A cable partial discharge type identification method based on ResNet

Abstract

The application relates to the technical field of power distribution switch control equipment manufacturing, and discloses a cable partial discharge type identification method based on ResNet, which comprises the following steps: synchronously acquiring an original partial discharge data array corresponding to a cable insulation state and power frequency voltage phase information; constructing a phase weight coefficient by calculating discharge probabilities corresponding to each instantaneous phase point in the power frequency voltage phase information, so as to locate a phase-sensitive feature area in the original partial discharge data array; inputting the original partial discharge data array into an identification model with a residual architecture, and calling the phase weight coefficient to perform weighted correction processing on intermediate feature variables output by each level when residual convolution mapping is performed; and the application applies constraints on the residual mapping process by using a physical distribution law, effectively suppresses random interference decoupled from the phase, and improves the stability of cable defect identification under a complex electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the field of power distribution switch control equipment manufacturing technology, specifically to a cable partial discharge type identification method based on ResNet. Background Technology

[0002] As a key facility in the power distribution system, the insulation condition of power cables directly affects the safe operation of the power grid. Partial discharge, as a physical phenomenon characterizing cable insulation defects, is a core element in achieving lean operation and maintenance of the power distribution network. Existing technologies use high-frequency current transformers or ultra-high-frequency sensors to monitor cable partial discharge signals, collect parameters such as the amplitude, phase, and repetition rate of discharge pulses, construct phase-resolved partial discharge maps based on the collected data, and use convolutional neural networks to extract texture features from the maps to achieve automatic identification of discharge types.

[0003] However, under the complex operating conditions of power distribution networks, the nonlinear characteristics of cable impedance changing with frequency and the frequency dispersion effect of pulse signals in the transmission path lead to features in the generated feature maps being diffused in the feature regions and key information points being severely interfered with by noise. The industry generally improves feature representation capabilities by increasing the depth of convolutional networks, which not only increases the computational load on the monitoring terminal but also, due to the lack of physical modeling of the partial discharge phase angle driving attributes, results in insufficient generalization ability of the model when processing data collected by heterogeneous sensors. For example, Chinese invention patent CN110381462B discloses an online monitoring system for partial discharge in power cables, which utilizes a clustered wireless sensor network to improve... To improve the reliability of monitoring data transmission, a secure network is established by calculating node reliability and risk indices. This allows for early warning based on the comparison between partial discharge quantity and safety threshold. Although this approach offers high flexibility at the data transmission level, its algorithm processing core still relies primarily on a single amplitude threshold criterion. It lacks physical prior modeling of the phase angle driving attributes of partial discharge, making it difficult to extract characteristic diffuse signals in high background noise environments. Furthermore, because it fails to deeply explore the profound semantic features of discharge pulses at different abstraction levels, its generalization ability is limited when facing non-stationary signals collected by heterogeneous sensors. It cannot provide accurate type identification support for complex insulation defects such as tip discharge and air gap discharge.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a deep learning model with phase sensing capability and high computational efficiency based on the physical characteristics of cable partial discharge signals, so as to improve the reliability of identification schemes in complex power distribution environments. Summary of the Invention

[0005] This invention proposes a ResNet-based method for identifying cable partial discharge types, comprising the following steps: Step S1: Simultaneously acquire the original partial discharge data array characterizing the insulation state of the tested power cable and the synchronous power frequency voltage phase information characterizing the power grid operation reference; the original partial discharge data array is a time series set containing the discharge pulse amplitude and its corresponding instantaneous phase angle acquired within a preset sampling time. Step S2: Based on the physical prior law that the distribution of partial discharge pulses varies with the amplitude fluctuation of electric field intensity within the power frequency voltage cycle, the probability of discharge occurrence corresponding to each instantaneous phase point in the synchronous power frequency voltage phase information is calculated. A phase weight coefficient is constructed to characterize the contribution of different phase intervals to the identification of partial discharge features, thereby locating the phase-sensitive feature region driven by electric field intensity in the original partial discharge data array. Step S3: Input the original partial discharge data array into the identification model built on the deep residual convolution architecture. The identification model performs multi-level residual mapping processing to generate intermediate feature variables corresponding to different abstract energy levels. In each level of residual mapping, the phase weight coefficient is called in real time to perform element-level weighted correction processing on the intermediate feature variables output at the current level. This is to force the identification model to enhance the signal gain in the phase-sensitive feature region through a physical constraint mechanism and suppress random background pulse interference decoupled from the power frequency phase. Step S4: The identification model performs feature vectorization and nonlinear classification mapping on the intermediate feature variables after weight correction, and outputs the partial discharge type identification result indicating the type of insulation defect existing inside the tested power cable.

[0006] Preferably, before performing step S1, the following steps are also included: Step S201, using a band-stop filter to perform filtering processing on the original cable partial discharge signal to filter out the 50Hz power frequency signal and its higher harmonic interference; Step S202, identifying transient burst pulses in the filtered signal based on adaptive amplitude threshold logic, and establishing the original partial discharge data array based on the amplitude of the transient burst pulses and their phase distribution within the power frequency cycle.

[0007] Preferably, step S2 includes the following steps: extracting the positive half-cycle center phase point and the negative half-cycle center phase point from the synchronous power frequency voltage phase information; using the positive half-cycle center phase point and the negative half-cycle center phase point as the reference center, constructing a probability density function with a bimodal symmetrical distribution according to a preset phase window width, and defining the output value of the probability density function as the phase weight coefficient.

[0008] Preferably, the phase weighting coefficient M(Φ) satisfies the following logical relationship: Where Φ is the real-time phase angle value in the synchronous power frequency voltage phase information. The preset phase constant for the center phase point of the positive half-cycle. σ is the preset phase constant of the negative half-cycle center phase point, and σ is the distribution coefficient characterizing the phase window width.

[0009] Preferably, the specific process of performing element-level weighted correction in step S3 is as follows: based on the channel dimension and spatial resolution of the intermediate feature variables, the phase weight coefficients are mapped to a weight matrix with the same dimension as the intermediate feature variables through linear interpolation logic; the weight matrix and the intermediate feature variables are multiplied element by element to guide the identification model to extract the energy distribution law in the process of discharge pulse evolution through phase physical constraints.

[0010] Preferably, the identification model includes a first convolutional layer, a multi-level residual module, and a global pooling layer. The identification model realizes long-distance transmission of partial discharge feature signals through the jump connection structure inside the multi-level residual module, and uses the first convolutional layer to perform spatial feature dimensionality reduction to extract deep semantic features about insulation defects in the original partial discharge data array.

[0011] Preferably, the partial discharge type identification result output in step S4 includes a discharge classification probability vector and an identification confidence quantification value.

[0012] Preferably, after outputting the partial discharge type identification result, the following steps are also included: retrieving the historical online monitoring archives of the target cable to obtain the corresponding historical identification dataset; performing trend correlation analysis between the partial discharge type identification result and the historical identification dataset; if the identification confidence quantification value is lower than the preset threshold of 0.85 and the identified insulation defect type changes, then generating an abnormal cable insulation status warning instruction.

[0013] Preferably, the partial discharge type identification results cover tip discharge, surface discharge, internal air gap discharge, and metal suspension discharge.

[0014] Preferably, the identification model is deployed in an edge computing terminal installed on the side of the distribution cabinet. The edge computing terminal performs real-time calculation of the partial discharge type identification results and conducts online assessment of the insulation safety level of the cable system.

[0015] The beneficial effects of this invention are: 1. In ResNet cable partial discharge type identification, a multi-dimensional feature representation system integrating time-domain, frequency-domain, and phase-domain information is constructed to solve the feature dispersion problem of cable partial discharge signals under complex working conditions. Partial discharge pulses are affected by impedance nonlinear changes during long-distance cable transmission, and features of a single metric dimension are easily distorted due to signal attenuation. This invention utilizes the collaborative modeling of multi-dimensional features to achieve feature complementarity of physical information of different dimensions in a deep network. This mechanism can capture the aggregation law and energy evolution trend of discharge pulses in the positive and negative half-cycle power frequency phases, so that the identification process no longer relies on single amplitude information. This ensures that the identification logic can still accurately lock the essential features of insulation defects in a heterogeneous sensor acquisition environment and avoid misjudgment caused by overlapping discharge modes.

[0016] 2. By introducing residual connection structures into the deep convolutional architecture, high-fidelity transmission of partial discharge features and deep semantic extraction are achieved. Cable partial discharge signals contain a large number of weak, high-frequency burst-like subtle textures. Traditional shallow networks or structures that directly increase the number of layers often lose this key information due to obstructed gradient transmission. This invention adopts a residual mapping mechanism to maintain the stability of feature gradients during the progressive abstraction process. This mechanism, combined with the phase-locked weight correction of partial discharge signals, can force the model to prioritize the feature gain of the phase point of abrupt change in electric field intensity, while suppressing non-phase-related random background noise. This deep coupling between physical laws and network structure enables the model to maintain computational efficiency while possessing the logical ability to eliminate complex background interference, thus improving the stability of the identification results.

[0017] 3. Through multi-level signal preprocessing and feature standardization processes, the identification model can be deployed in a lightweight manner at the edge terminal of the intelligent power distribution side. Due to the continuous power frequency and harmonic interference in the on-site electromagnetic environment, the signal-to-noise ratio of the original signal fluctuates drastically. This invention uses the linkage of band-stop filtering and adaptive threshold detection to pre-filter irrelevant fluctuations at the data input end, reducing the redundant computational load of the deep network on feature denoising. This preprocessing step works in conjunction with the normalization layer of the residual network, enabling the system to achieve the feature expression depth of the ultra-deep network using a network architecture with a small number of parameters. This reduces the monitoring system's dependence on high-performance computing resources and meets the requirements of edge devices such as power distribution cabinets for real-time and practical online identification of partial discharge. Attached Figure Description

[0018] Figure 1 This is a flowchart of the cable partial discharge type identification method that integrates phase weight correction mechanism according to the present invention; Figure 2 This is a schematic diagram of the overall system architecture for edge perception and remote decision-making collaboration of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] This invention relates to a ResNet-based method for identifying partial discharge types in cables. It utilizes a sensor array distributed in the grounding circuits of power distribution facilities such as switchgear to collect raw partial discharge data corresponding to the insulation state of the tested power cable. Simultaneously, a synchronous acquisition circuit is used to obtain synchronous power frequency voltage phase information characterizing the power grid's operating reference. The raw partial discharge data array consists of the discharge pulse amplitude and its corresponding instantaneous phase angle within the sampling period, with the sampling frequency set between 50MHz and 200MHz. The acquisition logic of this raw partial discharge data array is based on the grounding leads installed at cable terminals or joints. A high-frequency current transformer (HFCT) at the line captures weak charge-induced pulses in the nanosecond range. After conditioning, the signal is converted by a high-speed sample-and-hold circuit into a discrete amplitude stream reflecting the spatiotemporal evolution trend of discharge energy within the power frequency cycle. In the specific test group, 100MHz was selected as the sampling frequency to obtain ultra-high frequency partial discharge signals with pulse widths in the nanosecond range. After acquiring the raw partial discharge signal, the processor calls a band-stop filter to remove power frequency interference. The center frequency of the band-stop filter is set to 50Hz, and its bandwidth is set to 2Hz. The filtered signal is input to the pulse detection module, which calculates the signal mean μ and the signal standard deviation. And set the pulse detection threshold T to satisfy Before performing statistical calculations, the system allocates a linear circular buffer with a length of 2048 sampling points in memory and updates the underlying sampled values ​​in real time at an operating frequency of 100.0MHz. Every 1.0ms physical timing cycle, the processor extracts and aggregates the full charge sampled values ​​in the buffer, calculates the real-time mean and standard deviation over the past 20.0ms, and ensures that the threshold can adaptively track the random noise changes caused by cable load fluctuations, avoiding false pulse triggering due to sudden changes in environmental background noise. Here, μ is the signal mean. Where is the signal standard deviation and T is the pulse detection threshold. When the instantaneous amplitude of the signal exceeds the pulse detection threshold T, the system records the amplitude of the effective discharge pulse and its phase distribution within the power frequency cycle. This procedure filters out fluctuations with amplitudes lower than the mean of the background noise while retaining the characteristics of the discharge pulse.

[0021] To extract the physical constraint features of the discharge pulse in the power frequency phase space, the system calculates the discharge probability corresponding to each instantaneous phase point in the synchronous power frequency voltage phase information. Based on the physical property of the symmetrical distribution of partial discharge in the positive and negative half-cycles, a phase weighting coefficient M(Φ) is constructed, which is used to locate the phase-sensitive feature region. The calculation formula of the phase weighting coefficient M(Φ) is as follows: Where Φ is the phase angle of the synchronous power frequency voltage measured in real time, in degrees; The constant for the center phase point of the positive half-cycle is 90°. The negative half-cycle center phase point constant is 270°; σ is the distribution coefficient characterizing the phase window width, ranging from 15° to 30°. The system uses this nonlinear mapping to perform gain processing on features near the abrupt change in electric field intensity phase point, suppressing interference signals deviating from the center phase region. Regarding the procedure for determining key parameters in the phase weight coefficient M, at least 100 cycles of cable grounding loop current data are collected during the initial benchmark monitoring phase of equipment deployment. The distribution coefficient σ characterizing the phase window width is determined by calculating the standard deviation of the background noise energy distribution in the power frequency phase space. The specific calibration process is as follows: Initialization period of 2000.0ms after equipment startup... Within the system, 100 cycles of raw electromagnetic environment data are continuously captured at a sampling frequency of 200.0MHz. The phase offset of each sampling point relative to the 50Hz power frequency zero point is calculated, and the amplitude distribution statistical variance is extracted as the initial value of the standard deviation. The system automatically multiplies this initial standard deviation value by a mapping gain of 2.5 to generate a distribution coefficient value that dynamically adapts to the current electromagnetic shielding environment of the power distribution room. If the monitored ambient noise amplitude fluctuation deviation exceeds 15.0%, the system automatically resets and restarts the 2000.0ms parameter update procedure. The distribution coefficient σ is set to 2.5 times the energy distribution standard deviation to cover the discharge accumulation range, and the constant of the positive half-cycle center phase point is fixed. The angle is 90°, and the constant of the center phase point of the negative half-cycle is fixed. The phase weight distribution relationship is calculated based on the determined parameters to be 270°, and the sampling and calculation process is restarted to update the distribution coefficient σ when the electromagnetic characteristics of the monitored environment change. An adaptive feature filtering window is established at the feature extraction level using the prior laws of discharge physics.

[0022] The identification model is based on a deep residual convolutional architecture, which receives the original partial discharge data array and performs residual mapping processing. The model is customized and optimized based on the ResNet-18 skeleton, consisting of four stacked residual sub-blocks, each containing two 3×3 convolutional layers. It performs spatial feature dimensionality reduction using a downsampling strategy with a stride of 2, ensuring lossless fusion of low-order physical features and high-order semantic features through skip connections in the residual branches. In each residual convolutional mapping layer, the model calls phase weight coefficients to perform weighted correction on the intermediate feature variables of the output. A linear interpolation operator is used to map the scalar phase weight coefficients into a weight matrix with the same dimension as the intermediate feature variables, and then element-wise multiplication is performed between this matrix and the intermediate feature variables. Let x be the input feature of the residual block, and F(x) be the output feature of the residual branch. Then, the output feature y after phase correction satisfies the following condition: Through this feedback mechanism, the network strengthens high-dimensional semantic features consistent with the phase distribution pattern, reducing the feature diffusion effect caused by dispersion during long-distance signal transmission. It obtains the number of channels C, height dimension H, and width dimension W of the intermediate feature variables in the residual convolution mapping of the identification model. The 360° phase space within the power frequency cycle is divided into sampling grid points corresponding to the width dimension W. A linear interpolation function is used to resample the scalar phase weight coefficients M into a one-dimensional guiding vector of length W. This one-dimensional guiding vector is then broadcast and filled along the height dimension H and the channel dimension C to generate a weight matrix consistent with the dimension of the intermediate feature variables. The height dimension is defined as the spectral distribution features from 100.0 kHz to 50.0 MHz extracted after performing a discrete Fourier transform on the sampled signal. During the broadcast filling operation, the logic controller performs a dot product operation on the weight coefficients corresponding to the phase angle at the current moment for each spectral component point in the height dimension, ensuring that the energy characteristics of the high-frequency pulse at the 90° and 270° phase points receive a gain correction of 1.25 times. This deterministic gain value is obtained by the logic controller calling the mapping function in real time. Its technical origin lies in shifting and scaling the phase weights in the unit interval to the activation gain interval, forcibly enhancing the response weights in the feature channel that coincide with the physical mechanism, thereby suppressing the activation of noise features that deviate from the mechanism region. The physical phase domain distribution weights are converted into spatial gain control quantities by using the dot product of the weight matrix and the intermediate feature variables, and a feature enhancement channel modulated by the power frequency phase is constructed at the residual branch output.

[0023] The identification model performs global pooling and nonlinear mapping on the modified intermediate feature variables, outputting the corresponding partial discharge type identification results. The identification results include tip discharge, surface discharge, internal air gap discharge, and metal suspension discharge. During the identification process, the system outputs the classification probability vector and identification confidence quantification for each type. The identification confidence quantification is obtained by extracting the values ​​from the output layer of the identification model. The maximum value of the normalized probability vector is used to determine the threshold. If the identification confidence quantification value is lower than the preset discrimination threshold of 0.85, the historical identification dataset in the database is retrieved to perform correlation analysis. The dynamic calibration method of this discrimination threshold is as follows: during the 600.0s stable monitoring period after the initial energization of the cable, the system simulates the injection of a standard test pulse with the physical characteristics of tip discharge, extracts the normalized probability distribution of the classification layer output, and selects the minimum value under the 99.0% statistical confidence interval as the pre-judgment threshold. The deviation limit value of 0.15 of the Mahalanobis distance is obtained by calculating the projected Euclidean distance between the current real-time feature vector and the center point of the health dataset within the past 30 days. Each increase of 0.01 in this value corresponds to a 2.0% increase in the physical layer insulation degradation. This quantitative mapping is established because the Mahalanobis distance characterizes the projection deviation of the current signal pattern in the feature space relative to the healthy baseline distribution. Its physical basis stems from the dynamic law of the growth of microscopic carbonization channels in insulators caused by the accumulation of partial discharge energy. This proportional logic enables the abstract probabilistic algorithm to provide real-time feedback as a physical degradation indicator with practical guiding significance. If the alarm logic is triggered, the processor will send a hexadecimal control message 0xFB to the control register of the edge gateway, driving the alarm to execute intermittent audible and visual prompts at a frequency of 2.0Hz. The processor retrieves N sets of historical identification datasets stored within a preset time sliding window from the database and calculates the Mahalanobis distance between the currently output classification probability vector and the mean vector of the historical identification datasets. Where N is the number of historical sample groups, The Mahalanobis distance is a dimensionless parameter used to measure the deviation of a probability distribution; if the deviation occurs within three consecutive power frequency cycles... If all values ​​exceed the preset deviation threshold of 0.15 and are accompanied by a unidirectional decrease in the identification confidence quantification value, the system determines that there is a risk of evolutionary deterioration in the cable insulation condition.

[0024] Example 1: In an underground substation with multiple parallel-operating 10kV indoor cable junction boxes, due to the compact cable arrangement and proximity to high-power rectifier equipment, the monitoring environment experiences continuous power frequency harmonic interference and sudden pulse noise generated by frequent circuit breaker switching. When the system performs insulation status identification on one of the main cables, the aforementioned sensor array collects a raw partial discharge signal superimposed with background noise at the cable joint. In this scenario, due to the nonlinear characteristics of cable impedance changing with frequency, the two-dimensional spectrum acquired by the sensor is at risk of feature region dispersion and key information points being submerged by random noise. Traditional deep learning identification frameworks, lacking constraints on the physical distribution of discharge, are prone to misjudging randomly phase-distributed switching pulse noise as insulation defect signals, thus generating false alarms. Furthermore, the dispersion effect of the signal during long-distance transmission causes distortion of the extracted pulse features, making it difficult for the identification model to accurately obtain the physical features reflecting the essence of insulation defects.

[0025] To address the aforementioned electromagnetic interference and signal distortion, the technical solution of this invention implements a synchronous sensing procedure under this operating condition. It utilizes sensor arrays distributed at each end of the cable to acquire the original partial discharge data array corresponding to the insulation state of the tested power cable at a sampling frequency of 100MHz. Simultaneously, the acquisition circuit acquires synchronous power frequency voltage phase information characterizing the power grid's operating reference. The processor calls a band-stop filter with a center frequency of 50Hz and a bandwidth of 2Hz to eliminate power frequency interference, and then uses the mean μ of the filtered signal and the standard deviation of the signal... Set a pulse detection threshold T, which satisfies the following conditions: When the instantaneous amplitude of the signal exceeds the pulse detection threshold T, the system records the amplitude of the effective discharge pulse and its instantaneous phase angle distribution data within the power frequency cycle, thereby initially extracting the effective discharge pulse in the time domain. This is used to filter out fluctuations with amplitudes lower than the average background noise. After acquiring the original partial discharge data array containing phase information, the system constructs the phase weight coefficient M(Φ) according to the physical constraint logic of the aforementioned specific implementation method. At this time, the distribution coefficient σ is set to 22°, and the constant of the positive half-cycle center phase point is set... and the constant of the center phase point of the negative half-cycle Locked at 90° and 270° respectively; by executing the formula Through nonlinear mapping operations, the system generates a gain matrix that varies with the power frequency phase; where Φ is the real-time measured synchronous power frequency voltage phase angle value. The constant of the positive half-cycle center phase point. σ is the constant of the negative half-cycle center phase point, and σ is the distribution coefficient characterizing the phase window width. Under the action of this gain matrix, the pulse features near the point of maximum electric field strength gain weight, while the randomly distributed phase pulses generated by circuit breaker switching are logically suppressed because they deviate from the center phase region. Thus, the physical prior law is used to filter random noise interference in the feature extraction stage.

[0026] After receiving the processed data, the identification model, during the residual mapping process of each layer of its internal deep residual convolutional architecture, in real time calls the phase weight coefficients to perform weighted correction on the intermediate feature variables of the output. It uses a linear interpolation operator to map the scalar phase weight coefficients into a weight matrix with the same dimension as the intermediate feature variables, and applies it to the intermediate feature variables output by the residual branch to satisfy the correction relationship. Where y is the corrected output feature. Let x be the output feature of the residual branch, x be the input feature of the current residual block, and M(Φ) be the phase weight coefficient mapped to the feature space. This not only alleviates the problem of unstable discrimination logic caused by feature diffusion in complex working conditions, but also, through the synergy of phase-aware features and deep residual architecture, enables the model to offset signal distortion caused by cable impedance changes by compensating for feature gain at key physical phase points without increasing computational load. This achieves a balance between identification accuracy and computational efficiency. Finally, the identification model performs nonlinear classification mapping on the corrected feature vectors, outputting a discharge classification probability. The system outputs the insulation defect type as a point discharge, and its corresponding identification confidence quantification value is stable above 0.96, which is higher than the preset discrimination threshold of 0.85. At the same time, the edge computing terminal retrieves the historical identification dataset and performs trend correlation analysis to confirm that the current identification result matches the historical insulation degradation trajectory, thereby generating the corresponding cable insulation status assessment result and maintenance instructions. This procedure transforms the original physical parameters into the basis for operation and maintenance decisions, improving the reliability of cable status monitoring by the power distribution switch control equipment in a strong electromagnetic interference environment.

[0027] Example 2: The experiment uses a physical simulation platform to verify the extraction efficiency of phase-constrained logic for partial discharge pulse characteristics in cables under strong random electromagnetic interference. The platform includes a high-frequency current transformer and a sampling circuit with a sampling frequency of 100MHz. The sampling frequency of 100MHz is set to balance the waveform reconstruction accuracy of nanosecond-level pulses with the computational load of the backend. When the rise time of the measured signal is on the order of 10ns, to satisfy the Nyquist sampling law and leave a margin, the sampling frequency is set to be no less than 2.5 times the effective bandwidth of the signal. Thus, 100MHz is determined as the sampling reference under standard operating conditions. Gaussian white noise with a signal-to-noise ratio of 20dB is superimposed on the original partial discharge signal, and power frequency interference with a frequency of 50Hz and its harmonics is injected. The experiment sets up the sample group of the present invention, a control group 1 without phase weight correction, and a control group 2 with a sampling frequency of 10MHz and a distribution coefficient exceeding the upper limit. The system extracts the original partial discharge data array containing noise as the reference input. During the execution of the sample group of the present invention, the following is calculated: The phase weighting coefficient M(Φ) is 0.985 and 0.982 in the central sensitive areas of 90° and 270°, respectively, while it decays to 0.004 at the 180° phase point far from the discharge mechanism region. Here, M(Φ) is the phase weighting coefficient, and Φ is the real-time phase angle measured synchronously. Experimental observation data shows that the clarity of the cluster boundary of the intermediate feature variables after phase correction in the sample group of this invention is improved by 32.6% compared with the control group. For the tip discharge defect, the recognition accuracy of the sample group of this invention reaches 98.4%, while the recognition accuracy of the control group is 83.7% under the same noise level. This difference is due to the lack of phase physical constraints in the control group. In the residual iteration, the model misidentifies the randomly distributed noise features of the phase as the semantic components of the discharge signal. In the boundary verification of key parameters, when the distribution coefficient σ increases from 22° to 45°, the background noise in the non-discharge region enters the logic enhancement region due to the excessively large phase window width, and the recognition accuracy drops from 97.1% to 89.5%.

[0028] Control group 2, due to the use of a low sampling frequency of 10MHz, experienced signal aliasing and distortion, resulting in the loss of nanosecond-level details and a drop in recognition accuracy to 62.8%. Furthermore, in gradient tests targeting defect severity, when the discharge quantity increased from 50pC to 500pC, the recognition confidence quantification value of the proposed solution's sample increased from 0.88 to 0.99, while control group 1 had a misjudgment rate of 24.5% at weak discharge levels. This demonstrates that embedding phase physical constraints enhances the system's ability to capture subtle signs of insulation degradation. Here, pC stands for picoliter, representing the charge quantity. The proposed solution utilizes the coupling of phase weight coefficients and residual architecture to compensate for signal distortion caused by changes in cable impedance through characteristic gain compensation of physical phase points. The accuracy curves and confidence distribution patterns obtained from the experiments confirm the enhancing effect of physical distribution patterns on recognition stability. This method transforms the electric field distribution pattern into algorithm correction logic, solving the problem of evaluation failure caused by background noise masking key features during the operation of power distribution switch control equipment.

[0029] Example 3: This example combines Figures 1 to 2 This section describes a ResNet-based method for identifying cable partial discharge types, as follows: Figure 1 As shown, step S1 involves simultaneously acquiring the original partial discharge data array characterizing the insulation state of the tested power cable and the synchronous power frequency voltage phase information characterizing the power grid operation benchmark. Step S2 involves calculating the discharge probability of each instantaneous phase point in the synchronous power frequency voltage phase information based on physical prior laws, constructing phase weight coefficients to locate the phase-sensitive feature region driven by the electric field intensity in the original partial discharge data array. Step S3 involves inputting the original partial discharge data array into the identification model to perform multi-level residual mapping, and calling the phase weight coefficients to perform element-level weighted correction processing on the intermediate feature variables to enhance the signal gain in the phase-sensitive feature region and suppress random background pulse interference. Finally, step S4 involves the identification model performing feature vectorization processing and nonlinear classification mapping on the intermediate feature variables after weight correction, and outputting the partial discharge type identification result indicating the type of insulation defect present inside the tested power cable.

[0030] like Figure 2As shown, the system comprises a power distribution field monitoring area, an edge intelligent sensing area, and a remote operation and maintenance decision-making area. The power distribution field monitoring area houses the power cables under test within a 10kV power distribution switchgear. Signals collected by the high-frequency partial discharge sensor and the power frequency phase acquisition unit sensor, acting through inductive coupling, are transmitted via coaxial cable to the edge intelligent sensing area. This area is equipped with an edge computing terminal, i.e., an embedded GPU device. Internally, it is connected to a high-speed data acquisition module, a band-stop filter and preprocessing unit, and a deep residual identification engine. A phase-weighted controller performs parameter correction operations on the deep residual identification engine. The data output by the deep residual engine is transmitted to the remote operation and maintenance decision-making area via an encrypted communication network, i.e., 4G / 5G. This remote decision-making area is comprised of a monitoring center server. Its internal historical monitoring archive database transmits information to the trend correlation analysis module via data retrieval commands. Combined with data from the edge intelligent sensing area, the final output result is sent to the operation and maintenance early warning terminal.

[0031] Example 4: In a smart power distribution room with a voltage level of 35kV and containing cross-linked polyethylene cable joints, the signal captured by the monitoring terminal is interfered with by high-frequency pulse groups. In this case, the system faces the challenge of performing high-dimensional semantic mapping on the original partial discharge data array. Since the discharge pulses in the original partial discharge data array are discrete in the time domain, directly performing linear dimensionality reduction will result in the loss of phase texture details of the pulse waveform, making it impossible to establish stable weighted associations in the phase-sensitive feature region during subsequent residual mapping. To overcome the obstacle of unclear data representation, the system executes a tensor reconstruction procedure in data preprocessing. Specifically, the original partial discharge data array is converted into a 64×64 two-dimensional feature tensor using a sliding window slicing algorithm. The original discharge pulses exhibit high discreteness and randomness in the time domain. This reconstruction step compresses the three-dimensional discrete information of time, phase, and amplitude into a two-dimensional spatial pattern, inducing the subsequent residual convolution operator to extract texture features reflecting the essence of the defect using translation invariance. This allows the phase weighting coefficients mentioned earlier to directly perform physical guidance on key pixels of the pattern in the spatial dimension. The system uses the power frequency cycle as a reference, dividing the 360° phase space into 64 equally wide phase intervals, while simultaneously detecting... The measured pulse amplitude range was divided into 64 energy levels. By setting the physical width of the sliding window to 20.0 ms and the sliding step distance to a fixed 5.0 ms, the coverage of the power frequency waveform was ensured to reach 100%, and the feature overlap rate between adjacent windows was 75%. During the energy level division process, the maximum dynamic range of 500.0 mV obtained during the power-on self-calibration phase was used as the 100% calibration reference value, and 64-level linear step sampling was performed according to a fixed gain interval of 7.8125 mV. By statistically analyzing the frequency of each energy level in each phase interval, the system filled in two-dimensional features. The pixel values ​​in the tensor transform the discrete pulse sequence into a probability density spectrum with physical mechanism properties. Based on this, the value of the distribution coefficient σ is determined by calculating the root mean square value of the background noise, so that the phase mask function adaptively covers the electric field intensity change region. In the calibration process of this embodiment, the system detects that the standard deviation of the ambient background noise is 12.5μV. According to the monotonicity mapping rule of noise level and window width, the distribution coefficient σ is determined to be 25.6°. Here, σ is the distribution coefficient. The determination of this parameter enables the generated phase weight coefficient M(Φ) to be aligned with the charge accumulation phase point in the discharge mechanism.

[0032] When the identification model performs residual convolution mapping, its first residual block contains a 3×3 convolution kernel, and the number of convolution channels is set to 64. As data flows through this residual block, the system uses a linear interpolation operator to expand the scalar phase weight coefficients into a 64×64 mask matrix, and performs an element-wise weighted operation of y=F(x)⋅M(Φ)+x. Here, x is the feature tensor input to the current residual block, F(x) is the intermediate output after the 3×3 convolution operation, M(Φ) is the mask matrix corresponding to the current phase, and y is the output feature. The embedding of physical logic enables the model to suppress spurious features in the non-discharge phase region when processing heterogeneous sensor signals, solving the problem of deep networks grading in strong noise environments. To address the issue of degree disappearance, the system output partial discharge type identification result achieved a classification probability of 0.982 for metal suspension discharge defects after the aforementioned processing procedures. Furthermore, the identification confidence quantification value remained stable at 0.975, exceeding the preset discrimination threshold of 0.85. Because the two-dimensional energy feature map completely preserves the phase coupling pattern, the identification model exhibits stable semantic capture capabilities in low signal-to-noise ratio environments, ultimately guiding the edge computing terminal to generate accurate preventative maintenance work orders. This process demonstrates that through tensor reconstruction and adaptive parameter adjustment, the method of this invention can transform physical distribution characteristics into highly reliable algorithmic identification capabilities, meeting the engineering practical needs of intelligent power distribution systems for accurate identification of insulation defects.

[0033] Example 5: When the system faces a situation where the power frequency reference phase shift is caused by heterogeneous sensor deployment locations, the on-site phase zero-point calibration procedure is executed to align the phase weighting coefficient M(Φ) with the physical electric field period of the power cable. Specifically, the system receives a zero-crossing trigger signal from the power frequency voltage acquisition circuit and acquires the fundamental component of the cable grounding current under the same sampling clock. The processor uses discrete Fourier transform to calculate the initial phase difference between the fundamental component and the zero-crossing signal. The data is then stored in the field configuration register; in the subsequently generated raw partial discharge data array, the system uses the initial phase difference... The instantaneous phase angle of each pulse is calibrated so that the calibrated phase Φ is equal to the real-time phase angle value. Phase difference from the initial phase The difference; where M(Φ) is the phase weighting coefficient, and Φ is the calibrated real-time phase angle. The initial phase difference, This process provides real-time phase angle values, eliminating propagation delay errors introduced by sampling cables of varying lengths, and ensuring the constant phase point at the center of the positive half-cycle. Phase constant at the center point of the negative half-cycle The physical mapping provides a phase reference benchmark for cross-device deployment.

[0034] In a power distribution automation monitoring environment with random non-steady-state noise, the system performs baseline calibration of identification confidence quantification values ​​during the initial silent monitoring phase of deployment. Under conditions without partial discharge signal injection, samples are continuously collected for 1000 power frequency cycles, and the identification model calculates and outputs the background identification confidence distribution characteristics. The 99th percentile of this distribution is extracted as the environmental confidence noise floor. And select 0.85 as the environmental confidence floor noise. The maximum value among the results after increasing by 0.05 is used to correct the preset discrimination threshold; where, To ensure environmental confidence level, in specific test scenarios, when the mean value of the measured background noise confidence level fluctuation is 0.72, the system will set the discrimination threshold to 0.85 to ensure the statistical significance of the recognition results. By calculating the dynamic threshold through actual noise distribution, the parameter deviation caused by differences in sensor sensitivity and different electromagnetic shielding effectiveness on site is eliminated, so that the generated cable insulation status warning command has logical stability.

[0035] Example 6: In a deployment scenario involving an 800m long 10kV cross-linked polyethylene cable and its associated distribution switch control equipment, the system faces the obstacle of nonlinear phase shift in the pulses of the original partial discharge data array due to the high-frequency attenuation characteristics in the cable transmission path and the asynchronous nature of the sampling clocks between heterogeneous sensors. Directly calling the preset identification model would cause a misalignment between the gain region of the phase weight coefficient M(Φ) and the measured discharge mechanism phase, resulting in fluctuations in the identification confidence quantification value and interfering with the determination of the type of insulation defect. To eliminate parameter deviations caused by the aforementioned deployment environment, the system employs standardized pre-calibration and model parameter adaptation methods. During the equipment's no-load phase, synchronous power frequency voltage phase information is acquired and the phase reference of the power frequency cycle is determined. The processor uses a sliding sampling window to acquire 50 consecutive cycles of ground current signals and determines the mean μ and standard deviation of the amplitude. The pulse detection threshold T is calibrated as follows: The signal propagation delay was measured by injecting a standard pulse signal with a pulse width of 20 ns and a charge of 200 pC into the far end of the cable, and the initial phase difference was calculated. .

[0036] Based on this, the identification model performs residual mapping on the original partial discharge data array. The convolution kernel size of the first-level residual block is set to 3×3, the stride is 1, and the number of output channels is 64. The processor uses a linear interpolation operator to map the scalar phase weight coefficients M(Φ) into a 64×64 weight matrix and applies it to the intermediate feature variables of the residual branch output to satisfy the correction relationship. Where y is the corrected output feature, F(x) is the output feature of the residual branch, x is the input feature of the current residual block, M(Φ) is the phase weight coefficient mapped to the feature space, and μ is the mean amplitude. Where is the amplitude standard deviation, and T is the pulse detection threshold. The initial phase difference; after correction by the above calibration procedure, the identification confidence quantification value output by the system in the verification of metal suspension discharge defects is stable at 0.982, and the generated partial discharge type identification result has a consistency of no less than 99% with the manual analysis; since the on-site phase zero calibration procedure eliminates the time delay error in the signal transmission path, the phase weight coefficient is accurately aligned with the physical distribution law of partial discharge, enabling the edge computing terminal to output a determined insulation status assessment level and maintenance instructions.

[0037] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying cable partial discharge types based on ResNet, characterized in that, Includes the following steps: Step S1: Simultaneously acquire the original partial discharge data array characterizing the insulation state of the tested power cable and the synchronous power frequency voltage phase information characterizing the power grid operation reference; the original partial discharge data array is a time series set containing the discharge pulse amplitude and its corresponding instantaneous phase angle acquired within a preset sampling time. Step S2: Based on the physical prior law that the distribution of partial discharge pulses varies with the amplitude fluctuation of electric field intensity within the power frequency voltage cycle, the probability of discharge occurrence corresponding to each instantaneous phase point in the synchronous power frequency voltage phase information is calculated. A phase weight coefficient is constructed to characterize the contribution of different phase intervals to the identification of partial discharge features, thereby locating the phase-sensitive feature region driven by electric field intensity in the original partial discharge data array. Step S3: Input the original partial discharge data array into the identification model built on the deep residual convolution architecture. The identification model performs multi-level residual mapping processing to generate intermediate feature variables corresponding to different abstract energy levels. In each level of residual mapping, the phase weight coefficient is called in real time to perform element-level weighted correction processing on the intermediate feature variables output at the current level. This is to force the identification model to enhance the signal gain in the phase-sensitive feature region through a physical constraint mechanism and suppress random background pulse interference decoupled from the power frequency phase. Step S4: The identification model performs feature vectorization and nonlinear classification mapping on the intermediate feature variables after weight correction, and outputs the partial discharge type identification result indicating the type of insulation defect existing inside the tested power cable.

2. The cable partial discharge type identification method based on ResNet according to claim 1, characterized in that, Before performing step S1, the following steps are also included: Step S201, using a band-stop filter to perform filtering processing on the original cable partial discharge signal to filter out the 50Hz power frequency signal and its high-order harmonic interference; Step S202, identifying transient burst pulses in the filtered signal based on adaptive amplitude threshold logic, and establishing the original partial discharge data array based on the amplitude of the transient burst pulses and their phase distribution within the power frequency cycle.

3. The cable partial discharge type identification method based on ResNet according to claim 1, characterized in that, Step S2 includes the following steps: extracting the positive half-cycle center phase point and the negative half-cycle center phase point from the synchronous power frequency voltage phase information; using the positive half-cycle center phase point and the negative half-cycle center phase point as the reference center, constructing a probability density function with a bimodal symmetrical distribution according to the preset phase window width, and defining the output value of the probability density function as the phase weight coefficient.

4. The cable partial discharge type identification method based on ResNet according to claim 3, characterized in that, The phase weighting coefficient M(Φ) satisfies the following logical relationship: Where Φ is the real-time phase angle value in the synchronous power frequency voltage phase information. The preset phase constant for the center phase point of the positive half-cycle. σ is the preset phase constant of the negative half-cycle center phase point, and σ is the distribution coefficient characterizing the phase window width.

5. The cable partial discharge type identification method based on ResNet according to claim 1, characterized in that, The specific process of performing element-level weighted correction in step S3 is as follows: based on the channel dimension and spatial resolution of the intermediate feature variables, the phase weight coefficients are mapped to a weight matrix with the same dimension as the intermediate feature variables through linear interpolation logic; the weight matrix and the intermediate feature variables are multiplied element by element to guide the identification model to extract the energy distribution law in the discharge pulse evolution process through phase physical constraints.

6. The cable partial discharge type identification method based on ResNet according to claim 1, characterized in that, The identification model consists of a first convolutional layer, a multi-level residual module, and a global pooling layer. The identification model achieves long-distance transmission of partial discharge feature signals through the jump connection structure inside the multi-level residual module, and uses the first convolutional layer to perform spatial feature dimensionality reduction to extract deep semantic features about insulation defects in the original partial discharge data array.

7. The cable partial discharge type identification method based on ResNet according to claim 1, characterized in that, The partial discharge type identification results output in step S4 include the discharge classification probability vector and the identification confidence quantification value.

8. The cable partial discharge type identification method based on ResNet according to claim 7, characterized in that, After outputting the partial discharge type identification result, the following steps are also included: retrieving the historical online monitoring archives of the target cable to obtain the corresponding historical identification dataset; performing trend correlation analysis between the partial discharge type identification result and the historical identification dataset; if the identification confidence quantification value is lower than the preset threshold of 0.85 and the identified insulation defect type changes, an abnormal cable insulation status warning instruction is generated.

9. The cable partial discharge type identification method based on ResNet according to claim 1, characterized in that, The partial discharge type identification results cover tip discharge, surface discharge, internal air gap discharge, and metal suspension discharge.

10. The cable partial discharge type identification method based on ResNet according to claim 1, characterized in that, The identification model is deployed in an edge computing terminal installed on the side of the distribution cabinet. The edge computing terminal performs real-time calculation of the partial discharge type identification results and conducts online assessment of the insulation safety level of the cable system.

Images