evidence sources is discounted based on a correction factor, and the normalized dynamic BPA value of the non-conflicting independent evidence sources is proportionally adjusted according to the correction factor to increase the normalized dynamic BPA value of the non-conflicting independent evidence sources; finally, the normalized dynamic BPA values ​​of the three independent evidence sources are combined using the traditional DS synthesis rule to obtain the comprehensive BPA value.

[0042] If the conflict level is high, a weighted average fusion rule is used to synthesize the normalized dynamic BPA values ​​of the three independent evidence sources to obtain a comprehensive BPA value.

[0043] As an optimization, using the coordinates of the distributed array formed by the sensors, the time difference of arrival (TDOA) of the signal, and the signal propagation speed calibration data as inputs, a three-dimensional time difference positioning model is constructed and solved. The specific process of determining the specific coordinates of the defect inside the switchgear using particle swarm optimization-least squares method is as follows:

[0044] Obtain the three-dimensional coordinates of the sensors in the distributed array, denoted as... ;

[0045] The clock synchronization module acquires the arrival time of partial discharge signals at each sensor and calculates the time difference between any two sensors. ;

[0046] Determine the signal propagation speed v based on the current operating parameters;

[0047] A three-dimensional time-difference localization model is constructed based on the time difference, propagation speed, and defect coordinates. The three-dimensional time-difference localization model is as follows: ;

[0048] Where (x, y, z) are the defect coordinates;

[0049] The three-dimensional time difference positioning model is solved by particle swarm optimization-least squares method, thereby obtaining the defect coordinates.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] This invention integrates multiple source signals, including transient ground voltage (TEV), ultrasonic waves (AE), and ultra-high frequency (UHF). It first reduces the dimensionality through principal component analysis (PCA), and then uses a backpropagation neural network (BPNN) to screen key feature vectors, eliminate redundant features, retain key features that contribute highly to defect identification, reduce interference, improve the accuracy of defect identification, and reduce the complexity of subsequent data processing.

[0052] This invention dynamically adjusts the BPA value based on the historical identification accuracy of the defect database and in conjunction with real-time interference levels and sensor operating status, so that the BPA value can more accurately reflect the degree of trust of the current key feature vector in the defect type, providing a more reliable input for subsequent evidence fusion and enhancing the adaptability and accuracy of defect identification.

[0053] Based on the conflict coefficient k between evidence sources, this invention employs different fusion strategies (traditional DS, discounted DS, and weighted average) for low-conflict, medium-conflict, and high-conflict scenarios. This can reasonably integrate multi-source evidence, effectively handle conflicts between evidence, and improve the accuracy and reliability of defect type determination after multi-source information fusion.

[0054] This invention utilizes the time difference of arrival (TDOA) of distributed array sensors, combined with particle swarm optimization-least squares method to solve the three-dimensional positioning model, to achieve accurate location of internal defects in switchgear, providing accurate location information for maintenance.

[0055] This invention trains a Long Short-Term Memory (LSTM) network model based on time-series data of the entire defect development cycle in a defect database. It predicts the future rate of change of defect characteristics and risk level, and can predict the development trend of defects in advance. Combined with risk level standards, it outputs differentiated maintenance suggestions, which helps to take measures in advance and avoid more serious failures. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments. The illustrative embodiments and descriptions of this invention are only used to explain this invention and are not intended to limit this invention.

[0057] This embodiment 1 provides a method for diagnosing partial discharge in switchgear based on multi-information fusion and dynamic intelligent control, which includes a total of 5 steps.

[0058] Step 1: Collect theoretical partial discharge signals and corresponding operating condition parameters for each stage of typical insulation defects to construct a defect database. The theoretical partial discharge signals include transient ground voltage signals, ultrasonic signals, and ultra-high frequency signals, as well as corresponding defect type labels, corresponding signal propagation velocity calibration data, and time-series data of the entire defect development cycle.

[0059] Acquire basic data on typical insulation defects in switchgear and construct a defect database (hereinafter referred to as the database): For four typical insulation defects—metal protrusions, air gaps, suspended metal particles, and internal insulation discharge—through on-site measurements (collecting real partial discharge signals from actual operating switchgear throughout the entire cycle from minor to severe defects) or simulated operating conditions (reproducing the entire cycle evolution of defects from initial to worsening, collecting simulated partial discharge signals), simultaneously record the theoretical partial discharge signals (transient ground voltage signals, ultrasonic signals, and UHF signals) and corresponding operating parameters (ambient temperature and humidity, equipment operating load, and voltage level), with parameter recording accuracy meeting the following requirements: temperature ±0.5℃, humidity ±2%RH, load ±1% of rated load, and voltage ±0.5% of rated voltage). Based on the above data, construct the defect database. The theoretical partial discharge signals in this database are "theoretical supporting signals," containing time-series data of the entire defect development cycle (a sequence of multi-stage theoretical partial discharge signals continuously collected at time intervals of 5-10 minutes), which are used for subsequent model training (BPNN / LSTM), parameter calibration (signal propagation speed), and standard definition (feature validity, risk level).

[0060] In some embodiments, each stage includes a mild stage, a moderate stage, and a severe stage, wherein the mild stage is defined as a theoretical partial discharge signal with an average pulse amplitude of <100mV, the moderate stage is defined as a theoretical partial discharge signal with an average pulse amplitude of 100-300mV, and the severe stage is defined as a theoretical partial discharge signal with an average pulse amplitude of >300mV.

[0061] The acquisition methods for theoretical partial discharge signals are divided into field measurement and simulated operating conditions: Field measurement is conducted on operating switchgear with confirmed defect types, with a acquisition time of no less than 10 minutes per group, and the signal-to-noise ratio of each group of signals (theoretical partial discharge signals) is ≥3; Simulated operating conditions are used to reproduce the actual operating environment in the laboratory (temperature -20℃-60℃, humidity 30%-90%RH, load 30%-100% rated load), and the length of metal protrusions is set to 0.5-10mm, the thickness of air gaps to 0.1-5mm, the diameter of metal suspended objects to 0.1-2mm, and the size of pre-existing cracks / bubbles inside the insulation to 0.1-3mm through a customized defect module, and simulated signals (theoretical partial discharge signals) are acquired.

[0062] In the defect database, signal propagation speed calibration data are stored uniformly according to the combination of signal type (TEV / ultrasound / Ultra-high frequency) and propagation medium (air / insulating oil / solid insulating material).

[0063] For example, for the TEV signal (transient ground voltage signal) propagating in the air inside the switchgear, based on extensive laboratory tests and field verifications (such as the statistical analysis of simulated operating conditions and measured data in step 1), its propagation speed is determined to be v≈3×10.8 m / s (approximately the speed of light, considering that the air inside the switch cabinet has minimal impact on signal propagation), this calibration value is stored in the parameter field of the database "TEV signal - air medium".

[0064] For ultrasonic signals, if the propagation medium is insulating oil inside the switchgear, the propagation speed v≈1500m / s (typical value of sound speed in insulating oil) can be obtained through acoustic experiments (such as emitting and receiving ultrasonic waves in a simulated insulating oil environment, and measuring the propagation time and distance), and stored under the ultrasonic signal-insulating oil medium parameters.

[0065] The propagation speed calibration value in the database is not fixed. Instead, it is updated periodically (e.g., every 3 months) using data fitting algorithms such as the least squares method, based on the new data continuously collected in step 1 (including signal propagation tests under different temperature and humidity conditions) to ensure the consistency between the calibration value and the actual propagation characteristics. For example, if the results of 100 sets of ultrasonic signal tests under high temperature (50℃) and high humidity (80%RH) conditions show that the sound speed in insulating oil decreases by 5% due to changes in medium viscosity, the database will automatically update the ultrasonic signal-insulating oil medium propagation speed calibration value to 1425 m / s (1500 × (1-5%)).

[0066] In some embodiments, the defect type label adopts a structured labeling format, which is defect type-key parameter-severity level, such as "metal protrusion-5mm-severe" and "air gap-2mm-moderate", and the characteristic correlation between the label and the theoretical partial discharge signal is ≥98%; the key parameter here is the width of the defect type.

[0067] Signal propagation speed calibration data are stored according to signal type and propagation medium, with the calibration value for transient ground voltage signals in air medium being 3 × 10⁻⁶. 8 m / s ±5%, with a calibration value of 2.9 × 10 m / s in SF6 gas medium. 8 The calibration value of the ultrasonic signal is 340m / s±5% in air medium and 1500m / s±3% in insulating oil medium. For every 100 new theoretical partial discharge signals, the calibration data is iteratively updated using the least squares method.

[0068] The time series data of the entire defect development cycle are continuously collected at time intervals of 5-10 minutes, including transient ground voltage signals, ultrasonic signals and ultra-high frequency signals, as well as corresponding defect type labels and corresponding signal propagation speed calibration data. The duration of a single time series data sequence is not less than 72 hours and contains at least 3 complete characteristic change cycles (the characteristic change cycle is defined as the time span from the peak value to the valley value and back to the peak value).

[0069] Step 2, Dynamic Interference Adaptive Suppression and Multi-Method Collaborative Acquisition: Based on real-time interference monitoring results (electromagnetic interference, mechanical vibration), and referring to the operating condition parameters and signal validity standards in the defect database, the sensor parameters and signal acquisition strategy are dynamically adjusted. Based on the adjusted sensor parameters and signal acquisition strategy, the actual partial discharge signal of the switchgear to be diagnosed (this signal is the "actual diagnostic signal") is acquired through the collaborative acquisition of three technologies: transient ground voltage (TEV), ultrasonic (AE), and ultra-high frequency (UHF). At the same time, at least three TEV / ultrasonic sensors are arranged on each side of the switchgear shell to form a distributed array. The arrival timestamp of the signal from each sensor is acquired through a high-precision clock synchronization module (synchronization error ≤1μs), and the time difference of arrival (TDOA) is calculated. The validity of the acquired data is ensured through signal-to-noise ratio verification (SNR≥3, based on database standards), providing signal characteristics for subsequent defect identification and time difference data for location.

[0070] The specific implementation process of step 2 is as follows:

[0071] Step 2.1: Deploy the real-time interference monitoring system and collect data.

[0072] Sensor arrangement: Several interference monitoring sensors are set within a preset distance from the switch cabinet to be diagnosed. The interference detection sensors include electromagnetic interference sensors and mechanical vibration sensors. For example, the interference monitoring sensors are deployed according to the principle of uniform coverage and no blind spots: 3-4 electromagnetic interference sensors (frequency band 10kHz-1GHz, sampling frequency 5GHz) and 2-3 mechanical vibration sensors (range 0.1-1000Hz, sampling frequency 10MHz) are arranged within 3-5m around the switch cabinet to be diagnosed to monitor electromagnetic and vibration interference respectively.

[0073] Data Acquisition and Intensity Calculation: Interference data is collected by the interference monitoring sensor at preset time intervals. After continuous acquisition of data (As), the average value of the interference data is taken. The interference intensity I is calculated according to a set formula, and the interference level of the interference intensity I is classified with reference to the defect database. The interference data includes electromagnetic interference amplitude and mechanical interference amplitude. The interference intensity I = electromagnetic interference amplitude E × 0.65 + mechanical vibration amplitude V × 0.35. For example, interference data is collected once every 100ms, and the average value is taken for 5 consecutive seconds. The interference intensity I is calculated according to the formula: Interference intensity I = electromagnetic interference amplitude E × 0.65 + mechanical vibration amplitude V × 0.35 (weights are determined through training with 100 sets of historical data from the database). Referring to the database standard, I < 30dB is low interference, 30dB ≤ I < 60dB is medium interference, and I ≥ 60dB is high interference.

[0074] Step 2.2: Dynamic adjustment of sensor parameters and acquisition strategy:

[0075] Adjustment Basis: A mapping table based on interference level and optimal parameters is constructed using the theoretical partial discharge signals. The optimal parameters include sensor operating parameters and data acquisition parameters. Based on the mapping table, a combination scheme of sensor parameters and acquisition strategies adapted to the current interference level and operating conditions is determined in combination with the current operating conditions. This ensures that the signal-to-noise ratio (SNR) of the actual partial discharge signal corresponding to the combination scheme is greater than a set threshold. For example, based on the database interference level-optimal parameter mapping table (constructed from 200 sets of theoretical signals), a combination scheme is determined in combination with the current operating conditions (temperature, humidity, load, etc.). After adjustment, the SNR must be greater than or equal to 3 (database validity standard).

[0076] Combination scheme:

[0077] Low interference (I<30dB): TEV sampling frequency 200-300MHz (default 250MHz), ultrasonic sampling frequency 2-3MHz (default 2.5MHz), UHF sampling frequency 2-3GHz (default 2.5GHz), sampling interval 10min / group, duration 1min;

[0078] Interference in the medium range (30dB≤I<60dB): TEV sampling frequency increased to 300-400MHz (50Hz notch filter enabled), ultrasonic coupling agent thickness 0.5-1mm, UHF gain increased by 5-10dB (300MHz-2.5GHz bandpass enabled), sampling interval 5min / group, duration 1.5min;

[0079] High interference (I≥60dB): TEV dual-sensor differential acquisition (spacing 10-15cm), ultrasonic open center frequency ±5kHz narrowband filtering, UHF real-time interference avoidance frequency band, acquisition interval 2min / group, duration 2min.

[0080] Step 2.3, Distributed Array Deployment and TDOA Data Acquisition:

[0081] Array Arrangement: Based on the aforementioned combination scheme, several TEV sensors and ultrasonic sensors are arranged on the switchgear to be diagnosed. Actual partial discharge signals are acquired through these TEV and ultrasonic sensors. For example, M sensors (M=3-4) are arranged on each of the four sides of the switchgear (front / left / right / back, totaling 4 sides), with 2 TEV sensors and 1 ultrasonic sensor per side, arranged in an isosceles triangle. The coordinates are calibrated using a laser rangefinder (accuracy ±0.5cm).

[0082] Clock synchronization and timestamp acquisition: When the TEV sensor and ultrasonic sensor acquire the actual partial discharge signal, a timestamp is recorded; specifically, a Beidou + GPS dual-mode synchronization module is used (first threshold = 1μs, synchronization error ≤ 1μs). When the sensor detects a discharge signal (amplitude exceeding 80% of the minimum amplitude of similar defects in the database), a timestamp is recorded (accuracy ±0.1μs).

[0083] TDOA calculation: For the same discharge event, calculate the time difference between any two sensors of the same type based on the timestamp, bind the time difference with the corresponding sensor coordinates and store them.

[0084] For the same discharge event, extract the timestamps recorded by different sensors and calculate the time difference between any two sensors. Where i and j are sensor numbers, , These represent the timestamps of arrival at sensor i and sensor j, respectively. The TDOA represents the time difference between arrival at sensors i and j. The TDOA calculation accuracy is ±0.2μs. The calculation results are bound to the corresponding sensor coordinates and stored for subsequent solving of the three-dimensional time difference positioning model.

[0085] Here, any two sensors include sensors on the same housing (including TEV and ultrasonic sensors) as well as sensors on different housings, but the signal type matching principle must be followed (i.e., TEV sensor calculation with TEV sensor, ultrasonic sensor calculation with ultrasonic sensor).

[0086] Calculations between TEV sensors: Two TEV sensors on the same side (e.g., the front), or TEV sensors on different sides (e.g., the front and right sides), can calculate TDOA because both collect transient ground voltage signals and have consistent signal propagation characteristics (e.g., propagation speed is 3×10⁻⁶). 8 m / s (taken from the defect baseline database calibration value), the time difference can reflect the distance relationship between the defect location and the sensor.

[0087] Calculation between ultrasonic sensors: One ultrasonic sensor on the same side (such as the front) and another ultrasonic sensor on other sides (such as the left side) can calculate TDOA. Since both sensors collect ultrasonic signals and have the same propagation speed (such as 340m / s in air), the time difference also has positioning significance.

[0088] Same-plane sensors: primarily provide in-plane position information. For example, two TEV sensors on the front can use their TDOA to preliminarily determine the X-axis (length) and Z-axis (height) positions of defects on the front.

[0089] Different surface sensors: supplement spatial depth information. For example, the front TEV sensor and the right side TEV sensor can determine the Y-axis (width, i.e. front and back depth) position of the defect through their TDOA. Finally, by combining multiple sets of TDOA data to establish a three-dimensional positioning model, accurate positioning within ±3cm can be achieved.

[0090] The arrangement of 2 TEVs and 1 ultrasonic sensor on each side is precisely to provide a sufficient combination of similar sensors:

[0091] Based on a 4-sided shell, there are a total of 8 TEV sensors and 4 ultrasonic sensors, which can form 8×7 / 2=28 sets of TEV-TEVTDOA and 4×3 / 2=6 sets of ultrasonic-ultrasonic TDOA. A large amount of redundant data can be eliminated by subsequent particle swarm optimization-least square method (step 4) to remove errors and improve positioning accuracy.

[0092] Step 2.4, Multi-source signal feature set generation: The actual partial discharge signal is preprocessed and features are extracted to obtain a multi-source signal feature set.

[0093] Signal preprocessing: Wavelet threshold denoising (db4 basis, 3-level decomposition) is performed on the actual partial discharge signal collected, and effective pulses are extracted by adaptive threshold (removal ratio ≤10%).

[0094] Feature extraction and integration: Referring to the feature definition of the defect database, TEV (2 time-domain features), ultrasound (4 time-frequency features), and UHF (12 statistical features) are extracted and integrated into a multi-source signal feature set with 30-50 parameters. After labeling the association information, it is used for subsequent optimization.

[0095] TEV (Transient Ground Voltage) signal: 2 time-domain characteristics.

[0096] The time-domain characteristics of TEV signals directly reflect the amplitude variation of partial discharge pulses over time. The following two core characteristics are selected from the defect database:

[0097] Average pulse amplitude (unit: mV): The arithmetic mean of the peak amplitudes of all effective partial discharge pulses within a single acquisition duration (e.g., 1 minute under low interference). Calculated as: Average pulse amplitude = Σ (peak amplitude of a single effective pulse) / number of effective pulses. Database records show that the average TEV pulse amplitude of metallic protrusion defects is typically 30%-50% higher than that of air gap defects (e.g., 180mV for metallic protrusions vs. 120mV for air gaps), making it a key time-domain indicator for distinguishing defect types.

[0098] Maximum Pulse Amplitude (unit: mV): The maximum peak amplitude among all effective partial discharge pulses within a single acquisition duration. It is calculated by extracting the maximum value from the sequence of peak amplitudes of effective pulses. This feature reflects the "strongest discharge intensity" of partial discharge. In the database, the maximum amplitude of TEV pulses for severe defects (average pulse amplitude > 300 mV) is typically > 500 mV, which can help determine the severity of the defect.

[0099] Ultrasonic signal: 4 time-frequency characteristics.

[0100] The ultrasonic signal needs to be combined with both "time-domain amplitude variation" and "frequency-domain energy distribution" to reflect the discharge characteristics. The following four time-frequency features are selected from the database (covering two in the time domain and two in the frequency domain, taking into account both signal time and frequency dimensions):

[0101] Time-domain features (2):

[0102] Pulse duration (unit: μs): The time span from when the amplitude of a single effective ultrasonic discharge pulse reaches 10% of its peak value to when the amplitude drops to 10% of its peak value. It is calculated by thresholding the pulse waveform and then statistically analyzing the duration of this time interval. In the database, the ultrasonic pulse duration for metal suspension defects (15-25 μs) is significantly longer than that for internal insulation discharge (5-10 μs), due to the longer discharge process caused by the movement of the suspended matter.

[0103] Pulse amplitude standard deviation (unit: dB): The standard deviation of the peak amplitude of all effective ultrasonic pulses within a single acquisition duration. It is calculated as: Standard deviation = √[Σ(single pulse amplitude - average amplitude)² / (number of effective pulses - 1)]. This characteristic reflects discharge stability. In the database, the amplitude standard deviation of air gap defects (5-8 dB) is smaller than that of metal protrusions (10-15 dB) because air gap discharge is more regular.

[0104] Frequency domain features (2):

[0105] Center frequency (unit: kHz): The frequency corresponding to the maximum power spectral density of the ultrasonic signal, i.e., the frequency at which the signal energy is most concentrated. It is calculated by performing a Fourier transform on the ultrasonic signal to obtain the power spectrum, and then finding the frequency corresponding to the peak value of the power spectrum. In the database, the center frequencies of different defects vary. The center frequency of internal discharge in insulation (80-120kHz) is higher than that of metal protrusions (40-60kHz) because the discharge energy of internal defects is more concentrated in the high-frequency range.

[0106] Bandwidth (unit: kHz): The frequency range where the power spectral density is ≥ 50% of the maximum power spectral density (i.e., -3dB bandwidth). It is calculated by determining the upper and lower limits of the frequency range corresponding to "50% of the maximum power spectral density" on the power spectrum; bandwidth = upper limit frequency - lower limit frequency. In the database, the bandwidth of metal suspension defects (50-80kHz) is wider than that of air gaps (20-40kHz) because the ultrasonic frequency components generated by the discharge of suspended matter are more complex.

[0107] UHF (Ultra-High Frequency) signals: 12 statistical characteristics.

[0108] Due to the high frequency and large number of pulses in UHF signals, it is necessary to quantify their overall discharge patterns through statistical features. The 12 features selected in the database are all extracted based on PRPD (Phase-Resolved Pulse Sequence) maps (PRPD maps are plotted with "discharge phase" on the X-axis, "pulse amplitude" on the Y-axis, and "pulse count" on the Z-axis), as detailed below:

[0109] Amplitude-related statistical characteristics (3):

[0110] UHF pulse average amplitude (unit: mV): Same as the TEV pulse average amplitude definition, reflecting the overall discharge intensity of the UHF signal;

[0111] UHF pulse amplitude variance (unit: mV²): reflects the dispersion of UHF pulse amplitude; the more severe the defect, the larger the variance.

[0112] Percentage of pulses with amplitude exceeding the threshold (%): The proportion of pulses with amplitude greater than 1.5 times the average amplitude of similar defects in the database to the total effective pulses, with a higher proportion for severe defects.

[0113] Phase correlation statistical characteristics (5):

[0114] Positive half-cycle pulse count percentage (%): The proportion of pulses in the "0°-180° (positive half-cycle)" range in the PRPD spectrum to the total number of pulses. For metal protrusions, the positive half-cycle percentage is usually >60%.

[0115] Negative half-cycle pulse count percentage (%): Corresponding to the positive half-cycle, the negative half-cycle of the air gap has a slightly higher percentage (45%-55%).

[0116] Positive half-cycle phase distribution standard deviation (°): The standard deviation of all pulse phases within the positive half-cycle reflects the degree of phase concentration—the discharge phase inside the insulation is more concentrated, with a standard deviation <15°;

[0117] Negative half-cycle phase distribution standard deviation (°): Corresponding to the positive half-cycle, the phase distribution of metal suspensions is more dispersed, with a standard deviation >20°;

[0118] Phase interval (°) corresponding to the maximum pulse amplitude: Record the phase interval (e.g., 30°-60°) where the largest pulse amplitude is located in the PRPD spectrum. The interval varies for different defects.

[0119] Count-related statistical characteristics (4)

[0120] Total pulse count (number): The total number of valid UHF pulses within a single acquisition duration. The total count for severe defects is usually >500.

[0121] Positive half-cycle pulse count (number): The number of effective pulses within the positive half-cycle, which, together with the proportion of the positive half-cycle, reflects the phase characteristics;

[0122] Negative half-cycle pulse count (number): Corresponds to the positive half-cycle pulse count;

[0123] Pulse count change rate over time (pulse counts / min): The trend of pulse count change per unit time. When the defect worsens, the change rate is >10 pulse counts / min.

[0124] The above basic features total 2 (TEV) + 4 (ultrasound) + 12 (UHF) = 18 core features. The actual integration results in a multi-source signal feature set of 30-50 parameters because some features in the database have been subdivided and expanded. For example:

[0125] The average amplitude of a UHF pulse can be further divided into two parameters: the average amplitude of the positive half-cycle and the average amplitude of the negative half-cycle.

[0126] The center frequency of an ultrasound wave can be further divided into two parameters: the positive half-cycle center frequency and the negative half-cycle center frequency.

[0127] Some features need to be combined with related information such as acquisition duration and interference level to form derived parameters (such as the maximum amplitude of TEV pulses per unit time).

[0128] The final integrated basic feature set retains the core distinguishability and improves the accuracy of feature optimization in the subsequent four-level processing flow in step 3 by subdividing parameters, providing sufficient data dimensions for defect identification and localization.

[0129] Step 3: Partial discharge signal feature optimization processing: Based on the preset four-level processing flow, the feature set of the multi-source signal is optimized to obtain the key feature vector and dynamic BPA value.

[0130] The preset four-level processing flow is as follows:

[0131] The first level is multi-dimensional feature extraction, which includes using the multi-source signal feature set as a basis, calling the original pulse waveform data of transient ground voltage signal, ultrasonic signal, and UHF signal in the same batch of the original pulse waveform database by collecting batch index, extracting the depth features of the original pulse waveform data of the transient ground voltage signal, ultrasonic signal, and UHF signal, and merging the depth features into the multi-source signal feature set to obtain an initial extended feature set.

[0132] Based on the multi-source signal feature set generated in step 2, the transient ground voltage (TEV), ultrasonic, and ultra-high frequency (UHF) original pulse waveform data of the same batch in the original pulse waveform database are called by the batch index (which uniquely corresponds to the batch of the multi-source signal feature set). Combined with the feature definition of the defect basic database, the basic feature supplementation and deep feature calculation are completed:

[0133] TEV signal feature extension:

[0134] Based on the original TEV pulse waveform (each waveform contains ≥1000 time-amplitude data points), the peak factor (peak value of a single pulse ÷ effective value of a single pulse) and kurtosis factor (fourth-order central moment ÷ variance²) are calculated according to the database definition.

[0135] The average peak factor and standard deviation of peak factor of all valid TEV pulses in the same batch are statistically analyzed and then added to the average amplitude and maximum amplitude of TEV pulses in the original multi-source signal feature set to form a TEV signal feature subset (a total of 4 features).

[0136] Ultrasonic signal feature extension:

[0137] Based on the original ultrasonic pulse waveform, the power spectrum is obtained through Fourier transform, and the spectrum energy ratio (energy of the center frequency ± 10kHz band ÷ total band energy) and pulse rise time (time to rise from 10% to 90% of the amplitude) are calculated according to the database definition.

[0138] After supplementing the pulse duration, pulse amplitude standard deviation, center frequency, and bandwidth of the original multi-source signal feature set with the above two depth features, an ultrasonic signal feature subset (a total of 6 features) is formed.

[0139] UHF signal feature extension:

[0140] Based on the phase-amplitude data of the original UHF pulse, the phase entropy H (reflecting the uniformity of phase distribution, formula: ) is calculated according to the definition of the original pulse waveform database. , The percentage of pulses in each phase interval and the amplitude variation coefficient (amplitude standard deviation ÷ amplitude average).

[0141] After adding the above two deep features to the original multi-source signal feature set of 12 UHF statistical features, a UHF signal feature subset (a total of 14 features) is formed.

[0142] Feature integration: Merge the feature subsets of the three types of signals after expansion to form an initial extended feature set containing 4+6+14=24 features, ensuring that the feature dimensions cover the time domain, frequency domain, phase domain, and statistical domain.

[0143] The second stage is principal component dimensionality reduction. Using the theoretical partial discharge signals in the defect database as a reference, principal component analysis (PCA) is performed on the initial extended feature set. The covariance matrix of the initial extended feature set is calculated, the eigenvalues ​​and eigenvectors are solved, and the principal components with a cumulative contribution rate of ≥90% are selected. This reduces the feature dimension of the initial extended feature set to obtain the principal component feature set, reducing the feature dimension to 5-10 dimensions, and the information retention rate after dimensionality reduction is ≥95%.

[0144] Calculate the variance contribution rate of each feature in the initial extended feature set, select principal components (usually 8-12) with a cumulative variance contribution rate ≥90%, and remove redundant features with a variance contribution rate <1% (such as TEV pulse average amplitude being highly correlated with UHF pulse average amplitude, retaining UHF features with higher variance contribution rates).

[0145] Output principal component feature sets, compressing the feature dimensions to 10-15, ensuring that the data volume is reduced while no core information is lost.

[0146] Let the initial expanded feature set have 24 features (called the original feature values), denoted as . Assuming 100 valid samples are collected, denoted as... The specific calculation process of dimensionality reduction by principal component analysis (PCA) is as follows. The core objective is to screen principal components with a cumulative contribution rate of ≥90% by solving the covariance matrix and eigenvalues / eigenvectors, and compress the feature dimension from 24 dimensions to 10-12 dimensions.

[0147] The specific calculation process is as follows:

[0148] A1. Data preprocessing of the initial expanded feature set (eliminating the influence of dimensions):

[0149] Because the dimensions and numerical ranges of the 24 features differ significantly (e.g., the average amplitude of TEV pulses is in mV, the total UHF pulse count is in "pulses", and the center frequency of ultrasound is in kHz), direct calculation would lead to the features with larger dimensions dominating the analysis results. Therefore, standardization is required first, as shown in the following formula: (i=1,2,···,100; j=1,2,···,24), Let j be the j-th original feature value of the i-th sample group. The average value of 100 samples for the j-th original feature; Let be the standard deviation of 100 samples for the j-th original feature; The standardized feature value is the j-th original feature value of the i-th sample.

[0150] Taking two original features (hereinafter referred to as features) as an example:

[0151] Assuming the first feature Raw data for (TEV pulse average amplitude):

[0152] Average of 100 samples Standard deviation ;

[0153] Original values ​​of Group 1 sample After standardization ;

[0154] Original values ​​of Group 2 samples After standardization ;

[0155] All 24 features are standardized using this formula, resulting in a 100×24 standardized feature matrix. ; This represents the overall feature matrix used for principal component analysis after standardization. It describes the dimensions of the matrix, namely 100 rows (corresponding to 100 sets of samples) and 24 columns (corresponding to 24 features).

[0156] A2. Calculate the covariance matrix of the standardized feature matrix. The covariance matrix measures the degree of linear correlation between any two features. The covariance matrix of 24 features is a 24×24 symmetric matrix (denoted as Σ). Matrix elements... (The value in row j and column k) represents the covariance between the j-th feature and the k-th feature, as shown in the following formula:

[0157] n is the number of samples, which is 100 in this embodiment. Let be the average of all samples after standardization for the j-th feature. Let be the average of all samples after standardization for the k-th feature. It is the standardized value of the k-th feature of the i-th sample group.

[0158] Because the features have been standardized, (The average value of the standardized features is 0), the formula can be simplified to: .

[0159] Example of covariance matrix structure:

[0160] ;

[0161] diagonal elements The variance of the j-th feature (after standardization) (Since the variance of the standardized features is 1).

[0162] off-diagonal elements Let the covariance between the j-th and k-th features be, for example Let TEV pulse mean amplitude be the covariance between the average amplitude of the TEV pulse and the center frequency of the ultrasound. If the two are highly correlated, Close to 1 or -1.

[0163] In actual calculations, matrix operations can be used to simplify the process: , ( For the standardized matrix The transpose of the matrix (i.e., a 24×100 matrix) can be directly solved using tools such as MATLAB and Python (numpy library), avoiding errors from manual calculation.

[0164] A3. Find the eigenvalues ​​and eigenvectors of the covariance matrix, and the covariance matrix itself. eigenvalues ​​(denoted as) ) and the corresponding feature vector (denoted as ) The core of PCA is solving the characteristic equation. The specific steps are as follows:

[0165] A3.1 Solving for eigenvalues: Calculating the characteristic polynomial of the covariance matrix. (I is a 24th order identity matrix), resulting in 24 eigenvalues ​​(theoretically all of which are non-negative, since the covariance matrix is ​​a positive semi-definite matrix).

[0166] A3.2, Eigenvalue Sorting: Sort the 24 eigenvalues ​​in descending order, i.e. ( Corresponding to the first principal component, (Corresponding to the second principal component, and so on).

[0167] A3.3 Solving for the eigenvectors: For each sorted eigenvalue... (m=1,2,...,24), substitute into the equation Solving for the corresponding eigenvectors yields the results. (24-dimensional column vector, needs to be standardized, i.e.) . for The transpose of .

[0168] Quantization example (eigenvalue sorting results):

[0169] Assume that the first 12 eigenvalues ​​are obtained through the tool (since the cumulative contribution rate of subsequent values ​​is usually concentrated in the first 12):

[0170] Principal component number Eigenvalues 1 8.52 2 5.36 3 3.91 4 2.87 5 2.15 6 1.78 7 1.32 8 1.02 9 0.88 10 0.75 11 0.52 12 0.36

[0171] The last 12 eigenvalues ​​( The values ​​are all ≤0.25, and their contribution to the total variance is minimal.

[0172] A4. Calculate the principal component contribution rate and screen the principal components with a cumulative contribution rate ≥ 90%.

[0173] The "contribution rate" of a principal component reflects the degree to which it retains the total information of the original feature set. It is calculated using eigenvalues, and the specific formula is as follows:

[0174] Contribution rate of a single principal component: the contribution rate of the m-th principal component. The proportion of the principal component's eigenvalue to the sum of all eigenvalues:

[0175] ;

[0176] Cumulative contribution rate: The cumulative contribution rate of the first m principal components The proportion of the sum of the first m eigenvalues ​​to the total sum of all eigenvalues:

[0177] .

[0178] Key calculation steps:

[0179] Calculate the sum of all eigenvalues: sum the 24 eigenvalues, denoted as . Based on the example data from step three, we calculate that S = 29.07.

[0180] Calculate the individual and cumulative contribution rates: Taking the first 12 feature values ​​from step three as an example, the calculation results are shown in the table below:

[0181] Principal component number Eigenvalues Individual contribution rate (%) Cumulative contribution rate (%) 1 8.52 29.31（8.52 / 29.07） 29.31 2 5.36 18.44（5.36 / 29.07） 47.75（29.31+18.44） 3 3.91 13.45（3.91 / 29.07） 61.20 4 2.87 9.87（2.87 / 29.07） 71.07 5 2.15 7.39（2.15 / 29.07） 78.46 6 1.78 6.12（1.78 / 29.07） 84.58 7 1.32 4.54（1.32 / 29.07） 89.12 8 1.02 3.61（1.05 / 29.07） 92.73 (≥90%, stop screening) ··· ··· ··· ···

[0182] Principal component screening results: The cumulative contribution rate of the first 8 principal components has reached 92.73% (≥90%). Therefore, the first 8 principal components are selected as the features after dimensionality reduction. The feature dimension is compressed from 24 dimensions to 8 dimensions, which meets the requirement of "10-15 dimensions of the feature set after dimensionality reduction" (in actual data, due to sample differences, the number of principal components selected is usually between 8 and 12).

[0183] A5. Generate the dimensionality-reduced feature set (principal component score matrix).

[0184] After selecting the first m principal components (m=8 in this example), the principal component scores are calculated by multiplying the standardized feature matrix by the first m eigenvectors, resulting in the dimensionality-reduced feature set.

[0185] Y is the feature set after dimensionality reduction (100×8 matrix, 100 sets of samples, 8 principal component features). It is a matrix (24×8 matrix, each column corresponds to an eigenvector of a principal component) formed by the first m eigenvectors.

[0186] The entire process revolves around standardization to eliminate dimensions → covariance matrix to measure correlation → eigenvalue / eigenvector extraction of principal components → cumulative contribution rate to select dimensions. Through quantitative calculation, the 24-dimensional initial extended feature set is compressed to 8-12 dimensions, which not only solves the problem of overfitting BPNN training caused by excessive feature dimensions, but also retains the key information required for defect diagnosis to the greatest extent, which is in line with the technical goal of eliminating redundant features and reducing data volume in this application.

[0187] The third stage involves backpropagation neural network optimization. A backpropagation neural network model is constructed, consisting of an input layer (dimension = principal component count), two hidden layers (20-25 nodes per layer), and an output layer (4 nodes, corresponding to four types of defects). 70% of the theoretical partial discharge signals from the defect database are used as the training set, 15% as the validation set, and 15% as the test set. Redundant features with sensitivity <0.05 are removed through sensitivity analysis, and 5-7 key feature vectors are output.

[0188] More specifically, a 3-layer BPNN model is constructed, using theoretical partial discharge signal features and defect type samples (≥500 groups) from the defect database as the training set, and the principal component feature set is selected based on feature importance:

[0189] Model training: The number of input layer nodes equals the principal component feature dimension; the number of hidden layer nodes is determined by 2 × the number of input layer nodes + 1; the output layer contains defect type labels (metal protrusions, suspended metal objects, etc.); the training objective is for the model to achieve a recognition accuracy of ≥92%.

[0190] The number of hidden layer nodes is set in this way based on the dual goals of "model performance balance" and "engineering practicality" of BPNN (backpropagation neural network) in the partial discharge defect diagnosis scenario. This avoids "overfitting / underfitting" while ensuring that the model can accurately learn the mapping relationship between multi-source features and defect types.

[0191] Training phase: Using defect type labels to guide feature learning:

[0192] The training samples for BPNN come from a defect baseline database, formatted as principal component feature sets (e.g., 8-dimensional features after PCA dimensionality reduction) + defect type labels (e.g., metal protrusions, metal suspensions). It should be noted that the basic and deep features of theoretical partial discharge signals (including transient ground voltage signals, ultrasonic signals, and UHF signals, along with corresponding defect type labels and corresponding signal propagation velocity calibration data) contained in the defect baseline database can be extracted, and the principal component feature set of the theoretical partial discharge signal can be found based on the principal component feature set of the actual partial discharge signal.

[0193] During training, BPNN adjusts the weights from the input layer (dimensionality reduction features) to the hidden layer and from the hidden layer to the output layer through the backpropagation algorithm. The goal is to make the defect type prediction results of the output layer consistent with the true labels of the samples (i.e., recognition accuracy ≥ 92%).

[0194] The format of BPNN training samples is principal component feature set + defect type label (e.g., TEV average peak factor 3.3 + ... + defect type = metal suspension). The number of neurons in the output layer must correspond to the number of defect types (e.g., 4 output neurons for 4 types of defects) in order to adjust the weights by whether the defect type is accurately predicted.

[0195] Feature selection: The contribution of each feature to defect identification is calculated by the model weight coefficient. Features with a contribution of ≥80% (usually 6-10) are retained to form a key feature candidate set.

[0196] Screening phase: Weights reflect the contribution of features to defect identification.

[0197] After training, the weights from the input layer to the hidden layer directly reflect the contribution of each principal component feature to defect type identification. The larger the absolute value of the weight, the more effectively the principal component feature can distinguish defect types (e.g., a high weight for the TEV average peak factor indicates that it is crucial for identifying metal protrusions).

[0198] After the BPNN is trained, it does not directly output that a sample belongs to a metal suspension. Instead, it analyzes the weight relationship between input features and output labels to select features that contribute ≥80% to defect type identification (such as the average peak factor of TEV and the proportion of ultrasonic spectrum energy in the example).

[0199] After the BPNN is trained, the absolute value of the weight coefficients from the input layer to the hidden layer directly reflects the contribution of that feature to defect recognition. Specific calculation logic and examples are as follows:

[0200] Assuming a third-level setting:

[0201] Input layer node count = 10 (corresponding to 10-dimensional reduced features);

[0202] Number of hidden layer nodes = 2 × 10 + 1 = 21 (determined by "2 × number of input layer nodes + 1");

[0203] Output layer node count = 4 (corresponding to four types of defects: metal protrusions, metal suspensions, air gaps, and internal discharge of insulation);

[0204] Training set: 500 samples of "theoretical partial discharge signal characteristics - defect type" in the defect basic database (e.g., "TEV principal component 1=1.2+...+label=metal protrusion"). The training objective is to achieve an accuracy of ≥92% (the actual accuracy reached 93.5% after training).

[0205] From the trained BPNN model, extract the weight matrix of "10 features in the input layer → 21 nodes in the hidden layer" (a total of 10 × 21 = 210 weight values). Take the absolute value of the 21 weight values ​​of each feature and then calculate the average value to obtain the "feature average weight". Then, standardize it to the contribution of "0-100%" (standardization formula: contribution of a feature = average weight of the feature ÷ maximum average weight of all features × 100%).

[0206] High contribution features (≥80%): These all correspond to the basic features that are theoretically key to defect identification (such as TEV principal component 1 containing the TEV average peak factor, and the fourth level mentioning that when the TEV average peak factor is ≥3.2, it supports metal protrusions; UHF principal component 1 contains UHF phase entropy, which is theoretically the core feature for distinguishing metal suspensions).

[0207] Low contribution features (<80%): mostly redundant or interfering features (such as TEV principal component 2 containing the average amplitude of TEV pulses, which is mentioned in the second level as being highly correlated with the average amplitude of UHF pulses, and is a redundant feature; ultrasonic principal component 3 containing the duration of ultrasonic pulses, which theoretically has low discrimination ability for defect types).

[0208] Critical feature (UHF principal component 3, 78.16%): Because the standard deviation of the UHF negative half-cycle phase distribution is unstable in some defect types (such as air gaps), the contribution is slightly lower than the threshold and is not included in the candidate set for the time being.

[0209] The BPNN model does not directly output defect type judgment, but has the following effects:

[0210] 1. Reduce redundant feature interference and improve the mapping purity between features and defects:

[0211] During BPNN training, the original dimensionality-reduced feature set (obtained based on theoretical partial discharge data) may contain redundant features unrelated to the defect type (e.g., as mentioned in Level 2, "the average amplitude of TEV pulses is highly correlated with the average amplitude of UHF pulses; retain the UHF feature with a higher variance contribution rate"). Without filtering, redundant features will introduce interference during weight calculation, causing the model to misjudge the association between features and defects. For example, if both highly correlated TEV and UHF pulse average amplitudes are retained, their weights may cancel each other out or become confused, making it impossible for the BPNN to accurately determine whether a change in a set of feature values ​​is caused by a metal suspension defect or by feature redundancy. Filtering features with a contribution rate ≥80% ensures that the model only focuses on features truly useful for defect identification (e.g., in the Level 3 example, the average peak factor of TEV contributes significantly to the identification of metal protrusions), greatly improving the purity of the feature-defect mapping, reducing the probability of misjudgment, and directly improving the model's prediction accuracy (training objective: identification accuracy ≥92%).

[0212] Strengthen the causal logic between features and defects to improve the model's generalization ability:

[0213] The essence of selecting key features is to enable the BPNN to learn the causal logic of "feature value change → defect type change". For example, when it is found that the proportion of ultrasonic spectrum energy is ≥65%, it is highly correlated with metal suspension defects (contribution ≥80%). Based on this causal relationship, the BPNN can quickly associate the feature value of "68% of spectrum energy" in a new sample with metal suspension defects, rather than relying on a large number of feature-label memories. This reinforcement of causal logic enables the model to accurately determine the defect type based on causal relationships when faced with new feature combinations not covered by the training samples (such as small fluctuations in feature values ​​in field measurements), improving the model's generalization ability, avoiding overfitting (the third level verifies the generalization ability through test samples), and ensuring stable diagnostic accuracy in complex field environments.

[0214] Validation and optimization: Use 15% of the test samples in the database to verify the recognition accuracy of the key feature candidate set. If the accuracy is <90%, backtrack to the high-discrimination features (such as the standard deviation of the UHF negative half-cycle phase distribution) marked in the first-level supplementary defect base database, re-filter and output the key feature set, which consists of several key feature vectors.

[0215] Level 4: Dynamic BPA construction. The initial BPA value is set based on the historical identification accuracy of various detection technologies for different defects in the defect database (e.g., the initial BPA value of UHF for metal protrusions = historical accuracy / 100). It is then dynamically adjusted in combination with real-time interference level (low interference +0.05, medium interference 0, high interference -0.12) and sensor working status (normal +0.03, abnormal -0.1). The final BPA value is controlled within the range of 0.5-0.95.

[0216] More specifically, for each key feature vector, an initial BPA value is assigned according to the feature value-defect type matching degree in the defect base database (e.g., when the TEV average peak factor is ≥3.2, the initial BPA value supporting the metal protrusion defect is 0.6); a feature reliability coefficient is introduced (calculated based on the signal-to-noise ratio of the same batch of signals, the coefficient is 1.0 when SNR≥4, and the coefficient is 0.8 when SNR=3-4 (excluding)) to correct the initial BPA value (corrected BPA value = initial value × reliability coefficient); the final key feature vector and dynamic BPA value are output in the format of key feature set + dynamic BPA value, providing input for subsequent improvement of DS evidence theory fusion.

[0217] The BPA score is used to determine the level of trust in a defect type by analyzing key feature vectors. For example, the BPA value of TEV principal component 1 (eigenvalue 1.32) corresponding to the TEV signal represents the degree to which the defect corresponding to this feature is a metal protrusion / metal suspension / air gap / internal insulation discharge. For instance, if the BPA values ​​of TEV principal component 1 = 1.32 are allocated as follows: metal protrusion = 0.6, metal suspension = 0.1, air gap = 0.2, and internal insulation discharge = 0.1, it means that this feature supports the defect being a metal protrusion 60% (because TEV principal component 1 is composed of the TEV average peak factor 3.5mV, etc., and 3.5mV conforms to the characteristic pattern of metal protrusion defects); only 10% supports metal suspension, 20% supports air gap, and 10% supports internal insulation discharge (this feature has a low matching degree for these three types of defects); for the same key feature vector, the sum of the BPA values ​​of all defect types must be 1 (0.6 + 0.1 + 0.2 + 0.1 = 1), ensuring the rationality of the confidence allocation.

[0218] The key feature vector is a set of highly discriminative features filtered by BPNN. It is output in a structured manner according to feature name-feature value-feature type. Each feature is labeled with the corresponding signal type and physical meaning to ensure that the source of evidence can be clearly identified during DS fusion.

[0219] For example, a key feature vector can be represented as:

[0220] TEV principal component 1 (feature name) - 1.32 (eigenvalue) - TEV signal (generated by a linear combination of TEV average peak factor (3.5mV) and TEV peak factor standard deviation (0.8), reflecting the degree of electric field distortion of the metal protrusion defect. This is the feature type, including signal type and physical meaning).

[0221] This step employs a four-stage processing flow: multi-dimensional feature extraction, principal component analysis (PCA), backpropagation neural network (BPNN) optimization, and dynamic basic probability allocation (BPA). Based on the feature definitions and training samples (theoretical supporting signals) in the defect database, the feature set of the multi-source partial discharge signals (actual diagnostic signals) collected in step 2 is optimized (in the third stage, after training the backpropagation neural network with training samples from the defect database, the trained backpropagation neural network is used to filter the actual partial discharge signals), redundant features (sensitivity < 0.05) are removed, and key feature vectors and dynamic BPA values ​​are output (based on historical accuracy of the database and real-time operating condition adjustments).

[0222] Step 4, Defect Identification and Localization via Multi-Source Information Fusion: Using the key feature vector and dynamic BPA value output in Step 3 as evidence sources, and referencing the defect type labels in the defect database (based on theoretically supported signal annotations), the actual defect type is determined by fusing information from multiple detection technologies through improved DS evidence theory. Simultaneously, using the coordinates of the distributed array formed by the sensors, the time difference of arrival (TDOA) of the signal, and the signal propagation speed calibration data as inputs, a three-dimensional time difference localization model is constructed and solved. The specific coordinates of the defect inside the switchgear are determined using particle swarm optimization-least squares method, with the localization error controlled within ±3cm.

[0223] The specific process is as follows:

[0224] Step 4.1, Evidence Source Identification and Preprocessing:

[0225] The key feature vectors are split into three independent sources of evidence based on signal type:

[0226] Source of Evidence 1 (TEV Evidence Source): Includes two key feature vectors: the mean peak factor of TEV and the standard deviation of the peak factor of TEV, and the corresponding dynamic BPA value;

[0227] Evidence Source 2 (Ultrasonic Evidence Source): Includes two key feature vectors: ultrasonic spectrum energy ratio and ultrasonic pulse rise time, as well as the corresponding dynamic BPA value;

[0228] Evidence Source 3 (UHF Evidence Source): Includes four key feature vectors: UHF positive half-cycle pulse count ratio, UHF phase entropy, UHF amplitude variation coefficient, UHF total pulse count, and UHF pulse count change rate, as well as the corresponding dynamic BPA value.

[0229] The dynamic BPA value of each evidence source is normalized to ensure that the sum of the BPA values ​​of the four defect types in each evidence source is equal to 1, thus eliminating the interference of dimensional differences on the fusion results.

[0230] Step 4.2, Calculation of the coefficient of conflict of evidence:

[0231] Based on the theoretical partial discharge signal BPA value distribution (pre-set, i.e., the initial BPA mentioned earlier) in the defect database, the conflict coefficient k between the three independent evidence sources is calculated using the following formula:

[0232] ;

[0233] This indicates that the defect type A pointed to by evidence source 1, the defect type B pointed to by evidence source 2, and the defect type C pointed to by evidence source 3 have no intersection. If there is no overlap (i.e., the points point to different defect types), then it is counted as a conflict item.

[0234] , , These are the normalized dynamic BPA values ​​of evidence source 1 (TEV evidence source) for defect type A, reflecting the confidence level of TEV signal characteristics (such as TEV average peak factor) in supporting defect A. The larger the value, the higher the support. The normalized dynamic BPA value of evidence source 2 (ultrasonic evidence source) for defect type B reflects the confidence level that the ultrasonic signal characteristics (such as the proportion of ultrasonic spectrum energy) support defect B. ∑ represents the normalized dynamic BPA value of evidence source 3 (UHF evidence source) for defect type C, reflecting the confidence level that UHF signal characteristics (such as the proportion of UHF positive half-cycle pulse counts) support defect C, and ∑ is the summation of all combinations that satisfy "A∩B∩C=∅".

[0235] It should be noted that defects A, B, and C here all belong to any one of the defect type labels and satisfy the following conditions: .

[0236] Specifically, A: Evidence Source 1 (TEV Evidence Source) is the most likely defect type selected from four typical defects based on its own key feature vector (such as TEV average peak factor) (such as TEV average peak factor 3.3 → pointing to A = metal protrusion).

[0237] B: Evidence Source 2 (ultrasound evidence source) is selected from four typical defect types based on its own key characteristics (such as the proportion of ultrasonic spectrum energy of 69%). The most likely defect type is "high proportion of spectrum energy → pointing to B = metal suspension".

[0238] C: Evidence Source 3 (UHF Evidence Source) is the most likely defect type selected from four typical defects based on its own key characteristics (such as UHF phase entropy 1.3) (such as high phase entropy → pointing to C = metal suspension).

[0239] Divide the conflict level according to the value of k: when k ≤ 0.3, it is a low conflict; when 0.3 < k ≤ 0.6, it is a medium conflict; when k > 0.6, it is a high conflict.

[0240] Step 4.3. Execute the improved D-S fusion rule:

[0241] Adopt a differentiated fusion strategy for different conflict levels:

[0242] Low conflict (k ≤ 0.3): Directly adopt the traditional D-S synthesis rule, and the formula is:

[0243] ;

[0244] where X is the type of defect pointed to after fusion, and m(X) is the comprehensive BPA value of X after fusion;

[0245] Medium conflict (0.3 < k ≤ 0.6): Discount and correct the BPA values of the conflicting evidence sources. For example, reduce the weight of the conflicting evidence sources by 0.1, and proportionally increase the weights of the non-conflicting evidence sources, and then execute the traditional D-S synthesis rule after correction.

[0246] First, clarify the judgment criteria for the conflicting evidence sources.

[0247] Before performing the discount correction, it is necessary to first determine "which evidence source(s) is / are the conflicting evidence sources". The core judgment logic is that "the distribution of the BPA value of this evidence source has the greatest divergence from other evidence sources", which can be specifically judged through two dimensions:

[0248] 1. Based on the maximum pointing difference of the BPA values of the evidence sources:

[0249] Compare the types of defects with the largest BPA values (i.e., the types most supported by each evidence source) of the three evidence sources (TEV, ultrasonic wave, UHF):

[0250] If the maximum pointings of two evidence sources are the same (such as TEV and UHF both most support metal protrusions), and the maximum pointing of the third evidence source is completely different (such as ultrasonic wave most supports "internal insulation discharge"), then the third evidence source (ultrasonic wave) is the conflicting evidence source;

[0251] [[ID= thirty-seven]]2. Based on the contribution ratio of the conflict coefficient:

[0252] [[]]Further verification: Calculate the contribution ratio of each evidence source to the conflict coefficient k (i.e., the degree of conflict between this evidence source and the other two evidence sources), and the one with the largest contribution ratio is the conflicting evidence source:

[0253] Formula reference: The conflict contribution ratio of evidence source i = (the conflict value between evidence source i and j + the conflict value between evidence source i and k) / the total conflict coefficient k;

[0254] Using the previous example: if the total conflict coefficient k = 0.4 (medium conflict), where the conflict contribution of ultrasound and TEV is 0.25, the conflict contribution of ultrasound and UHF is 0.28, and the conflict contribution of TEV and UHF is only 0.07, then the conflict contribution ratio of ultrasound = (0.25 + 0.28) / 0.4 = 1.325 (dominant), further confirming ultrasound as the source of conflict evidence.

[0255] After identifying conflicting sources of evidence, the weight of conflicting sources is reduced by 0.1, while the weight of non-conflicting sources is increased proportionally. The core principle is to multiply all BPA values ​​of conflicting sources by a discount factor and the weight of non-conflicting sources by an increase factor. After the adjustment, the sum of the BPA values ​​of each source remains 1 (maintaining normalization). Specific steps are as follows:

[0256] For example, the default evidence source weights are: TEV (ω1=0.3), ultrasound (ω2=0.3), and UHF (ω3=0.4). The discount factor for the BPA value corresponding to the correction rule is the original weight - 0.1 (conflicting evidence source), and the enhancement factor for non-conflicting evidence source is the original weight + Δω (the reduced weight is allocated proportionally).

[0257] High conflict (k>0.6): A weighted average fusion rule is adopted, with the following formula:

[0258] ;

[0259] in, , , The weights of TEV, ultrasonic, and UHF evidence sources are determined based on the defect identification accuracy settings for each signal type in the defect database, with a default value. , , .

[0260] Step 4.4: After fusion is complete, output the results based on the "judgment threshold standard" of the defect base database:

[0261] The defect type with the largest overall BPA value is selected as the candidate type, and this BPA value must be "greater than 1.2 times the second largest BPA value" to avoid type confusion;

[0262] If the comprehensive BPA value of a candidate type is ≥0.5 (the database preset reliable judgment threshold), it is determined to be the actual defect type; if it is <0.5, it is backtracked to the third level of the four-level preprocessing process, and high-discrimination features such as "standard deviation of UHF negative half-cycle phase distribution" are added for re-optimization until the judgment conditions are met.

[0263] Step 4.5: Acquisition and calibration of positioning input parameters:

[0264] Sensor coordinates: Obtain the three-dimensional coordinates of 3 or more sensors in the distributed array (establish an x−y−z coordinate system with the lower left corner of the switch cabinet as the origin, unit: cm), the coordinates of the i-th sensor are: , ;

[0265] TDOA data: The arrival time of partial discharge signals at each sensor is acquired through a high-precision time synchronization module (synchronization accuracy ≤1ns). Calculate the time difference between any two sensors. ( );

[0266] Propagation speed calibration: Based on the signal propagation speed model under different gas media (air, SF6), temperature and humidity in the defect database, combined with the field environmental parameters (or operating parameters) (such as temperature 26℃, SF6 gas pressure 0.6MPa), the actual propagation speed v (unit: m / μs, default v=0.343 under air medium) is calibrated.

[0267] Step 4.6, Construction of 3D Time Difference Localization Model: Let the defect coordinates be (x, y, z). Based on "signal propagation distance difference = velocity × time difference", construct the localization equation set:

[0268] ;

[0269] The nonlinear equations are linearized (Taylor expansion to first-order terms), transforming them into matrix form: Where: A is the coefficient matrix (dimensions 1-2) , (The number of pairs of sensors), with elements being , These are the initial iteration coordinates; B is the defect coordinate correction value; B is a constant term matrix with elements of 1. .

[0270] Step 4.7: Particle Swarm Optimization - Least Squares Solution. A combined algorithm of "Particle Swarm Optimization (PSO) initialization + least squares iteration" is used to solve the model, avoiding the traditional least squares method from getting trapped in local optima.

[0271] Step 1 (PSO Initialization): Use the internal space of the switch cabinet as the search domain (e.g., ... cm cm (cm), randomly generate N particles (N=50), each particle corresponding to a set of candidate initial coordinates. Calculate the fitness value (positioning error) of each particle. (i.e., the square root of the sum of squared residuals), selecting the top 10% of particles with the best fitness as the initial iteration coordinate set;

[0272] Step 2 (Least Squares Iteration): For each optimal initial coordinate, solve using the weighted least squares method. (W is the weight matrix, set based on TDOA measurement accuracy; higher accuracy results in greater weights), update defect coordinates. ;

[0273] Step 3 (Convergence Verification): Repeat the iteration until the localization error is reached. (Database preset positioning accuracy threshold), output the final defect coordinates;

[0274] Step 4.8: Verify and correct the localization results, and verify the rationality of the localization by combining key feature vectors:

[0275] The final output is a structured result of the actual defect type and its three-dimensional coordinates, in the following format:

[0276] Actual defect type: [Metal protrusion / Metal suspension / Air gap / Internal discharge in insulation] (Comprehensive BPA value: XXX); 3D coordinates of the defect:

[0277] (x,y,z)=(XX.Xcm,YY.Ycm,ZZ.Zcm) (Positioning error: XX.Xcm);

[0278] Matching criteria: [Key feature vector matching degree, consistency description of positioning results and sensor signal distribution], providing direct basis for subsequent maintenance.

[0279] Step 5, Defect Development Trend Prediction: Based on the key feature vectors of the actual diagnostic signals output in Step 3, a Long Short-Term Memory (LSTM) time-series prediction model is constructed using the full-cycle time-series data of defect development (continuous time series of theoretical supporting signals, covering the mild-moderate-severe stages) pre-collected in the defect database as training samples. This model learns the correlation between the feature change rate and defect deterioration. The key feature vectors of the actual diagnostic signals are then input into the trained LSTM model to predict the feature change rate of the actual defects in the next 24-72 hours. Combined with the risk level standards defined in the defect database (e.g., a change rate > 15% / 24h is considered severe risk), the risk level and differentiated maintenance recommendations (e.g., severe risk requires immediate shutdown) are output.

[0280] In summary, this invention first collects multi-source partial discharge signals and constructs a defect database. Then, it performs multi-source feature generation, principal component dimensionality reduction, and BPNN key feature screening on the signals. Next, it constructs dynamic BPA values ​​based on the key features. Subsequently, it identifies defect types by fusing multi-source information through improved DS evidence theory and locates defects by combining TDOA and optimization algorithms. Finally, it uses an LSTM model to predict the development trend of defects.

[0281] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation, characterized in that, The method comprises the following steps: Collecting basic data of theoretical partial discharge signals and corresponding working condition parameters based on typical insulation defects at different stages to construct a defect basic database, wherein the theoretical partial discharge signals include transient ground voltage signals, ultrasonic signals and ultra-high frequency signals, and corresponding defect type labels and corresponding signal propagation speed calibration data; Based on real-time interference monitoring results, referring to the working condition parameters in the defect basic database and the signal effectiveness standard, dynamically adjusting the sensor parameters and signal acquisition strategy, and based on the adjusted sensor parameters and signal acquisition strategy, obtaining the actual partial discharge signal of the switch cabinet to be diagnosed, thereby obtaining a multi-source signal feature set, the actual partial discharge signal includes transient ground voltage signal, ultrasonic signal and ultra-high frequency signal; At the same time, at least M TEV and ultrasonic sensors are arranged on each surface of the switch cabinet shell to form a distributed array, and the actual partial discharge signal arrival time stamp of each sensor is collected through a clock synchronization module with higher precision than the first threshold, and the time difference TDOA of the actual partial discharge signal arrival is calculated; Based on the preset four-stage processing flow, the multi-source signal feature set is optimized to obtain a key feature vector and a dynamic BPA value; The key feature vector and dynamic BPA value are used as evidence sources, and the defect type labels in the defect basic database are used as references to fuse multi-detection technology information by improving the D-S evidence theory to determine the actual defect type of the actual partial discharge signal in the multi-source signal feature set; At the same time, taking the coordinates of the distributed array formed by the sensors, the signal time difference TDOA and the signal propagation speed calibration data as inputs, a three-dimensional time difference positioning model is constructed and solved, and the specific coordinates of the defect in the switch cabinet are determined by particle swarm optimization-least squares method.

2. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation according to claim 1, characterized in that, The typical insulation defects include insulation defects based on metal protrusions, insulation defects based on air gaps, insulation defects based on metal suspensions and insulation defects based on internal insulation discharge; Each stage includes a slight stage, a medium stage and a severe stage, wherein the slight stage is a theoretical partial discharge signal pulse average amplitude <100mV, the medium stage is a theoretical partial discharge signal pulse average amplitude 100-300mV, and the severe stage is a theoretical partial discharge signal pulse average amplitude >300mV; The working condition parameters include environmental temperature, environmental humidity, switch cabinet operating load and voltage level.

3. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation of claim 1, characterized in that, The defect type label adopts structured labeling, and the format is defect type-key parameter-severity level; The signal propagation speed calibration data is classified and stored according to signal type-propagation medium; The defect development full cycle timing data is composed of the pulse average amplitude of the theoretical partial discharge signal collected at a preset time interval, and the time length of a single defect development full cycle timing data is not less than 72h, and at least three complete characteristic change periods are included.

4. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation of claim 1, characterized in that, Based on the real-time interference monitoring results, the working condition parameters in the defect database and the signal validity standard are referred to, the sensor parameters and the signal acquisition strategy are dynamically adjusted, the actual partial discharge signals of the switch cabinet to be diagnosed are collected by using the transient ground voltage, ultrasonic wave and ultra-high frequency three technologies, so that a multi-source signal feature set is obtained, the actual partial discharge signals include transient ground voltage signals, ultrasonic wave signals and ultra-high frequency signals; meanwhile, at least M TEV sensors and ultrasonic wave sensors are arranged on each surface of the switch cabinet shell to form a distributed array, the signal arrival time stamps of each sensor are collected by a clock synchronization module with a high precision higher than a first threshold value, and the specific process of calculating the signal arrival time difference TDOA is as follows: A plurality of interference monitoring sensors are arranged within a preset range from the switch cabinet to be diagnosed, the interference detection sensors include electromagnetic interference sensors and mechanical vibration sensors; The interference data are collected by the interference monitoring sensors at a preset time interval, the average value of the interference data is taken after continuous collection for As, the interference intensity I is calculated according to a set formula, and the interference level of the interference intensity I is divided with reference to the defect database, wherein the interference data include electromagnetic interference amplitude and mechanical interference amplitude, the interference intensity I = electromagnetic interference amplitude E × 0.65 + mechanical vibration amplitude V × 0.35; A mapping table based on the interference level-optimal parameters is constructed by using the theoretical partial discharge signals, the optimal parameters include sensor working parameters and data acquisition parameters, the mapping table is taken as a reference, the combination scheme of the sensor parameters and the acquisition strategy adapting to the current interference level and working condition is determined in combination with the current working condition parameters; Based on the combination scheme, a plurality of TEV sensors and ultrasonic wave sensors are arranged on the switch cabinet to be diagnosed, the actual partial discharge signals are collected by the TEV sensors and the ultrasonic wave sensors, and the original pulse waveforms of the actual partial discharge signals are stored in an original pulse waveform database; and when the actual partial discharge signals are collected by the TEV sensors and the ultrasonic wave sensors, the time stamps are recorded; For the same discharge event, the time difference of any two sensors of the same type is calculated based on the time stamps, the time difference is bound with the corresponding sensor coordinates and stored; The actual partial discharge signals are preprocessed and features are extracted, so that a multi-source signal feature set is obtained.

5. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation according to claim 4, characterized in that, The basic features in the multi-source signal feature set include the pulse average amplitude and the pulse maximum amplitude of the transient ground voltage signals, the pulse duration, the pulse amplitude standard deviation, the center frequency and the frequency bandwidth of the ultrasonic wave signals, and the UHF pulse average amplitude, the UHF pulse amplitude variance, the pulse proportion exceeding the threshold value, the positive half-cycle pulse count proportion, the negative half-cycle pulse count proportion, the positive half-cycle phase distribution standard deviation, the negative half-cycle phase distribution standard deviation, the phase interval corresponding to the maximum pulse amplitude, the total pulse count, the positive half-cycle pulse count, the negative half-cycle pulse count and the pulse count change rate with time of the ultra-high frequency signals.

6. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation according to claim 1, characterized in that, The preset four-level processing procedure is as follows: The first stage is to extract multi-dimensional features, including calling the transient earth voltage signals, ultrasonic signals and ultra-high frequency signals of the same batch in the original pulse waveform database through the batch index based on the multi-source signal feature set, extracting the deep features of the transient earth voltage signals, ultrasonic signals and ultra-high frequency signals, and merging the deep features into the multi-source signal feature set to obtain an initial extended feature set; The second stage is to perform principal component dimension reduction, taking the theoretical partial discharge signals in the defect database as a reference, performing principal component analysis on the initial extended feature set, calculating the covariance matrix of the initial extended feature set, solving the eigenvalues and eigenvectors, and selecting the principal components with a cumulative contribution rate of ≥90% to reduce the feature dimension of the initial extended feature set and obtain a principal component feature set; The third stage is to optimize the back propagation neural network, build a back propagation neural network model, train the back propagation neural network model by taking 70% of the theoretical partial discharge signals in the defect database as a training set, 15% as a validation set and 15% as a test set, remove redundant features with a sensitivity of <0.05 through sensitivity analysis, and output 5-7 key feature vectors to obtain a key feature set. The back propagation neural network model includes an input layer, two hidden layers and an output layer. The node dimension of the input layer is the same as the dimension of the features in the reduced feature set. The output layer outputs a defect type label. The corresponding key feature vector is selected according to the defect type label output by the output layer. The fourth stage is to build a dynamic BPA. The initial BPA value is set based on the defect database, and the BPA value is assigned to each key feature. The initial BPA value is modified by introducing a feature reliability coefficient. The modified BPA value = initial BPA value × reliability coefficient.

7. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation of claim 6, characterized in that, The deep features include the average peak factor and peak factor standard deviation of the transient earth voltage signals, the spectral energy proportion and pulse rise time of the ultrasonic signals, and the phase entropy H and amplitude variation coefficient of the ultra-high frequency signals.

8. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation of claim 6, characterized in that, The key feature vectors and dynamic BPA values are used as evidence sources, and the defect type labels in the defect database are used as a reference to improve the D-S evidence theory to fuse multiple detection technology information, and the specific process of determining the actual defect type of the actual partial discharge signals in the multi-source signal feature set is as follows: The key feature vectors are divided into three independent evidence sources according to the signal type, which are TEV evidence source, ultrasonic evidence source and UHF evidence source. The dynamic BPA values corresponding to each independent evidence source are normalized. The conflict coefficient k between the three independent evidence sources is calculated, and the conflict level is divided according to the value of the conflict coefficient k. The formula is as follows: ； wherein, represents the defect type A pointed by TEV evidence source, the defect type B pointed by ultrasonic evidence source, the defect type C pointed by UHF evidence source have no intersection, , , respectively are the normalized dynamic BPA value of TEV evidence source to defect type A, reflecting the confidence degree that TEV signal feature supports the defect to be A, the greater the value, the higher the support degree; represents the normalized dynamic BPA value of ultrasonic evidence source to defect type B, reflecting the confidence degree that ultrasonic signal feature supports the defect to be B; represents the normalized dynamic BPA value of UHF evidence source to defect type C, reflecting the confidence degree that UHF signal feature supports the defect to be C, and ∑ is the summation of all combinations satisfying A∩B∩C=∅. Different fusion strategies are adopted for different conflict levels to fuse the BPA values and obtain a comprehensive BPA value. Select the defect type with the largest comprehensive BPA value as a candidate type, if the comprehensive BPA value of the candidate type is greater than or equal to a determination threshold, the candidate type is determined as the actual defect type, if the comprehensive BPA value of the candidate type is less than the determination threshold, the back propagation neural network is re-optimized until the determination condition is met.

9. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation of claim 8, characterized in that, The differentiated fusion strategy is specifically: If the conflict level is low conflict, the traditional D-S synthesis rule is directly used to synthesize the normalized dynamic BPA values of the three independent evidence sources to obtain a comprehensive BPA value; If the conflict level is medium conflict, the discounted modification is performed on the normalized dynamic BPA values of the conflicting independent evidence sources based on a modification coefficient, and the normalized dynamic BPA values of the non-conflicting independent evidence sources are proportionally increased by the modification coefficient; finally, the modified normalized dynamic BPA values of the three independent evidence sources are executed to obtain a comprehensive BPA value by using the traditional D-S synthesis rule; If the conflict level is high conflict, the weighted average fusion rule is used to synthesize the normalized dynamic BPA values of the three independent evidence sources to obtain a comprehensive BPA value.

10. The switch cabinet partial discharge diagnosis method based on multi-information fusion and dynamic intelligent regulation according to claim 1, characterized in that, The three-dimensional time difference positioning model is constructed and solved by taking the coordinates of the distributed array formed by the sensors, the signal time difference of arrival (TDOA) and the signal propagation speed calibration data as inputs, and the specific process of determining the specific coordinates of the defect inside the switch cabinet by using the particle swarm optimization-least squares method is as follows: obtaining three-dimensional coordinates of the sensors in the distributed array, denoted as ; The clock synchronization module collects the time of arrival of the partial discharge signal at each sensor and calculates the time difference between any two sensors ; Determine the signal propagation speed v based on the current working condition parameters; constructing a three-dimensional time-difference positioning model based on the time difference, the propagation speed and the defect coordinates, the three-dimensional time-difference positioning model being: ; Where (x, y, z) is the defect coordinate; Solve the three-dimensional time difference positioning model by using the particle swarm optimization-least squares method, thereby obtaining the defect coordinate.

Images