A method, system, and storage medium for partial discharge event-driven hierarchical discrimination of cables for edge terminals.

Abstract

This invention discloses a method, system, and storage medium for event-driven hierarchical discrimination of cable partial discharge for edge terminals. The method includes the following steps: S1: Acquire continuous stream data and extract candidate event windows from the continuous stream data according to edge-side event triggering rules; S2: Perform multi-step processing on the candidate event windows to form standardized samples; S3: Calculate compact descriptors by combining standardized samples; S4: Calculate resource status index and perform resource-aware scheduling among rapid screening mode, standard identification mode, and review mode; S5: Output partial discharge category, risk level, or review mark; S6: Perform online updates of the prototype center for samples that meet preset information conditions. This invention, by constructing an integrated mechanism of event triggering, compact descriptors, prototype matching, and resource-aware scheduling, reduces the computational and communication overhead at the edge while maintaining effective differentiation of typical internal partial discharge modes.

Description

Technical Field

[0001] This invention relates to the field of cable condition monitoring technology, specifically to a method, system, and storage medium for graded discrimination of partial discharge events in cables for edge terminals. Background Technology

[0002] During long-term operation, power cables are susceptible to factors such as insulation aging, moisture degradation, mechanical stress, construction process deviations, impurities, and complex electromagnetic environments. Partial discharge often precedes insulation breakdown. Therefore, partial discharge monitoring is not only a condition awareness issue but also a crucial step in early warning and maintenance decision-making. For urban power distribution networks, industrial park cable lines, and critical load corridors, if abnormal partial discharge is not identified in time, it may gradually develop into joint failures, terminal flashovers, or even line outages.

[0003] In existing engineering systems, one approach is to continuously upload raw monitoring data to a central server, where waveform analysis, spectrum identification, and risk assessment are performed. While this method facilitates unified modeling and subsequent management, raw pulse data is characterized by strong continuity, transient bursts, and sparse samples. Long-term full uploading not only consumes significant bandwidth but also increases energy consumption for edge node acquisition, caching, and transmission. Furthermore, it can lead to response delays when on-site network conditions are unstable.

[0004] Another approach is to set a fixed threshold or a few statistical rules at the edge to achieve partial discharge pulse counting and simple alarms at a lower cost. However, the partial discharge signal in the field varies with the site's background noise, seasonal humidity, equipment load, sensor sensitivity, and grounding environment. A fixed threshold is very likely to be effective for a period of time but become ineffective when the site is changed, resulting in an increase in false alarms or false alarms. Especially in the context of low signal-to-noise ratio, it is difficult for a single threshold to simultaneously ensure both sensitivity and robustness.

[0005] In recent years, there have also been solutions that utilize complex deep models to directly process partial discharge waveforms, spectra, or PRPD maps. These solutions can improve feature representation capabilities, but they generally rely on more parameters, larger model files, and continuous inference computation. For edge monitoring nodes that operate close to the equipment for extended periods and have limited power consumption budgets, directly copying large models from the central side often leads to problems such as long model loading times, high memory usage, unstable real-time inference, and impacts on edge-side heat generation and battery life.

[0006] Furthermore, cable partial discharge monitoring exhibits a distinct event-driven characteristic. That is, most of the time there are no effective discharge events, with diagnostically valuable pulses appearing only at certain short intervals. If a continuous identification link with full-time coverage, full resolution, and uniform complexity is still used, edge computing power is easily consumed on a large number of invalid background segments.

[0007] Meanwhile, although the partial discharge types at different sites share commonalities, there are significant deviations in specific amplitude, phase distribution, frequency band centroid, and waveform sharpness. Fixed-boundary models trained offline in a single instance often exhibit boundary shifts after deployment to new sites. If large-scale sample re-collection and overall retraining are required for each drift, not only will maintenance costs be high, but it will also fail to meet the requirements for continuous operation in the field.

[0008] Therefore, there is an urgent need for a partial discharge discrimination technology designed for edge computing scenarios, enabling the system to complete event triggering, low-cost representation, hierarchical identification, and limited online correction at the edge. For the few truly uncertain segments, verification can then be carried out by the upper-level node or the operation and maintenance side. This approach can reduce edge-side resource consumption and facilitate long-term stable operation in the field. Summary of the Invention

[0009] To address these issues, this invention provides a cable partial discharge event-driven hierarchical discrimination method, system, and storage medium for edge terminals, which solves the problems of high power consumption, large latency, and heavy communication burden caused by continuous operation of complex identification links on the edge side in existing technologies; insufficient stability of simple threshold methods under noise fluctuations and operating condition changes; and difficulty of fixed boundary models adapting to different sites, different sensors, and different operating seasons over a long period of time.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] Firstly, a hierarchical discrimination method for cable partial discharge events driven by edge terminals includes the following steps:

[0012] S1: Collect continuous stream data including continuous raw signals of partial discharge of cables, synchronous power frequency phase information, terminal resource status information and operation auxiliary information, and extract candidate event windows from the continuous stream data through edge-side event triggering rules;

[0013] S2: Perform baseline correction, background noise estimation, phase alignment, amplitude normalization, and segment quality assessment on the candidate event window to form standardized samples that can be processed in real time by the edge terminal;

[0014] S3: Extract the pulse micromorphology vector, phase density vector, and bandwidth proportion vector from the standardized samples, and calculate the compact descriptor;

[0015] S4: Calculate the resource status index based on the terminal's remaining power, processor utilization, available memory ratio, and event trigger frequency per unit time, and perform resource-aware scheduling between quick screening mode, standard identification mode, and review mode;

[0016] S5: Call the edge-side prototype memory library to perform category matching on the compact descriptor, and output the partial discharge category, risk level or mark to be reviewed by combining the dynamic threshold and confidence calibration results;

[0017] S6: For samples that meet the pre-set confidence conditions, perform online updates at the prototype center; for low-confidence samples, upload either a summary or a fragment. The update rules are as follows:

[0018]

[0019] in, For the first time before the update Class Prototype Center For the current sample's compact descriptor, For the projection matrix, To update the step size.

[0020] Furthermore, the edge-side event triggering rule is used to scan continuous stream data through a sliding window, and in each window, a trigger score is calculated based on amplitude abrupt change, window energy ratio, and pulse kurtosis. The calculation formula is as follows:

[0021]

[0022] in, Indicates the current time The instantaneous amplitude of the corresponding sampling point, Indicates the background reference amplitude. Indicates the energy of the event window. Indicates the background window energy. Indicates the pulse kurtosis index. , and These are the weighting coefficients. It is a stable term.

[0023] Furthermore, the specific steps of S2 are as follows:

[0024] S2.1: Perform baseline correction on the candidate event window;

[0025] S2.2: Perform background noise estimation on the candidate event window;

[0026] S2.3: Perform phase alignment on the candidate event window;

[0027] S2.4: Perform amplitude normalization on the candidate event window and calculate the first... Normalized amplitude of each sampling point ;

[0028] S2.5: Perform fragment quality assessment on candidate event windows and calculate sample quality scores for the same event fragment. ;

[0029] Among them, the Normalized amplitude of each sampling point The calculation formula is as follows:

[0030]

[0031] in, The mean of the window. For the window standard deviation, For the first The original amplitude of each sampling point It is a stable term;

[0032] Sample quality score The calculation formula is as follows:

[0033]

[0034] in, For signal-to-noise ratio, For phase concentration, This represents the stability of the shape of adjacent windows.

[0035] Furthermore, the pulse micromorphological vector It includes one or more of the following: peak value, rise time, decay time, half-maximum width, pulse asymmetry, and local energy density, used to represent the time-domain waveform geometry of a single discharge pulse;

[0036] The phase density encoding It is a phase density distribution vector within one or more power frequency cycles, used to represent the aggregation and migration characteristics of discharge events on the power frequency phase;

[0037] The frequency band proportion vector To divide the signal spectrum into multiple sub-bands and calculate the proportion of energy in each sub-band to the total energy of all sub-bands;

[0038] Based on pulse micromorphic vector Phase density vector and frequency band proportion vector Compute compact descriptor The calculation formula is as follows:

[0039]

[0040] in, , Harmony A mapping matrix is ​​used to project three types of vectors onto the same low-dimensional feature space through their respective mapping matrices.

[0041] Furthermore, the specific content of S4 is as follows:

[0042] First, calculate the resource status index. Resource Status Index This reflects the terminal's current computing and communication capabilities: the more remaining battery power, the lower the CPU utilization, the lower the event trigger frequency, and the higher the available memory ratio, the better. The larger the value, the more suitable the terminal is for performing high-precision discrimination;

[0043] Then based on the resource status index With preset threshold The comparison results determine whether to use the rapid screening mode, the standard identification mode, or the verification mode.

[0044] when Then it enters the standard recognition mode, and the edge terminal calls the complete compact descriptor. ;

[0045] when If it does, it enters the rapid screening mode, enables some low-cost features in the compact descriptor, and uses distance calculation or rules to perform coarse-grained discrimination, outputting a preliminary classification or anomaly label, without performing complete prototype matching;

[0046] when If the current sample is in the verification mode, the compact descriptor, phase information, and event fragment index of the current sample will be sent to the collaborative node for verification by the central side.

[0047] Furthermore, the resource status index The calculation formula is as follows:

[0048]

[0049] in, Remaining battery power For processor utilization, The frequency of event triggering per unit of time. To set the maximum event frequency, Available memory, Total memory, , , and These are the weighting coefficients.

[0050] Furthermore, the prototype memory includes prototype centers corresponding to multiple categories. And the corresponding reference phase density vector.

[0051] Furthermore, the specific content of S5 is as follows:

[0052] First, a projection mapping is performed on the compact descriptor to transform the original descriptor to the feature space where the prototype centers are located. Then, the matching distance between the projection vector and each prototype center is calculated. Take the matching distance The category corresponding to the smallest value is used as the candidate category, and the calculation formula is as follows:

[0053]

[0054] in, For the projection matrix, For cosine similarity, Encode the phase density of the sample to be judged. For the corresponding category, the reference phase density vector, , and These are preset weighting coefficients, which can be configured according to the reliability of each feature component;

[0055] After obtaining the candidate categories, let the optimal matching distance be... The second-best matching distance is Combined with sample quality score Calculate confidence level Confidence level The higher the value, the more credible the current judgment. The calculation formula is as follows.

[0056]

[0057] in, The preset adjustment coefficient, It is a stable term;

[0058] Calculate the dynamic threshold again and dynamic threshold With confidence level The comparison and calculation formula are as follows:

[0059]

[0060] in, Based on the threshold, This represents the percentage of remaining battery power. For processor utilization, The frequency of event triggering per unit of time. , , and The preset adjustment coefficient;

[0061] like If so, then output the candidate category and its corresponding risk level; if If so, then the "to be reviewed" flag will be output.

[0062] Secondly, a cable partial discharge event-driven hierarchical discrimination system for edge terminals includes an acquisition interface module, an event slicing module, a standardization processing module, a compact descriptor construction module, a sample quality assessment module, a resource-aware scheduling module, a prototype matching discrimination module, an online update module, a collaborative upload module, and an edge execution module.

[0063] The acquisition interface module is used to acquire continuous partial discharge signals and synchronization phase information;

[0064] The event slicing module is used to extract candidate event windows based on the trigger score;

[0065] The standardization processing module is used to perform baseline correction, phase alignment, amplitude normalization, and background estimation to form standardized samples.

[0066] The compact descriptor construction module is used to generate pulse micromorphology vectors, phase density codes, and bandwidth proportion vectors.

[0067] The sample quality assessment module is used to output a quality score;

[0068] The resource awareness and scheduling module is used to select a rapid screening mode, a standard identification mode, or a verification mode.

[0069] The prototype matching and discrimination module is used to call the prototype memory to complete category matching and confidence calibration;

[0070] The online update module is used to perform prototype center updates on high-confidence samples, so that the edge model slowly follows the site background, sensor drift and seasonal changes.

[0071] The collaborative upload module is used to perform summary upload or fragment upload for low-confidence samples, thereby reducing communication bandwidth usage.

[0072] The edge execution module is used to output category results, risk levels, and alarm information.

[0073] Thirdly, a cable partial discharge event-driven hierarchical discrimination storage medium for edge terminals is provided, which, when executed by a computer program by a processor, is used to implement a cable partial discharge event-driven hierarchical discrimination method for edge terminals.

[0074] The present invention has the following advantages: The present invention constructs an integrated mechanism of event triggering, compact descriptor, prototype matching and resource-aware scheduling around the edge-side partial discharge processing link. While reducing the computing and communication overhead at the edge, it maintains effective differentiation of typical internal partial discharge modes, which is different from the traditional schemes that rely on overall compression of large models or fixed boundary classification.

[0075] Meanwhile, by implementing on-demand computing through triggering and mode switching, it is more suitable for deployment on edge nodes with limited power, clock speed and memory. Combined with prioritizing the confirmation of most events on the edge side and uploading summaries or limited fragments only for low-confidence or drift samples, it can reduce the network burden caused by continuously uploading raw data.

[0076] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0077] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0078] Figure 1 This is a flowchart illustrating the implementation of a cable partial discharge event-driven hierarchical discrimination method for edge terminals according to the present invention.

[0079] Figure 2 This is a module architecture diagram of a cable partial discharge event-driven hierarchical discrimination system for edge terminals according to the present invention.

[0080] Figure 3 This is a scatter plot of the feature mapping of four types of partial discharge samples and their corresponding confusion matrix under the standard identification configuration in this embodiment of the invention.

[0081] Figure 4 This is a feature mapping scatter plot and its corresponding confusion matrix plot for rapidly distinguishing four types of partial discharge samples under the configuration in this embodiment of the invention. Detailed Implementation

[0082] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. 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.

[0083] Please see Figure 1 A cable partial discharge event-driven hierarchical discrimination method for edge terminals includes the following steps:

[0084] S1: Continuous stream data, including continuous raw signals of partial discharge in cables, synchronous power frequency phase information, terminal resource status information, and operation auxiliary information, is collected through the edge terminal. Candidate event windows are extracted from the continuous stream data according to the event triggering rules on the edge side. In each window, amplitude change, window energy ratio, and pulse kurtosis are calculated to generate a trigger score. This allows for the joint consideration of transient amplitude, local energy, and pulse morphology. Compared with triggering methods that rely on only a single threshold, it can maintain a high effective event extraction capability under strong background fluctuation conditions. This avoids the edge terminal performing subsequent analysis on a large number of invalid background segments, thereby reducing invalid calculations and power consumption, and realizing event-driven on-demand wake-up.

[0085] Edge-side event triggering rules are used to scan continuous stream data through a sliding window, calculating a trigger score in each window based on amplitude abrupt changes, window energy ratio, and pulse kurtosis. ,when Greater than the preset trigger threshold At that time, a candidate event window centered on the current moment is captured. The calculation formula is as follows:

[0086]

[0087] in, Indicates the current time The instantaneous amplitude of the corresponding sampling point, Indicates the background reference amplitude. Indicates the energy of the event window. Indicates the background window energy. Indicates the pulse kurtosis index. , and These are the weighting coefficients. It is a stable term;

[0088] S2: Perform baseline correction, background noise estimation, phase alignment, amplitude normalization, and segment quality assessment on the candidate event window to form a standardized sample with uniform amplitude scale, consistent phase reference, and noise level and quality labels. This sample can be directly used by the edge terminal for subsequent compact descriptor construction and resource-aware scheduling.

[0089] The specific steps of S2 are as follows:

[0090] S2.1: Perform baseline correction on the candidate event window: For the truncated candidate event window, first remove the DC component and slowly changing trend terms from the signal to eliminate baseline fluctuations caused by sensor drift or preamplifier zero-point offset. By estimating the background mean of the non-pulse region within the window, this mean is subtracted from the original signal to restore the event waveform to near the zero baseline, facilitating subsequent amplitude normalization and feature extraction.

[0091] S2.2: Perform background noise estimation on the candidate event window: When truncating the candidate event window, retain a length of background sampling points before and after the event pulse. Using these background sampling points, calculate the background window energy and the background reference amplitude. The background window energy is estimated by summing the squares of the background sampling points and represents the energy level of the background noise in the current environment; the background reference amplitude is the maximum absolute value of the background sampling points, representing the noise intensity level in the current environment. This estimation result is used in two denominators of the event trigger score: the amplitude ratio and the energy ratio, thus enabling adaptive adjustment of trigger sensitivity under strong background fluctuations. Furthermore, the background noise estimate is also used for the signal-to-noise ratio of subsequent samples. The calculation (the ratio of signal energy to background noise energy) can effectively distinguish between effective partial discharge pulses and random noise interference, avoiding misjudging noise segments as events or missing weak real discharges.

[0092] S2.3: Perform phase alignment on the candidate event window: using synchronously acquired power frequency phase information The partial discharge pulses within the candidate event window are mapped to a unified reference phase coordinate system. Specifically, a phase recalibration method is used to obtain the phase recalibration results. This makes event segments acquired from different power frequency cycles and different channels comparable, thereby supporting the accurate construction of phase density coding. The calculation formula is as follows:

[0093]

[0094] in, This is the phase compensation amount.

[0095] S2.4: Perform amplitude normalization on the candidate event window: calculate the first value based on the normalization formula. Normalized amplitude of each sampling point The normalized samples have a uniform amplitude scale, which facilitates subsequent compact descriptor construction and prototype matching. The calculation formula is as follows:

[0096]

[0097] in, The mean of the window. For the window standard deviation, For the first The original amplitude of each sampling point To stabilize the term and prevent small positive numbers with a denominator of zero, this processing reduces dimensional differences between different sampling links, different gains, and different stations.

[0098] S2.5: Perform fragment quality assessment on candidate event windows: Calculate sample quality scores for the same event fragment. This is used to indicate whether the event is suitable for entering the complete judgment process, and the calculation formula is as follows:

[0099]

[0100] in, For signal-to-noise ratio, For phase concentration, This refers to the stability of the shape of adjacent windows (the degree of fluctuation in background noise of adjacent windows).

[0101] when If the quality falls below a preset quality threshold, the segment is determined to be potentially affected by strong noise or occasional spikes. Only the time of the anomaly and the coarse-grained risk are recorded, without proceeding to the subsequent full prototype matching process.

[0102] S3. Extracting pulse micromorphic vectors from standardized samples Phase density encoding and frequency band proportion vector And compute compact descriptors This compact descriptor, while maintaining the ability to effectively distinguish typical modes such as internal discharge, surface discharge, corona discharge, and levitation discharge, has a fixed length and low dimension, making it suitable for caching, comparison, and frequent invocation on edge terminals, and providing input for subsequent resource-aware scheduling and prototype matching and discrimination.

[0103] Among them, pulse micromorphic vector Including one or more of peak value, rise time, decay time, half-maximum width at half maximum (FWHM), pulse asymmetry, and local energy density, used to represent the temporal waveform geometry of a single discharge pulse; phase density encoding. A phase density distribution vector within one or more power frequency cycles, used to represent the aggregation and migration characteristics of discharge events on the power frequency phase; bandwidth proportion vector. To divide the signal spectrum into multiple sub-bands and calculate the proportion of energy in each sub-band to the total energy of all sub-bands, the calculation formula is as follows:

[0104]

[0105] in, , Harmony The mapping matrix is ​​used to project the three types of vectors mentioned above onto the same low-dimensional feature space through the corresponding mapping matrix. The value of the mapping matrix is ​​obtained by linear discriminant analysis using a typical partial discharge sample set in the offline stage and remains fixed after deployment at the edge terminal. Its function is to unify the dimensions of different features and reduce the feature dimension. This indicates vector concatenation or weighted fusion.

[0106] S4: Based on the remaining battery power of the terminal Processor utilization Available memory ratio Frequency of event triggering per unit time Computational resource status index And perform resource-aware scheduling between rapid screening mode, standard identification mode and review mode;

[0107] First, calculate the resource status index. This index comprehensively reflects the terminal's current computing and communication capabilities: the more abundant the remaining battery power, the lower the CPU utilization, the lower the event trigger frequency, and the higher the available memory ratio, the better. A higher value indicates that the terminal is more suitable for performing high-precision discrimination. The calculation formula is as follows:

[0108]

[0109] in, To set the maximum event frequency, Available memory, Total memory, , , and These are the weighting coefficients.

[0110] Then based on the resource status index With preset threshold The comparison results determine the rapid screening mode, standard identification mode, or verification mode.

[0111] when If so, it enters the standard identification mode. In this mode, the edge terminal invokes the complete compact descriptor. (Including all components of pulse micromorphology, phase density encoding, and bandwidth proportion vector), perform all prototype matching and confidence calibration procedures, and output a high-confidence partial discharge category.

[0112] when If the condition is met, the system enters the rapid screening mode. In this mode, to reduce computation and power consumption, only some low-cost features in the compact descriptor are enabled (e.g., only low-dimensional subsets of pulse micromorphology vectors or frequency band proportion vectors are used), and coarse-grained discrimination is performed using distance calculation or rules to output preliminary classification or anomaly labels, without performing complete prototype matching.

[0113] when If the terminal battery is extremely low, the processor load is too high, or the current sample poses a high risk (such as the continuous appearance of abnormal fragments with blurred boundaries), then the system enters the verification mode. This mode is typically triggered when the terminal battery is extremely low, the processor load is too high, or the current sample poses a high risk (such as the continuous appearance of abnormal fragments with blurred boundaries). In this mode, the compact descriptor, phase information, and event fragment index of the sample are directly sent to the collaborative node for verification by the central side.

[0114] Based on the compact descriptor built by S3, the above-mentioned resource-aware scheduling mechanism realizes dynamic adjustment of the discrimination link complexity, enabling the edge terminal to adaptively compromise between recognition accuracy and energy consumption according to the real-time resource status, thereby extending the continuous operation time in a battery-powered environment.

[0115] S5: The edge-side prototype memory is invoked to perform category matching on the compact descriptor. Combined with dynamic thresholding and confidence calibration results, the partial discharge category, risk level, or verification label is output. This avoids maintaining a large-scale classifier on the edge side, achieving robust category discrimination through lightweight prototype matching and dynamic confidence calibration. Furthermore, it outputs a verification label when the boundary is ambiguous, rather than forcibly giving unstable conclusions. This reduces the storage and inference requirements of the edge model while improving the reliability and interpretability of the discrimination results.

[0116] After completing resource-aware scheduling, the edge terminal calls the locally stored prototype memory to perform category matching on the compact descriptor of the current sample according to the selected execution mode (fast screening mode, standard identification mode, or verification mode).

[0117] The prototype memory includes prototype centers for multiple categories (such as internal discharge, surface discharge, corona discharge, and levitation discharge). And the corresponding reference phase density vector.

[0118] First, a projection mapping is performed on the compact descriptor, and the projection matrix is... Obtained through offline linear discriminant analysis, it is used to transform the original descriptor to the feature space where the prototype centers reside. The matching distance between the projection vector and each prototype center is then calculated. Take the matching distance The category corresponding to the smallest value is selected as the candidate category. The calculation formula is as follows:

[0119]

[0120] in, For the projection matrix, For cosine similarity, Encode the phase density of the sample to be judged. For the corresponding category, the reference phase density vector, , and These are preset weighting coefficients, which can be configured according to the reliability of each feature component.

[0121] After obtaining the candidate categories, further confidence calibration is performed, and the optimal matching distance (minimum value) is set as follows. The second-best matching distance is Combined with sample quality score Calculate confidence level Confidence level The higher the value, the more credible the current judgment. The calculation formula is as follows:

[0122]

[0123] The first term represents the relative difference between the optimal and suboptimal matching distances; a larger difference indicates a clearer category distinction. The second term is adjusted using a preset coefficient. Sample quality score The higher the quality of the sample, the greater its contribution to the confidence score.

[0124] To avoid failure when using a fixed decision threshold under different load and power conditions, a dynamic threshold is calculated. When the remaining battery power is low, CPU utilization is high, or event frequency is high, the dynamic threshold is increased accordingly, making the judgment conditions more stringent and tending to output a "to be reviewed" flag to reduce the risk of misjudgment by the terminal when resources are scarce. When the sample quality score R is high, the dynamic threshold is decreased, making high-quality samples more likely to be accepted as valid judgment results. The calculation formula is as follows:

[0125]

[0126] in, Based on the threshold, This represents the percentage of remaining battery power. For processor utilization, The frequency of event triggering per unit of time. , , and This is the preset adjustment coefficient.

[0127] confidence level With dynamic threshold If a comparison is made, If so, then output the candidate categories and their corresponding risk levels. If the output is not cleared, a "to be verified" label will be output, indicating that the current sample has insufficient confidence on the edge side and needs to be sent to the higher-level node or maintenance personnel for further confirmation. The output of this step (partial discharge category, risk level, or "to be verified" label) is directly related to the execution mode selected in S4, the compact descriptor constructed in S3, and the quality score calculated in S2, forming a complete edge-side discrimination output and avoiding the forced conclusion of unstable conclusions under conditions of ambiguous boundaries.

[0128] S6. Based on the judgment results output by S5, the edge terminal performs differentiated processing on different types of samples. For samples that meet the preset confidence conditions, the prototype center is updated online. For low-confidence samples, the summary or fragment is uploaded to achieve continuous discrimination and collaborative review on the edge terminal.

[0129] When the confidence level of the sample Not lower than the dynamic threshold And it meets the high confidence condition (e.g.) ,in (Preset high confidence threshold) and sample quality score A preset quality score threshold is set, and if a sample is repeatedly classified as belonging to the same category, it is deemed suitable for updating the local prototype memory. The update rules are as follows:

[0130]

[0131] in, For the first time before the update Class Prototype Center For the current sample's compact descriptor, For the projection matrix, To update the step size.

[0132] This rule causes a weighted shift of the prototype center between its original location and the new sample projection location, resulting in a quality score. The higher the step size, the greater the contribution of new samples to the correction of the prototype; The overall update magnitude is controlled to avoid drastic perturbations to the prototype from a single sample. Through the above online updates, the edge terminal can gradually absorb site differences, sensor drift, and environmental changes, maintaining its discrimination performance without retraining.

[0133] For the samples in S5 that are to be verified (i.e. In addition to newly emerging drift candidate clusters that appear repeatedly, the edge terminal does not perform local prototype updates but instead initiates a collaborative upload strategy. Specific upload triggering conditions include: confidence level... At the preset threshold Quality rating If the quality score falls below a preset lower threshold, or if the same type of low-confidence drift candidate cluster is detected N times consecutively, the edge terminal only uploads a compact descriptor, local phase summary, event fragment index, and a raw pulse fragment of limited length, rather than the complete waveform, to reduce communication bandwidth consumption. Simultaneously, the edge terminal employs optimization strategies such as localization, operator merging, and event-driven wake-up for the partial discharge discrimination link, further reducing computational latency and standby power consumption. After receiving the uploaded information, the central side can perform more complex verification analysis and send back correction parameters to update the prototype memory or discrimination threshold on the edge side. This mechanism allows the edge terminal to collaboratively process only a small number of uncertain samples, thereby significantly reducing communication bandwidth consumption while maintaining long-term stability.

[0134] Please see Figure 2 A cable partial discharge event-driven hierarchical discrimination system for edge terminals includes an acquisition interface module, an event slicing module, a standardization processing module, a compact descriptor construction module, a sample quality assessment module, a resource-aware scheduling module, a prototype matching discrimination module, an online update module, a collaborative upload module, and an edge execution module.

[0135] The acquisition interface module is used to acquire continuous partial discharge signals and synchronization phase information, providing raw input for event-driven hierarchical discrimination and supporting low-power wake-up mechanism on the edge side;

[0136] The event slicing module is used to extract candidate event windows based on the trigger score, realize event-driven wake-up, and reduce the processing overhead of invalid background fragments on the client side;

[0137] The standardization processing module is used to perform baseline correction, phase alignment, amplitude normalization and background estimation to form standardized samples and eliminate dimensional differences between different sites and sensors.

[0138] The compact descriptor building module is used to generate pulse micromorphology vectors, phase density codes, and bandwidth allocation vectors, reducing end-side computation and storage overhead.

[0139] The sample quality assessment module is used to output quality scores, prevent low-quality segments from entering the complete discrimination chain, and improve edge adaptability.

[0140] The resource-aware scheduling module is used to select between rapid screening mode, standard identification mode, or verification mode to achieve on-demand computation.

[0141] The prototype matching and discrimination module is used to call the prototype memory to complete category matching and confidence calibration, thereby improving the credibility and interpretability of the discrimination results.

[0142] The online update module is used to perform prototype center updates on high-confidence samples, enabling the edge model to slowly follow site background, sensor drift and seasonal changes, thus enhancing long-term stability.

[0143] The collaborative upload module is used to upload summaries or fragments for low-confidence samples, reducing communication bandwidth usage and enabling edge-center collaborative review.

[0144] The edge execution module is used to output category results, risk levels, and alarm information.

[0145] The acquisition interface module of this invention is responsible for accessing the high-frequency sampling channel, the power frequency synchronization channel, and necessary operational auxiliary channels; the event slicing module is responsible for identifying short segments worthy of analysis in continuous streaming data; the standardization processing module is responsible for completing baseline correction, phase alignment, amplitude normalization, and background estimation; the compact descriptor construction module is responsible for converting time-domain pulses, phase statistics, and frequency band information into fixed-length representations; the sample quality assessment module is responsible for determining whether the current segment is suitable for entering the complete identification link; the resource-aware scheduling module is responsible for selecting the execution mode based on power consumption, computing power, memory, and event frequency; the prototype matching and discrimination module is responsible for completing category matching, confidence calculation, and threshold decision; the online update module is responsible for slowly correcting the prototype center under safe conditions; and the collaborative upload module is responsible for forwarding necessary summaries or segments to the upper-level node.

[0146] The above modules form a closed-loop structure with feedback. For example, the quality score not only affects whether to enter the complete recognition link, but also affects the confidence level and update step size; the resource status index not only determines the execution mode, but can also participate in the calculation of dynamic thresholds; the verification results returned by the collaborative uplink can also reversely correct the reference center and phase compensation amount in the prototype memory unit.

[0147] This invention can also adjust the descriptor configuration according to different equipment types. For terminal discharge monitoring scenarios, the weight of phase distribution characteristics can be increased; for joint partial discharge monitoring scenarios, the weight of pulse decay time and local energy density can be increased; for sites with complex background interference, the proportion of signal-to-noise ratio and stability terms in the quality score can be increased. The above adjustments do not change the overall principle of this invention, but are merely parameter optimizations based on specific operating conditions.

[0148] In practical applications, the adjustment coefficients in trigger thresholds, quality score thresholds, resource status index segmentation thresholds T1 and T2, and dynamic thresholds are used. All can be configured according to the hardware platform and business strategy.

[0149] For example, in scenarios where ultra-long battery life is the priority, the resource threshold for entering the standard identification mode can be appropriately increased, so that the system stays in the rapid screening mode more often; in scenarios where high-reliability alarms are the priority, the trigger threshold for the verification mode can be lowered to ensure more thorough confirmation of suspected high-risk samples.

[0150] This invention also allows the prototype memory to adopt a single-center or multi-center structure. For partial discharge modes with a relatively compact distribution within a class, only a single prototype center can be retained; for cases where the same discharge category exhibits significant sub-cluster differences under different load conditions and humidity levels, multiple sub-prototype centers can be configured for that category, and the nearest sub-prototype can be selected for discrimination during the matching phase. This configuration does not change the overall principle of this invention, but merely extends the form of the prototype memory.

[0151] For collaborative uploading strategies, this invention does not limit the uploading of the original waveform. In cases of extremely limited bandwidth, only a compact descriptor, phase summary, and a small amount of statistical information can be uploaded; in cases requiring manual image interpretation, a limited-length original pulse segment or a reconstructed simplified image can be uploaded. Therefore, this invention is compatible with different network conditions and different operational processes.

[0152] This invention constructs an integrated mechanism encompassing event triggering, compact descriptors, prototype matching, and resource-aware scheduling, centered around the edge-side partial discharge processing link. This differs from traditional schemes that rely on large-model overall compression or fixed-boundary classification. Furthermore, it exhibits strong edge adaptability: because the system does not perform high-complexity inference on continuous streaming data throughout the entire timeframe, but rather implements on-demand computation through triggering and mode switching, it is more suitable for deployment on edge nodes with limited power, clock speed, and memory.

[0153] This invention prioritizes confirming most events at the edge, uploading summaries or limited-length segments only for low-confidence or drifting samples, thus reducing the network burden caused by continuously uploading raw data. Simultaneously, through small-step online updates of high-confidence samples, the prototype center can slowly follow changes in site background, improving its continuous applicability across seasons, devices, and sensors.

[0154] Two more examples are given below.

[0155] Example 1: This example illustrates the specific implementation process of the present invention on a single edge terminal. The edge terminal can be installed in a cable termination, intermediate joint, ring main unit monitoring node, or portable partial discharge detection device to continuously receive high-frequency partial discharge raw signals and power frequency synchronization information.

[0156] 1) Data Acquisition and Event Triggering

[0157] The edge terminal continuously acquires raw partial discharge signals. Let the continuously sampled signal be... The power frequency synchronization phase is The terminal segments the continuous stream data according to a preset sliding window length. For the first... There are 1 window, and the corresponding sampling sequence is denoted as . .

[0158] To avoid performing a full analysis on all background fragments, this embodiment first calculates the trigger score for each window. Let the background reference amplitude for this window be [value missing]. Window energy is The background window energy is The pulse kurtosis index is Then its trigger score is defined as:

[0159]

[0160] when When the window contains candidate partial discharge events, the system determines that the window contains such events and extracts the preceding and following neighboring segments to form event samples.

[0161]

[0162] In a preferred embodiment, the sampling frequency can be set to 10MHz to 200MHz, the trigger window length can be set to 64 points to 2048 points, and the event segment length can be set to 128 points to 1024 points.

[0163] 2) Standardization Processing

[0164] After candidate event segments are formed, the system performs baseline correction, phase recalibration, and amplitude normalization on the event segment. Let the first... The mean of the event segments is The standard deviation is The standardized sample is then represented as:

[0165]

[0166] If the original power frequency phase corresponding to this event is Phase compensation amount is The phase recalibration result is:

[0167]

[0168] To ensure a uniform input format for event segments of different lengths on the edge terminal, this embodiment further employs a resampling method to unify each segment to a fixed length. , denoted as:

[0169]

[0170] 3) Sample quality assessment

[0171] To prevent noisy segments from entering the complete recognition process, this embodiment calculates a quality score for each event sample. Let the signal-to-noise ratio of this sample be... Phase concentration is morphological stability is The quality score is then defined as:

[0172]

[0173] when At that time, only anomaly registration or low-cost rapid screening is performed on the sample, without entering the full prototype matching process.

[0174] 4) Compact descriptor construction

[0175] For event samples that pass the quality screening, this embodiment calculates compact descriptors from three aspects: pulse micromorphology, phase distribution, and frequency band energy.

[0176] Let the first The pulse micromorphological vector of each sample is:

[0177]

[0178] The power frequency cycle is divided into The nth phase interval, then the nth The phase density encoding of a sample can be represented as:

[0179]

[0180] Furthermore, the frequency domain is divided into If there are energy sub-bands, then the frequency band proportion vector is expressed as:

[0181]

[0182] in,

[0183]

[0184] Finally, a mapping matrix is ​​used to uniformly project and concatenate the three types of features to form a compact descriptor:

[0185]

[0186] 5) Resource-aware scheduling

[0187] During long-term operation of an edge terminal, the remaining battery power, processor utilization, available memory, and event trigger frequency will change at different times. Therefore, this embodiment constructs a resource status index based on the terminal's current resource status:

[0188]

[0189] The system according to Determine the execution mode to be used for the current event:

[0190]

[0191] 6) Prototype matching discrimination

[0192] The system maintains one or more prototype centers for each type of partial discharge mode. Let the... Class prototype center is The sample descriptor is The projection matrix is Then the sample reaches the first The matching distance of a class prototype is defined as:

[0193]

[0194] The candidate categories are:

[0195]

[0196] Let the optimal matching distance and the second-best matching distance be respectively and Then its confidence level is defined as:

[0197]

[0198] The dynamic threshold is:

[0199]

[0200] The final judgment rule is:

[0201]

[0202] 7) Output Results

[0203] After the edge terminal completes the identification, it can output the partial discharge type, event timestamp, phase position, risk level, and execution mode label. The output format can be represented as follows:

[0204]

[0205] 8) Implementation effect illustration data

[0206] Please see Figure 3 and Figure 4 To illustrate the difference in performance of this embodiment under different edge detection paths, Figure 3 and Figure 4 These can be used as schematic diagrams of category distribution under standard recognition configuration and fast discrimination configuration, respectively.

[0207] Tables 1, 2, and 3 can be used as illustrative statistical results of the effects of this embodiment.

[0208] Table 1. Statistical Results of the Four-Class Confusion Matrix for Standard Identification Configuration (Unit: Classes)

[0209]

[0210] As shown in Table 1, under the standard identification configuration, the four types of partial discharge samples are mainly concentrated on the main diagonal, indicating that the present invention can reliably distinguish between internal discharge, surface discharge, corona discharge, and levitation discharge when the complete descriptor is involved in the discrimination. In particular, the internal discharge and levitation discharge categories show a more pronounced clustering on the main diagonal, indicating that this configuration has good discrimination stability in scenarios with clear inter-class boundaries. A small number of overlapping samples still exist between surface discharge and corona discharge, which is related to the similarity between the two modes in local phase distribution and some frequency band energy. However, the overall number of misclassifications is low and can meet the routine online identification needs of edge terminals.

[0211] Table 2. Statistical results of the quick discrimination configuration four-class confusion matrix (unit: cases)

[0212]

[0213] As shown in Table 2, under the fast discrimination configuration, the four types of samples still maintain a distribution characteristic dominated by the main diagonal, indicating that after compressing some features and reducing some computational steps, the present invention can still complete the basic classification of partial discharge events. Compared with the standard identification configuration, the number of off-diagonal elements in the fast discrimination configuration has increased, indicating that a certain degree of cross-discrimination will occur when the category boundaries are close or the background perturbation is strong. However, the purpose of this configuration is to conduct rapid screening under low power consumption and low latency conditions, so this moderate sacrifice of accuracy is acceptable, especially suitable for scenarios with low power consumption, high processor utilization, or a sudden increase in the frequency of event triggering per unit time.

[0214] Table 3 Comparison of target domain classification performance with different edge recognition configurations

[0215]

[0216] As shown in Table 3, the standard recognition configuration maintained higher accuracy and macro-average F1 on target domain samples, indicating that the complete descriptor and complete prototype matching strategies are more conducive to maintaining classification boundary stability in complex scenarios. Although the fast discrimination configuration had slightly lower overall performance, it still maintained a high level of usability, demonstrating that this invention does not simply abandon recognition capabilities under conditions of limited marginal resources, but rather achieves a dynamic trade-off between recognition performance and deployment costs through a hierarchical strategy. Therefore, Table 3 does not reflect the superiority or inferiority of two isolated models, but rather two implementation forms of the same technical solution under different resource constraints.

[0217] Example 2: Implementation of an edge-center coordinated partial discharge discrimination system

[0218] This embodiment illustrates the specific implementation of the present invention when the edge terminal and the central collaborative node are jointly deployed.

[0219] The system comprises an edge acquisition unit, an edge processing unit, a prototype memory unit, a collaborative upload unit, and a central verification unit. The edge acquisition unit is responsible for receiving raw partial discharge signals and power frequency synchronization information in real time; the edge processing unit is responsible for event triggering, standardization, descriptor construction, pattern scheduling, and local discrimination; the prototype memory unit is responsible for storing the edge-side prototype center; the collaborative upload unit is responsible for uploading summary information or limited-length fragments when necessary; and the central verification unit is responsible for receiving low-confidence samples and performing further verification and parameter feedback. Edge-side online updates are also supported.

[0220] In this embodiment, for samples that appear continuously and stably and meet the high confidence condition, the edge terminal is allowed to perform small-step online updates. Let the... The prototype center is at time The state is The update rule is as follows:

[0221]

[0222] To prevent abnormal samples from mistakenly entering the prototype center, the system preferably sets the following update constraints:

[0223]

[0224] 1) Collaborative upload strategy

[0225] For event samples with low confidence, isolation, suspected drift, or suspected unknown patterns, this embodiment does not immediately perform local prototype absorption, but instead performs summary upload or limited fragment upload. Let the upload criterion be... ,but:

[0226]

[0227] when At that time, the uploaded content can be defined as:

[0228]

[0229] 2) Central side verification and feedback

[0230] After receiving the samples uploaded from the edge, the central verification unit can use more complex analysis methods to make judgments and generate new parameter correction information, prototype center correction information, or summary templates based on the verification results. The returned content can be represented as:

[0231]

[0232] After receiving the returned information, the edge terminal updates its local prototype memory unit or discrimination parameters to improve its local discrimination capability for subsequent similar events.

[0233] 3) Operation closed loop

[0234] Therefore, in this embodiment, the system forms the following closed-loop operation:

[0235] Continuous monitoring Event triggered Standardized processing Descriptor Construction Pattern determination Local discrimination Online updates or collaborative submissions Central verification and parameter feedback.

[0236] 4) Implementation effect illustration data

[0237] To illustrate the differences in resource consumption and operating modes under different edge deployment configurations in this implementation method, Tables 4 and 5 can be used as illustrative implementation results.

[0238] Table 4 Comparison of complexity and edge deployment performance for different execution configurations

[0239]

[0240] As shown in Table 4, compared with the central reference configuration, the edge standard configuration shows significant reductions in parameter size, model size, single inference latency, and runtime memory. This indicates that the compact descriptor and prototype matching link constructed by this invention for edge terminals is more suitable for deployment in resource-constrained devices. Further adoption of edge fixed-point configuration further reduces model size and runtime memory, and average power consumption also decreases, demonstrating that without changing the overall discrimination logic, fixed-point configuration and operator merging can further improve the continuous operation capability of the edge.

[0241] Table 5 Performance Comparison under Different Operating Modes

[0242]

[0243] As shown in Table 5, different operating modes exhibit a clear division of labor in terms of latency, power consumption, and recognition capability. The standard recognition mode is suitable for routine daily operation, maintaining high accuracy while controlling resource consumption; the low-power fast mode further reduces average latency and power consumption, making it more suitable for continuous operation when the edge terminal's power is limited or the processor load is high; the high-risk verification mode, although slightly more resource-intensive, offers higher accuracy and is more suitable as an abnormal event confirmation path. Therefore, this invention does not employ a single fixed mode for long-term operation, but rather dynamically selects appropriate execution paths under different operating conditions through a resource-aware scheduling mechanism, thereby improving the overall applicability and long-term operational stability of the edge terminal.

[0244] As can be seen from the foregoing implementation methods, this invention does not simply interpret edge deployment as moving the central-side model to the terminal side. Instead, it simultaneously reconstructs the system from several levels: signal entry, event filtering, feature organization, discrimination methods, operating modes, and subsequent collaboration. The key is that only events that may genuinely contain partial discharge information are activated for analysis; only high-quality samples with sufficient resources are included in the complete identification process; and only truly necessary low-confidence samples are sent to the collaborative nodes. Therefore, with the same hardware budget, this invention can allocate more resources to diagnostically valuable segments rather than wasting them on large amounts of invalid background data.

[0245] From a maintainability perspective, this invention utilizes prototype memory and a small-step online update mechanism, enabling the system to slowly adapt to the field environment without frequent large-scale re-labeling and overall retraining. From an interpretability perspective, maintenance personnel can directly view the trigger score, quality score, resource status, prototype matching distance, and mode switching results corresponding to a specific alarm, facilitating the creation of closed-loop maintenance records. From an engineering application perspective, this invention is applicable to both fixed online monitoring nodes and portable inspection devices and edge gateways.

[0246] When the monitoring network of the edge terminal supports multi-node collaboration, this invention can also uniformly merge low-confidence summaries uploaded by different nodes to construct a regional-level drift map. This not only identifies novel samples that are difficult to judge from a single-node perspective, but also assists maintenance personnel in analyzing the discharge evolution relationship at different locations on the same line. This extension method is still based on edge computing and does not change the overall idea of ​​this invention, which focuses on rapid processing at the edge and minimal verification at the central side.

[0247] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cable partial discharge event-driven hierarchical discrimination method for edge terminals, characterized in that, Includes the following steps: S1: Collect continuous stream data including continuous raw signals of partial discharge of cables, synchronous power frequency phase information, terminal resource status information and operation auxiliary information, and extract candidate event windows from the continuous stream data through edge-side event triggering rules; S2: Perform baseline correction, background noise estimation, phase alignment, amplitude normalization, and segment quality assessment on the candidate event window to form standardized samples that can be processed in real time by the edge terminal; S3: Extract the pulse micromorphology vector, phase density vector, and bandwidth proportion vector from the standardized samples, and calculate the compact descriptor; S4: Calculate the resource status index based on the terminal's remaining power, processor utilization, available memory ratio, and event trigger frequency per unit time, and perform resource-aware scheduling between quick screening mode, standard identification mode, and review mode; S5: Call the edge-side prototype memory library to perform category matching on the compact descriptor, and output the partial discharge category, risk level or mark to be reviewed by combining the dynamic threshold and confidence calibration results; S6: For samples that meet the pre-set confidence conditions, perform online updates at the prototype center; for low-confidence samples, upload either a summary or a fragment. The update rules are as follows: in, For the first time before the update Class Prototype Center For the current sample's compact descriptor, For the projection matrix, To update the step size.

2. The cable partial discharge event-driven hierarchical discrimination method for edge terminals according to claim 1, characterized in that, The edge-side event triggering rules are used to scan continuous stream data through a sliding window, and in each window, a trigger score is calculated based on amplitude abrupt change, window energy ratio, and pulse kurtosis. The calculation formula is as follows: in, Indicates the current time The instantaneous amplitude of the corresponding sampling point, Indicates the background reference amplitude. Indicates the energy of the event window. Indicates the background window energy. Indicates the pulse kurtosis index, , and These are the weighting coefficients. It is a stable term.

3. The cable partial discharge event-driven hierarchical discrimination method for edge terminals according to claim 1, characterized in that, The specific steps of S2 are as follows: S2.1: Perform baseline correction on the candidate event window; S2.2: Perform background noise estimation on the candidate event window; S2.3: Perform phase alignment on the candidate event window; S2.4: Perform amplitude normalization on the candidate event window and calculate the first... Normalized amplitude of each sampling point ; S2.5: Perform fragment quality assessment on candidate event windows and calculate sample quality scores for the same event fragment. ; Among them, the Normalized amplitude of each sampling point The calculation formula is as follows: in, The mean of the window. For the window standard deviation, For the first The original amplitude of each sampling point It is a stable term; Sample quality score The calculation formula is as follows: in, For signal-to-noise ratio, For phase concentration, This represents the stability of the shape of adjacent windows.

4. The cable partial discharge event-driven hierarchical discrimination method for edge terminals according to claim 1, characterized in that, The pulse micromorphic vector It includes one or more of the following: peak value, rise time, decay time, half-maximum width, pulse asymmetry, and local energy density, used to represent the time-domain waveform geometry of a single discharge pulse; The phase density encoding It is a phase density distribution vector within one or more power frequency cycles, used to represent the aggregation and migration characteristics of discharge events on the power frequency phase; The frequency band proportion vector To divide the signal spectrum into multiple sub-bands and calculate the proportion of energy in each sub-band to the total energy of all sub-bands; Based on pulse micromorphic vector Phase density vector and frequency band proportion vector Compute compact descriptor The calculation formula is as follows: in, , Harmony A mapping matrix is ​​used to project three types of vectors onto the same low-dimensional feature space through their respective mapping matrices.

5. The cable partial discharge event-driven hierarchical discrimination method for edge terminals according to claim 1, characterized in that, The specific content of S4 is as follows: First, calculate the resource status index. Resource Status Index This reflects the terminal's current computing and communication capabilities: the more remaining battery power, the lower the CPU utilization, the lower the event trigger frequency, and the higher the available memory ratio, the better. The larger the value, the more suitable the terminal is for performing high-precision discrimination; Then based on the resource status index With preset threshold The comparison results determine whether to use the rapid screening mode, the standard identification mode, or the verification mode. when Then it enters the standard recognition mode, and the edge terminal calls the complete compact descriptor. ; when If it does, it enters the rapid screening mode, enables some low-cost features in the compact descriptor, and uses distance calculation or rules to perform coarse-grained discrimination, outputting a preliminary classification or anomaly label, without performing complete prototype matching; when If the current sample is in the verification mode, the compact descriptor, phase information, and event fragment index of the current sample will be sent to the collaborative node for verification by the central side.

6. The cable partial discharge event-driven hierarchical discrimination method for edge terminals according to claim 5, characterized in that, The resource status index The calculation formula is as follows: in, Remaining battery power For processor utilization, The frequency of event triggering per unit of time. To set the maximum event frequency, Available memory, Total memory, , , and These are the weighting coefficients.

7. The cable partial discharge event-driven hierarchical discrimination method for edge terminals according to claim 1, characterized in that, The prototype memory includes prototype centers corresponding to multiple categories. And the corresponding reference phase density vector.

8. The cable partial discharge event-driven hierarchical discrimination method for edge terminals according to claim 1, characterized in that, The specific content of S5 is as follows: First, a projection mapping is performed on the compact descriptor to transform the original descriptor to the feature space where the prototype centers are located. Then, the matching distance between the projection vector and each prototype center is calculated. Take the matching distance The category corresponding to the smallest value is used as the candidate category, and the calculation formula is as follows: in, For the projection matrix, For cosine similarity, Encode the phase density of the sample to be judged. For the corresponding category, the reference phase density vector, , and These are preset weighting coefficients, which can be configured according to the reliability of each feature component; After obtaining the candidate categories, let the optimal matching distance be... The second-best matching distance is Combined with sample quality score Calculate confidence level Confidence level The higher the value, the more credible the current judgment. The calculation formula is as follows. in, The preset adjustment coefficient, It is a stable term; Calculate the dynamic threshold again and dynamic threshold With confidence level The comparison and calculation formula are as follows: in, Based on the threshold, This represents the percentage of remaining battery power. For processor utilization, The frequency of event triggering per unit of time. , , and The preset adjustment coefficient; like If so, then output the candidate category and its corresponding risk level; if If so, then the "to be reviewed" flag will be output.

9. A cable partial discharge event-driven hierarchical discrimination system for edge terminals, characterized in that, It includes a data acquisition interface module, an event slicing module, a standardized processing module, a compact descriptor construction module, a sample quality assessment module, a resource-aware scheduling module, a prototype matching and discrimination module, an online update module, a collaborative upload module, and an edge execution module; The acquisition interface module is used to acquire continuous partial discharge signals and synchronization phase information; The event slicing module is used to extract candidate event windows based on the trigger score; The standardization processing module is used to perform baseline correction, phase alignment, amplitude normalization, and background estimation to form standardized samples. The compact descriptor construction module is used to generate pulse micromorphology vectors, phase density codes, and bandwidth proportion vectors. The sample quality assessment module is used to output a quality score; The resource awareness and scheduling module is used to select a rapid screening mode, a standard identification mode, or a verification mode. The prototype matching and discrimination module is used to call the prototype memory to complete category matching and confidence calibration; The online update module is used to perform prototype center updates on high-confidence samples, so that the edge model slowly follows the site background, sensor drift and seasonal changes. The collaborative upload module is used to perform summary upload or fragment upload for low-confidence samples, thereby reducing communication bandwidth usage. The edge execution module is used to output category results, risk levels, and alarm information.

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

Images