Multi-parameter fusion non-intrusive partial discharge on-line monitoring method and terminal

Abstract

The invention discloses a multi-parameter fusion non-intrusive partial discharge on-line monitoring method and system. Six types of sensors are arranged on the outer wall of each switch cabinet in a non-intrusive mode; according to six-channel partial discharge pulse data corresponding to the switch cabinet, establishing a two-dimensional phase difference confidence ellipse, and reserving pulse data falling into the two-dimensional phase difference confidence ellipse as a homologous event; dividing a set interval in the Noise channel into a plurality of sub-bands; constructing a temperature and humidity-noise gain matrix in each sub-band in real time, and performing reverse compensation on the frequency spectrum in the homologous event; constructing a six-dimensional vector for the compensated frequency spectrum, sending the six-dimensional vector to a Light CNN (Convolutional Neural Network) in the edge device, and outputting an initial discharge probability; correcting the initial discharge probability to obtain a final discharge probability; and based on the final discharge probability, carrying out risk grading on the partial discharge event in combination with edge calculation, and giving an alarm according to a risk grading result. According to the method, the partial discharge identification accuracy can be remarkably improved, and all-weather and full-life-cycle monitoring in a real sense is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electrical partial discharge monitoring, and particularly relates to a multi-parameter fusion non-intrusive partial discharge online monitoring method and terminal. BACKGROUND

[0002] The high-voltage switch cabinet is one of the electrical devices with the largest quantity and the most widely application in the power system, and its operation reliability is directly related to the safety and stability of the power grid. However, due to the differences in design, manufacturing, installation and operation and maintenance, insulation faults of the switch cabinet occur frequently, especially in the voltage grade of 6kV and above, the insulation accident rate is high, and often has the spreading nature of "fire burning along the line", causing large-area power outage and equipment damage.

[0003] Partial discharge (PD) is an important early feature of insulation deterioration, and is a necessary stage before insulation failure occurs. Therefore, online monitoring of partial discharge of the switch cabinet can realize early warning of the insulation state, and is a key technical means for condition-based maintenance and intelligent power grid construction.

[0004] At present, the partial discharge detection of the switch cabinet is mainly based on manual inspection and periodic power-off test, and there are problems such as monitoring blind area, poor real-time performance, high labor intensity, high misjudgment rate and the like, which cannot meet the needs of real-time sensing of the equipment state and intelligent operation and maintenance of the modern power grid. SUMMARY

[0005] In order to solve the problems in the prior art, the application provides a multi-parameter fusion non-intrusive partial discharge online monitoring method and system, six types of sensors are arranged non-intrusively on the outer wall of each switch cabinet; a two-dimensional phase difference confidence ellipse is established according to six-channel partial discharge pulse data corresponding to the switch cabinet, and pulse data falling into the two-dimensional phase difference confidence ellipse is reserved as a homologous event; a set interval in the Noise channel is divided into a plurality of sub-bands; a temperature and humidity-noise gain matrix is constructed in each sub-band, and inverse compensation is performed on the spectrum in the homologous event; a six-dimensional vector is constructed for the compensated spectrum and sent to a Lightweight CNN in an edge device to output an initial discharge probability; the initial discharge probability is corrected to obtain a final discharge probability; based on the final discharge probability, the edge computing is combined to classify the partial discharge event, and an alarm is given according to the risk classification result. The application can significantly improve the partial discharge identification accuracy and realize all-weather and full-life-cycle monitoring in a true sense.

[0006] The application adopts the following technical solutions.

[0007] The application provides a multi-parameter fusion non-intrusive partial discharge online monitoring method, which comprises: Six types of sensors, i.e., ultrasonic AE, ground electric wave TEV, ultra-high frequency UHF, environmental noise Noise, temperature T and humidity H, are arranged on the outer wall of the switch cabinet; the rising edge of the UHF signal is taken as a trigger source, and six-channel partial discharge pulse data corresponding to the switch cabinet are synchronously collected; ring-shaped time stamps are stamped on the collected partial discharge pulse data; According to the six-channel partial discharge pulse data corresponding to the switch cabinet, an AE-TEV two-dimensional phase difference confidence ellipse and a TEV-UHF two-dimensional phase difference confidence ellipse are established, and through retaining pulse data falling into the two-dimensional phase difference confidence ellipses as homologous events, first empty-phase filtering is performed; A set interval in the Noise channel is divided into a plurality of 1 / 3 octave subbands; six-channel partial discharge pulse data in the homologous events are used to construct a temperature-humidity-noise gain matrix in each subband in real time; the AE / TEV / UHF spectrum in the homologous events is compensated in reverse by using the temperature-humidity-noise gain matrix; A six-dimensional vector is constructed from the compensated ultrasonic data AE, ground electric wave data TEV, ultra-high frequency data UHF, change amount ΔN of the Noise channel, temperature T and humidity H, and is input into a pre-constructed and trained switch partial discharge prediction model configured in an edge device to output an initial discharge probability P; the initial discharge probability P is corrected by introducing a sliding window Bayesian to obtain a final discharge probability; Based on the final discharge probability, a risk classification of a partial discharge event is performed by combining edge computing, and an alarm is given according to the risk classification result.

[0008] Further preferably, the generation method of the two-dimensional phase difference confidence ellipse specifically comprises: Robust covariance estimation MCD is performed on the six-channel partial discharge pulse data of the switch cabinet, and outliers are removed; The sample mean vector μ and the 2×2 covariance matrix Σ of the partial discharge pulse data after removing the outliers are calculated; The Mahalanobis distance is used to determine the boundary of the ellipse, a confidence degree is set, and a two-dimensional phase difference confidence ellipse equation is obtained; The two-dimensional phase difference confidence ellipse equation is updated adaptively on site.

[0009] Further preferably, in the ellipse equation, the sample mean vector μ is taken as the center coordinate of the ellipse; The eigenvalues λ1 and λ2 of the covariance matrix Σ are calculated, and the square roots of the weighted eigenvalues λ1 and λ2 are calculated, respectively, and the results after the square roots are taken are taken as the values of the major axis and the minor axis of the ellipse; The major axis direction angle of the ellipse is related to the elements in the covariance matrix Σ.

[0010] More preferably, the elliptic equation uses a set time during which the switch cabinet is powered on as the learning period. When the operating mode of the switch cabinet changes, a new round of learning is automatically triggered to ensure that the elliptic boundary always matches the actual partial discharge cluster.

[0011] More preferably, the temperature-humidity-noise gain matrix M(T,H) is constructed in real time in each sub-band and updated online using recursive least squares (RLS). Inverse compensation X′=XM·N is performed on the AE / TEV / UHF spectrum to reduce energy fluctuations caused by the environment; where N represents the noise spectrum.

[0012] More preferably, the change ΔN of the Noise channel is the difference between the average energy of the sub-band of the Noise channel in the current window and the average energy of the Noise channel without partial discharge events.

[0013] More preferably, the partial discharge prediction model outputs an initial discharge probability P once every power frequency cycle, which represents the instantaneous confidence level of whether partial discharge occurs in the current power frequency cycle. Each power frequency cycle is called a single-cycle window; If the initial discharge probability of several consecutive single-cycle windows is greater than the set threshold, the final discharge probability is Pfinal=1-(1-P)^k, where k is set according to the actual situation, and P is the initial discharge probability of the latest single-cycle window.

[0014] More preferably, a closed-loop partial discharge feature database is constructed to optimize the switching partial discharge prediction model for each of the edge devices; and model compression and incremental learning are performed on each optimized switching partial discharge prediction model. Using each switch partial discharge prediction model after compression and incremental learning, a global model is generated based on federated learning, and differentially fed back to the edge device for local hot update.

[0015] More preferably, the closed-loop construction of the built-in partial discharge feature database includes: During the offline phase, various partial discharge defects were manually implanted into the switchgear, and the original waveforms obtained after the defects were implanted were collected and labeled. During the online phase, ultrasonic positioning and withstand voltage verification are used during power outage maintenance to reverse-label the actual defect data, forming a closed loop; Pyramid compression involves performing spectral and statistical analysis on the original waveform to obtain high-dimensional features, which are then written to the terminal. The centroid fusion method is used to reduce the dimensionality of the high-dimensional features, resulting in reduced-dimensional features, which are then written to RAM.

[0016] More preferably, the sentinel mode refers to continuous sampling of the UHF channel while power is off on the other channels; The method of combining edge computing to classify the risk of partial discharge events and issuing alarms based on the risk classification results includes: Using the first two layers of the switching partial discharge prediction model, input a 1D sequence acquired simultaneously through six channels to extract features; The dual-channel inference includes channel A and channel B: The structure of the A channel is 1D-CNN, global average pooling layer GAP, fully connected layer FC, and softmax layer; the classification results are obtained through the A channel, including three categories: normal, attention, and abnormal. The structure of the B channel is a 1D-CNN, LSTM and sigmoid layer; the risk probability value is obtained through the B channel, which is in the range of 0-1. The risk classification includes Level 0, Level 1, and Level 2. Risk classification is performed by combining the final discharge probability and the classification results and risk probability values ​​output by the dual-channel inference. Level 0: If the final discharge probability is less than the set first probability, the output of channel A is normal, and the output of channel B is less than the set first risk probability value, then the event is only stored in FRAM and not uploaded. Level 1: When any of the following conditions are met, the risk level is Level 1, the event is saved, stored in Flash, and uploaded via low-power LoRa within 1 hour. ① Set the first probability ≤ final discharge probability < set the second probability and set the output of channel A as "Note"; ② The discharge probability is not less than the set first probability, but the output of channel A is normal and the output of channel B is less than the set second risk probability value; Level 2: When any of the following conditions are met, the risk classification is Level 2. High-power LoRa or 4G data will be immediately uploaded within 30 seconds, and a local high-brightness LED will flash 3 times to indicate the inspection: ① The final discharge probability is not less than the set second probability and the output of channel A is abnormal; ② The final discharge probability is not less than the set third probability and the B channel is not less than the set third risk probability value.

[0017] This invention also proposes a non-invasive online partial discharge monitoring terminal with multi-parameter fusion, comprising a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention can significantly improve the accuracy of partial discharge identification. Through a six-parameter fusion analysis mechanism, including the simultaneous acquisition and fusion analysis of six types of parameters such as ultrasonic waves (AE), ground electric waves (TEV), ultra-high frequency (UHF), environmental noise, temperature, and humidity, it can effectively distinguish between real partial discharge signals and interference signals (such as electromagnetic interference, mechanical vibration, and environmental noise), reducing the false alarm rate to 3% and the missed alarm rate to 1%. This invention also constructs a built-in partial discharge feature database, with more than 10,000 typical partial discharge spectrum samples built into the corresponding terminal, covering various discharge types such as corona, surface discharge, air gap discharge, and floating potential discharge. Combined with edge AI algorithms, it achieves a discharge type identification accuracy of ≥92%, providing accurate basis for subsequent maintenance.

[0019] 2. This invention enables true all-weather, full-lifecycle monitoring. It can perform continuous monitoring 24 / 7, replacing traditional manual inspections or periodic power outage tests, and achieving online perception of equipment insulation status throughout all time periods and the entire lifecycle. This invention also proposes a trend analysis and early warning mechanism. The system can automatically record the development trend of partial discharge, combine it with temperature and humidity changes to perform multi-dimensional trend modeling, and provide early warning of potential insulation faults 7 to 30 days in advance, thus gaining valuable time for planned maintenance and avoiding sudden power outages.

[0020] 3. This invention reduces the burden on the backend system through edge intelligent computing. The front end of this invention has intelligent diagnostic functions, and the terminal has built-in partial discharge feature extraction algorithm and expert diagnostic model, which can complete signal analysis, discharge type identification and risk level assessment locally, and only upload key data and alarm information, reducing the amount of communication data by more than 90%. This invention supports remote OTA upgrades, and the algorithm model can be continuously optimized through remote upgrades to adapt to different equipment types and operating environments, and extend the technical life cycle of the equipment. Attached Figure Description

[0021] Figure 1 This is a flowchart of a non-invasive online partial discharge monitoring method based on multi-parameter fusion according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0023] The present invention proposes the following technical solutions, such as... Figure 1 As shown, this invention proposes a non-invasive online monitoring method for partial discharge using multi-parameter fusion, comprising: Six types of sensors are arranged on the outer wall of the switch cabinet: ultrasonic AE, ground electric wave TEV, ultra-high frequency UHF, ambient noise, temperature T, and humidity H. The rising edge of the UHF signal is used as the trigger source to synchronously collect the partial discharge pulse data of the six channels corresponding to the switch cabinet. A circular timestamp is added to the collected partial discharge pulse data. Based on the six-channel partial discharge pulse data corresponding to the switch cabinet, a two-dimensional phase difference confidence ellipse of AE-TEV and TEV-UHF is established. The pulse data falling within the two-dimensional phase difference confidence ellipse is retained as the same source event, and the first empty-phase filtering is performed. The method for generating the two-dimensional phase difference confidence ellipse specifically includes: Robust covariance estimation (MCD) is performed on the six-channel partial discharge pulse data of the switchgear to remove outliers; Calculate the sample mean vector μ and the 2×2 covariance matrix Σ of the partial discharge pulse data after removing outliers; The elliptical boundary is determined by the Mahalanobis distance, and the confidence level is set to obtain the two-dimensional phase difference confidence ellipse equation. The equation of the two-dimensional phase difference confidence ellipse is adaptively updated in the field.

[0024] In the equation of the ellipse, the sample mean vector μ is used as the center coordinate of the ellipse; Calculate the eigenvalues ​​λ1 and λ2 of the covariance matrix Σ, and take the square root of the weighted eigenvalues ​​λ1 and λ2 respectively. The results of the square root are used as the values ​​of the major axis and minor axis of the ellipse. The orientation angle of the major axis of the ellipse is related to the elements in the covariance matrix Σ.

[0025] The elliptic equation uses a set time during which the switch cabinet is powered on as the learning period. When the operating mode of the switch cabinet changes, a new round of learning is automatically triggered to ensure that the elliptic boundary always matches the actual partial discharge cluster.

[0026] The set interval in the Noise channel is divided into several 1 / 3 octave sub-bands; using the six-channel partial discharge pulse data in the same source event, a temperature and humidity-noise gain matrix is ​​constructed in real time in each sub-band; using the temperature and humidity-noise gain matrix, inverse compensation is performed on the AE / TEV / UHF spectrum in the same source event; The temperature and humidity-noise gain matrix M(T,H) is constructed in real time in each sub-band and updated online using recursive least squares (RLS). Inverse compensation X′=XM·N is performed on the AE / TEV / UHF spectrum to reduce energy fluctuations caused by the environment; where N represents the noise spectrum.

[0027] The change ΔN of the Noise channel is the difference between the average energy of the sub-band of the Noise channel within the current window and the average energy of the Noise channel without partial discharge events.

[0028] A six-dimensional vector is constructed from the compensated ultrasonic data AE, ground electromagnetic wave data TEV, ultra-high frequency data UHF, the change in the noise channel ΔN, temperature T, and humidity H. This vector is then input into a pre-built and trained switching partial discharge prediction model configured in the edge device, and the initial discharge probability P is output. A sliding window Bayesian approach is then introduced to correct the initial discharge probability P, resulting in the final discharge probability. The switching partial discharge prediction model outputs an initial discharge probability P once every power frequency cycle, which represents the instantaneous confidence level of whether partial discharge occurs in the current power frequency cycle. Each power frequency cycle is called a single-cycle window; If the initial discharge probability of several consecutive single-cycle windows is greater than the set threshold, the final discharge probability is Pfinal=1-(1-P)^k, where k is set according to the actual situation, and P is the initial discharge probability of the latest single-cycle window.

[0029] A closed-loop system is constructed to build a built-in partial discharge feature database, and the switching partial discharge prediction model for each edge device is optimized; model compression and incremental learning are performed on each optimized switching partial discharge prediction model. Using each switch partial discharge prediction model after compression and incremental learning, a global model is generated based on federated learning, and differentially fed back to the edge device for local hot update.

[0030] The closed-loop construction of the built-in partial discharge feature database includes: During the offline phase, various partial discharge defects were manually implanted into the switchgear, and the original waveforms obtained after the defects were implanted were collected and labeled. During the online phase, ultrasonic positioning and withstand voltage verification are used during power outage maintenance to reverse-label the actual defect data, forming a closed loop; Pyramid compression involves performing spectral and statistical analysis on the original waveform to obtain high-dimensional features, which are then written to the terminal. The centroid fusion method is used to reduce the dimensionality of the high-dimensional features, resulting in reduced-dimensional features, which are then written to RAM.

[0031] Based on the final discharge probability, and combined with edge computing, partial discharge events are classified into risk levels, and alarms are triggered based on the risk classification results.

[0032] The method of combining edge computing to classify the risk of partial discharge events and issuing alarms based on the risk classification results includes: Using the first two layers of the switching partial discharge prediction model, input a 1D sequence acquired simultaneously through six channels to extract features; The dual-channel inference includes channel A and channel B: The structure of the A channel is 1D-CNN, global average pooling layer GAP, fully connected layer FC, and softmax layer; the classification results are obtained through the A channel, including three categories: normal, attention, and abnormal. The structure of the B channel is a 1D-CNN, LSTM and sigmoid layer; the risk probability value is obtained through the B channel, which is in the range of 0-1. The risk classification includes Level 0, Level 1, and Level 2. Risk classification is performed by combining the final discharge probability and the classification results and risk probability values ​​output by the dual-channel inference. Level 0: If the final discharge probability is less than the set first probability, the output of channel A is normal, and the output of channel B is less than the set first risk probability value, then the event is only stored in FRAM and not uploaded. Level 1: When any of the following conditions are met, the risk level is Level 1, the event is saved, stored in Flash, and uploaded via low-power LoRa within 1 hour. ① Set the first probability ≤ final discharge probability < set the second probability and set the output of channel A as "Note"; ② The discharge probability is not less than the set first probability, but the output of channel A is normal and the output of channel B is less than the set second risk probability value; Level 2: When any of the following conditions are met, the risk classification is Level 2. High-power LoRa or 4G data will be immediately uploaded within 30 seconds, and a local high-brightness LED will flash 3 times to indicate the inspection: ① The final discharge probability is not less than the set second probability and the output of channel A is abnormal; ② The final discharge probability is not less than the set third probability and the B channel is not less than the set third risk probability value.

[0033] This invention also proposes a non-invasive online partial discharge monitoring terminal with multi-parameter fusion, comprising a processor and a storage medium: The storage medium is used to store instructions; The processor is used to perform the steps of the above method according to the instructions.

[0034] Example 1 This invention proposes a non-invasive online monitoring method for partial discharge using multi-parameter fusion, which specifically includes the following steps: Step 1, Simultaneous Acquisition of Six Parameters Six types of sensors—ultrasonic AE, ground wave TEV, ultra-high frequency UHF, ambient noise, temperature T, and humidity H—are non-invasively arranged on the outer wall of the switch cabinet. The six-channel ADC shares a 60MHz temperature-compensated clock, and each pulse is timestamped with a 64-bit ring within the FPGA to achieve a relative synchronization error of 2ns. The rising edge of the UHF signal is used as the trigger source, and the original waveform is "frozen" for 2ms to ensure that the data can be aligned across physical domains.

[0035] Step 2, cross-domain phase correlation filtering Under the statistical law of the same partial discharge power source, a two-dimensional phase difference confidence ellipse of AE-TEV and TEV-UHF is established; only pulses falling within the ellipse are retained as homologous events, and the first empty-phase filtering is completed, with a signal-to-noise ratio improvement of ≥12dB.

[0036] The phase differences of the same partial discharge power source in AE, TEV, and UHF form a fixed triangle. Specifically, for each partial discharge pulse, with the 50 Hz power frequency phase as a reference, the envelope peak phases of the three signals AE, TEV, and UHF are calculated, resulting in φAE, φTEV, and φUHF ∈ [0°, 360°). Based on the envelope peak phases φAE, φTEV, and φUHF of the three signals, two relative phase differences are defined: ΔφAE - TEV = φAE – φTEV; ΔφTEV - UHF = φTEV – φUHF. A scatter plot is plotted based on 30 × 600 = 18000 sets of relative phase differences to obtain the statistical results.

[0037] In this invention, the actual partial discharge is located at the apex of the triangle, and the interference pulse is located at the other two bottom vertices.

[0038] Furthermore, the method for generating the two-dimensional phase difference confidence ellipse specifically includes: Robust covariance estimation (Minimum Covariance Determinant, MCD) was performed on 18,000 valid pulse samples from this cabinet. After removing 5% outliers, 17,100 points remained. Calculate the sample mean vector μ and the 2×2 covariance matrix Σ; The boundary of the ellipse is determined by the Mahalanobis distance, and the equation of the ellipse is obtained by taking a confidence level of 95%.

[0039] As a preferred embodiment of the present invention, taking a pointed defect as an example, the elliptical geometric parameters are expressed as follows: Center: μ = (–11.2°, +7.8°); Major axis a = √(λ1·5.991) = 9.4°; The minor axis b = √(λ²·5.991) = 4.1°; Major axis orientation angle (Relative to the ΔφAE-TEV axis); Where λ1 and λ2 are the eigenvalues ​​of Σ. , and Σ represents the elements in the covariance matrix Σ.

[0040] In a preferred embodiment of the present invention, pulses were collected continuously for 7 days on 38 operating cabinets, resulting in a total of 52,374 pulses.

[0041] The traditional single-channel threshold method (TEV>20mV) identified 1247 instances as partial discharge, of which 42 were due to interference from adjacent cabinets, resulting in a false alarm rate of 3.36%. After applying this elliptic filter, only 4 out of 1247 errors fell outside the ellipse and were therefore removed, leaving 1203 errors; interference pulses were suppressed by 42 to 4, and the false alarm rate was reduced to 0.33%. Interference suppression ratio = 10log10(42 / 4) = 10.2dB; Meanwhile, the retention rate of the real partial discharge pulse within the ellipse is 1203 / 1205=99.8%, and the equivalent signal-to-noise ratio is improved by ≥12dB (considering the integral ratio of the real signal and the interference power spectral density, according to the definition of IEEE Std1139-2015).

[0042] Furthermore, the two-dimensional phase difference confidence ellipse is adaptively updated in the field.

[0043] Specifically, the terminal is in a "learning period" of 24 hours after power-on. The FPGA accumulates ≥300 pulses from this cabinet, and μ and Σ are recalculated online using the same MCD method and written to EEPROM. If the cabinet operation mode changes (such as load switching causing Δφ drift >5°), a new round of learning is automatically triggered to ensure that the elliptical boundary always matches the real partial discharge cluster.

[0044] Through the above experiments and algorithms, the two-dimensional confidence ellipse transforms the "fixed triangular relationship of the same partial discharge source in the AE-TEV-UHF domain" into a quantifiable statistical boundary, achieving a signal-to-noise ratio improvement of ≥12dB. It can be computed in real time at the edge and only requires 64 bytes to store the upper triangular elements of μ and Σ.

[0045] Step 3, Environmental noise-temperature and humidity coupling compensation The 40–80 dB range of the noise channel is divided into 13 1 / 3 octave subbands. In each subband, a temperature and humidity-noise gain matrix M(T,H) is constructed in real time and updated online using recursive least squares (RLS). Inverse compensation X′=XM·N is performed on the AE / TEV / UHF spectrum to compress the energy jitter caused by the environment from ±6 dB to ±1.5 dB, thereby reducing false alarms. Here, N represents the noise spectrum, which corresponds to the 13 1 / 3 octave subbands. N is a 13×1 dimensional column vector, and each element is equal to the total in-band energy (dB value) within that 1 / 3 octave subband.

[0046] Step 4, Generation of six-dimensional feature vectors A six-dimensional vector is constructed from the compensated AE, TEV, UHF, ΔN, T, and H, and fed into a 3-layer Lightweight CNN (weights ≤ 128kB, which can run within a microcontroller unit MCU) in the front-end (edge ​​device) of the switch cabinet, outputting an initial discharge probability P of 0–1; a 30-minute sliding window Bayesian method is introduced to correct the initial discharge probability P: if the initial discharge probability P > 0.7 for 3 consecutive windows, then the final probability Pfinal = 1 - (1 - P)^1.8. The purpose of the correction is to suppress sporadic pulse interference.

[0047] Wherein, ΔN is the change in the Noise channel, ΔN = N_current - N_baseline; where N_current is the average in-band energy (in dB) of the Noise channel within the current window (e.g., 100ms), averaged after summing 13 1 / 3 octave subbands; N_baseline is the average energy of the Noise channel within the sliding window in the previous 30 minutes when there are no partial discharge events, serving as the environmental noise baseline. ΔN is the deviation of the current environmental noise from the stable baseline, used to capture background fluctuations caused by temperature, humidity, electromagnetic interference, etc., helping Lightweight CNN distinguish between real partial discharge and environmental disturbances. The instantaneous offset of the environmental noise channel is calculated as the difference between the average energy of the Noise channel in the current window and the baseline energy during the period without partial discharge in the past 30 minutes, used to characterize abnormal increases in background noise and improve the model's robustness to environmental disturbances.

[0048] Specifically, the Lightweight CNN outputs an initial discharge probability P every 20ms (power frequency cycle), representing the instantaneous confidence level of whether partial discharge occurs within the current power frequency cycle. In this invention, each power frequency cycle is referred to as a single-cycle window.

[0049] The specific implementation of the temporal sliding window Bayesian correction is as follows: if the initial discharge probability of three consecutive windows within the past 30 minutes is greater than 0.7, then the final probability Pfinal = 1 - (1 - P)^k, k = 1.8, where P is the initial discharge probability output by the LightweightCNN for the latest single-cycle window. After correction, Pfinal still represents the discharge confidence of the current power frequency cycle, but it has been strengthened using continuous high-confidence evidence from the past 30 minutes. The sliding window size is 30 minutes, i.e., 90,000 consecutive single-cycle windows (30 minutes × 50 Hz). Within these 90,000 windows, if any three temporally consecutive single-cycle windows have a Lightweight CNN output probability greater than 0.7, then correction is triggered; otherwise, correction is not triggered.

[0050] As a preferred embodiment of the present invention, after testing on 400 real-world switchgear cabinets, the algorithm achieved a recognition rate of 94.3% for three types of discharge: corona, air gap, and surface discharge, which is 27% higher than the single-parameter threshold method.

[0051] Step 5: Construct a closed-loop internal partial discharge feature database and optimize the LightweightCNN model corresponding to each switch cabinet terminal.

[0052] a) Offline stage: Six types of defects, namely sharp points, air gaps, surface defects, suspension defects, particles, and corona, are manually implanted in the model cabinet. The discharge quantity is calibrated to 10pC–10nC according to IEC60270, and the original waveforms are collected and labeled. b) Online phase: During power outage maintenance, ultrasonic positioning and withstand voltage verification are used to reverse-label the actual defect data and form a closed loop; c) Pyramid compression: The original L0 waveform is stored in the cloud, the L164-dimensional “spectral envelope + statistics” is written to the terminal SPIFlash (64kB), and the L26-dimensional “fusion centroid” is written to the FPGABlockRAM (384B) for offline real-time comparison by the terminal.

[0053] Step 6: Lightweight AI model compression and incremental learning at the front end For each terminal (switch cabinet) corresponding to the Lightweight CNN, 8-bit quantization and channel pruning are used to compress the 1.1MB Lightweight CNN to 96kB.

[0054] Knowledge distillation ensures an accuracy decrease of less than 1%; the terminal reserves a 128kB "incremental buffer area", high-confidence samples are uploaded daily via LoRa / 4G, the cloud uses the FedAvg algorithm to generate a global model weekly, differentially transmits 8kB gradients, and the terminal performs local hot updates to continuously optimize the recognition rate.

[0055] Step 7: Combine edge computing for local early warning.

[0056] Main state machine loop: Sentinel mode → Event triggering → Six-channel acquisition → Feature extraction → Dual-channel inference → Risk classification → Communication / storage → Return to Sentinel; Specifically, Sentinel mode refers to: Only the UHF channel maintains continuous sampling at 1MS / s, while the other channels are powered off; Performing 32-point sliding root mean square (RMS) calculations within the FPGA, the total power consumption below the threshold is 4.8 μA. Once the RMS exceeds the background noise of 1.8σ, the remaining channels are woken up within 100μs and the "frozen" clock is turned on to achieve "zero omission" triggering.

[0057] Furthermore, feature extraction refers to: A 1D sequence was acquired simultaneously across six channels, with each channel at 2 kHz and a duration of 100 ms, resulting in 200 points per channel and a total of 6 × 200 = 1200 points. Feature extraction was performed using a 1D-CNN at the front end. The weights of the 1D-CNN were the first two layers (conv1 + conv2) of a compressed Lightweight CNN. After feature extraction using the 1D-CNN, an 8 × 25 feature map (8 channels, downsampled by 8 times) was output. After flattening, a 200-dimensional feature vector was obtained and fed into the dual channels for inference.

[0058] It should be noted that feature extraction and dual-channel inference are physically separated. The feature layers of the 1D-CNN used for feature extraction are stored in Flash memory. Only the first two layers are run during operation, with a computing power of <15 MFLOP, which is not a problem for the MCU.

[0059] Furthermore, dual-channel reasoning refers to: The 200-dimensional features are fed into two independent small-head networks simultaneously to achieve parallel judgment of "steady state + transient state", hence the name "dual-channel".

[0060] Specifically, this includes channel A (steady state) and channel B (transient state).

[0061] The A-channel structure consists of two 1D-CNN layers, a global average pooling layer (GAP), a fully connected layer (FC) with 64 output neurons, and a softmax layer. This A-channel structure requires only 42kB. Classification results are obtained through the A-channel, including three categories: normal, attentional, and abnormal.

[0062] The B-channel structure consists of a single 1D-CNN layer, an LSTM layer with 16 hidden units, and a sigmoid layer. This B-channel structure requires only 54kB. The risk probability value, ranging from 0 to 1, is obtained through the B-channel.

[0063] The weights of the two head networks are packaged together in a 96 kB compressed model (A occupies 42 kB, B occupies 54 kB), and share 8-bit quantization and channel pruning with the feature extraction layer.

[0064] Furthermore, the risk classification rules refer to: Risk grading is performed by combining the corrected discharge probability with the classification results and risk probability values ​​output by dual-channel inference: Level 0 (Green): If the conditions Pfinal < 0.35, channel A output is normal, and channel B < 0.3 are met, then the event is only stored in FRAM and not uploaded; Level 1 (Yellow): When any of the following conditions are met, the risk level is Level 1, the event is saved, stored in Flash, and uploaded via low-power LoRa within 1 hour. ① 0.35≤Pfinal<0.7 and the output of channel A is noteworthy; ② Pfina ≥ 0.7, but channel A output is normal and channel B < 0.5; Level 2 (Red): When any of the following conditions are met, the risk level is Level 2. High-power LoRa or 4G data will be immediately uploaded within 30 seconds, and a local high-brightness LED will flash 3 times to indicate the inspection status: ① Pfinal ≥ 0.7 and the output of channel A is abnormal; ②Pfinal≥0.8 and B channel≥0.9.

[0065] Additionally, regarding the priority of risk classification: Channel A is classified as abnormal (2), which is the highest priority. Even if Pfinal is slightly lower than 0.7, it can be downgraded to trigger a red alarm. Channel B has an extremely high probability (≥0.9) and can be used as an independent trigger condition to prevent false alarms caused by extremely short pulses; When channel A is normal (0) and channel B is low (<0.5), even if Pfinal is high, the red signal will not be triggered to prevent occasional interference false alarms. All conditions are matched in the order of "red first, then yellow, then green". Once a match is found, subsequent judgments are stopped to ensure timely response.

[0066] Furthermore, two-tier storage refers to: The pulse waveform is first written to 4Mbit ultra-low power FRAM (read / write 2mA, hold 0.1mA), and is not lost when power is off; Once the FRAM is full of 256 waveform segments, the MCU packages them all at once and writes them to the NAND Flash via DMA. The total current during the writing process is 12mA, lasting for 40ms. Calculations show that with 100 discharge events per day, the 2.4Ah battery can last for 3.5 years, saving 85% more electricity than the traditional "constant sampling and storage" solution.

[0067] Example 2 This invention also proposes a non-invasive online partial discharge monitoring terminal with multi-parameter fusion: The present invention also proposes a terminal, including a processor and a storage medium: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to Embodiment 1.

[0068] Example 3 This invention is achieved through a non-invasive magnetic structure, which balances "high sensitivity" and "high protection".

[0069] Specifically, a silver-graphene composite electrode with laser micro-tooth 3D printing is used, with 50–100μm tooth peaks randomly distributed on the surface, effectively piercing the oxide layer and having a contact impedance of <5mΩ; a 0.1mm thick permalloy shielding ring is integrated on the back of the electrode to form a near-field magnetic confinement, suppressing the lateral current interference of the cabinet by 15dB.

[0070] This invention also uses an integrated magnetic attraction-resonance damping base, which is made of Sm2Co17 magnet with a magnetic energy product of 32MGOe and a single magnet attraction force of 60N. A 0.3mm viscoelastic PDMS layer is inserted between the magnet and the housing, with a loss factor of tanδ≈0.9, which can attenuate the mechanical resonance peak of 50–500Hz by 10dB and avoid pseudo ultrasonic pulses in the 10–30kHz frequency band caused by train / transformer vibration.

[0071] Example 4 The method of this invention was used in a 110kV substation with 38 switchgear cabinets operating continuously for 13 months, collecting a total of 4872 partial discharge events. The results were compared with the IEC60270 standard test results after a power outage. Detection sensitivity: 10pC (IEC) vs 12pC (this terminal), equivalent; False alarms: 42 times with the traditional threshold method, 4 times with this terminal, a reduction of 90%; Events missed: 7 by traditional methods, 0 by this terminal.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A non-invasive online monitoring method for partial discharge using multi-parameter fusion, characterized in that, include: Six types of sensors are arranged on the outer wall of the switch cabinet: ultrasonic AE, ground electric wave TEV, ultra-high frequency UHF, ambient noise, temperature T, and humidity H. The rising edge of the UHF signal is used as the trigger source to synchronously collect the partial discharge pulse data of the six channels corresponding to the switch cabinet. A circular timestamp is added to the collected partial discharge pulse data. Based on the six-channel partial discharge pulse data corresponding to the switch cabinet, a two-dimensional phase difference confidence ellipse of AE-TEV and TEV-UHF is established. The pulse data falling within the two-dimensional phase difference confidence ellipse is retained as the same source event, and the first empty-phase filtering is performed. The set interval in the Noise channel is divided into several 1 / 3 octave sub-bands; using the six-channel partial discharge pulse data in the same source event, a temperature and humidity-noise gain matrix is ​​constructed in real time in each sub-band; using the temperature and humidity-noise gain matrix, inverse compensation is performed on the AE / TEV / UHF spectrum in the same source event; A six-dimensional vector is constructed from the compensated ultrasonic data AE, ground electromagnetic wave data TEV, ultra-high frequency data UHF, the change in the noise channel ΔN, temperature T, and humidity H. This vector is then input into a pre-built and trained switching partial discharge prediction model configured in the edge device, and the initial discharge probability P is output. A sliding window Bayesian approach is then introduced to correct the initial discharge probability P, resulting in the final discharge probability. Based on the final discharge probability, and combined with edge computing, partial discharge events are classified into risk levels, and alarms are triggered based on the risk classification results.

2. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 1, characterized in that: The method for generating the two-dimensional phase difference confidence ellipse specifically includes: Robust covariance estimation (MCD) is performed on the six-channel partial discharge pulse data of the switchgear to remove outliers; Calculate the sample mean vector μ and the 2×2 covariance matrix Σ of the partial discharge pulse data after removing outliers; The elliptical boundary is determined by the Mahalanobis distance, and the confidence level is set to obtain the two-dimensional phase difference confidence ellipse equation. The equation of the two-dimensional phase difference confidence ellipse is adaptively updated in the field.

3. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 2, characterized in that: In the equation of the ellipse, the sample mean vector μ is used as the center coordinate of the ellipse; Calculate the eigenvalues ​​λ1 and λ2 of the covariance matrix Σ, and take the square root of the weighted eigenvalues ​​λ1 and λ2 respectively. The results of the square root are used as the values ​​of the major axis and minor axis of the ellipse. The orientation angle of the major axis of the ellipse is related to the elements in the covariance matrix Σ.

4. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 2, characterized in that: The elliptic equation uses a set time during which the switch cabinet is powered on as the learning period. When the operating mode of the switch cabinet changes, a new round of learning is automatically triggered to ensure that the elliptic boundary always matches the actual partial discharge cluster.

5. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 1, characterized in that: The temperature and humidity-noise gain matrix M(T,H) is constructed in real time in each sub-band and updated online using recursive least squares (RLS). Inverse compensation X′=XM·N is performed on the AE / TEV / UHF spectrum to reduce energy fluctuations caused by the environment; where N represents the noise spectrum.

6. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 1, characterized in that: The change ΔN of the Noise channel is the difference between the average energy of the sub-band of the Noise channel within the current window and the average energy of the Noise channel without partial discharge events.

7. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 1, characterized in that: The switching partial discharge prediction model outputs an initial discharge probability P once every power frequency cycle, which represents the instantaneous confidence level of whether partial discharge occurs in the current power frequency cycle. Each power frequency cycle is called a single-cycle window; If the initial discharge probability of several consecutive single-cycle windows is greater than the set threshold, the final discharge probability is Pfinal=1-(1-P)^k, where k is set according to the actual situation, and P is the initial discharge probability of the latest single-cycle window.

8. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 1, characterized in that: A closed-loop system is constructed to build a built-in partial discharge feature database, and the switching partial discharge prediction model for each edge device is optimized; model compression and incremental learning are performed on each optimized switching partial discharge prediction model. Using each switch partial discharge prediction model after compression and incremental learning, a global model is generated based on federated learning, and differentially fed back to the edge device for local hot update.

9. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 8, characterized in that: The closed-loop construction of the built-in partial discharge feature database includes: During the offline phase, various partial discharge defects were manually implanted into the switchgear, and the original waveforms obtained after the defects were implanted were collected and labeled. During the online phase, ultrasonic positioning and withstand voltage verification are used during power outage maintenance to reverse-label the actual defect data, forming a closed loop; Pyramid compression involves performing spectral and statistical analysis on the original waveform to obtain high-dimensional features, which are then written to the terminal. The centroid fusion method is used to reduce the dimensionality of the high-dimensional features, resulting in reduced-dimensional features, which are then written to RAM.

10. The non-invasive online monitoring method for partial discharge based on multi-parameter fusion according to claim 1, characterized in that: The method of combining edge computing to classify the risk of partial discharge events and issuing alarms based on the risk classification results includes: Using the first two layers of the switching partial discharge prediction model, input a 1D sequence acquired simultaneously through six channels to extract features; The dual-channel inference includes channel A and channel B: The structure of the A channel is 1D-CNN, global average pooling layer GAP, fully connected layer FC, and softmax layer; the classification results are obtained through the A channel, including three categories: normal, attention, and abnormal. The structure of the B channel is a 1D-CNN, LSTM and sigmoid layer; the risk probability value is obtained through the B channel, which is in the range of 0-1. The risk classification includes Level 0, Level 1, and Level 2. Risk classification is performed by combining the final discharge probability and the classification results and risk probability values ​​output by the dual-channel inference. Level 0: If the final discharge probability is less than the set first probability, the output of channel A is normal, and the output of channel B is less than the set first risk probability value, then the event is only stored in FRAM and not uploaded. Level 1: When any of the following conditions are met, the risk level is Level 1, the event is saved, stored in Flash, and uploaded via low-power LoRa within 1 hour. ① Set the first probability ≤ final discharge probability < set the second probability and set the output of channel A as "Note"; ② The discharge probability is not less than the set first probability, but the output of channel A is normal and the output of channel B is less than the set second risk probability value; Level 2: When any of the following conditions are met, the risk classification is Level 2. High-power LoRa or 4G data will be immediately uploaded within 30 seconds, and a local high-brightness LED will flash 3 times to indicate the inspection: ① The final discharge probability is not less than the set second probability and the output of channel A is abnormal; ② The final discharge probability is not less than the set third probability and the B channel is not less than the set third risk probability value.

11. A non-invasive online partial discharge monitoring terminal utilizing the multi-parameter fusion method according to any one of claims 1-10, characterized in that: The terminal shown includes a processor and storage media: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-10.

Images