GIS equipment partial discharge mode identification method based on deep learning

Abstract

The invention proposes a GIS equipment partial discharge mode recognition method based on deep learning, and relates to the technical field of GIS partial discharge recognition, and the method comprises the steps: collecting a partial discharge signal of GIS equipment through an ultrahigh frequency sensor, and carrying out the signal conditioning; performing digital sampling and processing on the conditioned partial discharge signal to generate a phase-resolved pulse sequence (PRPS) map, and performing intelligent noise reduction processing on the PRPS map, the PRPS map including three-dimensional information of phase, amplitude and frequency; inputting the PRPS atlas into a preset neural network model, and extracting atlas features; and identifying and classifying the partial discharge type based on the extracted map features, and outputting an identification result. According to the method, automatic acquisition, intelligent processing and accurate identification of the partial discharge signals are realized, and compared with a traditional manual identification method, the identification efficiency and accuracy are greatly improved, and reliable technical support is provided for state monitoring and fault diagnosis of GIS equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of GIS partial discharge identification, and in particular to a GIS equipment partial discharge pattern identification method based on deep learning. Background Art

[0002] GIS (Gas Insulated Metal Enclosed Switchgear) plays an important role in the power system. During operation, GIS equipment will be affected by various factors such as electricity, heat, machinery, and environment, and complex chemical and physical changes will occur inside it, causing the performance to gradually deteriorate. This degradation phenomenon will affect the safe operation of the equipment and even cause accidents. In order to ensure the quality of the equipment and the reliability of the system power supply, strict quality inspection and daily maintenance of GIS equipment are required. At present, the maintenance of GIS equipment is changing from the traditional "mandatory maintenance when due" to the "repair when necessary" predictive maintenance mode. This change requires real-time monitoring of equipment status and fault diagnosis.

[0003] In the prior art, partial discharge detection is an effective means to detect early insulation defects inside GIS equipment. By installing UHF sensors on GIS equipment, the electromagnetic wave signals generated by partial discharge inside the equipment can be detected. These signals are usually in the ultra-high frequency range and need to be conditioned by signal amplification and filtering. The detection system usually adopts a distributed measurement structure, including UHF sensors, signal acquisition units, and monitoring hosts. During the signal acquisition process, phase synchronization is ensured through various synchronization methods, and phase-resolved pulse sequence spectra are generated through digital sampling and real-time data processing for analyzing partial discharge characteristics.

[0004] However, the existing partial discharge detection technology has the following problems: First, environmental noise interference seriously affects the detection accuracy. Although manual noise reduction, adaptive noise reduction and other methods are used, it is difficult to effectively distinguish environmental noise from real partial discharge signals, which can easily lead to misjudgment. Secondly, the existing signal processing methods mainly rely on traditional filtering and noise reduction technologies, and have limited ability to extract complex partial discharge features, making it difficult to accurately identify different types of discharge modes. Thirdly, the intelligence level of the detection system is insufficient, and fault diagnosis mainly relies on manual experience, lacking automated feature extraction and pattern recognition capabilities. In addition, the existing technology is inefficient in processing massive monitoring data, making it difficult to achieve real-time fault warning. These problems seriously restrict the accuracy and reliability of GIS equipment status monitoring and affect the implementation effect of predictive maintenance. Summary of the invention

[0005] In view of this, the present invention proposes a local discharge pattern recognition method for GIS equipment based on deep learning. The local discharge signal is collected by a UHF sensor, and combined with the PRPS spectrum technology and an improved three-dimensional convolutional neural network model to realize intelligent noise reduction and automatic recognition of the local discharge signal of GIS equipment. It can accurately distinguish five different types of discharge modes, such as suspended potential, metal particles, tip, surface and hollow conductor, thereby improving the accuracy and reliability of local discharge detection of GIS equipment and providing effective technical support for equipment status monitoring and fault diagnosis.

[0006] The technical solution of the present invention is achieved in this way:

[0007] The present invention provides a method for identifying partial discharge patterns of GIS equipment based on deep learning, comprising:

[0008] Collect partial discharge signals of GIS equipment through UHF sensors and perform signal conditioning;

[0009] Digitally sample and process the conditioned partial discharge signal to generate a phase-resolved pulse sequence PRPS spectrum, and perform intelligent noise reduction on the PRPS spectrum, wherein the PRPS spectrum contains three-dimensional information of phase, amplitude and frequency;

[0010] Input the PRPS spectrum into the preset neural network model to extract the spectrum features;

[0011] The partial discharge type is identified and classified based on the extracted spectrum features, and the identification results are output.

[0012] Based on the above solution, preferably, the signal conditioning includes:

[0013] Signal amplification through a low noise amplifier;

[0014] Filtering the signal by a combination filter, wherein the combination filter includes a high-pass filter, a low-pass filter and a band-stop filter;

[0015] The partial discharge pulse envelope signal is obtained through the detector.

[0016] Based on the above solution, preferably, the digital sampling and processing of the conditioned partial discharge signal includes:

[0017] The conditioned partial discharge signal is sampled at high speed using a 50MHz sampling rate and 12-bit ADC.

[0018] Real-time data processing is performed through FPGA, including: peak detection of sampled data; data calibration and preliminary feature extraction; generation of N-point PRPS raw data within 1 second;

[0019] Data analysis and calculation are performed through the ARM processor, including: obtaining the PRPS raw data processed by the FPGA; calculating the characteristic values ​​of partial discharge phase, frequency, and amplitude; generating real-time PRPS three-dimensional spectrum data, in which: the phase is aligned with the 50Hz synchronization signal; the amplitude reflects the discharge intensity; and the frequency represents the distribution of discharge times.

[0020] On the basis of the above scheme, preferably, the intelligent noise reduction processing adopts PRPS-based weighted adaptive noise reduction, and the noise reduction processing adopts the following formula:

[0021]

[0022] In the formula, represents phase, A represents amplitude, f represents frequency, is the PRPS data after noise reduction, is the original PRPS data, is the three-dimensional noise reduction factor, defined as:

[0023]

[0024]

[0025]

[0026]

[0027] in, is the phase weight coefficient, is the reference phase point, is the phase attenuation coefficient, is the phase response function, k2 is the response coefficient, which is used to adjust the phase sensitivity; K A (A) is the amplitude modulation coefficient, E(t) is the normalized value of the environmental noise sensor, k1 is the noise suppression coefficient, 0 <k1<1,A th is the amplitude judgment threshold, which is set based on the range of -80dBm to -10dBm, α is the amplitude modulation index, 0<α<1; K f (f) is the frequency weight coefficient, f min is the minimum effective frequency threshold, f max is the maximum reference frequency.

[0028] On the basis of the above scheme, preferably, the preset neural network model is an improved three-dimensional convolutional neural network model, which specifically includes:

[0029] Input layer, receiving PRPS spectrum data;

[0030] The phase block convolution layer includes three groups of 3×3×3 convolution kernels with a step size of 1, and each group of convolution kernels corresponds to a different phase region;

[0031] Batch normalization layer and ReLU activation function layer are used to standardize feature distribution and increase the nonlinear expression ability of the network;

[0032] The three-dimensional maximum pooling layer has a pooling kernel size of 2×2×2 and a step size of 2, which is used to extract local salient features;

[0033] Feature fusion layer, including attention mechanism, is used to fuse time domain features and frequency domain features;

[0034] The classification layer includes two fully connected layers and an output layer. The first fully connected layer has 2048 neurons, the second fully connected layer has 512 neurons, and the output layer has 5 neurons, corresponding to five types of local discharges: suspended potential discharge; metal particle discharge; tip discharge; surface discharge; and hollow conductor discharge.

[0035] Based on the above scheme, preferably, the phase block convolution layer includes a phase weight matrix, which divides the 360° phase range into a main discharge phase region, a secondary discharge phase region and other phase regions according to the importance of the discharge phase, and assigns different weight coefficients to each of them.

[0036] Based on the above solution, preferably, the expression of the phase weight matrix is ​​as follows:

[0037]

[0038]

[0039]

[0040] in, is the phase weight matrix, is the phase weight basis matrix, R(A,f) is the amplitude-frequency modulation matrix, represents a matrix multiplication operator; W1 is the basic weight of the main discharge phase region, W2 is the basic weight of the secondary discharge phase region, and W3 is the basic weight of other phase regions; It is the boundary of the main discharge phase area; is the boundary of the secondary discharge phase area; A ref is the reference amplitude; f ref is the reference frequency; σ A is the amplitude standard deviation; σ f is the frequency standard deviation.

[0041] On the basis of the above scheme, preferably, the method further includes a phase synchronization control step before collecting the local discharge signal of the GIS equipment through the UHF sensor to provide an accurate phase reference benchmark for the PRPS spectrum, and the phase synchronization control adopts one of PT synchronization, AC220V synchronization and external signal source synchronization.

[0042] On the basis of the above scheme, preferably, after performing intelligent noise reduction processing on the PRPS spectrum, the method further includes an event triggering step, screening interference signals through a dual threshold judgment mechanism, and the event triggering step is as follows:

[0043] Set event trigger amplitude threshold and trigger number threshold;

[0044] The amplitude of each point of the PRPS signal after noise reduction is compared with the event trigger amplitude threshold;

[0045] When the trigger times threshold is met, the partial discharge alarm is triggered.

[0046] On the basis of the above solution, preferably, the detection frequency range of the UHF sensor is 300MHz-1500MHz, and the dynamic range is -80dBm to -10dBm.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) The present invention collects partial discharge signals through ultra-high frequency sensors, combines PRPS spectrum technology and an improved three-dimensional convolutional neural network model, realizes intelligent noise reduction and automatic recognition of partial discharge signals of GIS equipment, and can accurately distinguish five different types of discharge modes: floating potential, metal particles, tip, surface and hollow conductor, thereby improving the accuracy and reliability of partial discharge detection;

[0049] (2) The present invention adopts an intelligent noise reduction algorithm based on PRPS weighted adaptation. Through the synergistic effect of phase weight coefficient, amplitude modulation coefficient and frequency weight coefficient, the PRPS spectrum is subjected to three-dimensional noise reduction processing, which effectively distinguishes the interference of environmental noise on the monitoring signal and reduces the misjudgment rate of partial discharge alarm.

[0050] (3) The improved three-dimensional convolutional neural network model designed by the present invention realizes the effective extraction and fusion of time domain features and frequency domain features in the PRPS spectrum through the combination of phase block convolution layer, batch normalization layer, ReLU activation function layer and three-dimensional maximum pooling layer, combined with the attention mechanism in the feature fusion layer, thereby improving the accuracy of partial discharge type recognition;

[0051] (4) The present invention introduces a phase weight matrix in the phase block convolution layer, divides the 360° phase range by weight according to the importance of the discharge phase, and enhances the features through the amplitude-frequency modulation matrix, thereby improving the recognition sensitivity of the discharge features in different phase regions;

[0052] (5) The present invention adopts a dual-threshold event trigger mechanism. By setting the event trigger amplitude threshold and the trigger number threshold, the noise-reduced PRPS signal is screened and judged, which effectively avoids false triggering caused by instantaneous interference and improves the credibility of the partial discharge alarm. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 is a flow chart of the method of the present invention;

[0055] Figure 2 It is a technical implementation diagram of the present invention;

[0056] Figure 3 It is a structural diagram of a neural network model of the present invention;

[0057] Figure 4 It is a phase synchronization control flow chart of the present invention;

[0058] Figure 5 Schematic diagram of the event triggering mechanism of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] like Figure 1 and Figure 2 As shown, the present invention provides a method for identifying partial discharge patterns of GIS equipment based on deep learning, comprising:

[0061] Collect partial discharge signals of GIS equipment through UHF sensors and perform signal conditioning;

[0062] Digitally sample and process the conditioned partial discharge signal to generate a phase-resolved pulse sequence PRPS spectrum, and perform intelligent noise reduction on the PRPS spectrum, wherein the PRPS spectrum contains three-dimensional information of phase, amplitude and frequency;

[0063] Input the PRPS spectrum into the preset neural network model to extract the spectrum features;

[0064] The partial discharge type is identified and classified based on the extracted spectrum features, and the identification results are output.

[0065] Specifically, in one embodiment of the present invention, the detection frequency range of the UHF sensor is 300MHz-1500MHz, and the dynamic range is -80dBm to -10dBm. Signal conditioning includes:

[0066] Signal amplification through a low noise amplifier;

[0067] Filtering the signal by a combination filter, wherein the combination filter includes a high-pass filter, a low-pass filter and a band-stop filter;

[0068] The partial discharge pulse envelope signal is obtained through the detector.

[0069] In one example, a UHF sensor is used to collect partial discharge signals of GIS equipment and perform signal conditioning, which specifically includes the following implementation steps:

[0070] 1. UHF sensor installation

[0071] (1) Built-in sensor installation: The disc structure sensor is pre-installed on the GIS equipment before it leaves the factory. The sensor detection frequency range is 300MHz-1500MHz, and the sensor dynamic range is -80dBm to -10dBm.

[0072] (2) External sensor installation: installed on the pot insulator outside the GIS to receive electromagnetic wave signals radiated from the inside of the GIS.

[0073] 2. Signal conditioning circuit design

[0074] (1) Low-noise amplifier unit: A low-noise amplifier is used to amplify the collected weak signal, and an amplifier circuit with clamping protection function is designed to prevent excessive signals from damaging the system.

[0075] (2) Combined filter unit: A combined filter including a high-pass filter, a low-pass filter and a band-stop filter is designed to realize the programmable switching function of a multi-band filter. The frequency band can be switched to narrowband, full-band, low-frequency, high-frequency, etc., effectively avoiding the noise interference frequency band.

[0076] (3) Detection unit: The filtered signal is processed by the detector to obtain the partial discharge pulse envelope signal.

[0077] Specifically, in one embodiment of the present invention, digital sampling and processing of the conditioned partial discharge signal includes:

[0078] The conditioned partial discharge signal is sampled at high speed using a 50MHz sampling rate and 12-bit ADC.

[0079] Real-time data processing is performed through FPGA, including: peak detection of sampled data; data calibration and preliminary feature extraction; generating N-point PRPS raw data within 1 second;

[0080] Data analysis and calculation are performed through the ARM processor, including: obtaining the PRPS raw data processed by the FPGA; calculating the characteristic values ​​of partial discharge phase, frequency, and amplitude; generating real-time PRPS three-dimensional spectrum data, in which: the phase is aligned with the 50Hz synchronization signal; the amplitude reflects the discharge intensity; and the frequency represents the distribution of discharge times.

[0081] In a specific example, the ADC sampling system uses a high-speed ADC chip AD9626 for sampling, with a sampling rate of 50MHz, a resolution of 12bit, a signal-to-noise ratio of 72dB, an effective number of bits of 11.7bit, and an input bandwidth of 250MHz. It is configured in differential input mode to improve anti-interference ability. A 50Ω impedance matching network is designed at the input end to ensure signal integrity. The sampling clock system uses a TCXO temperature-compensated crystal oscillator as the reference source, generates a 50MHz sampling clock through a phase-locked loop multiplication, and uses differential wiring transmission to reduce clock jitter. To ensure sampling accuracy, a star layout is used in the PCB design, the analog ground and digital ground are connected at a single point, and a shielding layer is added on the key signal path.

[0082] FPGA real-time data processing is implemented as follows:

[0083] 1. Peak detection module

[0084] The peak detection module uses an adaptive threshold detection algorithm based on a sliding window, which detects and extracts partial discharge pulses by analyzing signal characteristics in real time. The specific implementation process is as follows:

[0085] First, a 32-point data cache unit is implemented in the FPGA to store the continuously sampled data sequence. The cache unit adopts the First-In-First-Out (FIFO) structure to ensure real-time data update. When new sampled data arrives, the earliest data will be removed from the cache unit to maintain the sliding characteristics of the window. The selection of 32 points as the window length is based on the typical duration of the partial discharge pulse (about 0.64μs) and the system sampling rate (50MHz). It can not only ensure the complete capture of a single discharge pulse, but also effectively reduce the computational complexity.

[0086] In the peak detection process, the algorithm first calculates the dynamic threshold value of the data in the window. The dynamic threshold value V th The calculation method is as follows: first calculate the mean μ and standard deviation σ of the data in the window, and then use the formula V th =μ+k×σ to determine the threshold level, where k is an adjustable sensitivity coefficient, which is initially set to 3. This adaptive threshold setting method can effectively adapt to changes in background noise and improve detection reliability.

[0087] For each data point V in the window i , the algorithm determines whether it is a valid peak point through the following steps:

[0088] Calculate the local features of the 16 points before and after the point centered;

[0089] Determine whether the amplitude of the point is greater than the dynamic threshold V th ;

[0090] Verify whether the point is greater than the amplitude of the 16 points before and after it;

[0091] Check whether the time interval between this point and the previously detected peak point is greater than the minimum interval requirement (set to 20 sampling points, i.e. 0.4 μs).

[0092] When a data point meets all the above conditions at the same time, it is marked as a valid peak point. In order to improve the accuracy of detection, the algorithm also implements anti-interference processing: by calculating the signal morphology characteristics near the peak point (such as rise time, fall time, pulse width, etc.), the interference signal that does not meet the local discharge characteristics is filtered out.

[0093] For each valid peak point detected, the system needs to record its precise time information. The timestamp recording module is implemented using a 64-bit counter. The clock source of the counter comes from the system's 50MHz sampling clock. After frequency division, a timing reference with a time resolution of 20ns is obtained. The design of the 64-bit counter ensures that counting overflow will not occur during continuous operation. In order to handle possible system restart situations, a timestamp rollback processing mechanism is implemented: when it is detected that the counter is about to overflow, the upper system is notified through the software interface to ensure the continuity of time information.

[0094] The corresponding relationship between the output value of the counter and the absolute time of the system is established through the power-on initialization process. During initialization, the system writes the current real-time clock (RTC) time into a dedicated register as the reference point for the timestamp. All subsequent timestamps can be calculated from the offset from this reference point to obtain the corresponding absolute time, accurately recording the moment when each discharge pulse occurs.

[0095] 2. Data calibration and feature extraction module

[0096] The data calibration module mainly completes the linear correction and temperature compensation of the sampled data. The system implements a calibration algorithm based on piecewise linear interpolation in the FPGA. First, a calibration curve is established through a high-precision signal source, and the calibration data is stored in the Flash. During operation, the corresponding calibration parameters are dynamically selected according to the current temperature sensor reading. The calibration process is implemented using a pipeline structure to ensure real-time performance.

[0097] The specific calibration steps include:

[0098] Temperature compensation: Based on the real-time reading of the temperature sensor, the corresponding temperature compensation coefficient is read from the Flash;

[0099] Zero point correction: By collecting the output of the system in idle state, the zero point offset is calculated and compensated;

[0100] Gain correction: Perform nonlinear correction on the sampled data based on the piecewise linear interpolation algorithm.

[0101] The feature extraction module realizes the real-time feature calculation of the calibrated data. It mainly includes:

[0102] Amplitude characteristics: calculate the peak value, mean value, variance and other statistical characteristics of the signal;

[0103] Time characteristics: extract the time domain characteristics of the pulse such as rise time and duration;

[0104] Frequency Characteristics: Calculate the frequency spectrum characteristics of the signal through sliding FFT.

[0105] 3.PRPS raw data generation module

[0106] The PRPS raw data generation module is implemented in FPGA, and a double buffer mechanism is used to ensure the continuity of data processing. The system divides a power frequency cycle (20ms) into 360 phase points, and the amplitude and frequency information corresponding to each phase point are obtained through real-time statistics.

[0107] 4. ARM processor data analysis and calculation

[0108] The ARM processor is responsible for further processing and analysis of the PRPS raw data output by the FPGA. The system uses a dual-core ARM processor, one core is responsible for data reception and caching, and the other core is responsible for data processing and feature calculation.

[0109] Core 0 (data receiving core) is responsible for data reception and preprocessing: responding to DMA transfer completion events through interrupts; implementing a double buffering mechanism: setting up two 32MB data buffers, while one buffer receives data, the data in the other buffer is handed over to Core 1 for processing; performing data frame parsing: verifying the frame header identifier (0xAA55AA55), calculating the CRC checksum, and extracting timestamp information; implementing data decompression: using a fast decompression algorithm based on table lookup to restore the compressed data sent by the FPGA to its original format.

[0110] Core 1 (data processing core) performs specific data analysis and feature calculations:

[0111] Phase feature calculation: using the phase locking algorithm, establish an accurate phase reference by analyzing the zero-crossing point of the synchronization signal; realize adaptive phase compensation: automatically adjust the phase offset according to the system delay; reduce phase jitter through sliding average filtering, and the phase accuracy is better than 0.1 degree; establish a 360-degree phase mapping table to support fast phase search; discharge feature analysis: the amplitude feature adopts the histogram statistical method to update the discharge amplitude distribution in real time; the frequency feature uses the counting accumulation algorithm to count the discharge frequency in different phase intervals; the time feature calculates the discharge interval distribution through timestamp analysis.

[0112] PRPS 3D map generation: normalize the raw data to eliminate system gain differences; generate standard format output.

[0113] Specifically, in one embodiment of the present invention, the intelligent noise reduction process is performed on the PRPS spectrum using PRPS weighted adaptive noise reduction, and the noise reduction process uses the following formula:

[0114]

[0115] In the formula, represents phase, A represents amplitude, f represents frequency, is the PRPS data after noise reduction, is the original PRPS data, is the three-dimensional noise reduction factor, defined as:

[0116]

[0117]

[0118]

[0119]

[0120] in, is the phase weight coefficient, is the reference phase point, is the phase attenuation coefficient, is the phase response function, k2 is the response coefficient, which is used to adjust the phase sensitivity; K A (A) is the amplitude modulation coefficient, E(t) is the normalized value of the environmental noise sensor, k1 is the noise suppression coefficient, 0 <k1<1,A th is the amplitude judgment threshold, which is set based on the range of -80dBm to -10dBm, α is the amplitude modulation index, 0<α<1; K f (f) is the frequency weight coefficient, f min is the minimum effective frequency threshold, f max is the maximum reference frequency.

[0121] Specifically, the PRPS weighted adaptive intelligent noise reduction algorithm provided by the present invention realizes adaptive noise reduction processing of different dimensions (phase, amplitude, frequency) in the PRPS spectrum by constructing a three-dimensional weight matrix. The core idea of ​​the algorithm is to use the characteristic differences of the partial discharge signal in these three dimensions to selectively enhance or suppress the signal. Among them, the basic mathematical model of the algorithm adopts weight modulation in the form of multiplication, and the noise reduction factor adopts a decomposition form, allowing each dimension to be optimized independently.

[0122] Specifically, the phase weight coefficient adopts a dynamic reference phase, that is, Not fixed, can be updated based on historical data, for example, using a sliding window (typical window length 20 cycles) to calculate the most frequent discharge phase and introducing an adaptive attenuation coefficient Dynamically adjust according to the signal-to-noise ratio, increase in the case of low signal-to-noise ratio Expand the suppression range and reduce the Improve phase selectivity. In the calculation formula of amplitude modulation coefficient, A thSpecifically, it can be dynamically calculated through the CFAR (Constant False Alarm Rate) algorithm. The α value is dynamically adjusted according to the statistical characteristics of the signal. The α value is reduced in the strong signal area to avoid over-enhancement, and the α value is increased in the weak signal area to improve the detection sensitivity. In the calculation formula of the frequency weight coefficient, f min and f max Dynamically adjust according to historical statistical characteristics, in f min A steeper inhibition curve is used near f max A slower decay curve is used nearby.

[0123] The present invention adopts an intelligent noise reduction algorithm based on PRPS weighted adaptation, and realizes all-round noise reduction of PRPS spectrum through three-dimensional noise reduction factors. In the phase dimension, the algorithm introduces the phase weight coefficient K φ (φ), combined with the phase attenuation coefficient σ φ And phase response function Q(φ), to achieve adaptive noise reduction in different phase regions; in the amplitude dimension, through the amplitude modulation coefficient K A (A) In combination with the real-time data of the environmental noise sensor, the noise suppression coefficient k1 is dynamically adjusted to effectively suppress the interference of environmental noise; in the frequency dimension, the frequency weight coefficient K is used f (f) By setting the minimum effective frequency threshold and the maximum reference frequency, the random interference signal is filtered out. This three-dimensional collaborative noise reduction scheme not only maintains the characteristic integrity of the partial discharge signal, but also significantly improves the signal-to-noise ratio, laying a good foundation for subsequent feature extraction and pattern recognition.

[0124] Specifically, Figure 3 As shown, in one embodiment of the present invention, the preset neural network model is an improved three-dimensional convolutional neural network model, which specifically includes:

[0125] Input layer, receiving PRPS spectrum data;

[0126] The phase block convolution layer includes three groups of 3×3×3 convolution kernels with a step size of 1, and each group of convolution kernels corresponds to a different phase region;

[0127] Batch normalization layer and ReLU activation function layer are used to standardize feature distribution and increase the nonlinear expression ability of the network;

[0128] The three-dimensional maximum pooling layer has a pooling kernel size of 2×2×2 and a step size of 2, which is used to extract local salient features;

[0129] Feature fusion layer, including attention mechanism, is used to fuse time domain features and frequency domain features;

[0130] The classification layer includes two fully connected layers and an output layer. The first fully connected layer has 2048 neurons, the second fully connected layer has 512 neurons, and the output layer has 5 neurons, corresponding to five types of local discharges: suspended potential discharge; metal particle discharge; tip discharge; surface discharge; and hollow conductor discharge.

[0131] The phase block convolution layer includes a phase weight matrix, which divides the 360° phase range into a main discharge phase region, a secondary discharge phase region, and other phase regions according to the importance of the discharge phase, and assigns different weight coefficients to each of them.

[0132] The expression of the phase weight matrix is ​​as follows:

[0133]

[0134]

[0135]

[0136] in, is the phase weight matrix, is the phase weight basis matrix, R(A,f) is the amplitude-frequency modulation matrix, represents a matrix multiplication operator; W1 is the basic weight of the main discharge phase region, W2 is the basic weight of the secondary discharge phase region, and W3 is the basic weight of other phase regions; It is the boundary of the main discharge phase area; is the boundary of the secondary discharge phase area; A ref is the reference amplitude; f ref is the reference frequency; σ A is the amplitude standard deviation; σ f is the frequency standard deviation.

[0137] In this embodiment, the input layer receives the preprocessed PRPS spectrum data, and the input data contains information in three dimensions: phase, amplitude, and frequency. The phase block convolution layer contains three groups of 3×3×3 convolution kernels, with a step size of 1. Each group of convolution kernels corresponds to a different phase region. The phase weight matrix is ​​introduced to divide the 360° phase range into three regions: the main discharge phase region is assigned the highest weight W1, the secondary discharge phase region is assigned a medium weight W2, and other phase regions are assigned the lowest weight W3. This partition design can highlight the feature extraction of important phase regions.

[0138] In a specific example, the processing of the PRPS map is divided into four main stages:

[0139] The first stage is data input processing. The model first receives PRPS spectrum data containing three-dimensional information of phase, amplitude and frequency. After preprocessing, these data are input into the input layer of the network in a format suitable for neural network processing.

[0140] The second stage is feature extraction. This stage is first processed through a phase block convolution layer, which contains three groups of 3×3×3 convolution kernels with a step size of 1. Each group of convolution kernels corresponds to a different phase region, and a phase weight matrix is ​​introduced to process the features of different regions. The phase weight matrix divides the 360° phase range into three regions: the main discharge phase region, the secondary discharge phase region, and the other phase region, and assigns different weight coefficients (W1, W2, W3) to each region. The extracted features are then standardized by a batch normalization layer, and the nonlinear expression capability is increased by the ReLU activation function. Finally, a 2×2×2 three-dimensional maximum pooling layer (with a step size of 2) is used to extract local significant features.

[0141] The PRPS graph is essentially a three-dimensional data structure S(φ,A,f), which contains two important features:

[0142] Time domain characteristics: mainly reflected in the phase-amplitude (φ-A) plane, reflecting the distribution characteristics of the discharge pulse within the power frequency cycle, including: the phase position of the discharge, the amplitude distribution of the discharge pulse, the phase concentration of the discharge, and the amplitude change law of the discharge.

[0143] Frequency domain characteristics: mainly reflected in the amplitude-frequency (Af) plane, reflecting the statistical characteristics of discharge occurrence, including: discharge repetition rate, discharge continuity, statistical distribution of discharge amplitude, and discharge frequency distribution law.

[0144] The phase block convolution layer uses three groups of 3×3×3 convolution kernels to extract features in three dimensions:

[0145] In the phase dimension (φ): the 360° phase range is divided into three regions, and different weights are used for feature enhancement in each region to extract the concentration and distribution characteristics of the discharge phase.

[0146] In the amplitude dimension (A): extract the variation characteristics of the discharge amplitude and capture the statistical characteristics of the amplitude distribution.

[0147] In the frequency dimension (f): extract the statistical characteristics of the discharge frequency and capture the repetition rate and persistence characteristics.

[0148] Each convolution kernel slides simultaneously in these three dimensions to form a local receptive field and extract local features. Through the design of the phase weight matrix, the features of important phase areas are enhanced, so that the model can better focus on the features of key areas.

[0149] The third stage is feature fusion processing. In this stage, the model fuses the extracted features through the feature fusion layer. This layer contains an attention mechanism that can effectively fuse time domain features and frequency domain features, and can adaptively adjust the importance weights of different features.

[0150] The fourth stage is classification decision processing. The data after feature fusion is first processed by two fully connected layers, the first layer contains 2048 neurons, the second layer contains 512 neurons, and finally by the output layer containing 5 neurons, which outputs the recognition results of the five types of partial discharges: suspended potential discharge, metal particle discharge, tip discharge, surface discharge and hollow conductor discharge.

[0151] Specifically, the improved three-dimensional convolutional neural network model of the present invention further includes a training process before being actually deployed and used, as follows:

[0152] Step 1: Data collection and preprocessing

[0153] The partial discharge signal of GIS equipment is collected by UHF sensor, with the frequency range of 300MHz-1500MHz and the dynamic range of -80dBm to -10dBm. The collected original signal is amplified by low noise amplifier, then filtered by combined filter (including high pass, low pass and band stop filter), and finally the partial discharge pulse envelope signal is obtained by detector.

[0154] Step 2: PRPS spectrum generation

[0155] The pre-processed signal is sampled at high speed using a 50MHz sampling rate and a 12-bit ADC. The sampled data is first processed in real time by the FPGA, including peak detection, data calibration, and preliminary feature extraction, to generate N-point PRPS raw data within 1 second. The ARM processor then performs data analysis and calculations to calculate the PD phase, frequency, and amplitude characteristic values, and finally generates a real-time PRPS three-dimensional map containing three-dimensional information of phase, amplitude, and frequency.

[0156] Step 3: Intelligent noise reduction processing

[0157] The generated PRPS spectrum is subjected to intelligent noise reduction processing based on weighted adaptation.

[0158] Step 4: Dataset Division

[0159] The processed PRPS atlas dataset is divided into training set, validation set and test set according to reasonable proportions. Each data sample contains complete three-dimensional feature information and is labeled with the corresponding discharge type.

[0160] Step 5: Model construction and initialization

[0161] Build an improved 3D convolutional neural network model, including:

[0162] Input layer: receives PRPS atlas data; phase block convolution layer: three groups of 3×3×3 convolution kernels for different phase areas; batch normalization layer and ReLU activation function layer; three-dimensional maximum pooling layer: 2×2×2 pooling kernel; feature fusion layer: includes attention mechanism; classification layer: fully connected layer of 2048->512->5 neurons.

[0163] Step 6: Initialize the weight matrix

[0164] Initialize the phase weight matrix, divide the 360° phase range into primary, secondary and other phase regions, and set the corresponding weight coefficients. At the same time, initialize the amplitude-frequency modulation matrix to adjust the importance of different feature dimensions.

[0165] Step 7: Model training process

[0166] Perform iterative training, where each iteration includes: batch data input, forward propagation calculation, calculation of loss function, back propagation parameter update, and regular evaluation of model performance on the validation set.

[0167] Step 8: Model evaluation and optimization

[0168] Regularly evaluate the model performance on the validation set, including indicators such as recognition accuracy and loss value change. According to the evaluation results, adjust the network hyperparameters, optimize the phase weight matrix parameters, and adjust the network structure if necessary until the model performance meets the requirements.

[0169] Step 9: Model testing and deployment

[0170] The final evaluation is conducted on the test set to verify the model's ability to identify five types of discharges (suspension potential discharge, metal particle discharge, tip discharge, surface discharge, and hollow conductor discharge). After confirming that the model performance meets the requirements, the model is deployed in the actual application environment.

[0171] The main improvement of the improved three-dimensional convolutional neural network model provided by the present invention is that the design mechanism of phase block and weight matrix is ​​introduced. The method achieves accurate division and weighted processing of the importance of discharge phases, and combines the amplitude-frequency modulation matrix R(A,f) for feature modulation. This design significantly improves the model's ability to extract features of important phase regions. In addition, through the introduction of the attention mechanism, the effective fusion of time domain and frequency domain features is achieved, further improving the model's feature expression ability. These improvements enable the model to more accurately identify different types of partial discharges and improve the accuracy and reliability of classification.

[0172] Specifically, in one embodiment of the present invention, before collecting the local discharge signal of the GIS equipment through the UHF sensor, the method also includes a phase synchronization control step to provide an accurate phase reference benchmark for the PRPS spectrum, and the phase synchronization control adopts one of PT synchronization, AC220V synchronization and external signal source synchronization.

[0173] Phase synchronization control is the key prerequisite for the generation of PRPS spectra. Its purpose is to provide an accurate phase reference for the partial discharge signal, ensure that the discharge signal corresponds accurately to the power frequency cycle, and thus ensure the accuracy of the phase information in the PRPS spectrum.

[0174] like Figure 4 As shown in the figure, the complete process of phase synchronization control and signal acquisition is as follows: 1. First, select the appropriate synchronization mode. 2. Perform corresponding processing on the corresponding synchronization signal. 3. Generate a unified synchronization trigger signal. 4. Control the acquisition of partial discharge signals. 5. Finally, generate the PRPS spectrum.

[0175] Specifically, the PT synchronization method uses the power frequency voltage signal output by the voltage transformer (PT) as the synchronization reference. Implementation steps: obtain the power frequency voltage signal through the PT, step down and isolate the PT output signal, obtain the zero-crossing information through the zero-point detection circuit, and generate a synchronization trigger signal.

[0176] The AC220V synchronization mode directly uses the 220V industrial frequency power supply as the synchronization reference. Implementation steps: step down the 220V AC power, perform electrical isolation through an isolation transformer, use a shaping circuit to obtain a standard square wave, and detect the zero crossing point to generate a synchronization signal.

[0177] The external signal source synchronization method uses an external standard signal source to provide a synchronization reference. Implementation steps: Connect an external 50Hz standard signal source, process it through a signal conditioning circuit, and convert it into a standard trigger signal.

[0178] Specifically, in one embodiment of the present invention, after performing intelligent noise reduction processing on the PRPS spectrum, the method further includes an event triggering step, screening interference signals through a dual threshold judgment mechanism, and the event triggering step is as follows:

[0179] Set event trigger amplitude threshold and trigger number threshold;

[0180] The amplitude of each point of the PRPS signal after noise reduction is compared with the event trigger amplitude threshold;

[0181] When the trigger times threshold is met, the partial discharge alarm is triggered.

[0182] like Figure 5 As shown in the figure, the event triggering mechanism process based on double thresholds is as follows:

[0183] 1. First, set the two thresholds of amplitude and frequency.

[0184] Event trigger amplitude threshold setting: set based on the dynamic range of the device (-80dBm to -10dBm), consider the environmental noise level, and reserve a reasonable margin. Trigger count threshold setting: set a reasonable counting period based on actual discharge characteristics and consider the probability of false triggering.

[0185] 2. Determine the amplitude of the PRPS signal after noise reduction.

[0186] The amplitude of the PRPS signal after noise reduction is obtained, compared with the set amplitude threshold, and events exceeding the threshold are recorded.

[0187] 3. Determine the number of times.

[0188] Count the over-threshold events within the set time window and compare them with the trigger count threshold to determine whether the trigger condition is met.

[0189] 4. Alarm triggering.

[0190] When both the amplitude and frequency conditions are met, the partial discharge alarm signal is triggered and the alarm information is output.

[0191] The present invention adopts a dual-threshold event trigger mechanism, and performs multi-dimensional judgment on the noise-reduced PRPS signal by setting the event trigger amplitude threshold and the trigger number threshold at the same time. This mechanism first filters low-intensity signals through the amplitude threshold, and then confirms the continuous discharge phenomenon through the trigger number threshold. Only when these two conditions are met at the same time will the partial discharge alarm be triggered. This dual protection mechanism effectively avoids false triggering due to instantaneous interference or random noise, and significantly improves the credibility of the partial discharge alarm. At the same time, the mechanism also supports dynamic adjustment of the threshold, and can flexibly set the trigger conditions according to different operating environments and equipment states, realizing the intelligence and personalization of partial discharge monitoring.

[0192] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying partial discharge patterns of GIS equipment based on deep learning, characterized in that: include: Collect partial discharge signals of GIS equipment through UHF sensors and perform signal conditioning; Digitally sample and process the conditioned partial discharge signal to generate a phase-resolved pulse sequence PRPS spectrum, and perform intelligent noise reduction on the PRPS spectrum, wherein the PRPS spectrum contains three-dimensional information of phase, amplitude and frequency; Input the PRPS spectrum into the preset neural network model to extract the spectrum features; The partial discharge type is identified and classified based on the extracted spectrum features, and the identification results are output.

2. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 1, characterized in that: Signal conditioning includes: Signal amplification through a low noise amplifier; Filtering the signal by a combination filter, wherein the combination filter includes a high-pass filter, a low-pass filter and a band-stop filter; The partial discharge pulse envelope signal is obtained through the detector.

3. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 1, characterized in that: Digital sampling and processing of the conditioned partial discharge signal includes: The conditioned partial discharge signal is sampled at high speed using a 50MHz sampling rate and 12-bit ADC. Real-time data processing is performed through FPGA, including: peak detection of sampled data; data calibration and preliminary feature extraction; generation of N-point PRPS raw data within 1 second; Data analysis and calculation are performed through the ARM processor, including: obtaining the PRPS raw data processed by the FPGA; calculating the characteristic values ​​of partial discharge phase, frequency, and amplitude; generating real-time PRPS three-dimensional spectrum data, in which: the phase is aligned with the 50Hz synchronization signal; the amplitude reflects the discharge intensity; and the frequency represents the distribution of discharge times.

4. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 1, characterized in that: Intelligent noise reduction processing uses PRPS-based weighted adaptive noise reduction, and its noise reduction processing uses the following formula: In the formula, represents phase, A represents amplitude, f represents frequency, is the PRPS data after noise reduction, is the original PRPS data, is the three-dimensional noise reduction factor, defined as: in, is the phase weight coefficient, is the reference phase point, is the phase attenuation coefficient, is the phase response function, k2 is the response coefficient, which is used to adjust the phase sensitivity; K A (A) is the amplitude modulation coefficient, E(t) is the normalized value of the environmental noise sensor, k1 is the noise suppression coefficient, 0<k1<1, At h is the amplitude judgment threshold, which is set based on the range of -80dBm to -10dBm, α is the amplitude modulation index, 0<α<1; K f (f) is the frequency weight coefficient, f min is the minimum effective frequency threshold, f max is the maximum reference frequency.

5. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 1, characterized in that: The preset neural network model is an improved three-dimensional convolutional neural network model, which specifically includes: Input layer, receiving PRPS spectrum data; The phase block convolution layer includes three groups of 3×3×3 convolution kernels with a step size of 1, and each group of convolution kernels corresponds to a different phase region; Batch normalization layer and ReLU activation function layer are used to standardize feature distribution and increase the nonlinear expression ability of the network; The three-dimensional maximum pooling layer has a pooling kernel size of 2×2×2 and a step size of 2, which is used to extract local salient features; Feature fusion layer, including attention mechanism, is used to fuse time domain features and frequency domain features; The classification layer includes two fully connected layers and an output layer. The first fully connected layer has 2048 neurons, the second fully connected layer has 512 neurons, and the output layer has 5 neurons, corresponding to five types of local discharges: suspended potential discharge; metal particle discharge; tip discharge; surface discharge; and hollow conductor discharge.

6. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 5, characterized in that: The phase block convolution layer includes a phase weight matrix, which divides the 360° phase range into a main discharge phase region, a secondary discharge phase region, and other phase regions according to the importance of the discharge phase, and assigns different weight coefficients to each of them.

7. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 6, characterized in that: The expression of the phase weight matrix is ​​as follows: in, is the phase weight matrix, is the phase weight basic matrix, R(A,f) is the amplitude-frequency modulation matrix, represents a matrix multiplication operator; W1 is the basic weight of the main discharge phase region, W2 is the basic weight of the secondary discharge phase region, and W3 is the basic weight of other phase regions; It is the boundary of the main discharge phase area; is the boundary of the secondary discharge phase area; A ref is the reference amplitude; f ref is the reference frequency; σ A is the amplitude standard deviation; σ f is the frequency standard deviation.

8. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 1, characterized in that: Before collecting the partial discharge signal of the GIS equipment through the UHF sensor, the method also includes a phase synchronization control step to provide an accurate phase reference benchmark for the PRPS spectrum. The phase synchronization control adopts one of PT synchronization, AC220V synchronization and external signal source synchronization.

9. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 1, characterized in that: After performing intelligent noise reduction processing on the PRPS spectrum, the method further includes an event triggering step, in which interference signals are screened through a dual threshold judgment mechanism. The event triggering step is as follows: Set event trigger amplitude threshold and trigger number threshold; The amplitude of each point of the PRPS signal after noise reduction is compared with the event trigger amplitude threshold; When the trigger times threshold is met, the partial discharge alarm is triggered.

10. The method for identifying partial discharge patterns of GIS equipment based on deep learning according to claim 1, characterized in that: The detection frequency range of the UHF sensor is 300MHz-1500MHz, and the dynamic range is -80dBm to -10dBm.

Images