A method for intelligent diagnosis of partial discharge insulation defects of a cable

Abstract

The application belongs to a kind of cable partial discharge insulation defect intelligent diagnosis method, aiming at the problem of automatic identification and causal mechanism explanation of insulation defect under multi-source heterogeneous physical signal, first, four kinds of signals of ultra-high frequency electromagnetic, high-frequency current pulse, infrared heat and micro-vibration are preprocessed and multi-scale wavelet analysis is carried out, and key events such as partial discharge are automatically extracted;Subsequently, combined with the discharge mechanism of high voltage engineering, the causal relationship diagram of defect type and each physical response is constructed, the causal mapping of observation event and physical defect is realized;Further, through the light encoder and graph neural network, cross-modal feature fusion is realized under physical constraint, and model optimization is carried out by using causal consistency loss;Finally, the defect type is distinguished by using the fusion feature, and the diagnosis report with physical explanation is output. The scheme improves the accuracy and result explainability of insulation defect intelligent identification under multi-signal cooperation, and is convenient for engineering tracing and reliable application.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis of cable insulation defects and cross-modal feature modeling technology, and in particular to an intelligent diagnosis method for partial discharge insulation defects in cables. Background Technology

[0002] Currently, intelligent diagnosis and identification technology for cable insulation defects is a crucial support for online monitoring of power equipment status and high-reliability operation and maintenance. For cable insulation failures caused by partial discharge, academia and engineering have proposed various multi-source signal fusion and feature modeling methods to attempt accurate classification of typical defects such as air gap discharge, surface creepage, and metal tip corona. The mainstream technical solutions currently focus on modal feature fusion, primarily concentrating on synchronous acquisition and explicit alignment under multiple signal sources. They extract the spatiotemporal features of multimodal signals through deep learning models and employ end-to-end black-box neural networks for information integration and defect type discrimination. Typical methods include early feature-level / decision-level fusion based on time synchronization, multimodal feature weighting based on attention mechanisms, contrastive learning and multi-task assisted supervision, and representation mapping based on spectrograms, text anchoring, or pre-trained cross-modal large models (such as CLIP).

[0003] Typical industry applications are based on the monitoring of multiple physical quantities, such as ultra-high frequency (UHF) electromagnetic signals, high-frequency current pulses, surface infrared thermal response, and micro-vibration acceleration signals. Feature extraction and discrimination are performed after time-series alignment and synchronization processing. Some solutions employ multi-scale analysis and manual feature selection, supplemented by traditional machine learning classifiers. Recent developments focus on end-to-end deep networks, emphasizing automated feature learning and big data-driven complex pattern recognition. These methods are applicable to various scenarios, including high-voltage cable field testing, factory applications, and laboratory research environments, playing a role in improving fault diagnosis accuracy and operational assurance levels.

[0004] However, the aforementioned existing technologies all rely on a "black box" feature fusion paradigm, lacking physical interpretability for the interaction and alignment process of multimodal information. Specific problems include: First, reliance on temporal or spatial synchronous alignment makes it difficult to adapt to real-world engineering conditions such as asynchronous on-site acquisition, communication delays, or heterogeneous sampling rates. Second, deep fusion strategies fail to explicitly model the causal relationships between various physical signals; cross-modal semantic alignment mostly stops at surface-level similarity learning in the feature space, making it difficult to trace the physical mechanisms and evolution paths of multi-physics responses. Third, most existing methods only use physical laws as post-processing verification or dimensionality reduction priors, failing to implement physical logical constraints on the internal processes of the model, resulting in limited reliability and engineering usability of diagnostic results. When abnormal operating conditions or noise interference occur, the system is prone to high-confidence erroneous judgments or lack of interpretability, making it difficult to support the safe operation needs in complex real-world scenarios.

[0005] Therefore, the current field of cross-modal intelligent diagnosis of partial discharge insulation defects in cables urgently needs a modeling technology that possesses both multimodal adaptive alignment capabilities and the ability to deeply integrate physical causal mechanisms to generate traceable interpretable paths. An ideal solution should be able to automatically guide the implicit semantic fusion of multiple signal types based on the cable insulation failure mechanism without requiring time synchronization, and optimize feature integration through the principle of physical consistency, ensuring the causal interpretability and engineering practicality of classification decisions. Overcoming this technical bottleneck will significantly improve the robustness, reliability, and debugging efficiency of diagnostic systems in real engineering environments, helping to propel the power equipment health management system towards a new stage of high efficiency, intelligence, and interpretability, and providing a technical foundation for widespread deployment. Summary of the Invention

[0006] This application provides an intelligent diagnostic method for partial discharge insulation defects in cables, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0007] This application provides an intelligent diagnostic method for partial discharge insulation defects in cables, specifically including: S1: Acquire multi-source heterogeneous raw time-series data in the cable partial discharge monitoring scenario, and perform local abrupt event detection to extract discrete event label sequences; S2: Based on the discharge mechanism chain of electron avalanche to structural micro-strain accumulation in high-voltage engineering, a physical cause-effect graph of cable insulation defects is constructed using the discrete event labeling sequence; S3: Input the discrete event tag sequence into the lightweight encoder of the corresponding mode, perform feature mapping operation, and generate an initial mode feature vector that is mapped to a unified dimension implicit semantic space and serves as the potential representation of the observation node in the physical causal graph of the cable insulation defect; S4: Based on the topological constraints of the physical cause-effect graph of the cable insulation defect, perform graph neural network message passing processing on the initial modal feature vector to generate causal modulation fusion features; S5: Calculate the feature transformation deviation on each causal edge in the physical causal graph of the cable insulation defect according to the preset physical monotonicity constraint, construct the causal consistency loss function, and generate the physical constraint gradient; S6: Utilize the physical constraint gradient to perform backpropagation optimization on the parameters of the lightweight encoder, driving the initial modality feature vector to converge along the physical causal direction in the implicit semantic space, generating the final fused feature representation; S7: Perform a defect type discrimination operation on the final fused feature representation to generate a defect type confidence distribution; S8: Based on the activation intensity of the critical edge in the confidence distribution of the defect type, trace the decision path and generate a cable insulation defect diagnosis report.

[0008] The intelligent diagnostic method for partial discharge insulation defects in cables provided in this application has the following beneficial effects: (1) By deeply embedding the physical mechanism of cable insulation defects into the entire process of cross-modal modeling, this invention effectively overcomes the dependence of traditional multi-source signal fusion methods on strict time synchronization and high-precision sampling alignment. It achieves reliable correlation of heterogeneous modal data such as ultra-high frequency electromagnetic, high frequency current, infrared thermal imaging and micro-vibration acceleration without the need for complex time alignment. Unlike existing technologies that commonly use attention mechanisms, optimal transmission, or contrastive learning to forcibly shorten the representation distance between modes, this scheme constructs a directed acyclic graph model based on the physical causal chain of typical partial discharge types. The discrete event markers extracted from each mode (such as UHF pulse start point, di / dt zero crossing point, thermal gradient jump frame, etc.) are mapped to a unified semantic space as observation nodes. In the message transmission process of the graph neural network, the system is forced to follow a directional physical path such as "discharge intensity → electromagnetic radiation → conduction distortion → temperature rise rate → micro-strain" for feature modulation. This enables the information collected by different sensors to achieve implicit collaboration and dynamic verification under causal logic constraints, which significantly improves the robustness and generalization ability of the system under real-world conditions such as heterogeneous sampling rates, communication delay fluctuations, and severe field interference.

[0009] (2) By introducing a causal consistency loss function based on physical monotonicity constraints, the model training process is no longer limited to minimizing classification error or maximizing modal similarity, but actively optimizes the rationality of cross-modal feature transformation along preset causal edges. For example, it ensures that the infrared temperature rise trend corresponding to the increase in discharge intensity does not show an abnormal decrease, thereby driving each modal encoder to output an intrinsic representation that conforms to the basic laws of high-voltage engineering. This mechanism fundamentally avoids the semantic drift and false correlation problems caused by black-box fusion, and enables the network to learn a joint representation structure consistent with the actual physical evolution process. This not only greatly improves the accuracy and stability of defect identification, but also gives the final decision a clear physical interpretation path—the classification confidence can be traced back to the activation intensity of key causal edges, supporting maintenance personnel to understand "why it is judged as air gap discharge" or "which modal linkages the evidence of surface creep comes from". Compared with methods that rely on PRPD spectrum modeling, image conversion or large model pre-training, this solution does not require manual construction of two-dimensional feature maps or the introduction of external knowledge bases. It has low computational overhead and flexible deployment, and is particularly suitable for cable online monitoring scenarios where edge resources are limited but diagnostic reliability requirements are high.

[0010] (3) The overall architecture abandons the explicit alignment strategies (such as timestamp synchronization, attention-weighted fusion, and contrastive loss design) commonly used in traditional cross-modal learning, as well as the reliance on dedicated network structures (such as lightweight integrated networks and CLIP-like architectures). It achieves a fundamental innovation in the modeling paradigm for power equipment fault diagnosis tasks. The entire identification process organizes information flow around the "physical causal graph," forming a closed-loop, self-consistent multi-physics response reasoning system. This ensures the effective integration of multi-source heterogeneous signals and enhances the model's tolerance to external disturbances and sensor failures. For example, when a certain modal signal is lost due to interference, the remaining modal signals can still indirectly infer their possible states through causal paths, maintaining the system's continuous diagnostic capability. This feature is particularly suitable for real-world applications where underground cable channels are complex, sensor deployment conditions are limited, and local degradation is likely to occur during long-term operation. It provides a scalable technical framework for building a new generation of intelligent diagnostic systems with strong interpretability, high adaptability, and sustainable evolution capabilities.

[0011] The combined effect of the above-mentioned technical means has constructed a new pattern recognition paradigm that deeply integrates physical mechanisms and data-driven approaches, achieving a leap from "empirical signal splicing" to "causal state deduction". This significantly improves the relevance, reliability and interpretability of cable partial discharge defect diagnosis, while maintaining low computational complexity and engineering implementation threshold, and has outstanding practical value and promotion prospects. Attached Figure Description

[0012] Figure 1 This is the main flowchart of an intelligent diagnostic method for partial discharge insulation defects in cables. Figure 2 This is a sub-flowchart of an intelligent diagnostic method for partial discharge insulation defects in cables; Figure 3 This is another sub-flowchart of a method for intelligent diagnosis of partial discharge insulation defects in cables. Detailed Implementation

[0013] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0014] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0015] like Figure 1 As shown, this application provides an intelligent diagnostic method for partial discharge insulation defects in cables, specifically including: S1: Acquire multi-source heterogeneous raw time-series data in a cable partial discharge monitoring scenario, perform local abrupt event detection, and extract discrete event marker sequences. The multi-source heterogeneous raw time-series data includes: ultra-high frequency electromagnetic signals, high-frequency current pulse signals, surface infrared thermal response signals, and micro-vibration acceleration signals. The discrete event marker sequences include: UHF pulse cluster start points, di / dt zero-crossing points, thermal gradient jump frames, and vibration energy peak frames.

[0016] S2: Based on the discharge mechanism chain of electron avalanche to structural micro-strain accumulation in high-voltage engineering, the discrete event labeling sequence is used to construct a physical causal graph of cable insulation defects, generating a directed acyclic graph topology structure with typical defect types as root cause nodes and multi-physics response causal paths as causal edges.

[0017] S3: Input the discrete event tag sequence into the lightweight encoder of the corresponding modality, perform feature mapping operation, and generate an initial modal feature vector that is mapped to a unified dimension implicit semantic space and serves as the potential representation of the observation node in the physical causal graph of the cable insulation defect.

[0018] S4: Based on the topological constraints of the physical causal graph of the cable insulation defect, perform graph neural network message passing processing on the initial modal feature vector, specifically through cross-node feature updates via causal edges, to generate causal modulation fusion features that conform to physical consistency logic.

[0019] S5: Calculate the feature transformation deviation on each causal edge in the physical causal graph of the cable insulation defect according to the preset physical monotonicity constraint, construct a causal consistency loss function to reflect the logical relationship between physical quantities such as discharge intensity and infrared temperature rise rate, and generate a physical constraint gradient for reverse adjustment of encoder parameters.

[0020] S6: Utilize the physical constraint gradient to perform backpropagation optimization on the parameters of the lightweight encoder, driving the initial modal feature vector to converge along the physical causal direction in the implicit semantic space, generating a final fused feature representation with cross-modal semantic alignment characteristics.

[0021] S7: Perform a defect type discrimination operation on the final fused feature representation to generate a defect type confidence distribution. The defect type confidence distribution includes classification results for air gap discharge, surface creepage, or metal tip corona discharge.

[0022] S8: Based on the activation intensity of the key edges in the confidence distribution of the defect type, trace the decision path, generate a cable insulation defect diagnosis report with physical interpretability, and complete the closed-loop diagnosis process from multi-source signal acquisition to interpretable classification result output.

[0023] Step S1: Acquire ultra-high frequency electromagnetic signals, high-frequency current pulse signals, surface infrared thermal response signals, and micro-vibration acceleration signals under the cable partial discharge monitoring scenario. Perform local abrupt event detection on the multi-source heterogeneous raw time-series data, and extract a discrete event marker sequence containing the UHF pulse cluster start point, di / dt zero-crossing point (i.e., the point where the current change rate is zero), thermal gradient jump frames, and vibration energy peak frames. Specifically, this includes: S1.1: Simultaneously acquire and preprocess ultra-high frequency electromagnetic signals, high frequency current pulse signals, surface infrared thermal response signals, and micro-vibration acceleration signals to generate a multi-source heterogeneous original time-series data set containing noise interference, providing basic data input for subsequent local feature extraction.

[0024] In cable partial discharge monitoring scenarios, multi-source heterogeneous raw signals from ultra-high frequency electromagnetic sensors, high frequency current pulse sensors, surface infrared thermal imagers, and micro-vibration accelerometers are acquired as input. Signal sampling must be conducted under unified acquisition control parameters to avoid timing coverage errors caused by the randomness of sensor startup delays. For the raw data streams output by each modal sensor, bandwidth-matched filtering is performed to ensure that the sampling rate and frequency response range of different modal signals meet the physical requirements for capturing discharge-related transients. The cutoff frequency is determined based on the physical characteristics of each sensor by setting the parameters of the bandpass filter. For the bandwidth-matched signals, analog-to-digital conversion accuracy correction is performed, and the quantization bit width is adjusted to maintain effective resolution within the amplitude range of abrupt events, avoiding signal saturation or quantization noise affecting the sensitivity of partial discharge identification. The digitized signals of each modality are input into a unified time base index module, recording the acquisition timestamp and attaching a sensor identification code to construct a synchronous data structure that can be used for subsequent cross-modal correlation. In this process, strict time synchronization requirements are not imposed to maintain the independence of acquisition between modes. Background noise modeling and noise superposition simulation are performed on the signal sample set in the aforementioned synchronous data structure. A noise interference component consistent with the field is generated using a noise power spectral density model based on the sensor's operating environment, and linearly superimposed onto the original signal waveform to form a multi-source heterogeneous original time-series data set containing noise interference. Through this processing method, the result of the previous step is transformed into a data set with unified acquisition control, bandwidth matching, quantization accuracy correction, and noise interference characteristics, achieving input consistency and physical scene reproduction capabilities for subsequent local abrupt event detection.

[0025] S1.2: Based on wavelet transform processing, the original multi-source heterogeneous time series data set is decomposed into multiple scales to generate a multi-scale wavelet coefficient matrix that reflects the energy distribution characteristics of each modal signal at different frequency components, so as to eliminate background noise and highlight local transient features.

[0026] For the multi-source heterogeneous raw time series data set output by S1.1, a multi-scale analysis task is established to address the non-stationary characteristics of different modal signals. Ultra-high frequency electromagnetic signals, high frequency current pulse signals, surface infrared thermal response signals, and micro-vibration acceleration signals are sequentially input into the wavelet transform processing pipeline.

[0027] For each modal time series data, an orthogonal wavelet basis function matching its characteristic frequency band is selected, and the decomposition layer parameter is set according to the sampling frequency and the detection target to complete the calculation of the multi-scale decomposition coefficients.

[0028] In the wavelet decomposition process, the scaling function coefficients are used to characterize the low-frequency smoothing components, and the wavelet function coefficients are used to characterize the high-frequency transient disturbances, effectively distinguishing between background noise and discharge transient characteristics.

[0029] A threshold function is used to perform soft or hard thresholding denoising on the wavelet coefficients of each layer. The threshold calculation formula is as follows:

[0030] in The standard deviation of the noise estimate, Given the number of sample points, the standard deviation of the noise estimate is obtained by estimating the standard deviation of the high-frequency coefficients.

[0031] The wavelet coefficients after thresholding are reorganized according to scale and time position to form the coefficient matrix of energy distribution of this mode under different frequency components.

[0032] The modal coefficient matrices are indexed uniformly according to the time step and scale to ensure that the input space for subsequent mutation event detection is consistent.

[0033] Through the above multi-scale decomposition and denoising process, the background noise of the original time series signal is significantly suppressed, the discharge-related transient characteristics are preserved, and the local mutation events are highlighted in multiple frequency domains and multiple time scales.

[0034] S1.3: Adaptive threshold detection processing is used to perform abrupt change point identification operations on the multi-scale small s-wave coefficient matrix to generate a time sequence of candidate local abrupt change events that indicate drastic changes in signal amplitude or gradient, and to initially lock the potential discharge signal segments in each mode.

[0035] For the wavelet coefficient matrix of multi-source heterogeneous signals after multi-scale wavelet decomposition, an adaptive threshold function based on statistical characteristics and energy aggregation is constructed for the coefficient distribution of each modal signal to separate the background noise level from the abrupt change feature level. During the threshold function construction process, the basic threshold level is determined using statistical measures such as the mean, variance, and local kurtosis of the decomposition coefficients, and dynamically corrected in conjunction with the modal signal-to-noise ratio parameter, enabling the threshold parameters of different modal signals to have self-adjusting capabilities. For each coefficient matrix, element-wise threshold determination is performed based on the row and column indices, and a zero-to-one mask matrix is ​​generated to identify candidate positions of abrupt change points. A mask value of 1 indicates that the amplitude or gradient exceeds the threshold. Based on the localization of candidate abrupt change units, and combined with the cross-scale correlation of wavelet coefficients, a multi-scale consistency detection strategy is used to identify coefficients that simultaneously exhibit significant amplitude changes at multiple scales, thereby enhancing the robustness of transient event detection. Cross-scale consistent abrupt change points are aggregated by time index into a candidate local abrupt change event time point sequence to initially identify signal segments with cable partial discharge characteristics and provide the original event locations for subsequent morphological verification.

[0036] By using an adaptive threshold detection processing method, the wavelet coefficient matrix of the previous step is transformed into a time sequence of candidate local mutation events that indicate drastic changes in signal amplitude or gradient, thereby initially locking the potential discharge signal segments in each mode.

[0037] For example, in cable partial discharge monitoring, the UHF electromagnetic signal undergoes 5 layers of Daubechies-8 wavelet decomposition, with a measured signal-to-noise ratio (SNR) of 18.5. The basic threshold is determined by multiplying the square root of the coefficient variance by a scaling correction factor of 0.75, and dynamically corrected using an SNR normalization factor. The high-frequency current pulse signal has an SNR of 12.0, and the threshold correction factor is set to 0.85. The threshold determination is based on the abrupt change mask value calculated using the following formula:

[0038] Where M is the element of the mask matrix, and H is the Heaviside step function. This represents the coefficient value at the j-th time index of the i-th scale. The threshold is used for the corresponding parameters. In cross-scale consistency detection, when the mask values ​​of three adjacent scales are all 1, it is determined to be a mutation point, and a candidate mutation event time index list is generated. This list is output after statistical aggregation; for example, 12 candidate events are generated for the UHF mode and 9 candidate events are generated for the high-frequency current mode. The processing effect of this embodiment significantly improves the transient event detection capability in field verification and stably provides the candidate event location data required for subsequent morphological verification.

[0039] S1.4: Based on the waveform criteria defined by the physical mechanism of high voltage insulation defects, the time sequence of the candidate local mutation events is subjected to morphological verification processing to generate a list of valid local mutation events after artifact removal, ensuring that the extracted events conform to the characteristics of real physical processes such as air gap discharge or surface creepage.

[0040] The candidate local mutation event time point sequence output after adaptive threshold detection is set as the input object as a multimodal mutation candidate list containing ultra-high frequency electromagnetic signals, high frequency current pulse signals, surface infrared thermal response signals and micro-vibration acceleration signals. Based on the waveform criteria defined by the physical mechanism of high voltage insulation defects, modal morphological feature construction processing is performed to convert the original time segment corresponding to the candidate event into a criterion vector set containing morphological description parameters such as peak amplitude, waveform sharpness, half-peak width, and polarity sequence.

[0041] Physical mechanism matching analysis is performed on the criterion vector set. Typical waveform templates such as air gap discharge and surface creepage are called from the defect mechanism knowledge base. The similarity coefficient between morphological parameters and template features is calculated. The candidate event vectors that conform to the discharge physical mode are screened out by combining Euclidean distance calculation and normalized correlation coefficient.

[0042] In the event vector set after similarity filtering, multimodal cross-validation is performed. For the time points of abrupt events in different signal modes, the relative position stability in the physical mechanism chain is analyzed, and isolated event records that lack mutual corroboration in the multimodal response are removed to reduce the probability of misjudgment.

[0043] For events retained after cross-validation, noise artifact removal is performed. By utilizing the amplitude statistics of the time windows before and after and the difference in waveform spectrum energy distribution, non-discharge abrupt fragments originating from mechanical shock or environmental interference are identified and filtered out.

[0044] The filtered event set is sorted by time series and bound with modal tags to generate a valid local mutation event confirmation list, enabling accurate determination from candidate mutations to reliable physical discharge events.

[0045] By constructing morphological criteria, matching physical mechanisms, performing multimodal cross-validation, and removing artifacts, the time sequence of candidate local mutation events from the previous step is transformed into a list of valid local mutation events that conform to the real physical process characteristics of air gap discharge or surface creepage, thus achieving high accuracy and high reliability in partial discharge event extraction.

[0046] S1.5: Based on the list of confirmed effective local mutation events, extract key feature parameters including the starting point of UHF pulse clusters, di / dt zero-crossing points, thermal gradient jump frames, and vibration energy peak frames, and generate a discrete event tag sequence to characterize the multi-physics response state, thus completing the transformation from continuous time-series data to discrete semantic events.

[0047] After morphological verification and artifact removal, a list of valid local mutation events is confirmed. The multimodal event parsing module is called to classify and index the different modal events in the list, forming an index mapping table corresponding to UHF pulse clusters, di / dt zero crossings, thermal gradient jump frames, and vibration energy peak frames.

[0048] For the UHF pulse cluster events in the index mapping table, peak detection and time positioning processing are used to extract key parameters such as the start time, peak amplitude and pulse cluster duration, and record them in vector form.

[0049] For the di / dt zero-crossing event, perform first-order differential calculation and zero-crossing analysis to extract parameters such as the zero-point occurrence time, the change in slope before and after, and the corresponding waveform amplitude abrupt change rate to characterize the current waveform distortion features.

[0050] For the thermal gradient jump frame event, perform infrared thermal image frame difference operation and temperature rise rate calculation to extract parameters such as the gradient jump start frame number, temperature rise rate value, and maximum gradient value of thermal field distribution change.

[0051] For the vibration energy peak frame event, short-time energy integration of the acceleration signal and identification of the dominant frequency component are performed to extract parameters such as peak frame time position, total energy and frequency spectrum centroid position to characterize the micro-strain amplitude of the structure.

[0052] By merging the key feature parameters of each modal event and generating a single event tag for each event based on the discretization coding rules, a discrete event tag sequence covering four physical response modes is constructed.

[0053] By extracting key feature parameters and discretizing them, the list of valid local mutation events from the previous step is transformed into a discrete event label sequence that characterizes the response state of multiphysics fields. This enables the mapping from continuous time-series data to discrete semantic events, providing a physically semantic input for the subsequent construction of causal graphs.

[0054] For example, in a 10kV cable partial discharge online monitoring scenario, the list of valid local abrupt change events includes 8 UHF modal events, 5 di / dt zero-crossing events, 3 thermal gradient jump frame events, and 2 vibration energy peak frame events. The starting point of the UHF pulse cluster is extracted using the peak threshold method, with an amplitude threshold set to 0.8 mV and a starting time positioning error of less than 0.2 μs. The zero-crossing time of the di / dt events is obtained through gradient calculation, and the slope change is calculated using the following formula:

[0055] Where I(t+1) and I(t-1) are the current amplitudes at the next and previous sampling times, respectively, and Δt is the sampling interval; the formula for calculating the temperature rise rate of the thermal gradient jump frame is:

[0056] Where T(f+1) and T(f-1) are the temperature values ​​of the next and previous frequency points, respectively, and Δf is the frequency interval; the total energy of the vibration energy peak frame event is obtained through the integral formula.

[0057] The acceleration amplitude is calculated. In this scenario, after discretization and encoding of the modal feature parameters, a sequence containing 18 event markers is generated, which can significantly improve the matching accuracy between different modal events and the discharge physical process in the subsequent causal graph construction.

[0058] Step S2: Based on the discharge mechanism chain of electron avalanche to structural micro-strain accumulation in high-voltage engineering, a physical causal graph of cable insulation defects is constructed using the discrete event labeling sequence, generating a directed acyclic graph topology with typical defect types as root cause nodes and multi-physics response causal paths as causal edges. The physical causal graph of cable insulation defects is a directed graph, and its edges are directed edges with causal orientation, i.e., causal edges.

[0059] The discharge mechanism chain from electron avalanche to structural micro-strain accumulation in high-voltage engineering described above is a standardized physical causal knowledge framework based on high-voltage engineering theory and experimental observations, describing the entire discharge development process under cable insulation defects. It clarifies how the discharge begins with a microscopic electron avalanche—the initial stage of an avalanche-like multiplication of electrons in the insulating medium caused by a strong electric field—and through the accumulation and transformation of discharge energy, sequentially triggers multi-physical field responses such as electromagnetic radiation, conduction current distortion, and medium temperature rise, ultimately leading to the accumulation of structural micro-strain within the insulating material, i.e., microscopic deformation and damage of the material.

[0060] Specifically, it includes: S2.1: Obtain a knowledge graph of the standard discharge mechanism chain in the field of high voltage engineering regarding the development of electron avalanche to the accumulation of structural micro-strain, and perform logical parsing processing on the discharge mechanism chain to extract a set of typical defect types including air gap discharge, surface creepage and metal tip corona, as well as a corresponding multi-physics response causal path rule library.

[0061] The discrete event tag sequence output by step S1 is processed for input preparation. A knowledge graph file of the standard discharge mechanism chain in the field of high voltage engineering regarding the cumulative evolution of electron avalanche towards structural micro-strain is loaded to ensure that the file contains physical induction path information of multiple types of insulation defects.

[0062] The knowledge graph is subjected to node semantic parsing operation. Based on the physical attributes and discharge stage of the nodes, a set of defect type nodes corresponding to air gap discharge, surface creepage and metal tip corona are extracted, and a semantic index is established for each node in the set and its discharge stage.

[0063] The edge attribute logic is deconstructed for the entire path from electron avalanche to structural micro-strain accumulation in the knowledge graph. Based on the sequential relationship of the high-voltage discharge stages, the corresponding multi-physics response feature categories are extracted on each path edge, such as electromagnetic radiation spectrum, conduction current waveform distortion, medium temperature rise rate, and structural micro-strain amplitude. A mapping rule between edge type and response feature category is established.

[0064] The extracted set of edge type mapping rules is subjected to duplicate removal and redundant connection elimination. Through correlation analysis, the core causal path rules that can only reflect the defect type to the multi-physics response are retained, forming a multi-physics response causal path rule library with a concise structure and physical interpretability.

[0065] The causal path rule base is formatted and encoded to transform each rule into a triplet data structure consisting of a source defect node, a target observation attribute node, and a causal edge attribute, so that causal association matching operations based on discrete event tags can be implemented in subsequent steps.

[0066] Through the above knowledge graph parsing and rule extraction processing methods, the results of the previous step are transformed into structured data containing a set of typical defect types and a corresponding multi-physics response causal path rule base, thus achieving the expected technical effect of providing accurate physical semantics and rule basis for causal graph construction.

[0067] S2.2: Perform root cause node instantiation operation based on the set of typical defect types, and map air gap discharge, surface creepage and metal tip corona as the starting root cause nodes of the physical cause-effect graph of cable insulation defects, so as to generate a root cause node initialization list with clear physical semantics.

[0068] The typical defect type set output by S2.1 is received as input conditions to establish a root cause location mapping rule set for defect categories in the cause-effect graph, so as to ensure that the defect node initialization process has clear physical semantics.

[0069] The three defect types—air gap discharge, surface creepage, and metal tip corona—are categorized and encoded, and assigned unique identifiers and physical mechanism labels to form a defect identification data structure that can be used for graph structure reference.

[0070] Based on the cause-multi-physics response logic in the discharge mechanism chain, each identifier in the defect identifier data structure is mapped to the root cause node of the physical cause-effect graph of cable insulation defects. The node metadata of three types of starting nodes is constructed and their physical state attributes are defined.

[0071] The root cause node metadata is initialized using a node initialization function, which assigns attribute values ​​to the node location index, initial values ​​of state variables, and a set of potential connections with neighboring observation nodes, to ensure that the connection relationships can be directly referenced when the causal graph is generated later.

[0072] The root cause node set after attribute assignment is output in a structured list to form an initial list of root cause nodes with clear physical semantics, providing a precise starting point for causal association matching in S2.3.

[0073] By instantiating root cause nodes and binding them with physical semantics, the set of defect types output in the previous step is transformed into a list of node initializations that can be directly referenced in the cause-effect graph, thus achieving the technical effect of transforming abstract defect categories into a concrete graph structure.

[0074] S2.3: The discrete event marker sequence output by the previous step is processed by causal association matching using the multiphysics response causal path rule base. The starting point of the UHF pulse cluster, the di / dt zero crossing point, the thermal gradient jump frame and the vibration energy peak frame are mapped as observation variable nodes to generate an observation node sequence containing discharge intensity, electromagnetic radiation spectrum, conduction current waveform distortion, medium temperature rise rate and structural micro-strain amplitude.

[0075] Using the discrete event tag sequence output from the preceding steps as input, the event-to-physical quantity mapping rules established for different defect types in the multiphysics response causal path rule base are invoked. During the match-by-match process, the characteristic attributes of each discrete event are used as search conditions to retrieve the corresponding causal triggering path entries in the rule base. The matched UHF pulse cluster start point entries are parsed as discharge intensity observation variable nodes, and physical semantic labels are assigned to these nodes to indicate their positions in the causal graph. The matched di / dt zero-crossing points are mapped as conduction current waveform distortion observation variable nodes, and physical quantity encoding is completed based on the current pulse characteristic parameters defined in the rule base. The matched thermal gradient jump frames are mapped as medium temperature rise rate observation variable nodes, and the temperature rise rate parameterization is achieved by referencing the rate calculation rules within the heat conduction causal path. The matched vibration energy peak frame entries are mapped as structural micro-strain amplitude observation variable nodes, and the physical quantity description is completed using the strain amplitude calculation model defined in the mechanical response causal path rule base. By employing the electromagnetic radiation spectrum derivation path from the rule base, the discharge intensity observation nodes are extended and associated with the electromagnetic radiation spectrum observation nodes. Furthermore, a feature mapping matrix is ​​used to precisely bind the correspondence between multi-modal signals and physical semantic nodes. Through this matching and mapping process, the original discrete event label sequence is transformed into an observation node sequence containing discharge intensity, electromagnetic radiation spectrum, conduction current waveform distortion, dielectric temperature rise rate, and structural micro-strain amplitude, achieving semantic conversion from multi-source signals to physical causal observation variables.

[0076] For example, in a partial discharge monitoring scenario of a cable, the discrete event marker sequence includes a timestamp of the UHF pulse cluster start point at 12.45 μs with a peak amplitude of 0.85 V; a di / dt zero-crossing time of 12.48 μs, corresponding to a current change rate of 3.2 A / μs; a thermal gradient jump frame with a temperature difference of 1.8 K and an occurrence time of 12.51 μs; and an acceleration amplitude of 0.12 g for the vibration energy peak frame. The calculation method for the discharge intensity P is defined in the rule base as follows:

[0077] Where V is the peak amplitude at the start of the UHF pulse cluster, V ref Using a reference voltage, 1V can be selected. Substituting the peak voltage of 0.85V, the discharge intensity is negative, corresponding to a small discharge event. Conductive current waveform distortion. The calculation method is as follows:

[0078] Where k is the causality coefficient of the current conduction path, taken as 1.5, substituting the di / dt value of 3.2 A / μs, we get ΔI as 4.8 A / μs. The method for calculating the medium temperature rise rate R is as follows:

[0079] Where ΔT is the temperature difference of 1.8K, Δt is the time difference between the thermal gradient jump frame and the previous frame of 0.02ms, and R is 90K / s. The structural micro-strain amplitude ε is calculated as follows:

[0080] Where 'a' is the acceleration amplitude of 0.12g, and 'g' is the gravitational acceleration of 9.81m / s², ε is obtained as 1.1772m / s². By binding the above values ​​with each observation variable node, the resulting observation node sequence accurately characterizes the physical response state of the on-site discharge event, verifying that causal correlation matching processing can significantly improve the semantic alignment accuracy of multimodal signals.

[0081] S2.4: According to the sequential induction logic defined in the discharge mechanism chain, perform a causal edge construction operation between the root cause node initialization list and the observation node sequence to establish a unidirectional connection relationship from the typical defect type to the multi-physics response observation variable, so as to generate a preliminary causal graph topology prototype.

[0082] Given the root cause node initialization list and observation node sequence, the induction logic of each stage in the discharge mechanism chain is used as a constraint input to the construction module. Logical index matching is performed on the input data structure to instantiate the node pair combination relationship corresponding to each logical rule. Based on the node pair combination relationship, a topology mapping function is called to establish a directional connection table from the root cause node to the observation node. The direction attribute of the edges is determined by the chronological order of the physical causal chain, and physical semantic labels of the source node and target node are added to each causal edge in the connection table. Edge weight initialization operations are performed on the connection table. By numerically calculating the response sensitivity between discharge intensity and various multiphysics field observation variables, a preliminary edge weight coefficient matrix is ​​obtained, where the edge weight values ​​are calculated using the following formula:

[0083] in To observe the amplitude of the response change of the variable during the discharge process, This corresponds to the amplitude of the discharge intensity change at the root cause node. After initializing the edge weights, the directional connectivity relationships and their corresponding edge weight coefficients are mapped together into a graph data structure, forming a preliminary causal graph dataset containing node indices, multi-physics semantic labels, and edge weight matrices. The preliminary causal graph dataset undergoes structural serialization to provide a resolvable topological prototype for subsequent loop detection and acyclicity verification. Through this processing, the root cause node and observation node matching results from the previous step are transformed into preliminary connected causal graph topological data with physical mechanism constraints, achieving consistency between the temporal attributes and physical semantics of causal paths.

[0084] S2.5: Perform loop detection and acyclic verification on the preliminary causal graph topology prototype to eliminate redundant connections that violate the physical time arrow and causal transitivity, so as to finally generate a physical causal graph topology structure for cable insulation defects that strictly conforms to the characteristics of a directed acyclic graph.

[0085] like Figure 2 As shown, step S3: Input the discrete event tag sequence into the lightweight encoder of the corresponding modality, perform feature mapping operation, and generate an initial modal feature vector that is mapped to a unified-dimensional implicit semantic space and serves as the potential representation of the observation nodes in the physical causal graph of the cable insulation defect. Specifically, this includes: S3.1: Based on the observation node types defined in the physical cause-effect graph of cable insulation defects, an independent feature extraction channel is assigned to the discrete event marker sequence to construct a modality-specific lightweight encoder set. The modality-specific lightweight encoder set includes: an ultra-high frequency electromagnetic pulse encoder, a high frequency current pulse encoder, a surface infrared thermal response encoder, and a micro-vibration acceleration encoder.

[0086] Receive the topology of the physical cause-effect graph of cable insulation defects and the corresponding set of observation node types output by step S2, and determine the category of physical observation variables corresponding to each discrete event label sequence.

[0087] For UHF electromagnetic pulse observation nodes, a lightweight encoder structure with one-dimensional convolution and time-frequency domain embedding capabilities is constructed. The convolution kernel size and stride parameters are preset to adapt to the transient characteristic amplitude distribution of UHF pulse clusters.

[0088] For high-frequency current pulse observation nodes, a lightweight encoder channel capable of capturing the di / dt zero-crossing dynamic characteristics is constructed, and a high-bandwidth sampling window is set to ensure complete recording of rapid amplitude changes.

[0089] For surface infrared thermal response observation nodes, a lightweight encoder structure containing multi-scale spatial convolution and temperature gradient fitting units is constructed, and the filter scale is optimized to accurately extract the spectral amplitude pattern of thermal gradient jump frames.

[0090] For micro-vibration acceleration observation nodes, a lightweight encoder channel that integrates short-time Fourier transform preprocessing and time convolution layer is constructed to enhance the frequency domain stability identification capability of vibration energy peak frames.

[0091] The encoder structures for different observation nodes are encapsulated into a set of modality-specific lightweight encoders, and an independent mapping relationship is established between event tag sequences and corresponding encoder channels to achieve standardized configuration of modality separation and feature extraction entry points for the input data stream.

[0092] By constructing a set of modality-specific lightweight encoders, the physical feature information transmission path of the causal graph observation nodes is standardized into an independent feature extraction channel, realizing a conflict-free parallel mapping from discrete event label sequences to initial modal features, and providing accurate structured input for unified alignment of implicit semantic space.

[0093] S3.2: Perform local time-frequency domain slicing on the UHF pulse cluster start point, di / dt zero-crossing point, thermal gradient jump frame and vibration energy peak frame in the discrete event marker sequence, and use a sliding window to extract the original time series data segment containing complete mutation features to generate multimodal local time series data segment.

[0094] A modal classification index is established for each type of event in the discrete event tagging sequence, and the corresponding time-frequency analysis and processing strategy is determined based on the physical properties and waveform feature labels of the events.

[0095] The short-time Fourier transform analysis module is called on the original time-series data of the starting point of the UHF pulse cluster. The window length and step value are set to ensure complete coverage of the pulse cluster. The frequency domain energy integration result is calculated and a time-frequency domain matrix index is formed.

[0096] For di / dt zero-crossing events, continuous wavelet transform is used to perform waveform energy focusing processing in the high-frequency component range, extracting the high-frequency energy mutation mapping matrix in the region near the zero point, thereby enhancing the highlighting effect of the zero-crossing feature.

[0097] A two-dimensional discrete cosine transform is performed on the thermal gradient jump frame event to map the spatial temperature gradient change to a frequency domain representation, and the time domain segment containing the gradient jump is extracted while preserving the inter-frame difference spectrum distribution of thermal imaging.

[0098] The short-time energy detection and fast Fourier transform combined analysis module is invoked for the vibration energy peak frame event to extract the energy change curve and the main frequency offset within the window before and after the peak, forming the time-frequency response segment of the vibration mode.

[0099] A sliding window mechanism is used to perform window traversal on the time-frequency matrix of each mode, calling the energy aggregation and center frequency calculation formulas within the window. Obtain the characteristic frequencies within the window, where E is the frequency domain energy value, f is the corresponding frequency, and i is the frequency index.

[0100] Based on the window traversal results, select a window group that contains the entire process of event mutation, and output a set of multimodal local time series data segments according to modality classification.

[0101] By using local time-frequency domain slicing and sliding window truncation, the discrete event label sequence from the previous step is transformed into a multimodal local time-series data segment with complete mutation feature coverage, which can be used by the encoder to perform nonlinear feature mapping, thus realizing a unified slicing of time-series signals of different modes near the physical mutation point.

[0102] For example, in a 35kV cable partial discharge monitoring scenario, the collected UHF pulse cluster start point event corresponds to a signal sequence with a sampling rate of 100MHz in the original time series. The short-time Fourier transform window length is set to 1024 points, the step value is 256 points, and the output is a 512×256 time-frequency matrix. The high-frequency current pulse signal corresponding to the di / dt zero-crossing event has a sampling rate of 10MHz. The continuous wavelet transform uses the Daubechies-8 wavelet basis with a scale range of 1 to 128 to extract the high-frequency energy coefficients within ±5μs of the zero point. The thermal gradient jump frame event originates from a 320×240 infrared thermal response sequence. The two-dimensional discrete cosine transform extracts the difference spectral coefficients of the two frames before and after, retaining local segments with a gradient change rate greater than 2K / frame. The acceleration signal of the vibration energy peak frame event has a sampling rate of 50kHz. The short-time energy detection window length is 512 points, the FFT number is 1024, and the energy and frequency change curves before and after the peak are extracted for 20ms. In the calculation of the energy center frequency of the sliding window, within a certain window of the UHF mode, the energy value sequence E=[10,15,20] and the frequency sequence f=[5,10,15]MHz are applied. The formula yields a center frequency of 12MHz. The multimodal local time-series data fragments output in this step can significantly improve the stability and interpretability of cross-modal feature alignment after subsequent lightweight encoder processing.

[0103] S3.3: Input the multimodal local time-series data segment into the corresponding modality-specific lightweight encoder set, and use a one-dimensional convolutional neural network to perform multi-layer nonlinear transformation operations to extract high-dimensional abstract features reflecting the essence of the discharge mechanism from the original waveform to generate a high-dimensional abstract feature vector.

[0104] The input multimodal local temporal data segments are imported into the corresponding modality-specific lightweight encoder channels. The kernel size, stride, and padding strategy of the one-dimensional convolutional neural network are set to match the sampling rate and feature period of each modality data. Based on this setting, the first layer of convolution operation is performed to capture the local morphological changes of microsecond-level abrupt pulses or millisecond-level thermal responses on the time axis, forming the initial temporal feature map.

[0105] Based on the initial temporal feature mapping, batch normalization is applied to eliminate the amplitude scale differences between different modes and ensure the numerical stability of the convolution output. Then, a nonlinear activation function is applied to transform the linear patterns captured by the convolution layer into nonlinear feature representations through activation operations, so that the feature vectors can characterize the complex mechanisms of responses induced by different discharge types.

[0106] The feature maps after nonlinear activation are fed into the second layer of a one-dimensional convolutional neural network. The convolutional kernel span is adjusted to cover a longer time window, and cross-event correlation features are extracted, thereby revealing the implicit temporal coupling relationship between discharge intensity and dielectric temperature rise rate. Batch normalization and nonlinear activation are then performed on the output of this layer to enhance the extraction capability of cross-modal shared semantics.

[0107] Global average pooling is performed sequentially on the outputs of multi-layer convolutions to compress the time dimension and aggregate local features into a fixed-length vector over time. Feature-level concatenation is then performed on the pooled vectors to form an information matrix that covers multi-scale features of different modalities, providing a stable input for subsequent projection dimensionality reduction.

[0108] Through the above-mentioned chained processing of multi-layer convolution, normalization, activation and pooling, the multimodal local time-series data fragments of the previous step are transformed into high-dimensional abstract feature vectors that reflect the essence of the discharge mechanism, thereby realizing the initial feature representation of cross-modal data in the implicit semantic space.

[0109] For example, UHF pulse event segments with a sampling rate of 200MHz and high-frequency current segments with a sampling rate of 50kHz are input into the UHF encoder and high-frequency current encoder channels, respectively. The convolution kernel sizes are set to 5 and 15, respectively, with a stride of 1 for both, and zero padding is used to preserve the feature length. After the first convolution layer, batch normalization (momentum parameter 0.9) and ReLU activation are performed to generate initial feature maps. In the second convolution layer, the convolution kernel sizes are adjusted to 25 and 45 to cover a larger time window and capture waveform patterns across events. Batch normalization and ReLU activation are also performed to enhance the nonlinear representation capability. Global average pooling is performed on the outputs of each modality convolution to obtain a fixed vector of length 128, which is then concatenated at the feature level to form a 256-dimensional information matrix. After subsequent dimensionality reduction and projection processing, this matrix shows significantly improved defect type separability in the validation set, and a significantly enhanced implicit feature alignment between UHF and thermal response modes, ultimately achieving high-precision and robust initial modal feature vector extraction.

[0110] S3.4: Based on the preset unified implicit semantic space dimension, perform fully connected layer projection dimensionality reduction processing on the high-dimensional abstract feature vector, and force the feature distribution of different modalities to be aligned to the same vector space through a linear mapping matrix to generate a unified initial modal feature vector.

[0111] For high-dimensional abstract feature vectors generated by multi-layer nonlinear transformations of a one-dimensional convolutional neural network, a dimension reduction mapping process is completed by calling the projection operator of the fully connected layer based on the unified implicit semantic space dimension parameter setting.

[0112] The high-dimensional abstract feature vectors of different modalities are arranged in columns to form an input matrix, and a corresponding linear mapping matrix is ​​established. The weight dimension is determined by the number of high-dimensional features and the unified implicit space dimension.

[0113] Feature projection is performed using matrix multiplication to map the original high-dimensional feature distribution into the output space of the fully connected layer, ensuring that all modal features are completely consistent in dimension.

[0114] The projection results are batch normalized by adjusting the mean and variance to provide uniformity of the feature scale and eliminate amplitude shifts caused by differences in sensor sensitivity.

[0115] The normalized feature vectors are reorganized into a sequence of initial modal feature vectors of uniform dimension according to modal order, providing consistent feature input for subsequent observation node binding and causal graph message passing.

[0116] By using the dimensionality reduction and normalization methods described above, the high-dimensional abstract feature results from the previous step are transformed into unified dimensional initial feature data with forced alignment between modalities, achieving the expected technical effect that cross-modal features can be directly fused in the implicit semantic space.

[0117] For example, in the air gap discharge monitoring scenario, the high-dimensional abstract feature vector output by the UHF electromagnetic pulse encoder has a length of 128, the output length by the high-frequency current pulse encoder is 256, the output length by the surface infrared thermal response encoder is 96, and the output length by the micro-vibration acceleration encoder is 192. The unified implicit semantic space dimension is set to 64. When constructing the weight matrix, a fully connected mapping matrix from high dimension to 64 dimension is established separately for each mode. For example, the weight matrix dimension of the UHF electromagnetic pulse mode is 128×64. The projection operation is performed using the following mathematical form:

[0118] in This is the 64-dimensional vector after dimensionality reduction. The original high-dimensional feature vector, This is the mapping weight matrix for the corresponding mode. During batch normalization, the standardized vector components are calculated using the following formula:

[0119] in This is the batch average. The standard deviation is the batch size. In this embodiment, the output after projection of all modalities is 64-dimensional, and the distribution is consistent after normalization. The difference in cross-modal features is significantly reduced, which enables effective subsequent node binding and message passing in the causal graph, ultimately achieving a significant improvement in defect identification accuracy.

[0120] S3.5: Bind the unified dimension initial modal feature vector to the observation node in the physical cause-effect graph of the cable insulation defect, and assign each feature vector a clear physical semantic label to generate an initial modal feature vector with the potential representation attributes of the observation node in the physical cause-effect graph.

[0121] like Figure 3 As shown, step S4: Based on the topological constraints of the physical causal graph of the cable insulation defect, graph neural network message passing processing is performed on the initial modal feature vector, and cross-node feature updates are performed through causal edges to generate causal modulation fusion features that conform to physical consistency logic.

[0122] The topological constraints refer to the set of structural rules composed of the node connections and edge type attributes of the physical cause-effect graph of cable insulation defects. Physically, they define the propagation path, direction, and boundary conditions of discharge features during multiphysics evolution: the structure of the directed acyclic graph stipulates that features can only flow along a unidirectional causal chain of "root cause defect → multiphysics response," prohibiting reverse propagation or cross-link interference; simultaneously, through the adjacency matrix and edge type mapping table, it enforces that cross-modal features must follow the high-voltage discharge mechanism (such as the evolution logic from electron avalanche to structural micro-strain) during message transmission in the graph neural network.

[0123] Specifically, it includes: S4.1: Based on the initial modal feature vector and the node connection relationship of the physical cause-effect graph of the cable insulation defect, construct an adjacency matrix and edge type mapping table to generate graph structure constraint parameters for defining message passing paths.

[0124] Based on the directed acyclic graph topology of the physical causal graph of cable insulation defects output from the previous steps and the initial modal feature vectors binding the attributes of the observed nodes, the graph structure constraint construction process is executed. The unique identifiers of each node in the directed acyclic graph topology are matched and mapped with the index positions of the initial modal feature vectors to form a node index table to ensure a one-to-one correspondence between features and topology. The causal edge connection information of the topology is read, and adjacency matrix elements are established according to the combination of source and target nodes. The edge type encoding corresponding to different physical causal chains is explicitly distinguished in the matrix, and an edge type mapping table is constructed for subsequent message passing control based on physical mechanism chains. For the adjacency matrix, all observed nodes with an out-degree of zero are detected and marked as terminal node types in the edge type mapping table to provide boundary conditions for subsequent iterative message passing convergence determination. For nodes with multiple edge connections, a unique edge type ID is assigned based on the physical time arrow of the edge direction and the response variable category and written into the mapping table, synchronously updating the weight initialization state of the corresponding elements in the adjacency matrix. The node index table, adjacency matrix, and edge type mapping table are combined into a complete set of graph structure constraint parameters, which serve as the path control basis for subsequent node embedding initialization and directed message aggregation operations. This processing method transforms the topology and feature results of the previous step into computable and transferable structured parameters, thereby achieving physical consistency definition of the message transmission path.

[0125] S4.2: Utilize graph structure constraint parameters to perform node embedding initialization processing on the initial modal feature vector. Define the edge weight initialization strategy based on the discharge mechanism chain from discharge intensity to structural micro-strain accumulation, and generate node state representation vectors carrying physical prior knowledge.

[0126] The discharge mechanism chain from discharge intensity to structural micro-strain accumulation refers to the physical causal sequence describing the entire process of insulation defect discharge development in high-voltage engineering. It originates from the coupling mechanism of discharge physics and solid mechanics: the initial discharge intensity (energy release) first excites electromagnetic radiation and disturbs the conduction current; part of this electrical energy is converted into heat energy, causing the insulating medium to generate a temperature rise; the continuous thermal effect generates thermal stress inside the material, which eventually triggers and accumulates into macroscopic structural micro-strain.

[0127] The directed acyclic graph topology of the physical cause-effect graph of cable insulation defects and the graph structure constraint parameters composed of adjacency matrix and edge type mapping table are analyzed, and the connection relationship of each observation node and its associated causal edge is extracted as the index basis for node embedding.

[0128] The unified-dimensional initial modal feature vectors generated in the previous steps are mapped one-to-one with the node positions in the adjacency matrix to construct a feature position mapping matrix for initializing the embedding, and physical semantic labels on the discharge mechanism chain are pre-assigned to each node.

[0129] Based on the discharge mechanism chain sequence from discharge intensity to structural micro-strain accumulation, the edge weights are physically initialized a priori. The discharge intensity weight of the upstream node is set as the benchmark value, and the weights are decreased or increased along the path from electromagnetic radiation spectrum, conduction current waveform distortion, medium temperature rise rate to structural micro-strain amplitude.

[0130] The weights of different paths are adjusted proportionally using a weight normalization formula to ensure that the weights of each edge satisfy the energy conservation constraint of physical quantity transfer.

[0131] in, Let the initial weights of each causal edge on the path be denoted by the summation sign. This represents the summation of the weights of all edges.

[0132] The normalized edge weights are combined with the initial modal feature vectors of the corresponding nodes to construct a node state matrix represented by directional vectors, which reflects both the modal feature distribution and carries physical prior weight information.

[0133] By using the above processing method, the initial modal feature vector of the previous step is transformed into a node state representation vector carrying prior knowledge of physical causality, thereby achieving the expected technical effect that cross-modal features can be orderly transmitted along the causal chain in the graph neural network.

[0134] S4.3: Perform directional message aggregation operation based on the node state representation vector and edge type mapping table, extract the feature information of upstream neighbor nodes according to the causal path of the ultra-high frequency electromagnetic signal pointing to the surface infrared thermal response signal, and generate an aggregated message vector containing cross-modal physical association information.

[0135] Perform directional message aggregation operations on the node state representation vector initialized with physical prior knowledge and the constructed edge type mapping table, and select the upstream neighbor node feature information defined by the path from discharge intensity to infrared temperature rise rate in the causal graph topology.

[0136] Based on the directional connection relationships in the adjacency matrix, the physical weight coefficients for the edge type are called, and the upstream node state representation vector and the edge weight are multiplied element-wise to highlight the output of channels with high physical correlation.

[0137] Strict edge type filtering is performed on the obtained weighted node feature vectors, retaining only message components that conform to the path "electromagnetic radiation spectrum → dielectric temperature rise rate" in the high-voltage discharge mechanism chain.

[0138] Each retained component is summed and normalized according to a preset aggregation function to calculate the aggregated message vector of cross-modal physical association information, ensuring that the aggregation result does not deviate from the physical causal direction in the implicit semantic space.

[0139] Normalization is achieved using the following aggregation formula:

[0140] in, Let be the weight coefficient of the i-th causal edge. Let i be the feature component of the i-th upstream neighbor node. This serves as the association identifier (or index) for upstream neighbor nodes. Through weighted product and weighted sum normalization, it ensures that the proportions of different modal contributions conform to physical consistency logic.

[0141] Through the above-mentioned directional aggregation operation, the initial node state representation vector of the previous step is transformed into an aggregated message vector containing cross-modal physical association information, thereby realizing the construction of cross-node feature fusion channels under the physical consistency constraint in the causal graph.

[0142] S4.4: Perform nonlinear feature transformation and gating fusion processing on the aggregated message vector, update the local feature distribution by combining the current node's state representation vector, and generate the hidden state vector of the intermediate layer node that reflects the causal transmission effect of the multi-physics response.

[0143] After receiving the aggregated message vector obtained by the aggregation of directed messages and the state representation vector of the current node, the aggregated message vector is first normalized to eliminate the dimensional differences in amplitude of different modal inputs and ensure numerical stability, in order to meet the update requirements of the feature distribution within the node.

[0144] The normalized aggregated message vector is input into the nonlinear transformation unit, and a multilayer perceptron structure is used to realize feature mapping with activation function. Feature components from different sources on the physical causal chain are nonlinearly combined to enhance the ability to extract complex patterns.

[0145] A gating mechanism is introduced into the feature representation after nonlinear transformation. A gating fusion unit containing update gate and reset gate is constructed. The fusion ratio of the aggregated message vector and the current node state representation vector is controlled by the continuously adjustable gate value between 0 and 1.

[0146] The updated gate output is calculated using the following formula:

[0147] Where W is the weight matrix applied to the aggregated message vector, x is the normalized aggregated message vector, U is the weight matrix applied to the current node state representation vector, h is the current node state representation vector, and σ is the Sigmoid function.

[0148] The update gate output is multiplied with the current node state representation vector, and the reset gate output is multiplied with the nonlinearly transformed aggregated message vector to achieve selective retention and suppression of feature information.

[0149] The above product results are sequentially fed into the weighted summation module, and the fused feature vector is output according to the modulation ratio of the update gate and the reset gate, so as to ensure the cross-modal transmission of effective information and the shielding of irrelevant disturbances in the physical causal path.

[0150] The fused feature vectors are projected into the implicit semantic space through a fully connected mapping layer, and the physical consistency of edge connections is corrected by combining the topological constraints of the physical causal graph, so as to obtain the hidden state vector of the intermediate layer node that reflects the causal transmission effect of the multi-physics response.

[0151] Through the above-mentioned nonlinear transformation and gating fusion processing method, the aggregated message vector and node state representation vector of the previous step are transformed into intermediate hidden features with physical consistency logical constraints, thereby realizing the gradual convergence and information optimization of multimodal features on the physical causal chain.

[0152] For example, in a cable partial discharge detection task, the length of the aggregated message vector is set to 128 dimensions. After normalization, its amplitude range is stabilized in the range of [-1, 1]. The length of the current node state representation vector is also 128 dimensions. The nonlinear transformation unit adopts a two-layer fully connected network. The first layer weight matrix has a dimension of 128×256, and the activation function is ReLU. The second layer weight matrix has a dimension of 256×128, and the output length remains 128 dimensions. The update gate weight matrix W is set to 128×128, the reset gate weight matrix U is set to 128×128, and the gating function is Sigmoid. During the calculation of the update gate, when the 50th dimension of the aggregated message vector x is 0.8 and the 50th dimension of the node state representation vector h is 0.5, after the weight matrix multiplication and summation, the output of the update gate z in the 50th dimension is 0.73, achieving a high degree of new information introduction. During the fusion process, the reset gate outputs 0.25 in the 80th dimension, significantly suppressing the aggregated message vector in this dimension and reducing the propagation of irrelevant noise. Finally, after the fused features are mapped to the implicit semantic space via a fully connected layer, the relationship between discharge intensity and infrared temperature rise rate in the physical consistency check conforms to the monotonically non-decreasing constraint, verifying the effectiveness of this processing in improving cross-modal semantic alignment accuracy and causal consistency.

[0153] S4.5: Perform multi-round iterative message passing optimization based on the hidden state vector of intermediate layer nodes until the node features converge to a stable state that satisfies the physical consistency logic in the implicit semantic space, generating causal modulation fusion features with cross-modal semantic alignment characteristics.

[0154] Step S5: Calculate the feature transformation deviation on each causal edge of the physical causal graph of the cable insulation defect according to the preset physical monotonicity constraint, construct a causal consistency loss function to reflect the logical relationship between physical quantities such as discharge intensity and infrared temperature rise rate, and generate a physical constraint gradient for reverse adjustment of encoder parameters. Specifically, this includes: S5.1: Based on the causal edge connection relationship in the physical cause-effect graph of cable insulation defects, obtain the discharge intensity characterization vector output by the source node and the infrared temperature rise rate characterization vector input by the target node, and perform monotonic mapping matching processing on the discharge intensity characterization vector and the infrared temperature rise rate characterization vector to generate a sequence of feature pairs to be verified containing the logical correspondence of physical quantities.

[0155] S5.2: Using a preset physical monotonicity constraint, perform a numerical order relationship determination operation on each set of data in the sequence of features to be verified, so as to identify the set of abnormal feature deviation terms that violate the law of non-decreasing infrared temperature rise rate when the discharge intensity increases.

[0156] In this embodiment, the preset physical monotonicity constraint refers to the mathematical expression rule about the causal relationship between different physical quantities, derived from the energy conservation and conversion mechanism of high-voltage discharge. It explicitly stipulates that in the discharge mechanism chain of cable insulation defects, the discharge intensity (characterizing discharge energy) as the cause and the infrared temperature rise rate (characterizing thermal effect) as the result should satisfy a monotonically non-decreasing functional relationship. That is, when the discharge intensity increases, the infrared temperature rise rate caused by its conversion should not decrease. This physical monotonicity constraint originates from the physical principle (discharge energy must be partially converted into heat energy) and is expressed by mathematical inequalities (such as...). The form is preset to verify and constrain whether the relationships learned by the model in the feature space conform to the real physical laws.

[0157] When performing physical monotonicity constraint judgment on each pair of feature vectors in the sequence of features to be verified, the discharge intensity characterization data of the source node and the infrared temperature rise rate characterization data of the target node are used as input objects to ensure that both have been scaled and normalized in a unified implicit semantic space. Using an order relation comparison method, the measured values ​​corresponding to each pair of feature vectors are first differentially analyzed to obtain the changes in discharge intensity and infrared temperature rise rate, and a numerical correspondence matrix is ​​established. Based on a preset physical monotonicity rule, a sign judgment is performed on the differential values ​​of each row, and cases where the change in discharge intensity is positive and the change in infrared temperature rise rate is negative are marked as violations of the monotonicity constraint. Using mathematical logic operations, the judgment results are converted into a binary identifier matrix for subsequent statistical analysis of the number and location of abnormal deviation items. For detected abnormal items, their corresponding causal edge index numbers are recorded to establish a set of abnormal feature deviation items, providing index location for subsequent deviation amplitude calculation. By using the above-mentioned order determination and anomaly set construction processing method, the feature pair sequence of the previous step is transformed into a dataset with physical anomaly labels, thereby enabling data screening and location that violates the law of non-decreasing infrared temperature rise rate when the discharge intensity increases.

[0158] For example, in a cable partial discharge monitoring scenario, assuming the source node discharge intensity characteristic sequence of a causal edge is [0.52, 0.61, 0.74], and the target node infrared temperature rise rate characteristic sequence is [0.35, 0.33, 0.40], the changes are calculated using differential methods, resulting in a discharge intensity change matrix of [0.09, 0.13] and an infrared temperature rise rate change matrix of [-0.02, 0.07]. Using the physical monotonicity determination rule, the determination formula is constructed as follows:

[0159] Where ΔI is the change in discharge intensity, ΔT is the change in infrared temperature rise rate, and a value of 1 indicates a violation of monotonicity, while 0 indicates compliance. According to this formula, the first difference pair satisfying ΔI=0.09>0 and ΔT=-0.02<0 is marked as an anomalous deviation term; the second difference pair satisfying ΔI=0.13>0 and ΔT=0.07>0 satisfies monotonicity and is not marked as anomalous. The final set of anomalous feature deviation terms contains only the first pair of index numbers, serving as input for subsequent deviation value calculations. The application effect is that it significantly improves the accuracy of anomaly detection and ensures that the judgment process is consistent with the physical mechanism chain.

[0160] S5.3: Based on the deviation magnitude of each element in the set of abnormal feature deviation items, a piecewise linear penalty function is used to calculate the feature transformation deviation value on each causal edge, so as to quantify the degree of mismatch between the multimodal feature distribution and the high-voltage discharge mechanism chain in the current implicit semantic space and generate a physical inconsistency scalar.

[0161] The input condition is the set of abnormal feature deviation terms obtained in sub-step S5.2, which contains the source node and target node feature pairs that violate the physical monotonicity constraint and their numerical deviation magnitudes.

[0162] For each element in the set of abnormal feature deviation terms, amplitude normalization is performed to map the deviation values ​​of different dimensions to a unified numerical domain to facilitate the parameter setting of the subsequent penalty function.

[0163] Based on the normalized deviation values, a threshold range for the piecewise linear penalty function is defined. The penalty slopes for the slight, moderate, and severe deviation regions are determined according to the acceptable mismatch range of physical quantities in the high-voltage discharge mechanism chain. Each deviation value is input into the piecewise linear penalty function model, applying a lower slope penalty in the slight region, a moderate slope penalty in the moderate region, and a high slope penalty in the severe region, generating the corresponding quantified penalty values.

[0164] Specifically, based on the tolerance for the severity of discharge during condition-based maintenance of high-voltage equipment, a normalized range for the deviation amplitude is set: Slight mismatch zone [0, θ1]: a low slope k1 is set (e.g., k1=1), allowing for minor physical fluctuations; Medium mismatch zone [θ1, θ2]: a medium slope k2 is set (e.g., k2=5), indicating that the characteristics begin to deviate significantly; Severe mismatch zone [θ2, 1]: a high slope k3 is set (e.g., k3=10), corresponding to obvious physical contradictions or fault conditions.

[0165] The specific mathematical form of the piecewise linear penalty function is defined as follows:

[0166] Where d is the normalized deviation magnitude. This represents the feature transformation deviation value.

[0167] The penalty quantization value is bound to the corresponding causal path using a causal edge index, and the feature transformation bias value on each causal edge is calculated using the following formula:

[0168] in Here, abn represents the feature transformation deviation value corresponding to the causal edge, and abn is the outlier index. To normalize the deviation magnitude, The output value of the piecewise linear penalty function for the i-th normalized bias vi. This represents the total number of anomalies in the causal edge. To prevent division by zero of extremely small positive numbers, Let be the weight coefficient of the i-th outlier term within the causal edge.

[0169] Output the feature transformation bias values ​​on all causal edges as a physical inconsistency scalar for subsequent steps in S5.4 to construct the global loss function.

[0170] By using a piecewise linear penalty function to differentiate and quantify deviations of different magnitudes, the set of abnormal feature deviations from the previous step is transformed into a computable physical inconsistency metric, enabling an accurate assessment of the degree of mismatch between multimodal features and the discharge mechanism chain in the implicit semantic space.

[0171] S5.4: Based on the physical inconsistency scalar, perform a weighted aggregation operation on the feature transformation deviation values ​​of all causal edges to construct a causal consistency loss function that can globally reflect the quality of cross-modal semantic alignment and output the scalarized total loss error value.

[0172] Based on the scalar input of the physical inconsistency metric, the edge weight normalization coefficient matrix is ​​constructed by calling the causal edge and corresponding weight information in the causal graph topology to ensure that the contribution ratio of different physical edges to the loss value in the global aggregation process is controllable.

[0173] The characteristic transformation deviation value and the normalization coefficient of each causal edge are multiplied element by element to generate a weighted deviation value matrix, which is used to reflect the product effect of the physical significance and the deviation magnitude.

[0174] The weighted deviation value matrix is ​​subjected to similar edge grouping and aggregation processing using the physical path importance factor. Deviations belonging to the same physical process path (such as discharge intensity → infrared temperature rise) are weighted and summed to generate a path-level aggregated deviation value set.

[0175] A global causal consistency loss function is constructed using a set of path-level aggregated deviation values. A weighted average operator is used to calculate the global loss scalar value, where the weight factors come from the correspondence between the edge type mapping table and the physical mechanism chain.

[0176] The global loss scalar value is output to the subsequent backpropagation module as a direct input for calculating the physical constraint gradient, thereby enabling a quantitative evaluation of the alignment quality between multimodal feature distribution and physical causal chain.

[0177] By using weighted aggregation operations and loss function construction, the deviation value from the previous step is transformed into global loss error data, thereby achieving a unified metric for cross-modal semantic alignment quality.

[0178] For example, in a set of cable partial discharge monitoring data, the weight of the air gap discharge path edge is set to 0.35, the weight of the surface creepage path edge is set to 0.25, and the weight of the metal tip corona path edge is set to 0.40. With characteristic transformation deviations of 0.12, 0.08, and 0.15 for the three edges respectively, a normalized coefficient matrix {0.35, 0.25, 0.40} is first generated.

[0179] Element-wise multiplication of the deviation values ​​yields a weighted deviation matrix {0.042, 0.02, 0.06}. Weighted deviations belonging to the same physical process path are grouped together. For example, both the air gap discharge path and the metal tip corona path involve the discharge intensity → infrared temperature rise process. The grouped and summed deviations yield a path-level deviation value of 0.042 + 0.06 = 0.102, and the surface creepage path deviation value is 0.02.

[0180] Construct a global loss function and calculate it using a weighted average operator:

[0181] Where L is the causal consistency loss, and the denominator is the number of paths. The calculated result is 0.061, which is output as a scalar loss value to the backpropagation module.

[0182] In this embodiment, after the above processing, the global loss value is significantly reduced, which verifies the role of the weighted aggregation mechanism in improving the quality of cross-modal semantic alignment. Subsequent gradient updates can effectively drive the convergence of multimodal features along the physical causal direction.

[0183] S5.5: Perform automatic differentiation backpropagation operation on the causal consistency loss function to calculate the partial derivative of the total loss error value with respect to the lightweight encoder parameters, so as to generate a physical constraint gradient for correcting the feature map trajectory and driving the initial modal feature vector to converge along the physical causal direction.

[0184] To address the scalarized total loss error value output by the causal consistency loss function and the topological constraints of the constructed physical causal graph, a global feature transformation deviation parameter set containing the correspondence between physical quantities such as discharge intensity and infrared temperature rise rate is loaded as input conditions.

[0185] The total loss error value is input into the automatic differential calculation graph, and the gradient of the loss function with respect to the input features of each observation node is decomposed in layers according to the chain rule to generate the local gradient vector corresponding to each observation node.

[0186] By combining the local gradient vector with the edge weights and node state representations preserved during the message passing process of the graph neural network, the gradient vector is weighted and backpropagated using the edge weights to generate a backpropagation quantity that maps to the initial modality feature vector.

[0187] A unified dimensionality mapping transformation is performed on the reverse update quantity to ensure that the gradient components of different modal features can be directly aligned to the input distribution of the corresponding lightweight encoder in the implicit semantic space.

[0188] Based on the gradient mapping results, the gradient components of different modes are split according to the row and column structure of the encoder parameter matrix to form a parameter update gradient matrix with the same size as the weight matrix of each lightweight encoder.

[0189] The partial derivatives of the total loss error with respect to the encoder parameters are expressed mathematically:

[0190] Where L is the causal consistency loss function, For lightweight encoder parameters, Let be the predicted value for the i-th sample. Let N be the true value of the i-th sample, and N be the total number of samples. Describe the local gradient or sensitivity.

[0191] By using the gradient matrix mentioned above, the update direction and magnitude of the encoder parameters are established, and the causal consistency loss of the previous step is transformed into an executable physical constraint gradient, thereby achieving the convergence drive of the initial modality feature vector along the physical causal direction in the implicit semantic space.

[0192] Step S6: Utilize the physical constraint gradient to perform backpropagation optimization on the parameters of the lightweight encoder, driving the initial modal feature vector to converge along the physical causal direction in the implicit semantic space, generating a final fused feature representation with cross-modal semantic alignment properties. Specifically, this includes: S6.1: Obtain the scalar loss value and the corresponding physical constraint gradient tensor output by the causal consistency loss function, and perform inverse decomposition processing on the physical constraint gradient tensor based on the chain rule to calculate the parameter update gradient components mapped to the weight matrix of each modality lightweight encoder.

[0193] The scalar loss value and corresponding physical constraint gradient tensor output by the causal consistency loss function are obtained. A loss inverse mapping chain is established for the scalar loss value, and the loss is input into the automatic differential calculation framework as the global optimization objective. Gradient path tracing is performed on the physical constraint gradient tensor to extract the distribution characteristics of the gradient in each implicit semantic space coordinate to obtain a node mapping table of gradient flow directions. The gradient path is decomposed layer by layer according to the chain rule, and the global gradient tensor is divided into local gradient blocks corresponding to each lightweight encoder unit according to the network level and connection relationship. The node mapping table is used to match the binding relationship between the observation nodes in the physical causal graph and the lightweight encoder weight matrix, and the local gradient blocks are accurately mapped to the corresponding weight matrix index positions. The gradient components of each weight matrix element are calculated and updated through matrix multiplication and tensor broadcasting mechanism to ensure that the gradient components retain the directionality and magnitude information of the physical causal prior constraints.

[0194] Where g is the parameter update gradient component, L is the scalar value of the causal consistency loss function, and W is the weight matrix of the lightweight encoder. Through multi-level gradient decomposition and mapping, the physical constraint gradient output in the previous step is transformed into parameter update gradient components that can be directly applied to the weights of each lightweight encoder, thereby realizing the physical semantic orientation of the gradient signal.

[0195] S6.2: Use the parameters to update the gradient components and perform an adaptive moment estimation optimization update operation on the lightweight encoder weight parameters of the current iteration cycle to generate the next round of lightweight encoder weight parameter set containing physical causal prior knowledge.

[0196] S6.3: Based on the next round of lightweight encoder weight parameter set, re-execute the forward feature mapping inference process on the original discrete event label sequence to output the intermediate round modal feature vector after physical causality correction.

[0197] Based on the next-round lightweight encoder weight parameter set optimized in step S6.2, forward feature mapping inference is performed on the original discrete event label sequence output from step S1.5, serving as the input condition for this sub-step. The discrete event label sequence is then fed into the corresponding UHF electromagnetic pulse encoder, high-frequency current pulse encoder, surface infrared thermal response encoder, and micro-vibration accelerometer encoder according to modal category. Local temporal convolution operations are performed using the updated weight matrix to extract modal feature components corrected by physical causal gradients. Batch normalization is applied to the convolution output to ensure consistent numerical scales for each modality in the implicit semantic space, and a high-dimensional feature representation with physical consistency constraints is generated using a nonlinear activation function. Based on a unified implicit semantic space dimension setting, a fully connected projection transformation is performed on the high-dimensional features of each modality. A linear mapping matrix is ​​used to force the feature distribution to align to a set of intermediate-round modal feature vectors of a unified dimension. The aligned intermediate-round feature vectors are then bound one-to-one with the observation nodes in the physical causal graph of cable insulation defects, ensuring a strict correspondence between feature semantics and node physical meaning. By using the above processing method, the lightweight encoder weight optimization result of the previous step is transformed into intermediate round modal feature data containing physical causal correction effects, thereby achieving phased consistency of cross-modal feature distribution in the implicit semantic space.

[0198] For example, in the application of air gap discharge defect diagnosis, the intermediate round inference processing performs a one-dimensional convolution operation on the updated UHF electromagnetic pulse encoder weight matrix. The number of convolution kernels is set to 64, the kernel width to 5, and the stride to 1. After batch normalization, the ReLU activation function is used to output the feature matrix. The thermal gradient jump frame data is extracted by convolution of the infrared thermal response encoder and projected into a 128-dimensional unified implicit semantic space in the fully connected layer. The projection matrix is ​​set to 128×256 to ensure that the dimensionality is reduced from 256-dimensional convolution features to a 128-dimensional aligned space. When the intermediate feature vectors of each mode are bound to the observation nodes of the causal graph, the starting point of the UHF pulse cluster corresponds to the discharge intensity node, and the thermal gradient jump frame corresponds to the medium temperature rise rate node. The bound vectors maintain the distribution state after physical causal gradient correction in the implicit space. By comparing the feature Euclidean distance before and after physical correction, the distance value is significantly reduced, indicating that the cross-modal feature semantic consistency is improved. Finally, the output of this intermediate round feature vector provides a physical constraint basis for subsequent consistency verification.

[0199] S6.4: Perform a physical monotonicity consistency check operation on the distribution state of the intermediate round modal feature vectors in the unified dimension implicit semantic space to quantitatively evaluate the convergence degree of the logical deviation between the discharge intensity characterization and the infrared temperature rise rate characterization.

[0200] A feature matrix is ​​established based on the distribution of intermediate-round modal feature vectors generated after physical causality correction in a unified-dimensional implicit semantic space, serving as the input condition for the verification operation. The numerical distributions of the corresponding discharge intensity characterization vector and infrared temperature rise rate characterization vector in the feature matrix are extracted as a physical quantity comparison sequence to ensure a one-to-one correspondence between the indexes of each sample group, maintaining causal mapping consistency. A mapping judgment operator is constructed based on physical monotonicity constraints, and the physical logical consistency is detected by calculating the sign correlation coefficient between discharge intensity and infrared temperature rise rate on each sample. A logical deviation metric is constructed using sequence differencing, calculating the discharge intensity difference and infrared temperature rise rate difference for each data group, and performing directional consistency judgment to form a deviation label matrix. A quantization mapping formula is introduced to map the directional consistency judgment result into a numerical deviation index, for example, using the following physical monotonicity deviation calculation formula:

[0201] in, This represents the difference in the discharge intensity characterization vector between adjacent samples. This represents the difference in the infrared temperature rise rate characterization vector between adjacent samples. This represents the total number of samples. The deviation index is compared with a preset physical consistency threshold to generate a quantitative report on the convergence degree of logical deviation, which is used for subsequent termination condition determination. Through physical monotonicity consistency verification, the feature distribution of intermediate rounds is transformed into a precisely quantified logical deviation convergence degree, enabling a quantitative assessment of the quality of cross-modal semantic alignment.

[0202] S6.5: Determine whether the preset physical causal alignment termination condition is met based on the convergence degree of the logical deviation. If it is met, lock the current intermediate round modal feature vector as the final fused feature representation with cross-modal semantic alignment characteristics. Otherwise, return to perform parameter update operation until convergence.

[0203] Based on the quantified logical deviation convergence data in the unified-dimensional implicit semantic space, a preset set of physical causal alignment termination condition parameters is invoked. A threshold comparison operation is performed on the convergence state of the intermediate-round modal feature vectors generated in the current iteration cycle to form a convergence judgment signal. The convergence judgment signal is then Booleanized to generate a state flag for decision loop control. The state flag is input to the condition branch module. If the judgment value is satisfied, a feature locking operation is performed, directly binding the current intermediate-round modal feature vector to the final fused feature representation and writing it into the feature output buffer to form stable aligned feature data. If the judgment value is not satisfied, the lightweight encoder weight parameters of the current iteration round are sent to the gradient update module along with the logical deviation convergence data to return to the parameter update process defined in S6.1 to S6.4 according to the chained optimization path until the convergence judgment condition is met. Through threshold comparison, Boolean decision, feature locking, or iterative return processing, the logical deviation convergence result of the previous step is transformed into the final fused feature representation or a decision control index for further optimization, achieving the physical causal termination judgment effect for cross-modal semantic alignment.

[0204] For example, in a cable partial discharge monitoring scenario, to assess the convergence of the logical deviation between the discharge intensity representation vector and the infrared temperature rise rate representation vector in a unified-dimensional implicit semantic space, a physical causal alignment termination condition threshold of 0.05 is set. Specifically, the physical alignment termination condition is considered satisfied when the mean square value of the deviation is less than this threshold. The mean square value is calculated using the following formula during the convergence determination process: in Let be the square of the logical deviation on the i-th causal edge. This represents the total number of causal edges. In this embodiment, when the measured MSD value is 0.042, since it is less than the threshold of 0.05, the Boolean decision module outputs a satisfied signal, directly locking the modal feature vector of the current iteration as the final fused feature representation, ultimately achieving a stable output for cross-modal semantic alignment. In another case, when the measured MSD value is 0.061, which is greater than the threshold, the decision module returns a dissatisfied signal. The loop control logic sends the lightweight encoder weight parameters and this deviation value back to the gradient update module, continuing parameter optimization iterations until the deviation value drops below the threshold and the final feature locking is completed.

[0205] Step S7: Perform a defect type discrimination operation on the final fused feature representation to generate a defect type confidence distribution that includes classification results for air gap discharge, surface creepage, or metal tip corona discharge. Specifically, this includes: S7.1: Obtain the final fused feature representation after optimization by causal consistency loss, and perform linear projection transformation processing based on the fully connected layer architecture to map the causal modulation fused features in the unified dimension implicit semantic space to the high-level semantic discrimination space, generating a high-dimensional discriminative feature vector with class separability.

[0206] S7.2: Receive the high-dimensional discriminative feature vector, perform nonlinear activation and feature recombination processing based on a multilayer perceptron network to capture the complex boundary relationships of different defect types in the implicit semantic space, and generate a latent representation of defect patterns containing deep abstract information.

[0207] The high-dimensional discriminative feature vector generated by the linear projection transformation in step S7.1 is received as input. An initialization parameter set for the multilayer perceptron network is established based on the distribution of this vector in the implicit semantic space, including the number of layers, the number of neurons per layer, and the initial values ​​of the connection weights. Matrix multiplication is performed on the input high-dimensional discriminative feature vector, multiplying it with the weight matrix of the first layer of the multilayer perceptron and adding the corresponding bias vector to generate the first-layer linear transformation output. The first-layer linear transformation output is input to a nonlinear activation unit, employing activation functions such as the Modified Linear Unit (ReLU) or hyperbolic tangent (tanh) to enhance the nonlinear mapping capability, thereby generating nonlinearly separable features in the implicit semantic space. The activated feature vector is passed to subsequent hidden layers, sequentially performing a chain operation of weight matrix multiplication, bias summation, and nonlinear activation, forming a multilayer feature recombination process that allows discriminative features of different defect types to form complex boundary distributions in space. Residual connections or normalization are performed on the feature vectors in each hidden layer of the multilayer perceptron to suppress gradient vanishing and stabilize the feature distribution, ensuring the effective transmission of deep abstract information and the interpretability of feature patterns. Through the above feature recombination process, the high-dimensional discriminative feature vector is transformed into a latent defect pattern representation that contains deep abstract information and can characterize the differences in defect patterns such as air gap discharge, surface creepage, or metal tip corona discharge, thereby achieving accurate capture of complex boundary relationships.

[0208] S7.3: Utilizing the implicit representation of the defect pattern, perform normalized probability distribution calculation processing based on the soft maximum probability function to transform the deep abstract information into relative likelihood values ​​corresponding to various categories such as air gap discharge, surface creepage, and metal tip corona discharge, and generate an initial defect type probability score sequence.

[0209] The high-dimensional vector representing the implicit defect pattern is received as input to this sub-step. In the implicit semantic space, the target for discrimination of the three defect categories—air gap discharge, surface creepage, and metal tip corona—is clearly defined. For this high-dimensional implicit representation, a probability mapping mechanism including category output nodes is established to ensure that the deep abstract features in the vector are correctly mapped to the relative probability of occurrence for each category. During the mapping process, the implicit representation undergoes numerical normalization preprocessing to eliminate dimensional differences between different feature components and avoid bias effects from a single component on probability calculation. A normalization transformation formula constructed with an exponential function is used to transform the normalized category scores, amplifying the relative weight of high-scoring categories and attenuating the weight of low-scoring categories through the exponential value. The sum of the exponential values ​​of all category scores is calculated to generate a normalized denominator, ensuring that the final score for each category is within the range of 0 to 1 and the sum equals 1. The exponential score of each category is divided by this sum to obtain the normalized probability value for the corresponding category, forming the initial defect type probability score sequence. The above normalization calculation process is performed using a soft maximum probability function. By performing normalized probability distribution calculations based on this formula, deep abstract information is transformed into relative likelihood values ​​for three categories: air gap discharge, surface creepage, and metal tip corona, thereby outputting the initial defect type probability score sequence.

[0210] By using a soft maximum probability function transformation, the implicit representation of the defect pattern in the previous step is transformed into a probability index with clear physical meaning and numerical standardization, thereby realizing a quantitative expression of the probability of occurrence of each defect type.

[0211] S7.4: Based on the initial defect type probability score sequence, perform confidence correction processing based on temperature scaling calibration to eliminate model overconfidence bias and match the prior probability of defect occurrence in the actual physical scene, and generate a calibrated defect type confidence distribution.

[0212] Data extraction is performed on the initial defect type probability score sequence to obtain the normalized probability value vectors corresponding to air gap discharge, surface creepage, and metal tip corona as input data for calibration processing.

[0213] The probability value vector is initialized with a temperature scaling factor, and the optimal temperature parameter T is determined by minimizing the cross-entropy error between the predicted probability and the actual label on the validation set.

[0214] Temperature scaling is used to transform the probability vector, scaling the softmax input logits proportionally according to the temperature coefficient. The specific calculation formula is as follows:

[0215] Where z is the logit value corresponding to the category, T is the temperature parameter, p is the scaled probability value, and C is the total number of categories.

[0216] A distribution consistency check operation is performed on the probability value vector after temperature scaling transformation. By comparing the deviation between the actual defect occurrence frequency statistics and the scaled probability value distribution, the temperature coefficient is adjusted to match the prior probability of defects in the real scene.

[0217] The scaled probability vector after minimizing the bias is remapped to the calibrated defect type confidence distribution, ensuring that the sum of the probabilities of each category is always equal to 1.

[0218] By using temperature scaling calibration, the initial probability score sequence from the previous step is transformed into a defect type confidence distribution that matches the physical scene and eliminates overconfidence bias in prediction, thereby enhancing the credibility of the classification results in engineering applications.

[0219] S7.5: Integrate the calibrated defect type confidence distribution, perform final category locking processing based on threshold decision logic, and output the single defect type label with the highest confidence and its corresponding probability quantification index to generate a defect type confidence distribution that includes air gap discharge, surface creepage or metal tip corona classification results.

[0220] Step S8: Based on the activation intensity of key edges in the defect type confidence distribution, trace the decision path to generate a physically interpretable cable insulation defect diagnosis report, completing the closed-loop diagnosis process from multi-source signal acquisition to interpretable classification result output. Here, the key edges refer to those causal edges in the physical causal graph of cable insulation defects that make a decisive contribution to the final defect type classification decision, specifically edges with significant weights.

[0221] Specifically, it includes: S8.1: Perform threshold screening on the confidence distribution of defect types to extract candidate defect type nodes with activation intensity exceeding the preset confidence threshold and their associated high-weight causal edge set, generating an initial explanatory evidence chain containing potential fault modes such as air gap discharge or surface creepage.

[0222] S8.2: Based on the set of high-weight causal edges in the initial explanatory evidence chain, reverse path search processing is performed in the topology of the physical causal graph of cable insulation defects to locate the complete physical conduction sequence from the observation node back to the root cause node, and generate a discretized mechanism tracing path characterizing the process from electron avalanche to the accumulation of micro-strain in the structure.

[0223] The input conditions receive the initial explanatory evidence chain output by S8.1 and its corresponding set of high-weight causal edges, and load the complete directed acyclic topology of the physical causal graph of cable insulation defects.

[0224] For a set of high-weight causal edges, a reverse index table of nodes to edges is established, and the mapping relationship between the target node and the source node of each edge is stored as a lookup matrix that facilitates path backtracking.

[0225] Based on the lookup matrix, reverse path expansion processing is performed. Starting from each candidate observation node, the upstream source node is traced back along the reverse index. During each backtracking process, the physical semantics and signal mode labels corresponding to the node are recorded to form a progressively expanding sequence of path nodes.

[0226] During the path expansion process, physical mechanism chain constraints are introduced to limit the backtracking path to a set of nodes that conform to the sequential logic from electron avalanche to streamer, leader, main discharge, thermal effect to accumulation of structural micro-strain, and to eliminate path branches that do not conform to the arrow of time or unidirectional causality.

[0227] For the path node sequence filtered by the physical mechanism chain constraint, path termination detection is performed. When the upstream node belongs to the root cause node initialization list and its defect type label is consistent with the evidence chain defect type, the path is locked and archived as a complete physical transmission sequence.

[0228] All locked complete physical conduction sequences are sorted according to path length, node physical semantic coverage, and signal mode diversity to screen out highly interpretable discretization mechanism tracing paths that characterize the process from electron avalanche to structural micro-strain accumulation.

[0229] By using reverse path search and physical mechanism chain constraint processing, the initial explanatory evidence chain result of the previous step is transformed into a discretized mechanism tracing path with a complete cause-response chain, realizing the full traceability of the physical process from the observation layer to the root cause layer in the causal graph.

[0230] S8.3: Utilize the initial modal feature vectors corresponding to each node in the discretization mechanism tracing path to calculate the feature contribution gradient in the cross-modal semantic space, so as to quantify the semantic alignment deviation correction between the ultra-high frequency electromagnetic signal and the surface infrared thermal response signal on a specific causal edge, and generate a physical quantity logic verification index that reflects the consistency of multi-physics coupling.

[0231] S8.4: Based on the mapping relationship between the physical quantity logic verification index and the discretization mechanism tracing path, the high-voltage engineering discharge mechanism knowledge base is called to perform semantic matching operation, so as to translate the feature evolution trajectory in the implicit semantic space into natural language description text containing the relationship between the discharge intensity change rate and the dielectric temperature rise rate, and generate a defect evolution process explanation paragraph with engineering readability.

[0232] Receive the mapping relationship data between the physical quantity logic verification index and the discretization mechanism tracing path obtained by step S8.3, and establish a one-to-one correspondence index between physical causal nodes and feature evolution trajectories.

[0233] The system calls upon the high-voltage engineering discharge mechanism knowledge base to retrieve physical process entries that match the current source path node combination and its causal connectivity mode, and reads the mechanism description dataset containing the correlation between discharge intensity change and dielectric temperature rise rate.

[0234] Numerical trend analysis is performed on the feature evolution trajectories arranged sequentially along the source path in the implicit semantic space to calculate the rate of change of the discharge intensity characterization values ​​between consecutive nodes, using the difference operation formula:

[0235] Where P(t) represents the discharge intensity characterization value of the current state of the current node, ΔP is the discharge intensity change rate per unit time, and P(t-1) is the discharge intensity characterization value of the previous state of the current node.

[0236] The above discharge intensity change rate is logically matched with the corresponding node dielectric temperature rise rate to verify whether it satisfies the monotonically increasing constraint condition defined in the knowledge base.

[0237] Physical semantic mapping transformation is performed on node combinations that meet the constraints, transforming the numerical evolution patterns in the implicit space into quantitative insertion parameters in the natural language description template, forming a defect evolution process description statement with engineering readability.

[0238] The generated natural language description text must include the rate of change of discharge intensity, the rate of temperature rise of the medium, and their causal relationships, ensuring that the explanatory paragraphs can directly support the physical evidence section of the diagnostic report.

[0239] By using the above processing method, the results of the previous step are transformed into defect evolution process description data with readable physical mechanisms, thereby enhancing the engineering interpretability of the diagnostic results.

[0240] S8.5: Integrate the explanation paragraphs of the defect evolution process with the initial explanatory evidence chain, perform structured report assembly processing to generate a physically interpretable cable insulation defect diagnosis report that includes defect type determination conclusions, key physical evidence support, and credibility scoring dimensions, thereby achieving traceability and engineering applicability of the diagnosis results.

[0241] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0242] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0243] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered 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 intelligent diagnosis of partial discharge insulation defects in cables, characterized in that, include: S1: Acquire multi-source heterogeneous raw time-series data in the cable partial discharge monitoring scenario, and perform local abrupt event detection to extract discrete event label sequences; S2: Based on the discharge mechanism chain of electron avalanche to structural micro-strain accumulation in high-voltage engineering, a physical cause-effect graph of cable insulation defects is constructed using the discrete event labeling sequence; S3: Input the discrete event tag sequence into the lightweight encoder of the corresponding mode, perform feature mapping operation, and generate an initial mode feature vector that is mapped to a unified dimension implicit semantic space and serves as the potential representation of the observation node in the physical causal graph of the cable insulation defect; S4: Based on the topological constraints of the physical cause-effect graph of the cable insulation defect, perform graph neural network message passing processing on the initial modal feature vector to generate causal modulation fusion features; S5: Calculate the feature transformation deviation on each causal edge in the physical causal graph of the cable insulation defect according to the preset physical monotonicity constraint, construct the causal consistency loss function, and generate the physical constraint gradient; S6: Utilize the physical constraint gradient to perform backpropagation optimization on the parameters of the lightweight encoder, driving the initial modality feature vector to converge along the physical causal direction in the implicit semantic space, generating the final fused feature representation; S7: Perform a defect type discrimination operation on the final fused feature representation to generate a defect type confidence distribution; S8: Based on the activation intensity of the critical edge in the confidence distribution of the defect type, trace the decision path and generate a cable insulation defect diagnosis report.

2. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 1, characterized in that, The discrete event marker sequence includes: the UHF pulse cluster start point, the di / dt zero-crossing point, the thermal gradient jump frame, and the vibration energy peak frame.

3. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 2, characterized in that, Step S3 specifically includes: Based on the observation node type defined in the physical cause-effect graph of cable insulation defects, an independent feature extraction channel is assigned to the discrete event marker sequence to construct a set of modality-specific lightweight encoders. The starting point of the UHF pulse cluster, the di / dt zero-crossing point, the thermal gradient jump frame and the vibration energy peak frame in the discrete event marker sequence are respectively processed by local time-frequency domain slicing. The original time series data segment containing complete abrupt change features is extracted by using a sliding window to generate a multimodal local time series data segment. The multimodal local time-series data segments are input into the corresponding modality-specific lightweight encoder set, and multi-layer nonlinear transformation operations are performed using a one-dimensional convolutional neural network to generate high-dimensional abstract feature vectors. Based on the preset unified implicit semantic space dimension, the high-dimensional abstract feature vector is subjected to fully connected layer projection dimensionality reduction processing to generate a unified dimension initial modal feature vector; The initial modal feature vector of the unified dimension is bound to the observation node in the physical cause-effect graph of the cable insulation defect to generate the initial modal feature vector.

4. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 3, characterized in that, The modal-specific lightweight encoder set includes: an ultra-high frequency electromagnetic pulse encoder, a high frequency current pulse encoder, a surface infrared thermal response encoder, and a micro-vibration acceleration encoder.

5. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 1, characterized in that, The multi-source heterogeneous raw time-series data includes: ultra-high frequency electromagnetic signals, high frequency current pulse signals, surface infrared thermal response signals, and micro-vibration acceleration signals.

6. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 5, characterized in that, Step S4 specifically includes: Based on the initial modal feature vector and the node connection relationship of the physical cause-effect graph of cable insulation defects, an adjacency matrix and edge type mapping table are constructed to generate graph structure constraint parameters; The initial modal feature vector is initialized by performing node embedding using the graph structure constraint parameters. An edge weight initialization strategy is defined based on the discharge mechanism chain from discharge intensity to structural micro-strain accumulation to generate a node state representation vector. Based on the node state representation vector and the edge type mapping table, a directed message aggregation operation is performed. The feature information of the upstream neighbor nodes is extracted according to the causal path from the ultra-high frequency electromagnetic signal to the surface infrared thermal response signal, and an aggregated message vector is generated. The aggregated message vector is subjected to nonlinear feature transformation and gating fusion processing. The local feature distribution is updated in combination with the current node's state representation vector to generate the hidden state vector of the intermediate layer node. Based on the hidden state vector of the intermediate layer nodes, perform multi-round iterative message passing optimization until the node features converge to a stable state that satisfies the physical consistency logic in the implicit semantic space, and generate the causal modulation fusion features.

7. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 1, characterized in that, Step S5 specifically includes: Based on the causal edge connection relationship in the physical causal graph of the cable insulation defect, the discharge intensity characterization vector output by the source node and the infrared temperature rise rate characterization vector input by the target node are obtained. The discharge intensity characterization vector and the infrared temperature rise rate characterization vector are subjected to monotonic mapping matching processing to generate a sequence of feature pairs to be verified. Using a preset physical monotonicity constraint, a numerical order relationship determination operation is performed on each set of data in the sequence of features to be verified, and an abnormal feature deviation term set that violates the law that the infrared temperature rise rate does not decrease when the discharge intensity increases is identified. Based on the deviation magnitude of each element in the set of abnormal feature deviation items, a piecewise linear penalty function is used to calculate the feature transformation deviation value on each causal edge, generating a physical inconsistency scalar. Based on the physical inconsistency scalar, a weighted aggregation operation is performed on the feature transformation deviation values ​​of all causal edges to construct the causal consistency loss function and output the total loss error value. An automatic differentiation backpropagation operation is performed on the causal consistency loss function to calculate the partial derivative of the total loss error value with respect to the lightweight encoder parameters, in order to generate the physical constraint gradient.

8. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 7, characterized in that, The set of abnormal feature deviation terms includes each pair of source and target node features that violate the physical monotonicity constraint and their numerical deviation magnitude.

9. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 1, characterized in that, The confidence distribution of the defect type includes: air gap discharge, surface creepage, or metal tip corona classification results.

10. The intelligent diagnostic method for partial discharge insulation defects in cables according to claim 1, characterized in that, The cable insulation defect diagnosis report includes a defect type determination conclusion, key physical evidence support, and a credibility score dimension.

Images