Method and system for on-line monitoring and early warning of partial discharge of high-voltage switch cabinet

Abstract

This invention discloses a method and system for online monitoring and early warning of partial discharge (PD) in high-voltage switchgear, relating to the field of PD monitoring technology. The method includes: constructing a digital twin model to output a sensor health index and writing it into a shared cache; secondly, comparing the measured waveform with the twin waveform at edge nodes to generate a drift vector and encapsulating it to complete the trigger frame; when the drift exceeds a threshold or a node goes offline, using multi-scale diffusion generation combined with a peak template to reconstruct missing segments and simultaneously cache them; subsequently, a risk assessment process extracts the measured waveform, missing segments, and health index, and generates a PD confidence curve through cross-domain attention alignment and time-varying index weighting; finally, the early warning engine calculates the event severity based on the confidence curve and equipment files and generates a maintenance priority list to push work orders. This method can predict sensor degradation in advance, autonomously complete missing waveforms, form a closed loop for monitoring and maintenance, improve early warning accuracy and maintenance efficiency, and enhance operational safety.

Description

Technical Field

[0001] This invention relates to the field of partial discharge monitoring technology, specifically to a method and system for online monitoring and early warning of partial discharge in high-voltage switchgear. Background Technology

[0002] With the deepening of smart grid and distribution automation, high-voltage switchgear is gradually undertaking the mission of operating with multiple circuits, high loads, and high reliability. Partial discharge online monitoring has evolved from intermittent insulation testing to an all-weather, end-to-end data-driven mode. Currently, mainstream solutions generally employ various types of sensors, including ultra-high frequency, ultrasonic, and high-frequency current transformers. Waveforms are uploaded to a backend server via a monitoring terminal, and the discharge level is then determined using amplitude thresholds, pulse counts, or simple machine learning models. However, existing technologies mostly revolve around a single signal channel, lacking unified measurement across channels and multiple modes. At the data link level, the data acquisition and analysis ends still primarily rely on batch files or timed pushes, failing to meet millisecond-level diagnostic requirements. At the algorithm level, traditional threshold methods struggle to distinguish between electromagnetic interference and true discharge, while similar deep learning models often rely on high-performance cloud computing, making it difficult to deploy to the edge. At the equipment management level, partial discharge alarm results are often disconnected from asset ledgers and maintenance resource scheduling, leading to a passive and delayed decision-making chain.

[0003] As operating years increase, problems such as sensor power attenuation, antenna corrosion, connector oxidation, and communication network latency fluctuations frequently occur, causing data gaps in the originally seamless monitoring link and further amplifying the dependence of existing methods on complete timing sequences and stable thresholds. Therefore, in the high-load, strong electromagnetic, humid, and temperature-varying operating environment of high-voltage switchgear, existing partial discharge online monitoring technologies exhibit significant shortcomings in real-time performance, anti-interference capabilities, and closed-loop operation and maintenance. There is an urgent need for an innovative solution that can connect the entire process of data acquisition, edge processing, cloud decision-making, and resource scheduling.

[0004] In actual operation scenarios of high-voltage switchgear, key sensors are subjected to long-term stress from multiple factors such as humidity, heat, vibration, and electromagnetic coupling, which can lead to a gradual decrease in sensitivity, power fluctuations, or sudden offline phenomena. When the sensor degradation has not yet reached the complete failure threshold, its output waveform amplitude and shape have already undergone subtle shifts, and traditional amplitude thresholds or static models cannot detect this degradation process in real time. Once the degradation continues to deepen, short-term holes will appear in the real-time waveform, and current recognition algorithms based on complete rolling shutter timing cannot output reliable judgments due to the lack of continuous input, resulting in monitoring blind spots.

[0005] These blind spots coincide with the time window of the rapid rise phase of partial discharge. If the defect is not detected, the warning threshold will be misjudged as normal, making it difficult for the back-end decision-making system to issue maintenance instructions in a timely manner. This could ultimately lead to serious accidents such as insulation breakdown, phase-to-phase short circuits, or even busbar explosions. Conversely, if degradation is mistaken for discharge increments, it will trigger false alarms, causing unnecessary power outages for maintenance, increasing operation and maintenance costs, and reducing power supply reliability. Therefore, how to accurately quantify the degree of degradation in the early stages of sensor performance drift, and how to real-time supplement key waveforms and maintain the integrity of the diagnostic link after data gaps occur, and then seamlessly integrate the reliability assessment results into the asset management system to generate executable and resource-friendly maintenance priorities, are the core technical problems that urgently need to be solved in the field of online partial discharge monitoring. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides an online monitoring and early warning method and system for partial discharge (PD) in high-voltage switchgear. The method generates a drift vector by comparing measured waveforms with twin waveforms at edge nodes and encapsulates and completes the trigger frame. When the drift exceeds a threshold or a node goes offline, multi-scale diffusion generation combined with a peak template is used to reconstruct missing segments and simultaneously cache them. Subsequently, a risk assessment process extracts the measured waveform, missing segments, and health index, and generates a PD confidence curve through cross-domain attention alignment and time-varying index weighting. Finally, the early warning engine calculates the event severity based on the confidence curve and equipment records and generates a maintenance priority list to push work orders. This method can predict sensor degradation in advance, autonomously complete missing waveforms, form a closed loop for monitoring and maintenance, and solve the technical problems described in the background section.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] Online monitoring and early warning methods for partial discharge in high-voltage switchgear include:

[0011] A digital twin model is established, and the sensor health index is dynamically calculated based on historical waveforms and environmental parameters. The index, along with the real-time waveform, is written into the shared cache and marked with a version number for subsequent use.

[0012] The edge nodes continuously compare the measured waveform with the twin prediction, generate the drift vector and encapsulate the completion trigger frame, and upload the degradation direction and velocity to the shared cache for timing alignment during the completion process;

[0013] When the drift vector exceeds the threshold or a node goes offline, a multi-scale diffusion generation algorithm is triggered. The missing fragment is reconstructed by combining the twin prior and the peak template and written back to the shared cache in a synchronous manner to maintain temporal continuity.

[0014] The risk assessment process extracts real-time waveforms, missing segments, and sensor health indices. It outputs continuous partial discharge confidence curves through cross-domain attention mapping and time-varying weighted index and stores them in a versioned database.

[0015] The early warning engine calculates the severity of events based on the credibility curve and equipment operation records, and generates a maintenance priority list by combining the resource cost matching algorithm, and automatically pushes work orders to the operation and maintenance platform for execution.

[0016] Furthermore, the historical waveform matrix and the environmental state vector are aligned by timestamp, and the convolution kernel tensor and waveform-environment weight coefficients are simultaneously optimized using K-fold cross-validation maximum likelihood method. The optimization results are then solidified into the parameter library of the digital twin model.

[0017] Furthermore, information entropy-Wasserstein dual-distance calculation is performed on the real-time waveform; the health index gradient is obtained within the sliding window, and a residual vector is generated by shrinking the covariance matrix using Ledoit-Wolf; then the health index, residual vector, and version number are synchronously written into the shared cache.

[0018] Furthermore, after decomposing the real-time waveform and the twin predicted waveform using Morlet wavelet, the envelope phase difference and instantaneous energy residual are calculated to form a difference tensor, and the drift vector is generated through Riemannian manifold logarithmic mapping.

[0019] Furthermore, a comprehensive risk index is calculated based on the drift vector magnitude and information entropy density. The drift vector, peak template, and health index gradient are encapsulated into a completion trigger frame, which is then published to the message queue.

[0020] Furthermore, frequency domain weights are obtained by performing a Fourier transform on the drift vector based on the completion trigger frame, a noise scheduling sequence is generated according to the principle of high energy and low noise, and the digital twin reference waveform and peak template are used as the initial samples for diffusion.

[0021] Furthermore, a directional attention weight matrix is ​​introduced in the backdiffusion stage to perform directional correction on the noise residual output by the denoising network, and the sampling step size is adaptively adjusted according to the health index gradient to finally generate the missing fragment and write it back to the shared cache.

[0022] Furthermore, continuous wavelet transform is performed on the measured waveform and the missing segment to obtain a dual-domain feature tensor. Temperature-regulated soft attention weights are used to complete the global-local two-level mapping and output a cross-domain aligned tensor.

[0023] An exponential time-varying Dirichlet concentration vector is constructed based on the aligned tensor energy density, health index, and drift amplitude. After sampling to obtain the fusion weight, it is linearly combined with the normalized partial discharge intensity to generate a partial discharge confidence curve.

[0024] Furthermore, the background trend is obtained by fitting the partial release confidence curve with quantile regular splines, and then the short-term surge is corrected by the exponential drift coefficient to generate an adaptive threshold curve and extract the over-threshold segment to form a risk event set.

[0025] The severity of risk events and asset weights are combined to form an event-asset risk scoring matrix. After introducing maintenance resource vectors, a cost matrix is ​​constructed. An improved Hungarian algorithm is used to output a maintenance priority list and corresponding work orders.

[0026] High-voltage switchgear partial discharge online monitoring and early warning system, including,

[0027] The twin assessment module establishes a digital twin model, dynamically calculates the sensor health index based on historical waveforms and environmental parameters, and writes the index along with the real-time waveform into the shared cache, along with a version number for subsequent use.

[0028] The drift detection module continuously compares the measured waveform with the twin prediction at the edge nodes, generates the drift vector and encapsulates the completion trigger frame, and uploads the degradation direction and velocity to the shared cache for timing alignment during the completion process.

[0029] The data completion module triggers a multi-scale diffusion generation algorithm when the drift vector exceeds a threshold or a node goes offline. It combines twin priors and peak templates to reconstruct missing segments and writes them back to the shared cache simultaneously to maintain temporal continuity.

[0030] The risk fusion module extracts real-time waveforms, missing segments, and sensor health indices from the risk assessment process. It outputs continuous partial discharge confidence curves through cross-domain attention mapping and time-varying weighted index and stores them in a versioned database.

[0031] The early warning decision module uses the early warning engine to calculate the severity of events based on the credibility curve and equipment operation records, and generates a maintenance priority list by combining the resource cost matching algorithm, and automatically pushes work orders to the operation and maintenance platform for execution.

[0032] (III) Beneficial Effects

[0033] This invention provides a method and system for online monitoring and early warning of partial discharge in high-voltage switchgear, which has the following beneficial effects:

[0034] By generating a unified sensor health index in real time at the front end using a digital twin model, subtle signs of degradation can be accurately captured before the sensor fails, exposing potential blind spot risks in advance. This provides a single, reliable health measurement benchmark for all subsequent stages, enabling lossless information transfer from the hardware layer to the algorithm layer.

[0035] Edge nodes calculate drift vectors by comparing health indices with measured waveforms, and use completion trigger frames to encapsulate degradation direction, degradation rate and peak template at once and send them to the data completion process, which significantly reduces link bandwidth usage while maintaining spatiotemporal semantic integrity.

[0036] The multi-scale diffusion generation algorithm performs probabilistic reconstruction of missing segments with consistent direction under the guidance of the drift vector. Combined with peak template capping and health index gradient windowing, it ensures that the completed waveform conforms to both physical laws and statistical consistency, successfully sinking the data self-healing capability to the edge side and avoiding false alarms and missed alarms caused by monitoring interruption.

[0037] Cross-domain attention mapping first aligns the measured waveform with the missing segments, and then uses an exponential time-varying weighted strategy to generate a partial discharge confidence curve. This allows the real signal, the completed signal, and the sensor health to be dynamically integrated in the same semantic space. The curve is both sensitive to short-term sudden increases and robust in covering long-term slow changes, significantly improving the accuracy of early warning.

[0038] By leveraging adaptive threshold curves and topological persistent clustering technology, the partial discharge confidence curve is automatically discretized into risk events. Then, by combining asset weights and resource costs, a three-dimensional scoring matrix of events-asset-resources is constructed, which is matched to generate a maintenance priority list. This ensures that maintenance resources are accurately allocated to the switchgear with the highest risk, thereby achieving a dual improvement in operation and maintenance efficiency and safety margin. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the online monitoring and early warning method for partial discharge in high-voltage switchgear according to the present invention;

[0040] Figure 2 This is a schematic diagram of the structure of the high-voltage switchgear partial discharge online monitoring and early warning system of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1 This invention provides an online monitoring and early warning method for partial discharge in high-voltage switchgear, including:

[0043] High-voltage switchgear operates in a harsh environment characterized by high voltage, strong electromagnetic fields, and alternating humidity and heat. Partial discharge signals are the earliest perceptible indicator revealing insulation degradation trends. To achieve true early warning without power interruption, relying solely on the instantaneous output of front-end sensors is insufficient, as sensors themselves experience performance drift under multiple environmental stresses, weakening or even masking the true discharge characteristics. Traditional methods often trigger manual calibration only after significant distortion is detected, missing the optimal intervention window.

[0044] Step 1: Construct a digital twin model that integrates historical waveforms and environmental parameters, continuously calculate sensor health indices and write them into a shared cache with version numbers, providing a unified performance benchmark, data consistency and traceability for the entire chain, revealing micro-degradation trends and environmental coupling effects in real time, and laying a comprehensive health reference for subsequent drift capture, missing data completion, reliability fusion and maintenance sequencing.

[0045] Step 101: Calculate and update the sensor health index using a digital twin model under real-time conditions to provide a single, unified, and traceable health measurement benchmark for the entire system.

[0046] The partial discharge waveform inside the switch cabinet is extremely sparse and exhibits strong transient characteristics. It is impossible to distinguish between amplitude attenuation caused by sensor drift and signal changes caused by insulation degradation simply by relying on threshold judgment.

[0047] The digital twin model first inputs the historical waveform matrix and the environmental state vector, generates a reference spectrum through multidimensional convolution projection, then uses real-time waveform tensors for adaptive alignment, and finally outputs a sensor health index that is updated over time and written back to the shared cache. This index not only reflects the sensor sensitivity degradation, but also contains the instantaneous impact of environmental disturbances on the waveform, so it can become the direct input for subsequent drift vector calculation.

[0048] To accurately replicate physical degradation mechanisms in a virtual mapping, digital twin models First, construct the historical waveform matrix. With the environment state vector The joint feature tensor. Specifically,

[0049] The time series waveform is subjected to multi-resolution wavelet packet transform to obtain a three-dimensional frequency-energy-phase spectrum. This spectrum is then concatenated with time variables such as ambient temperature, relative humidity, and load factor after Gaussian kernel mapping to form a fourth-order tensor, and further processed by a multidimensional convolution kernel. Learn cross-domain coupling relationships to form a baseline spectrum. The projection process is characterized by the following equation:

[0050]

[0051] in, Represents the nonlinear mapping function of the convolution kernel to a multi-domain tensor; benchmark spectrum The ideal waveform spectrum presented under the current environmental assumptions is used as a twin alignment reference.

[0052] convolution kernel tensor : Higher-order weights used to capture waveform-environment coupling patterns, with values ​​in the range of real number tensors, applied to multi-domain fusion; historical waveform matrix A collection of multi-scale waveform slices from past monitoring periods, with the dimension being the number of samples. Number of sensor channels; environmental state vector : Corresponds to state variables such as temperature, humidity, and load rate at historical moments, with dimensions consistent with the time axis.

[0053] By simultaneously projecting waveforms and environmental information, the digital twin model can automatically construct ideal sensor outputs without requiring human experience, providing a quantifiable benchmark for subsequent health index calculations. Once the benchmark spectrum is generated, the digital twin model compares it with the real-time waveform tensor. Compared with the reference spectrum The sensor health index is defined using an exponential decay similarity function. .

[0054] Considering the randomness of the partial discharge waveform and the short-term interference from environmental fluctuations, a formula based on a dual weighting of information entropy distance and Wasserstein distance is adopted:

[0055]

[0056] Where: Sensor health index Real-time quantification of sensor performance integrity, value range The closer the value is to one, the healthier the performance.

[0057] Information entropy distance Kullback-Leibler divergence between the measured waveform probability distribution and the baseline distribution measures the difference in waveform morphology; Wasserstein distance. The optimal transmission distance between the real-time environmental state and the baseline environmental state measures the degree of environmental offset.

[0058] Weighting coefficients : Empirical correction parameters used to balance the contributions of waveform differences and environmental differences to the index; Observational probability distribution : From real-time waveform tensor The probability distribution obtained by kernel density estimation;

[0059] Baseline probability distribution Based on the benchmark spectrum Extracted probability distribution; baseline environmental state : The reference environment vector obtained through short-term mean smoothing is used to construct a stable control.

[0060] This similarity mapping uses an exponential decay form to transform waveform differences and environmental differences into a monotonic interpretable curve, which avoids the random errors caused by peak comparison and mitigates the instantaneous tearing of the exponent by extreme environmental fluctuations.

[0061] During the initial deployment of the digital twin, a ten-day "cold start calibration period" is introduced. This period involves adjusting the historical waveform matrix. With the environment state vector Perform K-fold cross-validation to obtain the convolution kernel tensor. With weighting coefficients Maximum likelihood estimation; after calibration, real-time output can begin. .

[0062] By solidifying the convolutional kernel tensor and similarity weights during the cold start calibration period, the digital twin model can output the sensor health index in real time according to environmental changes, thus providing a degradation warning before the sensor has completely failed. Maintenance personnel can obtain health metrics without disassembling the cabinet for inspection, enabling the monitoring system to establish a unified and reliable performance benchmark at the source, laying a traceable benchmark for subsequent drift capture and weight fusion.

[0063] Step 102: Generate a drift vector using the temporal changes of the health index to provide a priori information for subsequent reconstruction of missing segments and write the index into a shared cache.

[0064] The absolute value of the health index can measure the instantaneous state of the sensor, but it is still insufficient for judging the risk of future failure. By observing the trend of the index within the window and combining it with the real-time waveform residual, a drift vector can be constructed. This vector can not only warn of potential failures, but also provide physical constraints for data completion.

[0065] Edge nodes calculate the exponential gradient and residual covariance within the sliding time window, generate multi-component drift vectors, and write them into the shared cache. When the vector magnitude exceeds the threshold, it is determined that the sensor is about to leave the normal working boundary, thus injecting information on the degradation direction and uncertainty range into the subsequent diffusion model reconstruction data in advance, thereby reducing the generation space and improving the accuracy of the completion.

[0066] In length Within the sliding window, record the index sequence of the sensor's health index. And calculate its first-order time gradient. :

[0067]

[0068] Simultaneously, solve for the real-time waveform residuals:

[0069]

[0070] in: The twin prediction waveform output by the digital twin model is then used to calculate the residual covariance matrix. The drift vector is obtained by concatenating the matrix and vector of the two. :

[0071]

[0072] Where: exponential gradient The rate of change of the health index is used to determine the trend of decline; residual covariance matrix. Statistical distribution of quantized waveform prediction error, with the dimension being the number of channels. Number of channels;

[0073] Vectorization Operator Flatten the matrix into column vectors; coupling coefficient : Adjusting the contribution of residual statistics to the overall drift vector; drift vector A composite vector containing trend information and error propagation information, used to determine the sensor degradation level; sliding window length. Rounding is performed over a large range less than the number of sampling periods to ensure that partial discharge mutations can be captured.

[0074] By concatenating the exponential gradient and residual statistics, we can simultaneously measure the rate of continuous degradation and the degree of noise amplification, providing directional constraints for the subsequent data completion process.

[0075] To ensure data consistency under multi-process scheduling, edge nodes, after generating drift vectors, correlate them with health indices. Write it to the shared cache area together. Use a version number based on a logical clock. Perform atomic update:

[0076]

[0077] It also broadcasts the version number to the upper-level process to ensure that both the missing segment reconstruction and risk assessment processes read the most synchronized data view.

[0078] Where: Logic clock version number A monotonically increasing integer marker is used to ensure read / write consistency between concurrent processes; shared cache. : A shared memory space between edge nodes and upper-layer processes, storing real-time key variables; real-time waveform tensors The latest acquired raw waveform data is synchronously written and saved.

[0079] A logical clock mechanism is employed to avoid read / write conflicts, ensuring a strict correspondence between the health index, drift vector, and original waveform at the same time, preventing secondary misjudgments in subsequent processes due to timing misalignments. A Ledoit-Wolf mechanism for shrinking covariance and version number writing is used, enabling edge nodes to stably update residual statistics and generate drift vectors even in scenarios with insufficient samples or multi-threaded conflicts, avoiding misjudgments of degradation direction due to matrix singularities; simultaneously, the logical clock ensures cached views... Figure 1 This ensures that the completion trigger frame accurately carries the latest state, guaranteeing that the completion process is strictly aligned with real-time data.

[0080] Step one not only assigns a health index to each sensor that evolves over time, but also quantifies potential failure trends in advance using drift vectors, and synchronizes all key variables to the shared cache using version numbers. Thus, step two can directly use the sensor health index and drift vectors to determine whether to enter the data completion process when reading the cache; overcoming the multi-source data fragmentation obstacle that traditional distributed detection methods cannot avoid.

[0081] In step one, a sensor health index that evolves over time has been injected into each sensor. It is written to the shared cache in real time, but only when the micro fluctuations of the index are extracted into actionable information can they truly drive subsequent data completion and risk assessment.

[0082] Step 2: By continuously comparing the measured waveform with the twin prediction through edge nodes, a multi-component drift vector is generated and the completion trigger frame is encapsulated with a version number. The degradation direction, rate and peak template are compressed and uploaded to the shared cache. At the same time, shrinking covariance is used to ensure statistical stability, realizing low-bandwidth high semantic degradation description, and providing prior knowledge of time alignment, noise weight and morphological upper limit for multi-scale diffusion completion.

[0083] Step 201: Compare the real-time waveform with the twin prediction, and generate a high-resolution drift vector using multi-scale alignment and nonlinear mapping. This provides a unified coordinate system for the direction and magnitude of degradation. The ideal waveform of the digital twin output. Compared with the actual acquired waveform The differences between them contain two pieces of information, namely:

[0084] First, there is amplitude collapse caused by the degradation of sensor sensitivity itself; second, there is morphological distortion caused by environmental noise coupling. If only Euclidean distance is used as a measure, it is impossible to distinguish the weight of the contribution of the two to the degradation. Therefore, a multi-scale alignment mechanism is needed to correct the alignment deviation in the frequency domain, phase domain and time domain respectively. Then, the difference tensor is compressed into a single vector through nonlinear manifold mapping, so as to preserve the directionality and avoid information redundancy.

[0085] Therefore, the processing logic first applies wavelet basis sets to the ideal waveform. Compared with the actual acquired waveform A synchronous decomposition is performed, followed by the calculation of the phase difference spectrum and the energy residual spectrum. Then, the residual spectrum is mapped to a Riemannian manifold and an information geometry measure is applied, ultimately outputting the drift vector. And synchronize the version number and write it back to the shared cache.

[0086] Against a high-frequency pulse background, the partial discharge waveform exhibits strong transient and multi-harmonic characteristics. To avoid masking the high-frequency components by directly subtracting the original waveform, Morlet wavelet group decomposition is used to obtain the frequency-time spectrum:

[0087]

[0088] Next, the envelope phase difference is calculated for each frequency component. With instantaneous energy residual Combined into a difference tensor :

[0089]

[0090] in:

[0091]

[0092] Thus, multi-scale synchronous alignment decomposes the original differences into two complementary channels: phase and energy, which avoids high-frequency omissions and suppresses low-frequency drift errors.

[0093] Where: Morlet wavelet transform operator : A continuous wavelet operator for extracting transient features, used to generate a frequency-time spectrum; frequency index. : Wavelet center frequency, with a value range covering the sensor bandwidth; Delay index Partial discharge pulse arrival time window;

[0094] Envelope phase difference : Measures the phase mismatch of the waveform, with a value range Instantaneous energy residual : Measures amplitude attenuation or gain, with the same unit as waveform amplitude;

[0095] Difference tensor : Stores a two-dimensional tensor of phase difference and energy difference for subsequent mapping;

[0096] By aligning the frequency and time domain perspectives, the complex differences are decomposed into more interpretable parts, providing a fine-grained basis for subsequent mappings. This yields the difference tensor. Furthermore, directly calculating the distance in Euclidean space neglects the nonlinear coupling between the phase difference and the energy difference. Therefore, a Riemannian manifold mapping is introduced:

[0097]

[0098] in: To measure tensors The defined logarithmic map projects the difference tensor onto the tangent space; subsequently, adaptive weights are used. The phase and energy tangent vectors are concatenated and then broadcast to a unified dimension to form a drift vector:

[0099]

[0100] Where: Riemannian metric tensor The coupling weight matrix for measuring the phase channel and energy channel is positive definite, ensuring that the mapping is invertible.

[0101] Logarithmic mapping : The operation of mapping tensors to the tangent space; tangent vector : These represent the vectorized representations of the phase difference and energy difference in the tangent space, respectively;

[0102] splicing operator : Concatenate the two sets of tangent vectors element-wise along each dimension; weight coefficients Used to balance the contributions of two types of differences, with a range of values. ,and .

[0103] By using the logarithmic mapping of the Riemannian manifold, the drift vector compresses high-dimensional differences into a measurable space while preserving nonlinear coupling information, laying the coordinate foundation for estimating degradation trends. Transient differences are then transformed into unique drift vector objects through multi-scale alignment and manifold mapping, providing a unified input for step 202.

[0104] After vectorizing the health index difference and residual covariance, the system concatenates them into a multi-component drift vector. This not only captures the trend of sensor sensitivity decline but also simultaneously quantifies the waveform error distribution, forming a degradation description with directionality and amplitude. When uploading the completion trigger frame, only a compact vector needs to be sent to fully express the degradation status, significantly reducing the link load.

[0105] Step 202: Adaptively classify according to the magnitude and direction of the drift vector, and generate a completion trigger frame under bandwidth-limited conditions to drive the next process to accurately reconstruct the missing segment.

[0106] Wherein, the drift vector With diverse scales and frequent updates, uploading them as is would not only occupy the link but also easily cause congestion in upper-layer processing. Therefore, a lightweight decision-making mechanism is needed to evaluate the importance of drift vectors in real time and perform semantic compression. When the drift amplitude reaches a suspicious threshold, a missing data warning is triggered, and the necessary prior parameters are sent to the data completion process in the format of a completion trigger frame. When the amplitude is within a safe range, only the contour index is recorded locally for long-term trend analysis.

[0107] First, an entropy-energy spectrum dual-threshold strategy is used to analyze the drift vector. The high-risk drift vectors are then classified and encapsulated into complete trigger frames, along with the partial discharge morphological peak vector from the twin prior. and health index gradient This ensures that the completion algorithm has sufficient directionality and constraints. Specifically, the drift amplitude is defined. :

[0108]

[0109] And define information entropy density :

[0110]

[0111] A hazard index is constructed by linearly combining the two. :

[0112]

[0113] If the risk indicator threshold If the drift is identified as being in a danger zone, the completion process is triggered.

[0114] Where: drift amplitude The L2 norm of the drift vector quantifies the overall degradation magnitude; information entropy density. Uncertainty in the distribution of drift vector elements; combination coefficients : Used to balance the contribution of amplitude and entropy, with a value of and ;

[0115] threshold Dynamically adjusted hazard assessment threshold; hazard indicators Comprehensive judgment indicators.

[0116] The dual-threshold grading system balances amplitude magnitude and distribution complexity, avoiding misjudgments caused by simply having large amplitudes concentrated in a single dimension. Therefore, when the triggering condition is met, a complete trigger frame is constructed. And write it to the message queue, where:

[0117]

[0118] Where: complete trigger frame Message unit containing elements such as version number, drift vector, peak vector, and health gradient; peak vector : Partial discharge peak template derived from the twin model, providing shape priors for diffusion generation; Message queue: an asynchronous communication channel between edge nodes and the completion process; Version number : The logical clock marker defined above; health index gradient Used to constrain the time span of the completion process.

[0119] The queue consumer is responsible for the data completion process, which reads the completion trigger frame. Then you can use the drift vector. Specify the noise guiding direction of the diffusion model, using the peak vector. Set a rebuild limit and use a health index gradient. Limit the width of the time window.

[0120] By encapsulating key priors and scheduling them asynchronously, the completion process can obtain complete constraints without occupying real-time channels, thus shortening the generation latency. Step 202 transforms the drift vector into a triggering decision and packages the prior information and degradation trajectory together for transmission, laying the data and temporal foundation for the multi-scale diffusion reconstruction in step three.

[0121] By adding a peak template and a health index gradient to the completion trigger frame, the multi-scale diffusion generation algorithm obtains the upper limit of shape and the time window limit before entering the completion process. This avoids waveform drift caused by unrestricted noise diffusion and ensures that the reconstructed segment matches the real device state, achieving the synergy of data compression, prior injection and timing scheduling.

[0122] Step two condenses the difference between the real-time waveform and the twin prediction into a high-resolution but low-bandwidth drift vector, and ensures through a strict grading mechanism that only degraded events that truly threaten the integrity of monitoring can trigger the data completion process.

[0123] The drift vector It not only reflects the degradation depth in amplitude, but also provides noise guidance parameters in direction; together with the peak vector and health index gradient, the completion process can directly enter the constrained diffusion generation stage after receiving the completion trigger frame, shortening the reconstruction time and improving the realism.

[0124] With the combined effect of the first two steps, the real-time drift vector has been obtained. and its encapsulated completion trigger frame This trigger frame not only indicates the amplitude and direction of sensor degradation, but also carries the partial discharge template peak vector given by the digital twin. and health index gradient This provides boundary conditions for the temporal range and amplitude upper limit of missing data.

[0125] Step 3: When the drift vector exceeds the threshold or a node goes offline, the multi-scale diffusion generation algorithm is invoked. Based on the frequency domain noise weight mapping, peak template and health index gradient, the missing segment is reconstructed within a limited window. The reverse denoising convergence is accelerated by directional attention guidance. At the same time, the noise intensity range is limited to balance detail and smoothness. Then, it is written to the shared buffer synchronously with the real-time waveform to restore continuous timing and eliminate monitoring blind spots.

[0126] Step 301: Initialize the multi-scale diffusion model using the prior information in the completion trigger frame to construct a stochastic process with clear direction and boundaries for the reconstruction of missing segments.

[0127] Traditional diffusion models in electronic noise or image denoising scenarios often use a fixed noise schedule to diffuse forward and reconstruct samples in reverse iteration. However, the energy of the partial discharge waveform is concentrated in the ultra-high frequency short pulse region. When the sensor bandwidth is limited or the drift deteriorates, the signal-to-noise ratio of certain frequency bands drops sharply. If uniform noise injection is used directly, it will cause information burying.

[0128] Therefore, based on the drift vector The frequency domain components are dynamically adjusted to adjust the noise injection intensity, and the peak vector is used. Set a generation limit; then combine it with the health index gradient. By pruning the time window, the diffusion process can cover the missing intervals while avoiding overlap with known data.

[0129] First, decompose the drift vector across multiple frequency bands to obtain the weighting coefficient vector. Subsequently, according to Generate noisy scheduling sequence Next, the digital twin reference waveform segment... The forward diffusion is injected as an initial condition, mapping the noise level to each frequency band and forming the target probability manifold, thus laying the foundation for the segmented prior for reverse reconstruction.

[0130] First, in completing the time window The frequency domain amplitude spectrum is obtained by performing a Fourier transform on the drift vector. :

[0131]

[0132] Subsequently, the amplitude spectrum is normalized and mapped to a noise weight vector. :

[0133]

[0134] Construct a step-by-step noise sequence based on noise scheduling rules. :

[0135]

[0136] in, Cut to based on sensor bandwidth limitations ;

[0137] For diffusion step index, For the total diffusion steps, the Fourier operator Transform a time-domain vector into the frequency domain;

[0138] Amplitude spectrum : The energy distribution of the drift vector at various frequencies, used to indicate changes in the signal-to-noise ratio; noise weights Normalized energy percentage, range of values Basic noise intensity : Upper limit of initial noise injection for the model; noise scheduling sequence : No. Step in frequency The noise standard deviation on Decreasing.

[0139] Weight mapping ensures lower noise in the high-energy drift band to preserve the degradation direction; more noise is injected in the low-energy band to reduce the impact of measurement errors, thereby accurately locating missing features and extracting the digital twin reference waveform within the completion time window. :

[0140]

[0141] And combine it with the peak vector to form the initial sample. :

[0142]

[0143] Then, forward diffusion is performed according to the noise scheduling sequence to obtain forward diffusion samples. :

[0144]

[0145] Where: unit window function : Indicates the existence of an interval in the signal; peak vector The twin model provides a partial diffusion template, which is used as a shape prior; forward diffusion samples. : No. The noisy waveform of the step; noise tensor The mean is zero and the covariance varies with... Varying Gaussian noise; attenuation coefficient : Control signal hold rate, which is related to noise intensity.

[0146] While ensuring the shape boundary and amplitude upper limit, noise is injected in a frequency band sensitive manner to make backward sampling traceable and avoid unconstrained generation that deviates from physical laws. Step 301 establishes a forward process with clear direction and scale adaptation for multi-scale diffusion through weight mapping and probabilistic manifold construction, and prepares for noise-signal decoupling for backward reconstruction.

[0147] Based on the adaptive generation of noise weight mapping using the energy spectrum of the drift vector, the multi-scale diffusion algorithm can retain more of the original structure in the high-energy frequency band and inject stronger denoising randomness in the low-energy frequency band, thus balancing detail fidelity and overall smoothness in a single reverse sampling. This differentiated denoising strategy enhances the credibility and interpretability of the completed waveform.

[0148] Step 302: Perform backdiffusion sampling guided by the drift vector to generate missing segments. and version number Synchronously write back to the shared cache.

[0149] The key to backdiffusion lies in estimating the noise residual and progressively denoising and reconstructing the original sample. If only a general denoising network is used, the generated result will lack the sensor-specific degradation characteristics. Therefore, a directional attention guide is introduced, embedding the drift vector tangent vector shape into the multi-head attention weights of the denoising network to achieve directional weighting of the noise estimation at each step; at the same time, the sampling step size is dynamically tightened through the health exponential gradient to avoid the elongation of pseudo-waveforms caused by excessive iteration.

[0150] First, calculate the directional weight matrix. Then, the noise residual is oriented during each denoising step; finally, the missing segment is obtained. and update the shared cache. The waveform sequence in the code allows step four to instantly read the complete waveform. The noise estimate of the denoising network output is defined. for:

[0151]

[0152] Introducing a directional weight matrix In calculation First, the drift vector is aligned in time and then increased in dimension using a 1×1 convolution to make its key dimension consistent with the attention key vector of the denoising network; then, exponential decay is calculated:

[0153]

[0154] in The attention key vector within the network. Update the denoised output as an aligned version of the time-expanded drift vector:

[0155]

[0156] Perform reverse diffusion again:

[0157]

[0158] Where: denoising network : Parameters are The input consists of noisy samples and the number of steps; the directional weight matrix. Attention weights are exponentially decayed based on the drift vector direction; noise-gated output. Noise estimation after adding orientation correction; Hadamard multiplication Element-wise multiplication; pointing to the decay factor : Control the strength of drift consistency constraints.

[0159] Directional attention steers the denoising process inversely along the degradation direction, accelerating convergence while preserving degradation features and preventing the generation of smooth but physically unreliable pseudo-waveforms. The health index gradient is evaluated in real-time during the denoising iteration. :

[0160]

[0161] If the health index gradient threshold At that time, the sampling step size will be shortened to Otherwise, maintain the default step size. The missing fragment is obtained after the iteration is completed. :

[0162]

[0163] And perform a shared cache update:

[0164]

[0165] Where: threshold : Safety threshold for the rate of change of health index, the value is determined based on experience; step length Minimum and default iteration step size for backdivergence; missing fragments : The final waveform completion result; ← indicates an atomic replacement operation to ensure cache consistency.

[0166] The iteration step size is controlled by the health index gradient to match the generation rate with the sensor degradation rate; atomic update of the cache allows step four to read the first-hand complete waveform in real time, ensuring the continuity of the risk assessment timeline. Step 302 completes the reconstruction of missing segments under the dual constraints of direction guidance and step size adaptation, and immediately synchronizes it to the shared cache, providing sufficient and reliable data support for subsequent dynamic weight fusion.

[0167] Directional attention weights inject the spatiotemporal features of the drift vector into the denoising network, enabling the backdiffusion process to converge rapidly along the degradation direction; the health index gradient controls the step size to prevent excessive iteration from generating pseudo-signals; finally, the missing segments are seamlessly spliced ​​with the measured waveform to eliminate monitoring blind spots and improve the continuity and stability of subsequent credibility assessment.

[0168] If step three fails to output the missing fragment within five attempts Missing windows will be marked as "unfillable" and only the health index component will be used in the credibility calculation in step four to ensure that the link is not interrupted.

[0169] Step 3 upgrades the missing data completion process into an interpretable, schedulable, and traceable generation process: drift vector Provides direction, peak vector Set amplitude limits, health index gradient Shrinking time window, noisy scheduling sequence Imparting multi-band precision, the final combined effect generates missing segments consistent with the sensor degradation trajectory. The monitoring chain completes a closed loop from degradation detection to data self-healing, providing the early warning engine with a guarantee of the continuity and reliability of basic data.

[0170] In step three, the missing fragments have been obtained using a multi-scale diffusion generation algorithm. and real-time waveform and version number Synchronous writing to the shared cache; simultaneously, sensor health index With drift vector Continuous updates characterize the sensor's reliability and degradation trend at every moment. Step four, after reading these three types of information, outputs a continuous, quantifiable, and evolvable partial discharge reliability curve according to a dynamic weight fusion strategy. This curve directly determines the alarm level and maintenance priority of the early warning engine.

[0171] Step 4: Risk assessment process. Extract the measured waveform, complete segment, and health index. After using the temperature-regulated soft attention alignment tensor, dynamically allocate the weights of the measured, complete, and health sources according to the exponential time-varying Dirichlet model. Generate a sensitive yet robust partial discharge confidence curve and store it in a versioned database to provide a unified risk quantification for subsequent event extraction and priority determination.

[0172] Step 401: After reading the triples from the shared cache, cross-domain attention mapping is used to achieve a unified representation of the measured waveform and the completed waveform, thus preparing a homogeneous data base for dynamic weight fusion.

[0173] Measured waveform With missing fragments Although they are on the same time axis, the former originates from actual acquisition, while the latter is generated by diffusion, resulting in significantly different frequency domain smoothness and phase noise statistics. Direct splicing would cause the downstream statistical learning model to exhibit abrupt changes at the boundaries, leading to misjudgment of partial discharge intensity. Therefore, bidirectional attention is used to embed the two types of waveforms into a unified phase-energy manifold space, eliminating distribution differences and outputting an aligned tensor. .

[0174] First, time-frequency dual-domain feature tensors of the measured and completed waveforms are extracted within a sliding window. Then, an adaptive temperature factor is used to adjust the soft attention weights to balance the information density of the two domains. Finally, tensor alignment is completed under two layers of projection from the global scale and the local scale, providing a scale-consistent input for subsequent weight fusion.

[0175] In the window Within this process, continuous wavelet transforms are performed on both the measured waveform and the completed waveform:

[0176]

[0177] Then concatenate them into a two-domain feature tensor :

[0178]

[0179] Next, construct soft attention weights. :

[0180]

[0181] Where: Morlet wavelet operator Extracting transient spectral features; dual-domain feature tensor : The four-dimensional tensor obtained by splicing measured and completed features; To calculate the summation index, iterate through all frequencies and delays within the analysis window; soft attention weights. : Probability weights that measure the degree of difference between two domains; temperature factor : Control the sharpness of attention distribution, and set a value .

[0182] Among them, temperature-regulated soft attention can adaptively amplify or shrink the attention concentration according to frequency domain differences, enabling higher attention to large-difference frequency bands and providing sufficient gradient information for alignment mapping. Global-local hierarchical projection alignment is performed, employing bilinear mapping to obtain the alignment tensor. :

[0183]

[0184] in:

[0185]

[0186] Where: global mapping matrix : Learn the weight matrix of long-range coupling; The soft attention weight tensor is based on probability weights calculated from the measured-complete difference. It is a global mapping function, usually implemented as a set of one-dimensional or two-dimensional convolutional kernels with the same scale as the time-frequency grid;

[0187] Local mapping matrix Weighted local difference weight matrix; splicing operator : Concatenate tensors along the channel dimension; Align tensors : Cross-domain unified representation result, input for the next step.

[0188] The global-local two-level mapping ensures high-level semantic consistency while preserving local differences, enabling subsequent weight fusion to accurately compare the importance of the two waveforms in the isomorphic space. The alignment tensor output in step 401... Provides input from the same domain for the weight calculation in step 402, avoiding weight drift caused by different statistical distributions of waveforms from different sources.

[0189] Cross-domain attention mapping first uses temperature-regulated soft attention to perceive the differences between the measured and completed waveforms, and then outputs an alignment tensor through global and local double-layer projection to ensure that the two types of signals maintain multidimensional consistency in phase, energy and coherence in the same feature space; this can avoid misalignment of the completed segment and the real segment in frequency or phase.

[0190] Step 402: Based on the alignment tensor, sensor health index, and drift vector, dynamically calculate the fusion weights and output the partial discharge confidence curve. .

[0191] After unifying the representation, the key to evaluating the model's reliability lies in assigning weights that best reflect the true partial discharge risk to the measured signal, the completed fragment, and the health index. Using the health index as the prior concentration vector, the concentration-temperature is adjusted using the amplitude-entropy joint index of the drift vector. Then, the weights are dynamically allocated by combining the energy density estimation of the aligned tensor. Finally, a weight sequence evolving over time is formed at the output of the Dirichlet sampler. This is used to linearly integrate risk indicators and generate a continuous credibility curve.

[0192] To perform exponential time-varying Dirichlet concentration estimation, the energy density is first calculated. :

[0193]

[0194] Redefine the initial concentration vector:

[0195]

[0196] Where the coefficient Then, the drift amplitude and entropy metrics were used:

[0197]

[0198] Adjusting concentration and temperature :

[0199]

[0200] Finally, the weights are obtained by sampling from the Dirichlet distribution:

[0201]

[0202] Where: energy density : Global energy of the alignment tensor, used to characterize waveform intensity;

[0203] This is the drift amplitude weighting coefficient, with a value between 0.3 and 0.7. For information entropy weighting coefficients, The concentration-temperature amplification factor; the initial concentration vector. Prior weights for fusing information from three sources; concentration adjustment function : Amplify or shrink the weighted dispersion according to the degree of degradation; Dirichlet distribution Multidimensional probability distribution, used to generate weight vectors; weight fusion .

[0204] When the comprehensive drift intensity index for three consecutive days When the trigger rate exceeds twice the historical average, the system is automatically marked as "model aging," and data from the past week is collected into the offline training set to trigger the cloud-based retraining process.

[0205] The exponential time-varying concentration makes the weights adaptively sensitive to the degree of degradation; the more severe the drift, the sharper the distribution, thus tilting the weights towards the true waveform during abnormal time periods. Linear fusion and reliability curve generation are performed, where: a partial discharge intensity index is defined. :

[0206]

[0207]

[0208] Final credibility curve :

[0209]

[0210] in:

[0211]

[0212] Where: Partial discharge intensity index : Alignment tensor energy normalization result;

[0213] Complete the strength index Complete the waveform energy normalization result; The waveform is completed at time [time]. frequency With delay Below, the coefficient vector obtained through continuous wavelet transform; the credibility curve. The higher the comprehensive index, the greater the risk of partial discharge.

[0214] Fusion weights of each component The weights corresponding to the measured waveform, the completed waveform, and the health index are respectively determined to satisfy the following conditions: .

[0215] Linear fusion converges the three sources of information into a continuous curve. In areas of severe degradation, the measured waveform has a dominant weight; in areas with missing data, the weight of the completed waveform increases; and in areas where sensor degradation is significant but the waveform has not yet changed drastically, the health index has an increased weight, reflecting the overall risk. This is the partial discharge confidence curve output in step 402. This will be directly used as input for generating the maintenance priority list in step five, and will be written into the historical database over time to support long-term trend analysis.

[0216] The exponential time-varying Dirichlet weighting mechanism automatically adjusts the weights of the three sources based on energy density, drift amplitude, and information entropy. This makes the partial discharge confidence curve highlight measured data when degradation is severe, emphasize data completion during gaps, and introduce health index weighting during early minor changes, thus presenting a sensitive but not excessively oscillating risk expression.

[0217] Step four transforms the heterogeneous waveform and health information into a homogeneous tensor. Then, the sampling weights are adaptively selected based on the degree of degradation. The final output is the partial discharge confidence curve. This curve integrates physical waveforms, statistical energy, and sensor health into a single indicator, providing the early warning engine with a sensitive yet robust decision-making basis. It enables parameter integration from the original signal layer to the operation and maintenance layer, and ensures seamless connection of subsequent steps in terms of data semantics and timing.

[0218] Through the continuous action of the first four steps, a progressive relationship has been established around the front-end to the middle of the monitoring chain: data quality → degradation direction → missing data completion → credibility assessment. Among them, the partial discharge credibility curve... As the final comprehensive time-series indicator, it accurately carries three pieces of information: partial discharge intensity, sensor health, and data reliability. However, what the maintenance department is truly concerned with is not the curve itself, but when and what level of maintenance measures to take for which device. When reading the partial discharge reliability curve... After processing the equipment operation records, the system automatically translates continuous risk signals into discrete work orders: First, it identifies confidence peaks and continuously rising segments through an adaptive threshold curve. Then, it combines multi-dimensional metadata such as equipment alarm history, switch cabinet load level, and on-site accessibility to generate a multi-dimensional risk scoring matrix. Subsequently, it outputs a stable and resource-friendly maintenance priority list based on this matrix and transcribes it into a standard format work order to push to the operation and maintenance platform.

[0219] Step 5: The early warning engine generates an adaptive threshold curve by combining quantile splines with exponential drift. It then performs topological persistence and health gradient clustering on the threshold-exceeding segments of the confidence curve to form risk events. Finally, it constructs a risk scoring matrix based on event severity and asset weight. With the help of a resource cost matching algorithm, it outputs a stable list of maintenance priorities and automatically issues work orders to the operation and maintenance platform, thus realizing a data-driven operation and maintenance closed loop.

[0220] Step 501: Based on the partial discharge confidence curve An adaptive threshold curve is constructed and event semantic clustering is performed to provide discretized and interpretable alarm units for the subsequent risk scoring matrix.

[0221] Partial discharge confidence curve Although the performance is consistently smooth, operational decisions require clearly defined above-threshold events and sustained temperature rise periods. Using fixed thresholds can lead to frequent false alarms due to seasonal load fluctuations or changes in ambient humidity; relying solely on experience curves will prevent timely responses to sudden degradation.

[0222] Therefore, quantile splines are first used to analyze the recent partial discharge confidence curve. The background trend is obtained by fitting the curve, and then the sudden increase segment is corrected by the exponential drift coefficient to generate an adaptive threshold curve. Subsequently, the partial discharge confidence curve was... The portion exceeding the threshold is mapped to an alarm segment. Semantic clustering is then performed on adjacent alarm segments using a co-embedding space of topological persistent descriptors and sensor health gradients to obtain several risk events. Each event will carry characteristics such as peak amplitude, duration, and health decay rate, laying the foundation for the scoring matrix.

[0223] Let the length of the near window be... Take the local amplifier confidence curve. of The baseline curve is obtained by fitting quantile splines:

[0224]

[0225] in For B-spline basis functions, define the exponential drift coefficient:

[0226]

[0227] Final threshold curve :

[0228]

[0229] Where: quantile coefficient : Value This determines the degree of conservatism of the background curve; This represents the total number of B-spline basis functions involved in constructing the quantile spline curve, and its magnitude is determined by the selected node sequence.

[0230] Basis function weights : Spline fitting coefficients, obtained by minimizing the weighted absolute deviation;

[0231] Exponential drift coefficient The threshold sensitivity is adjusted by the drift amplitude; the amplification factor of the exponential drift coefficient. ;

[0232] scaling factor : Static correction factor considering seasonal loads.

[0233] Quantile splines ensure that the threshold moves gradually with medium- to long-term trends, while exponential drift responds quickly to short-term degradation, thus balancing robustness and sensitivity. Furthermore, topological persistence-healthy gradient semantic clustering is performed, first calculating the set exceeding the threshold. ,in:

[0234]

[0235] For the set of thresholds Persistent bar charts are constructed from each continuous segment using Vietoris-Rips reconstruction, and 0-1 sustaining long-scale vectors are extracted. Then, calculate the average health gradient for the corresponding segment:

[0236]

[0237] Concatenate into cluster vectors The DBSCAN output event set is adopted using spectral radius adaptive method.

[0238]

[0239] In the formula: The time span of the integration segment is represented by the difference between the start and end times of the current alarm segment, and the segment length is usually expressed in absolute value form.

[0240] Persistent Scale Vector Topological features describing peak persistence and intermittency; average health gradient : Rate of change of sensor health index within the segment; Radius-adaptive DBSCAN: Density clustering algorithm that adjusts neighborhood size based on spectral radius.

[0241] Topological persistence captures persistent patterns within events, health gradients characterize degradation rates, and joint clustering can group multi-morphological alarm fragments into physically consistent risk events. The event set output in step 501... Provides risk units with consistent granularity for the next step of assessing severity and ranking.

[0242] Quantile regularized splines combined with exponential drift coefficients generate adaptive threshold curves that can slowly shift with seasonal loads and quickly adjust for short-term degradation. Topological persistence and health gradient dual-feature clustering merge continuous out-of-threshold segments into risk events, avoiding fragment alarms and ensuring that major hidden dangers are not covered up.

[0243] Step 502: Construct an event-asset risk scoring matrix and perform resource matching to generate a stable maintenance priority list and push it to the operation and maintenance platform.

[0244] Risk events only have actionable value when mapped to specific assets and combined with on-site resource constraints. First, each risk event... Asset vectors in equipment operation files Make connections to construct a multidimensional scoring matrix. Then, the maintenance resource vector will be adjusted. Introducing and using improved Hungarian matching to solve task allocation under the dual objectives of risk-reward and resource cost, outputting a priority list. .

[0245] This list is constructed by considering event urgency, asset importance, and the availability of maintenance resources simultaneously, ensuring stable and operational ranking. An event-asset risk scoring matrix is ​​constructed by first applying dimensionless scaling to the peak and average health gradient terms, and then defining event severity. :

[0246]

[0247] Asset weight :

[0248]

[0249] Risk scoring matrix :

[0250]

[0251] Where: Severity : Combined peak value, duration, and rate of healthy decay; asset weighting : Combining asset criticality, service life, and load factor; : that is , , , , and , where is the positive weight coefficient, normalized to ; matrix elements : Indicates an event Acting on assets Risk score.

[0252] The matrix multiplicative structure strengthens the coupling weight between high-severity events and high-weight assets, highlighting truly critical maintenance targets. Furthermore, a cost matrix is ​​constructed. :

[0253]

[0254] in For resource group The set of assets that can be served is represented. Solve for:

[0255]

[0256] Obtain the matching matrix Based on the matching matrix Generate a priority list:

[0257]

[0258] Where: cost matrix Combining the inverse of risk-return with resource costs; weighting :satisfy ;

[0259] Trip Costs : Estimated man-hours from resource team to asset site; skills gap Errors in resource group skill coverage and task requirements; capacity. The number of tasks a resource group can handle simultaneously; The value is set to 0.001 to prevent the denominator from being zero;

[0260] Matching matrix The optimal allocation output by the Hungarian algorithm; This indicates that during the maintenance resource matching process, the event... With resource group The allocation decision variable between them is a binary element; priority list : A sequence of maintenance tasks to be performed, arranged in ascending order by column. The dimension specification of the matrix is ​​given by, where This represents the total number of risk events obtained through topological clustering. The number of schedulable maintenance resource groups;

[0261] Dual-objective matching ensures that high-risk events receive priority access to resources with high skill matching and suitable distance, significantly improving maintenance efficiency and risk reduction rate. Step 502 outputs a priority list. It will be serialized into a JSON ticket format that is recognized by CMMS, and will include a version number. With timestamps, the system can be tracked and audited by the higher-level system.

[0262] The event-asset risk scoring matrix multiplicatively couples severity with asset weights, and then uses a dual-objective matching algorithm to integrate resource costs, producing a stable maintenance priority list. The maintenance team can then prioritize high-risk, high-critical, and easily accessible cabinets, achieving a closed loop of intelligent monitoring and resource scheduling, and providing quantitative support for operation and maintenance decisions.

[0263] Step 5 successfully maps continuous credibility information into a discrete, executable priority list. Once the list is pushed out, it immediately triggers the generation of work orders and resource scheduling on the operation and maintenance platform, realizing data-driven optimal maintenance selection;

[0264] Simultaneously, the event-asset-resource ternary interaction information is written back to the database, providing real feedback for future model retraining. The monitoring-diagnosis-decision process is a complete closed loop, achieving intelligent early warning and optimal resource allocation for partial discharge risks in high-voltage switchgear without relying on manual intervention.

[0265] Please see Figure 2 This invention provides an online monitoring and early warning system for partial discharge in high-voltage switchgear, comprising:

[0266] The twin assessment module establishes a digital twin model, dynamically calculates the sensor health index based on historical waveforms and environmental parameters, and writes the index along with the real-time waveform into the shared cache, along with a version number for subsequent use.

[0267] The drift detection module continuously compares the measured waveform with the twin prediction at the edge nodes, generates the drift vector and encapsulates the completion trigger frame, and uploads the degradation direction and velocity to the shared cache for timing alignment during the completion process.

[0268] The data completion module triggers a multi-scale diffusion generation algorithm when the drift vector exceeds a threshold or a node goes offline. It combines twin priors and peak templates to reconstruct missing segments and writes them back to the shared cache simultaneously to maintain temporal continuity.

[0269] The risk fusion module extracts real-time waveforms, missing segments, and sensor health indices from the risk assessment process. It outputs continuous partial discharge confidence curves through cross-domain attention mapping and time-varying weighted index and stores them in a versioned database.

[0270] The early warning decision module uses the early warning engine to calculate the severity of events based on the credibility curve and equipment operation records, and generates a maintenance priority list by combining the resource cost matching algorithm, and automatically pushes work orders to the operation and maintenance platform for execution.

[0271] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0272] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0273] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0274] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0275] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for online monitoring and early warning of partial discharge in high-voltage switchgear, characterized in that: include, Establish a digital twin model, dynamically calculate the sensor health index based on historical waveforms and environmental parameters, and write the index along with the real-time waveform into a shared cache, along with a version number for subsequent use; Edge nodes continuously compare the measured waveform with twin predictions, generate drift vectors and encapsulate completion trigger frames, and upload the degradation direction and velocity to the shared cache for timing alignment during the completion process; When the drift vector exceeds the threshold or a node goes offline, a multi-scale diffusion generation algorithm is triggered. The missing fragment is reconstructed by combining the twin prior and the peak template and written back to the shared cache in a synchronous manner to maintain temporal continuity. The risk assessment process extracts real-time waveforms, missing segments, and sensor health indices. It outputs continuous partial discharge confidence curves through cross-domain attention mapping and time-varying weighted index and stores them in a versioned database. The early warning engine calculates the severity of events based on the credibility curve and equipment operation records, and generates a maintenance priority list by combining the resource cost matching algorithm, and automatically pushes work orders to the operation and maintenance platform for execution.

2. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 1, characterized in that: The historical waveform matrix and the environment state vector are aligned by timestamp, and the convolution kernel tensor and waveform-environment weight coefficients are simultaneously optimized using K-fold cross-validation maximum likelihood method. The optimization results are then solidified into the parameter library of the digital twin model.

3. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 2, characterized in that: The information entropy-Wasserstein dual distance calculation is performed on the real-time waveform; the health index gradient is obtained within the sliding window, and the residual vector is generated by shrinking the covariance matrix using Ledoit-Wolf. Then, the health index, residual vector, and version number are synchronously written into the shared cache.

4. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 3, characterized in that: After decomposing the real-time waveform and the twin predicted waveform using Morlet wavelet, the envelope phase difference and instantaneous energy residual are calculated to form a difference tensor, and the drift vector is generated through Riemannian manifold logarithmic mapping.

5. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 4, characterized in that: A comprehensive risk index is calculated based on the drift vector magnitude and information entropy density. The drift vector, peak template, and health index gradient are encapsulated into a completion trigger frame, which is then published to the message queue.

6. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 5, characterized in that: The frequency domain weights are obtained by performing a Fourier transform on the drift vector based on the completion trigger frame. A noise scheduling sequence is generated according to the principle of high energy and low noise, and the digital twin reference waveform and peak template are used as the initial samples for diffusion.

7. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 6, characterized in that: In the backdiffusion stage, a directional attention weight matrix is ​​introduced to correct the direction of the noise residual output by the denoising network, and the sampling step size is adaptively adjusted according to the health index gradient. Finally, the missing fragment is generated and written back to the shared cache.

8. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 7, characterized in that: A dual-domain feature tensor is obtained by performing continuous wavelet transform on the measured waveform and the missing segment. A global-local two-level mapping is completed by using temperature-regulated soft attention weights, and a cross-domain aligned tensor is output. An exponential time-varying Dirichlet concentration vector is constructed based on the aligned tensor energy density, health index, and drift amplitude. After sampling to obtain the fusion weight, it is linearly combined with the normalized partial discharge intensity to generate a partial discharge confidence curve.

9. The online monitoring and early warning method for partial discharge in high-voltage switchgear according to claim 8, characterized in that: The background trend is obtained by fitting the partial release confidence curve with quantile regular splines, and then the short-term surge is corrected by the exponential drift coefficient. An adaptive threshold curve is generated and the over-threshold segment is extracted to form a risk event set. The severity of risk events and asset weights are combined to form an event-asset risk scoring matrix. After introducing maintenance resource vectors, a cost matrix is ​​constructed. An improved Hungarian algorithm is used to output a maintenance priority list and corresponding work orders.

10. A high-voltage switchgear partial discharge online monitoring and early warning system, characterized in that: include, The twin assessment module establishes a digital twin model, dynamically calculates the sensor health index based on historical waveforms and environmental parameters, and writes the index along with the real-time waveform into a shared cache, along with a version number for subsequent use. The drift detection module continuously compares the measured waveform with the twin prediction at the edge nodes, generates a drift vector and encapsulates the completion trigger frame, and uploads the degradation direction and velocity to the shared cache for timing alignment during the completion process; The data completion module triggers a multi-scale diffusion generation algorithm when the drift vector exceeds a threshold or a node goes offline. It combines twin priors and peak templates to reconstruct missing segments and writes them back to the shared cache simultaneously to maintain temporal continuity. The risk fusion module extracts real-time waveforms, missing segments, and sensor health indices from the risk assessment process. It outputs continuous partial discharge confidence curves through cross-domain attention mapping and time-varying weighted index and stores them in a versioned database. The early warning decision module uses the early warning engine to calculate the severity of events based on the credibility curve and equipment operation records, and generates a maintenance priority list by combining the resource cost matching algorithm, and automatically pushes work orders to the operation and maintenance platform for execution.

Images