A high-voltage equipment partial discharge signal real-time detection method and system

Abstract

This invention relates to the field of power systems and high voltage technology, and particularly to a method and system for real-time detection of partial discharge signals in high-voltage equipment. The method includes: preprocessing a high-frequency pulse sequence based on the power frequency phase signal to obtain a standardized discharge pulse sequence; extracting a time-domain pulse feature map and a frequency-domain resonance feature map from the standardized discharge pulse sequence to generate a dual-channel discharge feature map; inputting the dual-channel discharge feature map into a lightweight convolutional neural network to obtain a preliminary discharge type probability distribution; and generating a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the signal-to-noise ratio parameter. This invention enables high-precision, high-real-time detection of partial discharge signals in high-voltage equipment on resource-constrained edge devices, ultimately accurately outputting the discharge type identification result.

Description

Technical Field

[0001] This invention relates to the field of power systems and high voltage technology, and in particular to a method and system for real-time detection of partial discharge signals in high-voltage equipment. Background Technology

[0002] With the development of smart devices, partial discharge pattern recognition technology based on deep learning has become a research hotspot. Existing technologies typically employ the following process: high-frequency current or electromagnetic wave signals are collected through sensors, and after preprocessing such as noise reduction and segmentation, they are fed into a convolutional neural network for end-to-end classification.

[0003] Partial discharge signals from high-voltage equipment inherently contain multiple feature scales: they include microsecond-level fine pulse waveform features reflecting the physical process of discharge, as well as millisecond-level macroscopic pulse sequence patterns reflecting the repetitive patterns of discharge. Existing lightweight convolutional neural networks mainly balance accuracy and speed by stacking small-sized convolutional kernels. However, under the constraints of parameters and computational cost, their effective receptive field is usually shallow, making it difficult to capture long-range temporal dependencies and macroscopic discharge cycle characteristics in the signal. This causes the model to focus more on learning the local morphological features of the signal, while ignoring more discriminative macroscopic patterns such as power frequency phase correlation spanning hundreds or even thousands of sampling points. Therefore, when faced with complex field noise backgrounds or similar discharge types, the model's discriminative ability and robustness still need to be improved. How to enable lightweight convolutional neural networks to fully perceive the multi-scale features of partial discharge on resource-constrained edge devices and provide reliable confidence assessments of their recognition results has become a core problem that urgently needs to be solved in the practical application of this technology. Summary of the Invention

[0004] The main objective of this invention is to provide a method and system for real-time detection of partial discharge signals in high-voltage equipment, aiming to solve the technical problems mentioned in the background art.

[0005] This invention proposes a method for real-time detection of partial discharge signals in high-voltage equipment, comprising:

[0006] The high-frequency pulse sequence and power frequency phase signal of partial discharge of high-voltage equipment are acquired synchronously. The high-frequency pulse sequence is subjected to adaptive noise reduction and pulse alignment preprocessing based on the power frequency phase signal to obtain a standardized discharge pulse sequence.

[0007] Based on the standardized discharge pulse sequence, time-domain pulse feature map and frequency-domain resonance feature map are extracted and fused to generate a dual-channel discharge feature map;

[0008] The dual-channel discharge feature map is input into a lightweight convolutional neural network, which includes a parallel multi-branch structure and a channel-time dual attention mechanism. The parallel multi-branch structure synchronously extracts the micro-waveform features and macro-sequence pattern features of the discharge pulse, and performs adaptive fusion through the channel-time dual attention mechanism. Then, it is processed by its classifier to obtain the preliminary discharge type probability distribution.

[0009] Obtain the signal-to-noise ratio (SNR) parameter of the standardized discharge pulse sequence, and generate a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the SNR parameter;

[0010] The confidence level of the comprehensive type determination is compared with a preset confidence threshold to determine and output the final discharge type identification result;

[0011] For the final discharge type identification result where the confidence level of the comprehensive type determination is lower than the confidence threshold, the corresponding high-frequency pulse sequence is marked as a suspected discharge sample and uploaded to the cloud analysis platform for verification.

[0012] Preferably, the step of synchronously acquiring the high-frequency pulse sequence and power frequency phase signal of partial discharge from high-voltage equipment, and performing adaptive noise reduction and pulse alignment preprocessing on the high-frequency pulse sequence based on the power frequency phase signal to obtain a standardized discharge pulse sequence includes:

[0013] The high-frequency pulse sequence of partial discharge of high-voltage equipment and the power frequency phase signal are acquired synchronously. The phase reference of the power frequency period is determined according to the zero crossing point. Each pulse in the high-frequency pulse sequence is mapped to a complete phase interval that matches the power frequency period according to the phase reference to generate a phase-resolved pulse sequence.

[0014] Based on the phase-resolved pulse sequence, the statistical distribution of pulse amplitude is calculated within a preset power frequency cycle characteristic phase segment, and a dynamic noise threshold synchronized with the power frequency cycle is generated based on the statistical distribution.

[0015] Based on the dynamic noise threshold, the phase-resolved pulse sequence is decomposed into wavelet packets to obtain wavelet coefficients for multiple sub-bands.

[0016] Calculate the corresponding coefficient screening threshold based on the energy proportion of each sub-band, and perform soft threshold denoising on the wavelet coefficients based on the coefficient screening thresholds to obtain denoised wavelet coefficients. Then, perform wavelet packet reconstruction on the denoised wavelet coefficients to obtain a denoised pulse sequence.

[0017] The pulse peak detection is performed on the denoised pulse sequence to obtain the position coordinates of all pulse peak points;

[0018] The statistical characteristics of the pulse interval are calculated based on the position coordinates of all pulse peak points. The pulse alignment reference point is determined based on the statistical characteristics. A signal segment of a preset length is extracted with each pulse alignment reference point as the center to obtain the standardized discharge pulse sequence.

[0019] Preferably, the step of extracting the time-domain pulse feature map and frequency-domain resonance feature map from the standardized discharge pulse sequence, and performing feature fusion to generate a dual-channel discharge feature map includes:

[0020] The standardized discharge pulse sequence is segmented into power frequency cycles, and the maximum amplitude, average amplitude, pulse repetition rate, and pulse phase distribution of the discharge pulses within each power frequency cycle are statistically analyzed and constructed into a time-domain statistical feature vector.

[0021] Obtain a preset typical discharge model library, which contains standard time-domain statistical features corresponding to various insulation defect types. Calculate the Mahalanobis distance between the time-domain statistical feature vector and each of the standard time-domain statistical features, and reconstruct all Mahalanobis distance values ​​into a time-domain pulse feature map.

[0022] The standardized discharge pulse sequence is subjected to S-transform to obtain its high-resolution time spectrum, and the energy distribution of the partial discharge characteristic frequency band is extracted from the high-resolution time spectrum according to the preset insulation defect sensitive frequency band range.

[0023] Calculate the quality factor Q, bandwidth, and center frequency drift of the partial discharge characteristic frequency band, construct the frequency domain resonance characteristic vector, and reconstruct it into a frequency domain resonance characteristic map;

[0024] The time-domain pulse feature map and the frequency-domain resonance feature map are spliced ​​and normalized to generate a dual-channel discharge feature map.

[0025] Preferably, the step of inputting the dual-channel discharge feature map into a lightweight convolutional neural network, which includes a parallel multi-branch structure and a channel-time dual attention mechanism, and simultaneously extracting the microscopic waveform features and macroscopic sequence pattern features of the discharge pulse through the parallel multi-branch structure, and adaptively fusing them through the channel-time dual attention mechanism to obtain a preliminary discharge type probability distribution includes:

[0026] The dual-channel discharge feature map is input into the micro-feature branch of the parallel multi-branch structure, and the micro-waveform features of the discharge pulse are extracted by the depth of the first void ratio in the micro-feature branch through separable convolution.

[0027] The dual-channel discharge feature map is synchronously input to the macroscopic feature branch of the parallel multi-branch structure. Macroscopic sequence pattern features spanning multiple power frequency cycles are extracted by depth-separable convolution of the second hole rate in the macroscopic feature branch, wherein the second hole rate is greater than the first hole rate.

[0028] The microscopic waveform features are concatenated with the macroscopic sequence pattern features to obtain multi-scale fusion features;

[0029] The multi-scale fusion features are input into the channel-time dual attention module. The channel attention submodule in the channel-time dual attention module evaluates the relative importance of the micro waveform features and the macro sequence pattern features. The time attention submodule focuses on the power frequency cycle feature phase segment where the key discharge event occurs, and obtains the attention-weighted optimized features.

[0030] The optimized features are passed to a preset classifier, which then outputs the initial discharge type probability distribution.

[0031] Preferably, the step of obtaining the signal-to-noise ratio (SNR) parameter of the standardized discharge pulse sequence and generating a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the SNR parameter includes:

[0032] Extract the maximum probability value and the second highest probability value from the preliminary discharge type probability distribution, and calculate the difference between the maximum probability value and the second highest probability value as the probability confidence factor;

[0033] Based on the discharge type with the highest probability in the preliminary discharge type probability distribution, query the preset discharge type-feature frequency band mapping library to determine the main discharge frequency band corresponding to the discharge type;

[0034] The ratio of signal energy within the main discharge frequency band to the total frequency band energy is calculated, and the weighted signal-to-noise ratio parameter is obtained by combining the concentration of the standardized discharge pulse sequence in the characteristic phase segment of the power frequency cycle.

[0035] The confidence level of the comprehensive type determination is obtained based on the signal-to-noise ratio parameter and the probability confidence factor.

[0036] Preferably, the step of calculating the ratio of signal energy within the main discharge frequency band to the total frequency band energy, and combining this with the concentration of the standardized discharge pulse sequence in the characteristic phase segment of the power frequency cycle, to calculate the weighted signal-to-noise ratio parameter includes:

[0037] Perform a Fast Fourier Transform on the standardized discharge pulse sequence to obtain the frequency amplitude spectrum;

[0038] In the frequency amplitude spectrum, multiple frequency band regions outside the main discharge frequency band and with uniform energy distribution are defined as reference noise frequency bands;

[0039] The sum of squares of the signal amplitudes within the main discharge frequency band is calculated as the main discharge frequency band energy, and the average value of the sum of squares of the signal amplitudes within the reference noise frequency band is calculated as the reference noise energy.

[0040] The reference signal-to-noise ratio is calculated based on the ratio of the main discharge frequency band energy to the reference noise energy.

[0041] The proportion of the number of pulses in the characteristic phase segment of the power frequency cycle of the standardized discharge pulse sequence to the total number of pulses is calculated as the phase concentration.

[0042] The weighted signal-to-noise ratio parameter is obtained based on the reference signal-to-noise ratio and the phase convergence.

[0043] The present invention also provides a real-time detection system for partial discharge signals of high-voltage equipment, comprising multiple modules, which are used to implement the steps of a real-time detection method for partial discharge signals of high-voltage equipment.

[0044] Preferably, the module includes multiple units, which are used to implement the steps of a method for real-time detection of partial discharge signals in high-voltage equipment.

[0045] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a method for real-time detection of partial discharge signals of high-voltage equipment.

[0046] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for real-time detection of partial discharge signals in high-voltage equipment.

[0047] The beneficial effects of this invention are as follows: By synchronously acquiring high-frequency pulse sequences of partial discharge from high-voltage equipment and power frequency phase signals, and through adaptive noise reduction, feature extraction, lightweight neural network inference, confidence assessment, and verification of suspected samples, this invention not only fully utilizes the physical characteristics of the discharge signal to improve feature discrimination, but also achieves multi-scale feature capture under the premise of lightweight through network structure optimization. On resource-constrained edge devices, it can achieve high-precision and high-real-time detection of partial discharge signals from high-voltage equipment, and finally accurately output the discharge type identification result. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0049] Figure 2This is a schematic diagram of the system structure according to an embodiment of this application.

[0050] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0051] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0052] like Figure 1 As shown, this application provides a method for real-time detection of partial discharge signals in high-voltage equipment, including:

[0053] S1, synchronously acquire the high-frequency pulse sequence and power frequency phase signal of partial discharge of high-voltage equipment, and perform adaptive noise reduction and pulse alignment preprocessing on the high-frequency pulse sequence according to the power frequency phase signal to obtain a standardized discharge pulse sequence;

[0054] S2, extract time-domain pulse feature map and frequency-domain resonance feature map according to the standardized discharge pulse sequence, and perform feature fusion to generate dual-channel discharge feature map;

[0055] S3, the dual-channel discharge feature map is input into a lightweight convolutional neural network. This lightweight convolutional neural network includes a parallel multi-branch structure and a channel-time dual attention mechanism. The parallel multi-branch structure synchronously extracts the microscopic waveform features and macroscopic sequence pattern features of the discharge pulse, and these are adaptively fused via the channel-time dual attention mechanism. The resulting data is then processed by its classifier to obtain a preliminary discharge type probability distribution. The lightweight convolutional neural network is deployed and inferred through the following steps: the trained network model parameters are converted to TensorFlow Lite or ONNX format and deployed on an edge computing device (such as an embedded GPU or high-performance MCU). During inference, the dual-channel discharge feature map is used as input, and the neural network inference engine is called to perform forward computation to obtain the preliminary discharge type probability distribution. The edge computing device acquires signals in real time through its high-speed data interface and utilizes its parallel computing capabilities to ensure the real-time performance of the detection process.

[0056] S4, obtain the signal-to-noise ratio (SNR) parameter of the standardized discharge pulse sequence, and generate a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the SNR parameter;

[0057] S5, compare the comprehensive type determination confidence level with the preset confidence threshold, determine and output the final discharge type identification result;

[0058] S6. For the final discharge type identification result where the confidence level of the comprehensive type determination is lower than the confidence threshold, the corresponding high-frequency pulse sequence is marked as a suspected discharge sample and uploaded to the cloud analysis platform for verification.

[0059] As described in steps S1-S6 above, the partial discharge signal of high-voltage equipment has significant multi-scale characteristics and strong phase correlation: on the one hand, the pulse signal generated during the discharge process can reach the microsecond to nanosecond level in time domain width, and its micro-waveform details such as rising edge, peak value, and falling edge directly reflect the physical characteristics of the discharge source. For example, the pulse rising edge of internal air gap discharge is relatively slow, and the pulse width of surface discharge is relatively narrow. On the other hand, the discharge pulse exhibits a periodic distribution within the power frequency cycle. The discharge of different types of insulation defects will be concentrated in specific power frequency phase segments, and the macroscopic sequence patterns such as pulse repetition rate and phase concentration are closely related to the defect type and severity. For example, floating potential discharge is mostly concentrated near the voltage peak (around 90° and 270°), while corona discharge is mostly distributed at the voltage rising and falling edges. At the same time, the operating environment of high-voltage equipment is complex, and there are a large number of periodic interferences (such as power electronic switch noise) and random noises on site. This can mask the true discharge signal, resulting in a low signal-to-noise ratio of the original acquired signal, directly affecting detection accuracy. Furthermore, edge-side devices are limited by hardware resources and cannot support the high computing power requirements of large neural networks. Traditional lightweight models, due to their limited receptive field, struggle to simultaneously capture the microscopic waveforms and macroscopic sequence features of the discharge signal. This makes achieving high-precision partial discharge detection at the edge a pressing technical problem. This invention, by simultaneously acquiring high-frequency pulse sequences of partial discharge from high-voltage equipment and power frequency phase signals, and through adaptive noise reduction, feature extraction, lightweight neural network inference, confidence assessment, and verification of suspected samples, fully utilizes the physical characteristics of the discharge signal to enhance feature discrimination. It also achieves multi-scale feature capture under lightweight conditions through network structure optimization. On resource-constrained edge-side devices, it can achieve high-precision and high-real-time detection of partial discharge signals from high-voltage equipment, ultimately outputting accurate discharge type identification results.

[0060] The steps for processing low-confidence identification results include: determining that identification results with a comprehensive type determination confidence level below a preset threshold are low-confidence results; obtaining the original high-frequency pulse sequence corresponding to the result, along with its timestamp and device identification information; packaging the original high-frequency pulse sequence, timestamp, device identification information, and the corresponding preliminary discharge type probability distribution and comprehensive type determination confidence level to generate a suspected discharge sample data package; uploading the suspected discharge sample data package to a cloud analysis platform via a communication network; the cloud platform can call more complex models or experts for verification and use confirmation tags for incremental model learning; by setting a confidence threshold and performing special processing on low-confidence results, it is possible to intercept and process identification results with high uncertainty, prevent edge devices from making incorrect decisions when information is insufficient, and improve the reliability of the system in long-term operation of unattended substations.

[0061] In one embodiment of the present invention, the step of synchronously acquiring the high-frequency pulse sequence and power frequency phase signal of partial discharge of high-voltage equipment, and performing adaptive noise reduction and pulse alignment preprocessing on the high-frequency pulse sequence based on the power frequency phase signal to obtain a standardized discharge pulse sequence includes:

[0062] S11, synchronously acquire the high-frequency pulse sequence of partial discharge of high-voltage equipment and the power frequency phase signal, determine the phase reference of the power frequency cycle based on the zero crossing point, and map each pulse in the high-frequency pulse sequence to a complete phase interval (i.e., the power frequency cycle phase interval) that matches the power frequency cycle based on the phase reference, thereby generating a phase-resolved pulse sequence.

[0063] S12, Based on the phase-resolved pulse sequence, calculate the statistical distribution of pulse amplitude within a preset power frequency cycle characteristic phase segment, and generate a dynamic noise threshold synchronized with the power frequency cycle based on the statistical distribution;

[0064] S13, perform wavelet packet decomposition on the phase-resolved pulse sequence based on the dynamic noise threshold to obtain wavelet coefficients of multiple sub-bands;

[0065] S14, calculate the corresponding coefficient screening threshold according to the energy ratio of each sub-band, and perform soft threshold denoising on the wavelet coefficients based on the coefficient screening thresholds to obtain denoised wavelet coefficients, and perform wavelet packet reconstruction on the denoised wavelet coefficients to obtain a denoised pulse sequence.

[0066] S15, perform pulse peak detection on the noise-reduced pulse sequence to obtain the position coordinates of all pulse peak points;

[0067] S16. Calculate the statistical characteristics of the pulse interval based on the position coordinates of all pulse peak points, determine the pulse alignment reference point based on the statistical characteristics, and extract a signal segment of a preset length centered on each pulse alignment reference point to obtain the standardized discharge pulse sequence.

[0068] As described in steps S11-S16 above, since high-voltage equipment operates in a power frequency AC environment, the occurrence of partial discharge has a clear dependence on the phase change of the power frequency voltage: different types of insulation defects (such as floating potential and surface discharge) will have discharge pulses that stably appear in specific phase intervals within the power frequency cycle. For example, corona discharge is mostly concentrated near the voltage peak (around 90° and 270°), while surface discharge is more likely to occur at the voltage rising and falling edges. At the same time, there are a large number of periodic interferences (such as switching noise of power electronic equipment) and random noise in the field environment. These noises may overlap with the real discharge pulses in the time domain, but they differ from the real discharge in phase distribution. The phase distribution of the real discharge pulses is clustered, while the phase distribution of the interference signal is more random. Based on this physical characteristic, if only high-frequency pulse sequences are processed in a single time domain, it is difficult to effectively distinguish between real discharge and interference. Moreover, the pulses at different acquisition times have differences in time position and length, which will lead to inconsistent benchmarks for subsequent feature extraction and affect the detection accuracy.

[0069] This invention elevates partial discharge analysis from a purely time-domain analysis to a two-dimensional analysis level correlated with the power frequency voltage phase by synchronously acquiring power frequency phase signals and generating phase-resolved pulse sequences. It directly addresses the core physical mechanism that partial discharge is generated under power frequency voltage excitation. By utilizing the physical characteristics of partial discharge occurring within specific phase intervals of the power frequency voltage, it provides a physically meaningful phase reference for subsequent processing, effectively distinguishing between periodic interference and genuine discharge. Furthermore, by generating dynamic noise thresholds based on the statistical characteristics of the phase-resolved pulse sequences within characteristic phase segments, it achieves localized adaptive noise assessment. This allows for the use of lenient thresholds to protect the true signal in high-discharge-occurrence areas and strict thresholds to suppress noise in non-critical areas, demonstrating adaptability to complex substation electromagnetic environments. Better; by performing frequency domain decomposition and adaptive threshold denoising through wavelet packet decomposition, soft threshold denoising, and reconstruction steps, fine noise separation can be achieved across the entire frequency band, maximizing the preservation of high-frequency resonant components of discharge pulses characterizing different insulation defects, thus avoiding the loss of feature information that may occur with general filtering methods; by calculating the statistical characteristics of pulse intervals, determining alignment reference points, and truncating signal segments for pulse alignment and truncation, random phase jitter of pulses is eliminated, making the subsequently extracted pulse repetition rate, interval, and other time-domain features stable and comparable, laying a solid foundation for neural network recognition of the periodic pattern of discharges; compared with traditional preprocessing methods, this scheme can improve the signal-to-noise ratio of the effective signal through the above steps, providing a cleaner and more regular data input for high-precision recognition.

[0070] The preset power frequency cycle characteristic phase segment is pre-set according to the physical characteristics of different types of partial discharge, such as the phase interval near the power frequency voltage peak such as 0°-90° and 180°-270°; the calculation formula of the dynamic noise threshold is: take the mean of the pulse amplitude within the characteristic phase segment, add a coefficient adjusted according to the on-site signal-to-noise ratio level and the product of the standard deviation of the pulse amplitude within the phase segment; the method for determining the pulse alignment reference point is: calculate the mode or average of the time interval of all pulse peak points, and adjust the time position of each pulse peak point to the grid point based on its integer multiple of the theoretical interval time, so as to achieve phase synchronization alignment of the pulse sequence;

[0071] In one embodiment of the present invention, the step of extracting the time-domain pulse feature map and the frequency-domain resonance feature map from the standardized discharge pulse sequence, and performing feature fusion to generate a dual-channel discharge feature map includes:

[0072] S21, the standardized discharge pulse sequence is segmented into power frequency cycles, and the maximum amplitude, average amplitude, pulse repetition rate and pulse phase distribution of the discharge pulse in each power frequency cycle are statistically analyzed and constructed into a time-domain statistical feature vector.

[0073] S22, Obtain a preset typical discharge model library, which contains standard time-domain statistical features corresponding to various insulation defect types, calculate the Mahalanobis distance between the time-domain statistical feature vector and each of the standard time-domain statistical features, and reconstruct all Mahalanobis distance values ​​into a time-domain pulse feature map;

[0074] S23, Perform an S-transform on the standardized discharge pulse sequence to obtain its high-resolution time spectrum, and extract the energy distribution of the partial discharge characteristic frequency band from the high-resolution time spectrum according to the preset insulation defect sensitive frequency band range;

[0075] S24, calculate the quality factor Q, bandwidth and center frequency drift of the partial discharge characteristic frequency band, construct the frequency domain resonance characteristic vector, and reconstruct it into a frequency domain resonance characteristic map;

[0076] S25, the time-domain pulse feature map and the frequency-domain resonance feature map are spliced ​​and normalized to generate a dual-channel discharge feature map.

[0077] As described in steps S21-S25 above, the physical processes of partial discharge caused by different types of insulation defects (such as floating potential, surface discharge, and internal air gaps) differ significantly, and these differences are directly reflected in the time and frequency domain characteristics of the discharge signal. In the time domain, the amplitude distribution, repetition frequency, and phase concentration of the discharge pulse are strongly correlated with the defect type: for example, due to the stability of air gap breakdown, internal air gap discharge has a high pulse repetition rate and its phase is concentrated in the 0-90° and 180-270° range of the power frequency cycle; surface discharge is affected by the uneven distribution of the surface electric field, resulting in large pulse amplitude fluctuations and a dispersed phase distribution. In the frequency domain, the electromagnetic wave signal generated during the discharge process will form resonance peaks in specific frequency bands due to differences in the properties of the insulating material (such as dielectric constant and loss angle) and the defect morphology (such as gap size and surface roughness). For example, the resonance energy of floating potential discharge is mostly concentrated in the 500MHz-800MHz range, while the resonance frequency band of internal discharge caused by insulation aging will shift to above 1GHz, and the quality factor Q (reflecting the sharpness of the resonance) will significantly decrease. Based on this physical law, it is difficult to fully characterize the defect characteristics by extracting only single-dimensional features. It is necessary to construct a feature system that is directly related to the type of insulation defect through the coordinated extraction and fusion of time-domain and frequency-domain features.

[0078] Therefore, this invention effectively captures the amplitude variation, repetition characteristics, and phase distribution patterns of discharge pulses under power frequency voltage excitation by segmenting statistical time-domain features according to the power frequency cycle. These statistics are key time-domain fingerprints for distinguishing different types of defects such as floating potential discharge, surface discharge, and internal air gap discharge. By introducing a pre-set library of typical discharge models and calculating Mahalanobis distance, the time-domain features of the signal under test are quantitatively compared with the physical models of known typical defects. This not only transforms abstract statistics into feature maps with spatial structures to adapt to convolutional neural network processing, but also deeply encodes domain expert knowledge of high-voltage insulation diagnosis into the features, thereby enhancing the discriminative power and physical interpretability of the features. This solves the problem of general CNNs automatically learning feature discriminative power from raw signals. The system addresses shortcomings and poor interpretability. By obtaining high-resolution time-frequency spectra through S-transform and focusing on the sensitive frequency band of insulation defects, it can reveal the frequency domain resonance characteristics of discharge pulses. These characteristics are closely related to the physical structure of the discharge source and the aging state of the insulation material, providing another dimension of identification basis that is different from time-domain information. By calculating deep-level frequency domain parameters such as quality factor and frequency drift, the sharpness and stability of resonance modes can be quantified, which is particularly sensitive to identifying progressive defects such as insulation aging and carbonization. By splicing and fusing the time-domain pulse feature map and the frequency-domain resonance feature map, a complementary dual-channel input is created, enabling the neural network to reason from both statistical patterns and physical mechanisms, resulting in higher accuracy in distinguishing common similar discharge types in the field.

[0079] The typical discharge model library is pre-established through statistical analysis of a large number of known partial discharge samples. The library stores standard time-domain statistical feature vectors of typical defects such as floating potential discharge, surface discharge, and internal air gap discharge. The Mahalanobis distance is calculated by weighting the covariance matrix of the standard feature vectors in the library with the time-domain statistical feature vector of the signal under test. The sensitive frequency band for insulation defects is pre-defined based on the electromagnetic wave propagation characteristics of different types of insulation defects. For example, for discharges in transformer oil-paper insulation, the frequency band from 300MHz to 1.5GHz can be focused on. The quality factor Q is calculated by dividing the resonant center frequency by the -3dB bandwidth.

[0080] In one embodiment of the present invention, the step of inputting the dual-channel discharge feature map into a lightweight convolutional neural network, the lightweight convolutional neural network including a parallel multi-branch structure and a channel-temporal dual attention mechanism, synchronously extracting the microscopic waveform features and macroscopic sequence pattern features of the discharge pulse through the parallel multi-branch structure, and adaptively fusing them through the channel-temporal dual attention mechanism to obtain a preliminary discharge type probability distribution includes:

[0081] S31, the dual-channel discharge feature map is input to the micro-feature branch of the parallel multi-branch structure, and the micro-waveform features of the discharge pulse are extracted by the depth of the first void ratio in the micro-feature branch through separable convolution.

[0082] S32, the dual-channel discharge feature map is synchronously input to the macroscopic feature branch of the parallel multi-branch structure, and macroscopic sequence pattern features spanning multiple power frequency cycles are extracted by depth-separable convolution of the second hole rate in the macroscopic feature branch, wherein the second hole rate is greater than the first hole rate;

[0083] S33, the micro-waveform features are spliced ​​with the macro-sequence pattern features to obtain multi-scale fusion features;

[0084] S34, the multi-scale fusion features are input to the channel-time dual attention module. The channel attention submodule in the channel-time dual attention module evaluates the relative importance of the micro waveform features and the macro sequence pattern features. The time-series attention submodule focuses on the power frequency cycle feature phase segment where the key discharge event occurs to obtain the attention-weighted optimized features. The key discharge event includes the initial pulse, high amplitude pulse, feature pattern pulse group, abnormal evolution pulse, and high signal-to-noise ratio pulse interval.

[0085] S35, the optimized features are passed to a preset classifier, and the preliminary discharge type probability distribution is output through the preset classifier.

[0086] As described in steps S31-S35 above, the defect identification value of partial discharge signals in high-voltage equipment is contained in both microscopic and macroscopic dimensions. At the microscopic level, details such as the rising edge slope of a single discharge pulse and the degree of waveform distortion near the peak directly correspond to the physical form of the discharge source (e.g., the size of the internal air gap and the gap distance of the floating potential). For example, the rising edge of a 10μm air gap discharge pulse is approximately 0.5μs, while the rising edge of a 50μm air gap discharge is extended to 1.2μs. At the macroscopic level, patterns such as pulse repetition rate changes across multiple power frequency cycles and phase concentration interval migration reflect the activity and development trend of defects. For example, internal discharge caused by insulation aging will gradually expand from the initial "0-90° phase concentration". The development is characterized by "0-120° phase dispersion," and the repetition rate increases from 20 times / cycle to 50 times / cycle. These two dimensions of features complement each other and are indispensable: relying solely on microscopic features can easily lead to misjudgment of defect types due to similar waveforms, while relying solely on macroscopic features makes it difficult to distinguish between early minor defects and interference signals. Furthermore, the computing power limitations of edge devices (such as embedded terminals, whose computing power is usually less than 1 TOPS) require strict control over the model's parameter scale and computational load. Traditional large-scale neural networks (such as ResNet50 with over 25 million parameters) cannot meet real-time requirements, while conventional lightweight models (such as MobileNetV2), although having fewer parameters, have fixed receptive fields for a single convolutional path, making it difficult to simultaneously cover the scale requirements of both microscopic and macroscopic features.

[0087] Therefore, this invention addresses the core technical contradiction between the diversity of feature scales in partial discharge signals and the limited receptive field of lightweight models by designing a parallel multi-branch structure and employing depthwise separable convolutions with different hole rates. The microscopic feature branches focus on capturing nanosecond-level pulse waveform details characterizing the physical properties of the discharge source, while the macroscopic feature branches utilize the sparse sampling characteristics of dilated convolutions to expand the receptive field, specifically for capturing long-range temporal patterns such as pulse repetition rate and phase clustering, spanning hundreds or even thousands of sampling points necessary for different discharge types. This solves the problem that lightweight models, due to their shallow network depth, struggle to fully understand the multi-scale characteristics of partial discharges. Through a channel attention mechanism, the network can adaptively learn and weight the importance of different feature channels, dynamically... The optimized feature combination determines whether to rely more on waveform details or sequence patterns in specific scenarios, thereby improving feature utilization efficiency. Through a temporal attention mechanism, the network can automatically focus on the phase intervals of key discharge events such as the initial pulse, high-amplitude pulses, and characteristic pattern pulse groups within the power frequency cycle, effectively suppressing the interference of noise in non-critical phase intervals and enhancing the model's ability to perceive the phase characteristics of discharge patterns. The lightweight convolutional neural network design of this invention enables a lightweight model with very few parameters to simultaneously observe the microscopic waveforms and macroscopic patterns of discharge. Under the same computational complexity, its ability to capture macroscopic sequence patterns is higher than that of standard lightweight CNNs, thus meeting the dual requirements of real-time performance and accuracy for high-voltage equipment condition monitoring.

[0088] The first dilatation rate is set to 1 (i.e., standard convolution), focusing on extracting local details such as pulse rising edge and pulse width; the second dilatation rate is set to a value greater than 1 (e.g., 3 or 5), expanding the receptive field without increasing the number of parameters, and is used to capture long-range patterns such as pulse repetition rate and phase clustering; the channel attention submodule is implemented through compression-excitation operation, first performing global average pooling on the input features, then learning the weights of each channel through two fully connected layers and the ReLU activation function between them, and finally normalizing the weights to between 0 and 1 through the Sigmoid function, and multiplying them with the original features channel by channel; the temporal attention submodule is implemented through one-dimensional convolution and the Softmax function, its function is to generate weights along the time dimension, so that the network focuses on the key phase intervals where the discharge activity is intense; the preset classifier consists of a global average pooling layer and a fully connected layer, and finally outputs the probability of belonging to each category through the Softmax function.

[0089] The steps for evaluating the relative importance of the microscopic waveform features and the macroscopic sequence pattern features through the channel attention submodule in the channel-time dual attention module, and focusing on the power frequency cycle feature phase segment where the critical discharge event occurs through the time-time attention submodule, are as follows: First, the channel attention submodule takes the multi-scale fused features (including microscopic and macroscopic feature channels) as input, compresses the time-domain information of each channel through global average pooling to obtain the channel feature vector, then learns the channel weights (weight values ​​0-1, reflecting feature importance) through two layers of fully connected networks, and finally multiplies the channel weights with the multi-scale fused features channel by channel to complete the process. The importance of micro and macro features is evaluated and optimized. Then, the temporal attention submodule takes the channel-weighted features as input, compresses the channel information of each time point through global average pooling to obtain the temporal feature vector, and then sets a temporal feature threshold based on dynamic noise threshold to initially screen candidate key time intervals in combination with the definition of key discharge events (first pulse, high amplitude pulse, etc.). Subsequently, the temporal weights are learned through lightweight gated recurrent unit (GRU) (the weight of candidate intervals approaches 1, and the weight of non-candidate intervals approaches 0). Finally, the temporal weights are multiplied with the channel-weighted features time-by-time to focus on the power frequency cycle feature phase segment of the key discharge event.

[0090] In one embodiment of the present invention, the step of obtaining the signal-to-noise ratio (SNR) parameter of the standardized discharge pulse sequence and generating a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the SNR parameter includes:

[0091] S41, extract the maximum probability value and the second largest probability value from the preliminary discharge type probability distribution, and calculate the difference between the maximum probability value and the second largest probability value as the probability confidence factor.

[0092] S42, based on the discharge type corresponding to the highest probability in the preliminary discharge type probability distribution, query the preset discharge type-feature frequency band mapping library to determine the main discharge frequency band corresponding to the discharge type;

[0093] S43, calculate the ratio of signal energy in the main discharge frequency band to the total frequency band energy, and combine it with the concentration of the standardized discharge pulse sequence in the characteristic phase segment of the power frequency cycle to calculate the weighted signal-to-noise ratio parameter;

[0094] S44, Obtain the comprehensive type determination confidence level based on the signal-to-noise ratio parameter and the probability confidence factor, using the following formula:

[0095]

[0096] In the formula, This indicates the confidence level of the comprehensive type determination. This represents the probability confidence factor. This represents the preset balance weight coefficient. This represents the weighted signal-to-noise ratio parameter. This represents the normalization constant.

[0097] As described in steps S41-S44 above, since the preliminary discharge type probability distribution only reflects the degree of feature learning and matching of the lightweight convolutional neural network, and in the field detection of high-voltage equipment, signal quality (signal-to-noise ratio) directly affects the accuracy of feature extraction, even if the model outputs a high probability of a certain type of discharge, if the original signal is interfered with by strong noise (such as high-frequency interference from power electronic equipment in substations), resulting in a low proportion of effective signal in the standardized discharge pulse sequence, the reliability of the probability result is still questionable. Conversely, if the signal signal-to-noise ratio is high and the features are clear, even if the model probability is slightly lower, the result may be more credible. At the same time, different discharge types have different signal-to-noise ratio performance: for example, the signal amplitude of floating potential discharge is stable and the signal-to-noise ratio is high (usually >30dB), while the signal signal-to-noise ratio of surface discharge is affected by surface state fluctuations and fluctuates greatly (15-25dB). Relying solely on probability cannot distinguish between "low-probability credible results under high signal-to-noise ratio" and "high-probability unreliable results under low signal-to-noise ratio".

[0098] Therefore, this invention quantifies the uncertainty of the neural network model's own judgment by calculating the probability confidence factor, avoiding the risk of false alarms or missed alarms that arises easily when the signal quality is poor, as is the case with existing technologies that directly use the maximum probability of the network output. By querying the discharge type-feature frequency band mapping library to dynamically determine the main discharge frequency band, the signal-to-noise ratio (SNR) assessment is no longer global and blind, but rather a targeted assessment related to the currently identified discharge type. For example, the energy of corona discharge may be concentrated in the lower frequency band, while surface discharge may have more obvious resonance in the higher frequency band, which can improve the effectiveness of the SNR parameter in the specific task of partial discharge detection. By combining phase convergence to calculate the weighted SNR, the essential difference between real discharge and noise is utilized, allowing the SNR parameter to better distinguish between signal and noise, improving the robustness of the assessment in strong interference environments. By linearly weighting and fusing the probability confidence factor and the weighted SNR parameter to generate a comprehensive type determination confidence, the invention combines the model's judgment and the quality of the signal itself to form a more reliable evaluation index, resulting in a lower false alarm rate at the edge.

[0099] The probability confidence factor reflects the degree of certainty of the model's judgment; the larger the difference, the more confident the model is. The discharge type-feature frequency band mapping library is established in advance based on the typical spectral characteristics of various discharge types. The balance weight coefficient is used to adjust the relative proportion of the model probability confidence and signal quality in the final comprehensive confidence, and can be adjusted according to the actual application scenario.

[0100] In one embodiment of the present invention, the step of calculating the ratio of signal energy within the main discharge frequency band to the total frequency band energy, and combining this with the concentration of the standardized discharge pulse sequence in the characteristic phase segment of the power frequency cycle, to calculate the weighted signal-to-noise ratio parameter includes:

[0101] S431, Perform a fast Fourier transform on the standardized discharge pulse sequence to obtain the frequency amplitude spectrum;

[0102] S432, in the frequency amplitude spectrum, multiple frequency band regions outside the main discharge frequency band and with uniform energy distribution are determined as reference noise frequency bands;

[0103] S433, calculate the sum of squares of the signal amplitudes within the main discharge frequency band as the main discharge frequency band energy, and calculate the average value of the sum of squares of the signal amplitudes within the reference noise frequency band as the reference noise energy;

[0104] S434, Calculate the reference signal-to-noise ratio based on the ratio of the main discharge frequency band energy to the reference noise energy;

[0105] S435, calculate the proportion of the number of pulses in the characteristic phase segment of the power frequency cycle of the standardized discharge pulse sequence to the total number of pulses, as the phase concentration degree;

[0106] S436, Obtain the weighted signal-to-noise ratio parameter based on the reference signal-to-noise ratio and the phase convergence, using the following formula:

[0107]

[0108] In the formula, This represents the weighted signal-to-noise ratio parameter. Indicates the reference signal-to-noise ratio (SNR) ,in, Indicates the energy of the main discharge frequency band. (Represents reference noise energy). Indicates phase cohesion ( ,in, This indicates the number of pulses in the characteristic phase segment of the power frequency cycle. (Indicates the total number of pulses). and These represent two preset weighting coefficients (the sum of which equals 1, representing the importance of the reference signal-to-noise ratio and phase cohesion, respectively).

[0109] As described in steps S431-S436 above, the signal-to-noise ratio (SNR) assessment of partial discharge signals from high-voltage equipment needs to consider both "frequency domain effectiveness" and "temporal domain regularity": In the frequency domain, the effective signal energy of different types of discharges is concentrated in specific main discharge frequency bands (e.g., floating potential discharge is concentrated in 500MHz-800MHz, and surface discharge is concentrated in 300MHz-600MHz), while noise energy is mostly distributed in low-frequency or cluttered frequency bands outside the main discharge frequency band. Only by calculating the energy ratio for the main discharge frequency band can the intensity difference between the effective signal and noise be accurately reflected; in the time domain, the true... Actual discharges are controlled by the phase of the power frequency voltage and will stably concentrate in a specific phase segment (such as internal air gap discharges concentrating in 0-90° and 180-270°). In contrast, the phase distribution of random noise is irregular. By statistically analyzing the phase concentration, it is possible to further verify whether the signal is a real discharge and avoid misjudging "periodic interference within the main discharge frequency band" as a valid signal. For example, interference generated by power electronic switches in a substation may form energy peaks within the main discharge frequency band (leading to a misjudgment of the global signal-to-noise ratio as high), but because its phase is random (low concentration), its phase characteristics can be used to distinguish it from a real discharge.

[0110] Therefore, this invention combines frequency domain energy ratio with phase convergence to achieve a more comprehensive and accurate assessment of signal quality. Specifically, this scheme obtains a complete spectral distribution including partial discharge characteristic frequency bands and noise frequency bands by performing a Fast Fourier Transform on a standardized discharge pulse sequence, providing a comprehensive frequency domain data foundation for subsequent accurate calculation of the signal-to-noise ratio. By combining a preset discharge type-characteristic frequency band mapping library, the main discharge frequency band matching the currently identified discharge type is intelligently selected from the spectrum. Simultaneously, a frequency band with uniform energy distribution outside this band is selected as a reference noise frequency band to avoid misjudging the resonant frequencies of other types of discharges or inherent harmonics of the power system as background noise, ensuring the accuracy and specificity of noise energy assessment. By calculating the sum of squares of signal energy within the main discharge frequency band and the average value of signal energy within the reference noise frequency band, the signal energy of a specific discharge type is quantified. The comparison between signal energy and background noise energy provides a reliable quantitative basis for signal-to-noise ratio (SNR) calculation. By introducing a phase clustering parameter based on power frequency phase analysis, the fundamental difference that real partial discharge signals have stable clustering characteristics in a specific phase interval of power frequency voltage, while random noise and interference signals do not possess this characteristic, allows SNR evaluation to consider not only frequency domain energy intensity but also the key feature of time domain phase distribution. By weighting and fusing the benchmark SNR and phase clustering according to preset weights, a more comprehensive and robust weighted SNR parameter is formed. The weighted SNR parameter reflects both the absolute strength and quality of the signal in the frequency domain and the distribution regularity of the signal in the time domain phase dimension. The final weighted SNR parameter can more accurately characterize the comprehensive quality level of partial discharge signals in complex substation environments.

[0111] The parameters of the Fast Fourier Transform must be set to ensure sufficient frequency resolution to accurately separate the main discharge frequency band from the noise frequency band. The selection principle of the reference noise frequency band is to avoid all known partial discharge characteristic frequency bands and inherent harmonic frequencies of the power system, and to select a region with relatively stable background noise in the low-frequency band. The calculation of the phase clustering depends on the power frequency cycle characteristic phase segment determined in the early stage. This phase segment is the most active phase range of various discharges, which is statistically derived from a large amount of historical data.

[0112] like Figure 2 As shown, the present invention also provides a real-time detection system for partial discharge signals of high-voltage equipment, comprising:

[0113] The signal processing module is used to synchronously acquire the high-frequency pulse sequence and power frequency phase signal of partial discharge of high-voltage equipment, and perform adaptive noise reduction and pulse alignment preprocessing on the high-frequency pulse sequence based on the power frequency phase signal to obtain a standardized discharge pulse sequence.

[0114] The extraction and fusion module is used to extract time-domain pulse feature maps and frequency-domain resonance feature maps based on the standardized discharge pulse sequence, and perform feature fusion to generate a dual-channel discharge feature map;

[0115] The network inference module is used to input the dual-channel discharge feature map into a lightweight convolutional neural network. The lightweight convolutional neural network includes a parallel multi-branch structure and a channel-time dual attention mechanism. The parallel multi-branch structure synchronously extracts the micro-waveform features and macro-sequence pattern features of the discharge pulse, and performs adaptive fusion through the channel-time dual attention mechanism to obtain a preliminary discharge type probability distribution.

[0116] The result processing module is used to obtain the signal-to-noise ratio (SNR) parameter of the standardized discharge pulse sequence and generate a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the SNR parameter.

[0117] The comparison and determination module is used to compare the comprehensive type determination confidence level with a preset confidence threshold, determine and output the final discharge type identification result;

[0118] The verification module is used to mark the corresponding high-frequency pulse sequence as a suspected discharge sample and upload it to the cloud analysis platform for verification if the confidence level of the comprehensive type determination is lower than the confidence threshold.

[0119] The network inference module includes:

[0120] The micro-extraction unit is used to input the dual-channel discharge feature map into the micro-feature branch of the parallel multi-branch structure, and extract the micro-waveform features of the discharge pulse by the depth of the first void ratio in the micro-feature branch through separable convolution.

[0121] The macroscopic extraction unit is used to synchronously input the dual-channel discharge feature map to the macroscopic feature branch of the parallel multi-branch structure, and extract macroscopic sequence pattern features spanning multiple power frequency cycles through depth-separable convolution of the second hole rate in the macroscopic feature branch, wherein the second hole rate is greater than the first hole rate.

[0122] The fusion unit is used to splice the micro-waveform features with the macro-sequence pattern features to obtain multi-scale fusion features;

[0123] An attention weighting unit is used to input the multi-scale fused features into a channel-time dual attention module. The channel attention submodule in the channel-time dual attention module evaluates the relative importance of the micro-waveform features and the macro-sequence pattern features, and the time attention submodule focuses on the power frequency cycle feature phase segment where the key discharge event occurs, to obtain the attention-weighted optimized features.

[0124] The probability output unit is used to pass the optimized features to a preset classifier, and output the preliminary discharge type probability distribution through the preset classifier.

[0125] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a method for real-time detection of partial discharge signals of high-voltage equipment.

[0126] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for real-time detection of partial discharge signals in high-voltage equipment.

[0127] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0128] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for real-time detection of partial discharge signals in high-voltage equipment, characterized in that, include: The high-frequency pulse sequence and power frequency phase signal of partial discharge of high-voltage equipment are acquired synchronously. The high-frequency pulse sequence is subjected to adaptive noise reduction and pulse alignment preprocessing based on the power frequency phase signal to obtain a standardized discharge pulse sequence. Based on the standardized discharge pulse sequence, time-domain pulse feature map and frequency-domain resonance feature map are extracted and fused to generate a dual-channel discharge feature map; The dual-channel discharge feature map is input into a lightweight convolutional neural network. The lightweight convolutional neural network simultaneously extracts the micro-waveform features and macro-sequence pattern features of the discharge pulse, and performs feature fusion through an attention mechanism. Then, it is processed by its classifier to obtain a preliminary discharge type probability distribution. Obtain the signal-to-noise ratio (SNR) parameter of the standardized discharge pulse sequence, and generate a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the SNR parameter; The confidence level of the comprehensive type determination is compared with the preset confidence threshold to determine and output the final discharge type identification result.

2. The method for real-time detection of partial discharge signals in high-voltage equipment according to claim 1, characterized in that, The steps of synchronously acquiring the high-frequency pulse sequence and power frequency phase signal of partial discharge from high-voltage equipment, and performing adaptive noise reduction and pulse alignment preprocessing on the high-frequency pulse sequence based on the power frequency phase signal to obtain a standardized discharge pulse sequence include: The high-frequency pulse sequence is mapped to the power frequency periodic phase interval based on the power frequency phase signal to generate a phase-resolved pulse sequence. Based on the statistical characteristics of the phase-resolved pulse sequence in the characteristic phase segment of the power frequency cycle, a dynamic noise threshold is generated. Based on the dynamic noise threshold, the phase-resolved pulse sequence is decomposed in the frequency domain and subjected to adaptive threshold denoising to obtain a denoised pulse sequence. The denoised pulse sequence is pulse aligned and truncated to obtain the standardized discharge pulse sequence.

3. The method for real-time detection of partial discharge signals in high-voltage equipment according to claim 1, characterized in that, The steps of extracting the time-domain pulse feature map and frequency-domain resonance feature map from the standardized discharge pulse sequence, and performing feature fusion to generate a dual-channel discharge feature map include: The standardized discharge pulse sequence is segmented into power frequency cycles, and its time-domain statistical features are extracted and a time-domain statistical feature vector is constructed. Calculate the Mahalanobis distance between the time-domain feature vector and various standard time-domain statistical features in the preset typical discharge model library, and reconstruct it into a time-domain pulse feature map; The standardized discharge pulse sequence is subjected to S-transform to obtain its high-resolution time spectrum, and the energy distribution of the partial discharge characteristic frequency band is extracted from the high-resolution time spectrum according to the preset insulation defect sensitive frequency band range. Calculate the quality factor Q, bandwidth, and center frequency drift of the partial discharge characteristic frequency band, construct the frequency domain resonance characteristic vector, and reconstruct it into a frequency domain resonance characteristic map; The time-domain pulse feature map and the frequency-domain resonance feature map are spliced ​​and normalized to generate a dual-channel discharge feature map.

4. The method for real-time detection of partial discharge signals in high-voltage equipment according to claim 1, characterized in that, The lightweight convolutional neural network includes a parallel multi-branch structure and a channel-temporal dual attention mechanism. The parallel multi-branch structure extracts microscopic waveform features and macroscopic sequence pattern features respectively, and then performs adaptive fusion through the channel-temporal dual attention mechanism.

5. The method for real-time detection of partial discharge signals in high-voltage equipment according to claim 4, characterized in that, The parallel multi-branch structure includes a micro-feature branch and a macro-feature branch, and each branch uses convolution operations with different receptive fields to extract features. The micro-feature branch uses a convolution operation with a first dilation rate, and the macro-feature branch uses a convolution operation with a second dilation rate, wherein the second dilation rate is greater than the first dilation rate. The convolution operation is a depthwise separable convolution.

6. The method for real-time detection of partial discharge signals in high-voltage equipment according to claim 1, characterized in that, The steps of obtaining the signal-to-noise ratio (SNR) parameter of the standardized discharge pulse sequence and generating a comprehensive type determination confidence level based on the preliminary discharge type probability distribution and the SNR parameter include: Extract the maximum probability value and the second highest probability value from the preliminary discharge type probability distribution, and calculate the difference between the maximum probability value and the second highest probability value as the probability confidence factor; Based on the discharge type with the highest probability in the preliminary discharge type probability distribution, query the preset discharge type-feature frequency band mapping library to determine the main discharge frequency band corresponding to the discharge type; The ratio of signal energy within the main discharge frequency band to the total frequency band energy is calculated, and the weighted signal-to-noise ratio parameter is obtained by combining the concentration of the standardized discharge pulse sequence in the characteristic phase segment of the power frequency cycle. The confidence level of the comprehensive type determination is obtained based on the signal-to-noise ratio parameter and the probability confidence factor.

7. The method for real-time detection of partial discharge signals in high-voltage equipment according to claim 6, characterized in that, The steps of calculating the ratio of signal energy within the main discharge frequency band to the total frequency band energy, and combining this with the concentration of the standardized discharge pulse sequence in the characteristic phase segment of the power frequency cycle, to calculate the weighted signal-to-noise ratio parameter include: The frequency domain transformation of the standardized discharge pulse sequence is performed to obtain the frequency amplitude spectrum; In the frequency amplitude spectrum, multiple frequency band regions outside the main discharge frequency band and with uniform energy distribution are defined as reference noise frequency bands; Calculate the main discharge frequency band energy and the reference noise energy of the reference noise band, and calculate the reference signal-to-noise ratio based on the ratio of the main discharge frequency band energy to the reference noise energy; The proportion of the number of pulses in the characteristic phase segment of the power frequency cycle of the standardized discharge pulse sequence to the total number of pulses is calculated as the phase concentration. The weighted signal-to-noise ratio parameter is obtained based on the reference signal-to-noise ratio and the phase convergence.

8. A real-time detection system for partial discharge signals in high-voltage equipment, characterized in that, It includes multiple modules for implementing the steps of the method according to any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images