Method for predicting performance of cable support based on environmentally friendly insulation composite

Abstract

This invention relates to a performance prediction method for cable supports based on environmentally friendly insulating composite materials. Addressing the difficulty in accurately assessing the performance degradation of existing materials, this invention proposes a multi-modal characterization method integrating high-resolution X-ray tomography and in-situ Raman spectroscopy, combined with graph neural networks and physical constraint correction mechanisms, to dynamically predict the microstructure evolution trajectory and insulation performance degradation. Its core technical solutions include accelerated aging sample multi-field coupling processing, simultaneous acquisition of three-dimensional microstructure and chemical information, multi-scale feature extraction and heterogeneous graph modeling, performance prediction based on structure-performance mapping, and online adaptive model updating. This solution enables cross-scale correlation between microscopic mechanisms and macroscopic properties, improving the accuracy and adaptability of insulation performance degradation prediction, and providing intelligent decision-making basis for the quantitative assessment of the remaining life and safety management of environmentally friendly insulating composite material cable supports.

Description

Technical Field

[0001] This invention relates to the field of insulation performance prediction and material aging modeling technology, and in particular to a method for predicting the performance of cable supports based on environmentally friendly insulating composite materials. Background Technology

[0002] Currently, environmentally friendly insulating composite materials have received widespread attention in engineering applications such as cable supports in power systems. Predicting insulation performance degradation and assessing material lifespan have become core technological requirements in this field. Under existing technologies, insulation performance prediction methods are mainly divided into two categories: one is based on physical mechanisms and material aging evolution models, typically using equations such as Fick diffusion and Arrhenius thermal activation to describe the microscopic aging reaction rate and macroscopic performance changes; the other relies on data-driven statistical models or machine learning algorithms, such as Weibull distribution fitting, Wiener process degradation modeling, SVM regression, and LSTM neural networks, to quantitatively determine lifespan by analyzing macroscopic performance test sequences.

[0003] Physical mechanism modeling techniques offer strong interpretability, reflecting the intrinsic influence of material composition, microstructure, and environmental conditions on the aging process. However, the acquisition of actual parameters and modeling complexity are high, making it difficult to accurately capture the nonlinear effects of multi-field coupling and complex microstructure evolution on insulation performance. Statistical modeling and data-driven methods focus on mining evolutionary patterns from historical data, offering fast parameter training and strong adaptability. However, they often lack sufficient physical constraints and microstructure interpretation capabilities, significantly limiting predictive generalization and stability when operating conditions vary greatly or stress fields are complex.

[0004] In representative technical applications, pure physical models are mainly used for material formulation optimization and experimental comparative analysis, focusing on performance evolution under a single stress field or steady-state aging scenario. Data-driven models, on the other hand, are more common in existing power operation and maintenance decision-making systems, focusing on fitting macroscopic performance degradation trends. They are difficult to perform personalized life assessments based on the internal evolution dynamics of materials, especially in the context of novel environmentally friendly composite materials with diverse microstructures, intertwined aging mechanisms, and limited data. Existing solutions cannot achieve physical consistency tracing and highly reliable generalized prediction from microstructural changes to macroscopic functional degradation.

[0005] Existing technologies generally suffer from the following shortcomings and technical defects: Existing methods for predicting insulation performance degradation typically separate physical mechanisms from data-driven approaches, lacking organic integration. Purely physical models face challenges such as high-dimensional and complex parameters, limited inference scale, and inability to adapt to heterogeneous microstructures and complex environments. Purely data-driven modeling struggles to explain the root causes of performance degradation from a microscopic perspective and is prone to prediction drift and misjudgments in situations with insufficient macroscopic sampling or novel material applications.

[0006] The industry commonly uses degenerate models based on Weibull distributions and Wiener stochastic processes, which are mainly designed for objects with static parameters collected over long periods and lack the ability to adaptively model changes across microstructures. While the introduction of deep learning methods such as LSTM and digital twins has improved time series prediction capabilities, it is prone to falling into the trap of substituting data correlation for physical causality, making it difficult to guarantee the interpretability and accuracy of physical mechanisms.

[0007] Environmentally friendly insulating composite materials undergoing complex and multi-stress synergistic aging exhibit prominent dynamic evolution characteristics of their microstructure and chemical composition. Existing technical approaches based on characterization hierarchical data or black-box models lack the ability to integrate multiple scales to understand the actual aging path of materials. The generalization and robustness of model output results to new working conditions, new materials, or microstructure disturbances are limited.

[0008] In the process of cross-scenario deployment and promotion of new composite materials, the model needs to be compatible with different aging environments and structural types. Relying on manual screening and statistical factor adjustment has practical bottlenecks such as high manpower, low efficiency and high migration costs, which cannot meet the actual needs of rapid adaptation and personalized life prediction in engineering applications.

[0009] In summary, the field urgently needs a technology for predicting insulation performance degradation that can deeply integrate the microscopic physical mechanisms of material aging with cross-scale data-driven characteristics. This technology should be intrinsically derived from the data flow, using the dynamic evolution of microstructures as a key carrier to achieve physically interpretable predictions of macroscopic performance degradation. It should also possess small-sample adaptive generalization and intelligent correction capabilities, thereby overcoming the current technical bottlenecks such as the disconnect between physical and statistical models, insufficient adaptability to engineering deployment, and imperfect mapping from micro to macro scales. This need provides a clear practical direction and theoretical foundation for the proposal and development of this invention. Summary of the Invention

[0010] This application provides a performance prediction method for cable supports based on environmentally friendly insulating composite materials, aiming to solve one of the problems or issues of the prior art mentioned in the background section above.

[0011] The performance prediction method for cable supports based on environmentally friendly insulating composite materials provided in this application specifically includes: S1: For the aging scenario of environmentally friendly insulated composite material cable brackets under composite stress environment, the accelerated aging process is collected synchronously to obtain the original multimodal characterization dataset. S2: Based on the original multimodal characterization dataset, perform image segmentation and topological analysis to extract phase interface area, pore connectivity, microcrack fractal dimension and polar group density gradient, and generate a microstructure evolution map library. S3: Based on the topological connections in the microstructure evolution map library, local phase regions are abstracted into nodes, interface coupling strength is abstracted into edges, and chemical and geometric features are abstracted into attributes to construct a dynamic heterogeneous graph data model with the current aging stage graph data model snapshot as input. S4: Input the dynamic heterogeneous graph data model into the graph neural network architecture to perform forward propagation calculation and output the set of predicted microstructure parameters for the next stage; S5: Based on the predicted set of microstructure parameters for the next stage, the deviation is compared with the calculation results of the preset physical aging model. When the deviation exceeds the preset threshold, the physical constraint reweighting mechanism is triggered to generate the corrected microstructure evolution path. S6: Using the modified microstructure evolution path as feature input, a nonlinear mapping transformation is performed through the microstructure-performance correlation decoder to calculate the insulation performance attenuation rate feature vector characterizing the intrinsic degradation logic of the environmentally friendly insulating composite material.

[0012] The performance prediction method for cable supports based on environmentally friendly insulating composite materials provided in this application has the following beneficial effects: (1) To address the problem that traditional methods for predicting the performance degradation of insulating materials generally rely on pure statistical modeling or empirical distribution fitting, which are difficult to characterize the complex microscopic evolution mechanism under multi-field coupling, this solution constructs a "microstructure evolution map library" and uses it as the core prediction carrier, achieving a fundamental shift from being driven by the appearance of macroscopic performance degradation to being driven by the dynamic evolution of microstructure. Existing technologies such as Weibull distribution fitting and Wiener process modeling usually simplify the aging process into a time series evolution of a single variable, ignoring the nonlinear coupling behavior of phase interface degradation, microcrack propagation and chemical bond breaking under the interaction of multiple stresses such as electro-thermal-humidity-mechanical, resulting in poor generalization ability and high extrapolation risk of prediction results under varying operating conditions. This invention acquires three-dimensional microstructure images and local chemical information simultaneously through high-resolution X-ray tomography and in-situ Raman spectroscopy, and extracts 12 types of data, including pore connectivity, microcrack fractal dimension and polar group density gradient, by combining image segmentation and topology analysis algorithms. The system constructs a microstructure evolution map covering the entire life cycle, enabling a refined characterization of the internal degradation path of the material. This map not only records the historical trajectory of structural evolution, but also forms a closed-loop feedback label system by marking measured insulation resistivity and dielectric loss factor, providing dual support for subsequent models with both physical meaning and data accuracy. This significantly improves the predictive model's ability to analyze complex degradation mechanisms and the reliability of cross-stage extrapolation.

[0013] (2) To address the problem that existing data-driven models are prone to deviating from the intrinsic physical laws of materials and have a tendency to be "black boxed," resulting in insufficient engineering credibility, this solution innovatively designs a lightweight graph neural network and a physical-data dual-path feedback mechanism, achieving for the first time the collaborative optimization of the computable propagation of the microstructure evolution process and physical consistency verification. Traditional LSTM, SVM regression, or digital twin threshold generation methods often abstract the input-output relationship into an end-to-end mapping, lacking explicit constraints on the evolution logic of intermediate states, and are prone to non-physical interpretation under small sample or stress mutation conditions. This invention abstracts the microstructure map into a dynamic heterogeneous graph, with local phase regions as nodes, interface coupling strength as edges, and geometric and chemical features as attributes, enabling the graph network to explicitly model the interactions between multiphase media and their topological evolution with the aging process. More importantly, a physical constraint reweighting module is introduced, which automatically adjusts the edge weights and node update rules when the graph network prediction path deviates from the prior physical models such as Fick diffusion-controlled hydrolysis and Arrhenius-type thermally induced crosslinking fracture, forcing the prediction trajectory to regress to the intrinsic degradation logic of environmentally friendly insulating composite materials. This fusion architecture, which is "data-driven as the main approach and physics-guided as the auxiliary approach," retains the powerful fitting ability of deep learning to high-dimensional nonlinear relationships while ensuring that the prediction process conforms to basic physical laws. It significantly improves the robustness and engineering interpretability of the model under unknown conditions and effectively overcomes the technical bottlenecks of traditional methods in predicting instability and mechanism decoupling under strong disturbances.

[0014] (3) To address the issues of high migration and adaptation costs and difficulty in quickly adapting to new batches of materials or changes in service environment in traditional prediction models, this solution proposes an online incremental learning mechanism based on a microstructure-performance correlation decoder, achieving efficient model adaptive updates driven by small sample field data. Unlike the traditional approach of directly fitting the insulation resistivity time series, this invention trains the decoder to learn only the nonlinear mapping relationship between the microstructure spectrum feature vector and the insulation performance decay rate, removing the influence of specific time scales and enhancing the model's generalization ability to different initial states and loading processes. When a new batch of environmentally friendly composite material supports is put into actual operation, only a small amount of field monitoring data needs to be transmitted back, and the model can be quickly adapted by fine-tuning the parameters of the last layer of the decoder without retraining the entire network, significantly reducing computational overhead and deployment delay. This mechanism is particularly suitable for environmentally friendly material systems with large fluctuations in composition, such as bio-based epoxy / nanocellulose, and can maintain stable prediction accuracy under real-world interferences such as raw material batch differences and process fine-tuning, fully ensuring the continuity of safety assessment and the timeliness of maintenance decisions for cable supports during long-term service. Overall, this solution constructs a closed-loop prediction system of "microscopic perception - structural modeling - physical calibration - dynamic adaptation", truly realizing the leap from static experience model to dynamic and evolvable intelligent prediction system, and providing strong technical support for the high-reliability application of environmentally friendly insulation materials in power infrastructure. Attached Figure Description

[0015] Figure 1 This is the main flowchart of a performance prediction method for cable supports based on environmentally friendly insulating composite materials; Figure 2 This is a sub-flowchart of a performance prediction method for cable supports based on environmentally friendly insulating composite materials; Figure 3 This is another sub-flowchart of the performance prediction method for cable supports based on environmentally friendly insulating composite materials. Detailed Implementation

[0016] 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.

[0017] 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.

[0018] like Figure 1 As shown, this application provides a performance prediction method for cable supports based on environmentally friendly insulating composite materials, specifically including: S1: For the aging scenario of environmentally friendly insulated composite material cable supports under a combined electro-thermal-humidity-mechanical stress environment, the accelerated aging process is synchronously acquired to obtain a raw multimodal characterization dataset. Specifically, the raw multimodal characterization dataset includes three-dimensional microstructure sequence images and local chemical bond evolution spectral data. The synchronous acquisition of the accelerated aging process includes: using high-resolution X-ray tomography to acquire the three-dimensional microstructure sequence images of the environmentally friendly insulated composite material cable supports during the aging process; and using in-situ Raman spectroscopy to acquire the local chemical bond evolution spectral data of the environmentally friendly insulated composite material cable supports during the aging process.

[0019] S2: Based on the original multimodal characterization dataset, perform image segmentation and topological analysis to extract the phase interface area, pore connectivity, microcrack fractal dimension and polar group density gradient, and generate a microstructure evolution map library covering the entire life cycle and its corresponding macroscopic insulation performance degradation tag set.

[0020] S3: Based on the topological connections in the microstructure evolution graph library, local phase regions are abstracted into nodes, interface coupling strength is abstracted into edges, and chemical and geometric features are abstracted into attributes. A dynamic heterogeneous graph data model is constructed with the current aging stage graph data model snapshot as input.

[0021] S4: Input the dynamic heterogeneous graph data model into the graph neural network architecture to perform forward propagation calculation, and output the set of predicted microstructure parameters for the next stage and the confidence interval vector representing the uncertainty of the model.

[0022] S5: Based on the predicted set of microstructure parameters for the next stage, the deviation is compared with the calculation results of the preset physical aging model. When the deviation exceeds the preset threshold, the physical constraint reweighting mechanism is triggered to generate a corrected microstructure evolution path that includes dynamically adjusted edge weights and node update rules.

[0023] S6: Using the modified microstructure evolution path as feature input, a nonlinear mapping transformation is performed through the microstructure-performance correlation decoder to calculate the insulation performance attenuation rate feature vector characterizing the intrinsic degradation logic of the environmentally friendly insulating composite material.

[0024] Step S1: For the aging scenario of environmentally friendly insulated composite material cable supports under a combined electro-thermal-humidity-mechanical stress environment, the accelerated aging process is synchronously acquired to obtain the original multimodal characterization dataset. Specifically, the original multimodal characterization dataset includes three-dimensional microstructure sequence images and local chemical bond evolution spectral data; synchronous acquisition of the accelerated aging process includes: using high-resolution X-ray tomography to acquire the three-dimensional microstructure sequence images of the environmentally friendly insulated composite material cable supports during the aging process; and using in-situ Raman spectroscopy to acquire the local chemical bond evolution spectral data of the environmentally friendly insulated composite material cable supports during the aging process. Specifically, this includes: S1.1: Perform multi-field coupling accelerated aging treatment on environmentally friendly insulating composite material cable bracket samples placed in an electro-thermal-humidity-mechanical composite stress loading device to generate a dynamic aging sample set at different aging stages, which serves as the physical object for subsequent in-situ characterization operations.

[0025] Accelerated aging of environmentally friendly insulated composite cable bracket samples requires placing the test samples inside a dedicated electro-thermal-humidity-mechanical composite stress loading device. Electric field strength, ambient temperature, relative humidity, and mechanical load parameters are set, and the combined stress level is selected based on the material's design life requirements and aging acceleration factor. A constant or periodically varying voltage is applied using a precision electric field generator to ensure the insulation layer is in the electric field distribution state under actual operating conditions. The chamber temperature is adjusted to the preset value using a high-precision temperature control system, and the target relative humidity is stably maintained using a humidity control module. A servo controller applies programmable tensile, compressive, or bending loads at the mechanical loading end, achieving the coupling and superposition of multiple mechanical stress modes. Samples are sampled in batches according to predetermined aging time intervals, and the corresponding stress parameters, aging timestamps, and initial microstructure state information are recorded for each batch of samples. For each sample, its surface temperature, current response, and deformation are monitored in real time to ensure that all parameters fluctuate within a safe range, and equipment errors are automatically calibrated. The aging acceleration factor is calculated using the following formula: ; Wherein, AF is the aging acceleration factor, t ref t is the time required for a material to reach a specific degradation level under natural service conditions. test To accelerate the time to reach the same decay level under experimental conditions.

[0026] A multi-field coupled loading scheme was adopted, and an automatic control system was used to synchronously apply and dynamically adjust four types of stress—electric, thermal, wet, and mechanical—in both spatial and temporal dimensions. After each aging cycle, the samples were removed from the loading device, numbered, and archived, forming a dynamic aging sample set at different aging stages, providing physical objects for subsequent in-situ characterization operations.

[0027] S1.2: Based on the dynamic aging sample set, a three-dimensional spatial tomographic imaging operation is performed using a high-resolution X-ray tomography device to obtain three-dimensional microstructure sequence image data reflecting the internal phase region distribution, pore morphology and microcrack topology of the material.

[0028] For a dynamic aging sample set of environmentally friendly insulating composite cable brackets that have undergone multi-field coupling accelerated aging treatment, spatial acquisition parameters of a high-resolution X-ray tomography scanner were set, including voxel size, scan step size, and energy window. The samples were sequentially loaded onto the scanning platform according to their aging stage numbers. A helical scanning mode was used to perform full-volume tomography imaging on each sample, acquiring a grayscale projection dataset covering the three-dimensional space inside the material. A reconstruction module based on a filtered back-projection reconstruction algorithm was used to convert the original projection data into a three-dimensional voxel matrix, achieving a three-dimensional reconstruction of the material's internal structure. For the reconstructed three-dimensional voxel matrix, multi-scale edge enhancement and noise suppression filtering were applied to improve the resolution of microcracks, pores, and phase interface regions. Combined with automatic partitioning based on threshold segmentation and region growing, spatial annotations were performed on the matrix phase region, reinforcement phase region, and pore defect region, generating a sequence of three-dimensional microstructure images with physical attribute labels. By continuously scanning samples at different aging stages, the three-dimensional microstructure sequence images of each stage were encapsulated into a dynamic evolution sequence according to a time index, enabling data tracking of microstructure changes as the aging process progresses. Through the above processing method, the dynamic aging sample set is transformed into three-dimensional microstructure sequence image data reflecting the internal phase region distribution, pore morphology and microcrack topology of the material, providing a high-precision spatial data foundation for subsequent multimodal characterization fusion and topology analysis.

[0029] S1.3: For the same aging time point corresponding to the three-dimensional microstructure sequence image data, perform molecular vibrational spectral scanning operation using an in-situ Raman spectrometer to extract local chemical bond evolution spectral data characterizing changes in polar group concentration, degree of chemical bond breaking, and interfacial hydrolysis reaction process.

[0030] For the same aging time point corresponding to the three-dimensional microstructure sequence image data, environmentally friendly insulating composite cable bracket samples after multi-field coupling accelerated aging treatment were selected as the characterization object. Molecular vibrational spectroscopy scanning was performed at precisely located spatial coordinates using an in-situ Raman spectrometer. The excitation wavelength was set to 532 nm, and the power was controlled below 2 mW to avoid thermal damage. An automatic focusing system ensured spatial consistency with the three-dimensional tomographic imaging results. Raman spectral signals were acquired for each scanning point, and vibrational mode intensity changes were recorded using a high-sensitivity CCD detector. Baseline correction and normalization were performed on the raw Raman spectral data to eliminate background noise and instrument drift. Principal component analysis (PCA) was used to perform feature reduction on the normalized Raman spectra, extracting characteristic peak regions reflecting changes in polar group concentration (such as –OH, –COOH, –NH2, etc.), and calculating the integral area of ​​each peak region to quantify the polar group content. Peak position fitting was used to quantitatively analyze the degree of breakage of key chemical bonds such as C–O and C=C. Combined with Lorentz / Gauss hybrid curve fitting, the breakage rate was expressed by the following formula: ; Where A(t) is the target bond peak area at aging time t, and A(0) is the target bond peak area in the initial unaged state. This ratio is used to characterize the degree of fracture.

[0031] By utilizing the dynamic evolution of characteristic peaks of hydrolysis reactions (such as water molecule bending vibrations and carboxyl group formation) in Raman spectra, combined with time-series analysis, the rate of interfacial hydrolysis reaction is fitted. The following Arrhenius-type rate equation is used: ; Where k is the hydrolysis reaction rate constant, A is the frequency factor, and E a R is the activation energy, T is the gas constant, and T is the absolute temperature. These parameters are obtained by fitting experimental data.

[0032] The results of the above-mentioned changes in polar group concentration, degree of chemical bond breaking, and interfacial hydrolysis reaction rate are registered with the spatial coordinates of the three-dimensional microstructure image data to output a characterization matrix containing information on the evolution of local chemical components at multiple points.

[0033] S1.4: Perform spatiotemporal coordinate registration and multimodal fusion processing on the three-dimensional microstructure sequence image data and the local chemical bond evolution spectral data to generate a multimodal synchronous characterization data stream with strict spatial correspondence and timestamp alignment.

[0034] Three-dimensional microstructure sequence image data and local chemical bond evolution spectral data are synchronously processed to obtain the spatial coordinate information and timestamp attributes corresponding to each aging stage. A feature point matching algorithm is used to spatially register the physical phase region boundary points in the three-dimensional microstructure sequence images with the Raman spectral acquisition positions, establishing a one-to-one mapping relationship between image voxels and spectral acquisition points. Using a timestamp alignment mechanism, each set of three-dimensional microstructure sequence images and its corresponding local chemical bond evolution spectral data are associated according to the aging process time sequence index, ensuring strict time synchronization of multimodal data within the same aging stage. For the spatial registration results, a multimodal fusion algorithm is used to jointly encode image voxel attributes (such as phase region type, pore morphology, and microcrack distribution) and spectral feature parameters (such as polar group concentration, chemical bond breakage rate, and interface hydrolysis rate), generating a multimodal feature vector set that integrates geometric structure and chemical composition. By constructing a multimodal synchronous characterization data stream, the above jointly encoded features are encapsulated into structured data packets according to spatial coordinate index and timestamp order, achieving high-precision synchronous recording of the three-dimensional microstructure state and local chemical reaction process. By using spatiotemporal coordinate registration and multimodal fusion processing, the three-dimensional microstructure sequence image data and local chemical bond evolution spectral data from the previous step are transformed into a multimodal synchronous characterization data stream with strict spatial correspondence and time stamp alignment. This achieves a high degree of information coupling between the microstructure state and the material aging reaction process, providing a unified data input basis for subsequent standardized packaging and metadata annotation operations.

[0035] For example, accelerated aging of bio-based epoxy / nanocellulose reinforced scaffold samples was carried out under a combined electro-thermal-humidity-mechanical stress environment. The acquisition cycle was 24 hours, and each cycle acquired a 128×128×64 voxel resolution three-dimensional X-ray tomographic image and 10 corresponding Raman spectral acquisition points. The SURF algorithm was used to locate the phase interface edges in the tomographic image, and a spatial registration threshold of 0.5 mm was set. The Raman spectral acquisition points were projected onto the tomographic image coordinate system to achieve a one-to-one mapping between physical phase regions and chemical reaction regions. Using a timestamp synchronization mechanism, all images and spectral data within the same cycle were sorted according to the aging stage number. Principal component analysis (PCA) was used to perform joint dimensionality reduction on the fused voxel properties and spectral parameters to generate a 32-dimensional multimodal feature vector set, which was then encapsulated into a standardized synchronous characterization data stream according to spatial index and timestamp. In this scenario, the multimodal synchronous characterization data stream output contains 128×128×64 spatial nodes, with each node having an additional 32-dimensional joint feature, covering all aging stages throughout the entire life cycle. This enables efficient coupling and recording of the microstructure evolution trajectory and the interface hydrolysis reaction process, significantly improving the data consistency and model generalization ability of subsequent evolution map library construction.

[0036] S1.5: Perform standardized encapsulation and metadata annotation operations based on the multimodal synchronous characterization data stream to output the original multimodal characterization dataset containing complete three-dimensional microstructure sequence images and local chemical bond evolution spectral data, which serves as the direct input source for constructing the microstructure evolution map library.

[0037] Step S2: Based on the original multimodal characterization dataset, perform image segmentation and topological analysis to extract phase interface area, pore connectivity, microcrack fractal dimension, and polar group density gradient, generating a microstructure evolution map library covering the entire life cycle and its corresponding macroscopic insulation performance degradation tag set. Specifically, this includes: S2.1: Perform multi-scale adaptive threshold segmentation processing on the three-dimensional microstructure sequence images contained in the original multimodal characterization dataset to separate the matrix phase region, reinforcement phase region and pore defect region in the environmentally friendly insulating composite material, and generate a binary phase region mask matrix with clear physical boundaries.

[0038] For the three-dimensional microstructure sequence images contained in the original multimodal characterization dataset, the voxel-level spatial distribution matrix output by a high-resolution tomographic scanner was used as the input. Based on the material aging stages and differences in physical properties, multi-scale adaptive threshold segmentation was performed, dynamically adjusting the threshold parameters to adapt to the grayscale distribution characteristics of the matrix phase region, the reinforcement phase region, and the porosity / defect region. Using local statistical analysis methods, the mean, standard deviation, and kurtosis were calculated for each spatial window, and the adaptive threshold function was set as follows: ; Where T is the local threshold, μ is the mean gray level of the current window, σ is the standard deviation of gray level, and k is the adjustment coefficient. The above threshold segmentation operation is performed on spatial windows at different scales to classify the matrix phase region, reinforcement phase region, and porosity defect region according to their gray level characteristics, achieving multi-scale segmentation. Morphological closing operations and connectivity filtering are applied to the preliminary segmentation results to eliminate isolated noise points and enhance the continuity of physical boundaries. Furthermore, combining prior material knowledge, gradient enhancement processing is performed at the boundary between the matrix phase region and the reinforcement phase region to improve the physical accuracy of the mask edges. Finally, all segmentation results are integrated into a three-dimensional binary phase region mask matrix, where each voxel is labeled as a matrix, reinforcement, or porosity category according to its physical properties.

[0039] Through the above chain processing method, the original three-dimensional microstructure sequence image is transformed into a binary phase region mask matrix with clear physical boundaries, providing standardized input for subsequent geometric topological feature quantization and microstructure evolution map construction.

[0040] S2.2: Based on the binary phase region mask matrix, perform three-dimensional skeleton extraction and Euler characteristic number topology analysis to quantify the phase interface area density, pore connectivity coefficient and microcrack fractal dimension, and generate a set of multi-dimensional geometric feature vectors characterizing the evolution of micro-geometric morphology.

[0041] The obtained three-dimensional binary phase region mask matrix is ​​processed. The input object is a three-dimensional spatial voxel matrix after multi-scale adaptive thresholding, where each voxel has been labeled as matrix, enhancement or pore category.

[0042] A 3D skeleton extraction algorithm is employed to simplify the spatial structure of pore defect and microcrack regions, transforming the original voxel set into a skeleton line set that retains only the main connected paths. During skeleton extraction, distance transformation and morphological refinement operations are used to ensure that the skeleton structure fully reflects pore connectivity and microcrack propagation direction.

[0043] Euler characteristic topological analysis is performed based on the skeleton line set. The Euler characteristic of each spatial sub-region is calculated, and the topological complexity is quantified by the following formula: ; Where χ is the Euler characteristic number, V is the number of nodes (skeleton endpoints), E is the number of edges (skeleton connecting lines), and F is the number of faces (surfaces formed by closed loops).

[0044] A quantitative analysis of the interfacial area density was performed. Surface area calculation was used to scan the mask at the boundary between the matrix and the reinforced phase region point by point. The total area of ​​all boundary voxels was calculated and normalized to a unit volume to obtain the interfacial area density index. The specific calculation formula is as follows: ; Among them, S d Let S be the interfacial area density, S be the total interfacial area, and V be the volume of the analysis region.

[0045] To determine the pore connectivity coefficient, a connected graph is constructed based on the skeleton extraction results. A breadth-first search is performed on all pore nodes, and the ratio of the largest connected subset to all subsets is used as the connectivity coefficient. The formula is expressed as follows: ; Among them, C p N is the pore connectivity coefficient. max N represents the maximum number of nodes in a connected subset. total This represents the total number of pore nodes.

[0046] The fractal dimension of microcracks was quantitatively calculated by statistically analyzing the minimum number of boxes covering the microcrack skeleton at different spatial scales using the box-counting method, and then fitting the fractal dimension D. The formula for calculating the fractal dimension is as follows: ; Where D is the fractal dimension, N b For the box count result (the number of boxes covering the microcrack skeleton), s b This refers to the box dimensions.

[0047] S2.3: Spatial registration and gradient field reconstruction of the multidimensional geometric feature vector set are performed using the local chemical bond evolution spectral data obtained by in-situ Raman spectroscopy to calculate the polar group density gradient distribution map and generate a microstructure state descriptor tensor that integrates geometric topological information and chemical composition information.

[0048] A spatial correspondence is established between the multidimensional geometric feature vector set output from the previous steps and the local chemical bond evolution spectral data acquired by in-situ Raman spectroscopy. A mapping mechanism based on spatial coordinate indexing is used to register the geometric feature points extracted from the 3D skeleton with the Raman spectral acquisition locations one-to-one. Spatial interpolation algorithms are used to complete the chemical composition information for uncovered points, ensuring that each geometric feature point has the corresponding chemical bond evolution parameters. Using the gradient field reconstruction method, a density distribution function of polar group concentration in three-dimensional space is constructed. The finite difference method is used to calculate the density gradient of polar groups, with the specific formula as follows: ; in, ρ(x,y,z) represents the polar group density gradient vector, ρ(x,y,z) represents the polar group concentration at a spatial point, and Δx represents the spatial step size. The gradient calculation is performed on all spatial nodes, and combined with the Euler characteristic number topological analysis results, geometric features and chemical composition information are fused at the tensor level. Using a tensor splicing method, geometric indices such as interface area, pore connectivity, and microcrack fractal dimension are encapsulated with chemical indices such as polar group density gradient and chemical bond breaking rate into a high-dimensional microstructure state descriptor tensor. Through normalization, various features are made comparable at the same numerical scale, forming standardized inputs for subsequent time-series correlation mapping and spectral library construction.

[0049] S2.4: Perform temporal correlation mapping construction processing based on the continuous change trajectory of the microstructure state descriptor tensor on the time axis to establish a microstructure evolution path index throughout the entire life cycle and generate a microstructure evolution map library covering different aging stages.

[0050] For the microstructure state descriptor tensor, which integrates geometric topological and chemical composition information, a time axis index is set as the reference benchmark for continuous evolution trajectory. A temporal data structure is used to organize the descriptor tensors of each aging stage into a dynamic feature stream according to timestamp order. A sliding window mechanism is used to perform difference operations on the descriptor tensor between adjacent time points to extract the temporal gradient vector reflecting the rate of microstructure change. The evolution amplitude of the microstructure state between each stage is calculated using Euclidean distance or cosine similarity indices. A temporal correlation mapping process based on the Dynamic Time Warping (DTW) algorithm is employed to perform path comparison and synchronization calibration of the multidimensional feature trajectory of the descriptor tensor throughout its entire lifecycle, generating a set of continuous evolution paths characterizing the degradation trend of the microstructure. For key nodes in the multidimensional feature trajectory, cluster analysis and mutation detection are used to identify aging stages where significant transitions in the microstructure state occur, and these key nodes are labeled as path index anchors. By recursively encapsulating all aging stage descriptor tensors and their corresponding evolution path indices, a microstructure evolution map library covering the entire life cycle is integrated, enabling data archiving and index management of the dynamic evolution of material microstructures during the aging process. Through the aforementioned chain-like derivation process, the microstructure state descriptor tensor generated in the previous step, which integrates geometric topological information and chemical composition information, is transformed into a microstructure evolution map library with temporal correlation, key node annotation, and path indexing functions, realizing a dynamic, traceable, and computable data foundation for the material degradation process throughout its entire life cycle.

[0051] For example, in an accelerated aging scenario of a bio-based epoxy / nanocellulose reinforced environmentally friendly insulating composite cable support, 10 descriptor tensors for different aging stages were collected. Each tensor contains 128-dimensional spatial topological features and 32-dimensional local chemical composition indicators. The time axis index was set to 0 to 9, and a sliding window width of 2 was used to calculate the Euclidean distance for each pair of adjacent stages. ; Where D is the descriptor difference measure between the two stages, x i The difference between each feature component is represented by a DTW (Dynamic Transformation Wave) method to compare the trajectories of 10 sets of descriptors throughout their entire lifecycle, outputting a set of continuous evolution paths. K-means clustering analysis was used to identify stages 3 and 7 as key nodes for microstructural abrupt changes. These nodes were labeled and indexed in a microstructure evolution library, resulting in a microstructure evolution library outputting a dynamic tracking capability covering the entire lifecycle. In practical applications, this library accurately reflects the changing trends of multidimensional features such as phase regions, porosity, and polar groups within materials as they age, providing high-precision time-series input for subsequent insulation performance degradation prediction models and effectively improving the model's data consistency and generalization ability under complex environments.

[0052] S2.5: Based on the measured macroscopic insulation resistivity and dielectric loss factor data synchronously associated with the node attributes of each stage in the microstructure evolution map library, perform label alignment and annotation processing to form supervised learning sample pairs of input features and output targets, and generate a standardized microstructure evolution map library with a set of macroscopic insulation performance attenuation labels.

[0053] like Figure 2 As shown, step S3: Based on the topological connections in the microstructure evolution map library, local phase regions are abstracted into nodes, interface coupling strengths are abstracted into edges, and chemical and geometric features are abstracted into attributes, constructing a dynamic heterogeneous graph data model based on the current aging stage graph data model snapshot as input. Specifically, this includes: S3.1: Perform region clustering on the three-dimensional spatial coordinate data in the microstructure evolution map library to identify and delineate independent local phase region entities, and generate a set of local phase region nodes containing unique identifiers.

[0054] The three-dimensional spatial coordinate data refers to the set of voxel coordinates used to describe the precise position of each local phase region (such as reinforcing phase, matrix, and pores) within the material in three-dimensional space. This three-dimensional spatial coordinate data originates from the three-dimensional binary phase region mask matrix generated after image segmentation and feature extraction following high-resolution three-dimensional imaging (such as X-ray tomography) of the accelerated aging sample in the preceding step (S2.4). Each effective voxel point in this matrix carries its corresponding three-dimensional coordinates (x, y, z), collectively forming the basic data used to characterize the spatial topology of the microstructure.

[0055] Region clustering is performed on the 3D spatial coordinate data in the microstructure evolution map library to obtain the 3D binary phase region mask matrix and its corresponding spatial coordinate information as input. Using voxel connectivity and spatial neighborhood features, a spatial scan is performed on the mask matrix for each aging stage. Based on the Euclidean distance between voxels and physical attribute labels, a spatial adjacency matrix is ​​constructed. Clustering based on connected component extraction is applied to the spatial adjacency matrix to identify all non-overlapping local phase region entities with clearly defined boundaries. A unique identifier is assigned to each local phase region entity, and its basic geometric parameters such as 3D centroid coordinates, volume, and surface area are recorded. All local phase region entities are integrated into a node set according to their identifier numbers, and a structured data table of the node set is output, including node ID, spatial coordinates, phase region category, and preliminary geometric attributes. Through the above chained processing method, the binary phase region mask matrix generated in the previous step is transformed into a data structure containing independent local phase region node sets, realizing the standardized transformation of microstructure states into heterogeneous graph model input.

[0056] S3.2: Calculate the interface coupling strength based on the spatial adjacency relationship of the local phase region node set, determine the connection weight between nodes using the phase interface area and pore connectivity, and generate a heterogeneous graph edge set representing the physical contact state.

[0057] Based on the set of local phase region nodes and their spatial 3D coordinate data output from step S3.1, and relying on the physical spatial adjacency relationship between nodes as the basis for judgment, a calculation process for interface coupling strength is constructed. For each pair of spatially adjacent local phase region nodes, the set of boundary voxels in the binary phase region mask matrix is ​​retrieved, and the corresponding phase interface area index is extracted. The quantized porosity connectivity coefficient is used as an auxiliary parameter to correct and weight the potential physical contact state between nodes. The following weight calculation formula is adopted: ; Where w represents the connection weight between nodes, A represents the interface area between the two phase regions, and C represents the porosity connectivity coefficient. The above weight calculation is performed on all adjacent node pairs, mapping each weight result to heterogeneous graph edge attributes. Further deduplication and sparsification processing is applied to the edge set, removing weakly connected edges with weights below a preset threshold to improve the physical accuracy of the graph structure and the lightweight nature of the model. The remaining high-weight edge set is structured and encapsulated according to node identifiers to generate a heterogeneous graph edge set representing the physical contact state, and the interface coupling strength value and spatial location index corresponding to each edge are recorded simultaneously.

[0058] Through the above chain-like derivation process, the set of local phase region nodes and their spatial adjacency relationships are transformed into a set of heterogeneous graph edges with clear physical meanings. This realizes the transformation of microscopic structural topological relationships into computable graph structure inputs, providing a standardized and traceable data foundation for subsequent node attribute feature encoding and dynamic heterogeneous graph model assembly.

[0059] S3.3: Perform feature vector encoding processing on the fractal dimension of microcracks and the density gradient of polar groups, map chemical and geometric features into high-dimensional attribute vectors, and generate a node attribute feature matrix attached to the node set of local phase regions.

[0060] The set of identified local phase region nodes and their spatial adjacency relationships in the microstructure evolution map library are used as input. The multidimensional geometric feature vector set and polar group density gradient distribution map output from the previous step are retrieved as the basic data source for feature encoding in this step. A feature normalization method is used to standardize geometric indices such as microcrack fractal dimension, phase interface area density, and pore connectivity coefficient to ensure consistency of feature scales across different physical quantities. Principal component analysis (PCA) is used to calculate the covariance matrix of the normalized geometric feature matrix, extracting principal component vectors to reduce redundancy while retaining the core information of microstructure evolution. For the polar group density gradient distribution map, a spatial gradient field encoding method is used to embed the polar group concentration change rate corresponding to each local phase region node into the node attribute space in vector form, achieving multimodal fusion of chemical composition information and geometric topological information. A high-dimensional attribute vector concatenation operation is performed on the fused feature set, merging the geometric feature principal component vector and chemical gradient field vector corresponding to each node into a set of high-dimensional attribute vectors. The following attribute encoding formula is used: ; Where F is the node attribute feature vector, P is the geometric feature vector after principal component analysis, G is the spatial gradient field chemical feature vector, and α and β are weighting coefficients, dynamically adjusted according to the material aging stage and stress environment. The above attribute encoding process is sequentially performed on all local phase region nodes to generate a complete node attribute feature matrix, which is then attached to the local phase region node set, achieving a high-dimensional fusion expression of the physical and chemical states of each node in the heterogeneous graph structure. Through chain-like feature normalization, principal component dimensionality reduction, spatial gradient field encoding, high-dimensional attribute concatenation, and weighting coefficient adjustment, the multimodal raw indices output in the previous step are transformed into a high-dimensional node attribute feature matrix conforming to the input specifications of graph neural networks, achieving a refined representation of the microstructural state in the heterogeneous graph model.

[0061] S3.4: Based on the timestamp index of the current aging stage, perform topological assembly of the local phase region node set, heterogeneous graph edge set, and node attribute feature matrix to generate a snapshot of the dynamic heterogeneous graph data model that reflects the microstructure state at a specific moment.

[0062] Based on the extracted local phase region node set, heterogeneous graph edge set, and node attribute feature matrix, all node entities and associated edge weights at the corresponding time are selected as assembly objects according to the timestamp index of the current aging stage. For each local phase region node, index matching is performed using a unique identifier and spatial coordinate information, binding its attribute feature vector to the corresponding physical location to achieve synchronous mapping between node spatial state and chemical / geometric features. For the heterogeneous graph edge set, based on the interface coupling strength calculation results, the connection weights between each node are arranged in a structured manner according to the topological adjacency relationship, forming an edge weight matrix reflecting the physical contact state. Topological assembly processing is performed, based on the node set, traversing all edge sets step by step, connecting each edge to the corresponding node pair according to the adjacency list method, and writing its weight parameters into the dynamic graph structure. For the node attribute feature matrix, dimensionality consistency verification is performed, unifying and standardizing all high-dimensional feature vectors to the preset input specifications to ensure compatibility with subsequent neural network processing. Combining the above assembly process, a snapshot of a dynamic heterogeneous graph data model is generated, which includes a collaborative description of node spatial distribution, edge weighted connections, and attribute feature encoding. Its structure is as follows: each node is represented by spatial coordinates and a unique identifier; each edge is weighted by its coupling strength; and each attribute vector uses chemical / geometric indices as elements, achieving a full digital representation of the microstructure state at a specific aging time. Through topological assembly and structural specification processing, the local phase regions, interface coupling relationships, and multimodal attribute features output from the previous step are transformed into a snapshot of a dynamic heterogeneous graph data model with spatiotemporal consistency. This achieves a high-precision digital mapping of the microstructure evolution state on the time axis, providing a standardized data foundation for subsequent graph neural network inputs.

[0063] For example, in the dynamic heterogeneity diagram assembly process of aging samples of bio-based epoxy / nanocellulose composite cable supports, the aging stage timestamp t=120h is selected. The corresponding node set contains 128 local phase region entities, each entity having three-dimensional spatial coordinates (e.g., x=45, y=67, z=22) and a unique number (e.g., NID=34). The edge set consists of 512 effective connections, with interface coupling strength ranging from 0.15 to 0.92, calculated using the following weight formula: Where w is the edge weight, S is the interface area, and C is the porosity. For a typical connection (NID=12 and NID=47), S=2.8mm², C=0.35, then The node attribute feature matrix consists of 128×24 high-dimensional vectors. Each vector includes indicators such as microcrack fractal dimension (e.g., 1.42), polar group density gradient (e.g., 0.18), and local dielectric constant. After normalization, these vectors are standardized to 24 dimensions. Topology assembly traverses all 512 edges, writing their weights into an adjacency list to achieve 128×128 sparse matrix storage. The final output dynamic heterogeneous graph snapshot includes 128 nodes, 512 weighted edges, and a 128×24 attribute matrix. After standardization and verification, its structure fully meets the input requirements of lightweight graph neural networks. In practical applications, this snapshot accurately reflects the microstructural state during the 120-hour aging stage, providing highly reliable input data for subsequent performance degradation prediction.

[0064] S3.5: Based on the snapshot of the dynamic heterogeneous graph data model, perform graph structure standardization verification, check the consistency of node connectivity and attribute dimensions, and output a dynamic heterogeneous graph data model that conforms to the input specification of lightweight graph neural networks.

[0065] like Figure 3 As shown, step S4: Input the dynamic heterogeneous graph data model into the graph neural network architecture to perform forward propagation calculation, and output the set of predicted microstructure parameters for the next stage and a confidence interval vector representing the uncertainty of the model. Specifically, this includes: S4.1: Perform linear projection transformation on the node attribute feature matrix contained in the dynamic heterogeneous graph data model to map the high-dimensional chemical geometric features into low-dimensional latent space embedding vectors, generating an initial node state representation set with uniform dimensional specifications.

[0066] The snapshot of the input node attribute feature matrix of the standardized and validated dynamic heterogeneous graph data model is used as the processing object in this step. A linear projection transformation method is employed to perform weight matrix multiplication on the high-dimensional chemical and geometric feature vectors, mapping the original attribute matrix to a predefined low-dimensional latent space. Let the node attribute feature matrix be , the weight projection matrix be , and the low-dimensional embedding vector set be , then the linear projection formula is as follows: ; in, The input matrix is ​​the number of nodes multiplied by the number of high-dimensional features. The weight matrix is ​​calculated as high-dimensional feature count × low-dimensional embedding dimension. Normalization is performed on the projection results to ensure that all initial node state representation sets have uniform dimensionality and numerical distribution. Furthermore, batch regularization is applied to the low-dimensional embedding vector set to suppress outliers and improve model robustness. An activation function (such as ReLU or LeakyReLU) is used to perform a nonlinear transformation on the normalized embedding vectors to enhance the discriminative power of the node state representations. Finally, the processed initial node state representation set is output for subsequent information fusion and temporal inference in graph convolution aggregation operators.

[0067] By using linear projection, normalization, regularization, and activation function processing, the high-dimensional node attribute feature matrix generated in the previous step is transformed into a unified initial node state representation set that conforms to the lightweight graph neural network input specification, thereby realizing the effective expression of microstructure multimodal information in low-dimensional space and improving computational efficiency.

[0068] S4.2: Based on the topological adjacency relationship defined by the initial node state representation set and the heterogeneous graph edge set, a multi-hop neighbor information fusion operation is performed using the graph convolution aggregation operator to capture the interface coupling effect of local phase intervals and generate an enhanced node feature tensor that fuses spatial context information.

[0069] The initial set of node state representations after linear projection transformation and the topological adjacency relationships defined by the heterogeneous graph edge set are given structured input to form node embedding vectors and sparse adjacency matrices with uniform dimensionality. Using the graph convolution aggregation operator, multi-hop neighbor information fusion is performed on the state representation of each node according to the connection weights in the adjacency matrix, weighting and aggregating the features of spatially directly and indirectly connected local phase region nodes according to their physical coupling strength. The graph convolution calculation formula is as follows: ; Among them, h i Let x be the enhanced feature representation of the i-th node, σ be the nonlinear activation function, and x be the value of x. i Let N be the initial state vector of the i-th node. i Let A be the set of neighbors of the i-th node. ij The connection weight (interface coupling strength) between nodes i and j in the adjacency matrix is ​​1 / D. i D is the degree normalization factor. iLet Wi be the degree value of the i-th node, and W0 be a trainable weight matrix. The multi-hop information aggregation process described above is performed on all nodes to achieve spatial context feature fusion, ensuring that each node not only contains its own attributes but also incorporates information from its physical neighbors and indirect coupling regions. For multi-hop aggregation scenarios, graph convolution operations are iteratively executed up to a preset number of hops K. For example, when K=3, each node will integrate third-order neighbor information to capture cross-interface coupling effects during microstructure evolution. The fused enhanced node feature tensor is normalized and its dimensions are verified to ensure it conforms to the input specifications of subsequent time-series inference units. Through the multi-hop neighbor information fusion method of graph convolution aggregation operators, the initial node state representation set and the heterogeneous graph edge set are transformed into an enhanced node feature tensor containing local phase interval interface coupling effects and spatial context information, realizing a digital representation of the physical coupling mechanism of the microstructure evolution path.

[0070] S4.3: The enhanced node feature tensor is input to the gated cyclic update unit to perform time-series state deduction calculation to simulate the dynamic evolution logic of the microstructure on the time axis and generate a set of predicted microstructure parameters that characterize the micromorphological changes in the next stage. For example, in a scenario where the aging stage (t=120h) of a bio-based epoxy / nanocellulose composite cable support is tagged with 128 initial node state vectors, each with a dimension of 16. These vectors are topologically assembled into a 128×128 sparse adjacency matrix, where the weights of highly coupled edges are distributed in the range of 0.65 to 0.92. A graph convolution aggregation operator is used with a hop count of K=2, performing one first-order neighbor aggregation and one second-order neighbor aggregation on each node. During the first-order aggregation process, each node is processed according to the formula: ;

[0071] Among them, D i Update the gate vector for the degree (number of neighbors) of node i.

[0072] Update gate (z): ; σ is the Sigmoid activation function, W z To update the gate weight matrix, x t U is the input for the current enhanced node features. z Let h be the hidden state weight matrix. t-1 The state that was hidden in the previous moment.

[0073] Reset door (r): ; Among them, W r To reset the gate weight matrix, U r This is the hidden state weight matrix.

[0074] Candidate hidden state: ; Among them, h t ' represents the candidate hidden state vector, φ is the Tanh activation function, and W... h U h These are the weight matrix of the candidate input state and the cyclic weight matrix of the candidate state, respectively.

[0075] The final output hidden state is synthesized using the following formula: ; Batch processing is performed on the cyclic update process of all nodes along the time axis. The hidden state vector output in each round is mapped to the microstructure parameter space. A set of predicted microstructure parameters for the next stage is generated through fully connected layers or linear transformations, including core indicators such as phase interface area, pore connectivity, microcrack fractal dimension, and polar group density gradient. Through the above chain-like processing method, the enhanced node feature tensor is transformed into a set of predicted microstructure parameters characterizing the microstructure evolution trend of the next stage through time-series deduction by gated cyclic units, thus achieving high-precision simulation of the dynamic evolution logic of the microstructure.

[0076] S4.4: Using the set of predicted microstructure parameters, a probability density function fitting operation is performed through a parallel distributed estimation head to quantify the predicted fluctuation range of the data-driven model under complex stress conditions, and to generate a confidence interval vector characterizing the uncertainty of the model.

[0077] Distribution features are extracted from the set of predicted microstructure parameters output by the gated cyclic update unit to obtain the numerical distribution of each node and its attributes in the prediction space. A parallel distribution estimation head is used, taking the set of predicted microstructure parameters as input, to construct multiple probability density function fitting models, performing independent probability distribution fitting for different types of microstructure parameters (such as interface area, polar group density gradient, pore connectivity, etc.). For each type of microstructure parameter, the maximum likelihood estimation method is used to fit its probability density function based on the statistical relationship between historical data samples and current predicted values. For normally distributed parameters, the mean μ and standard deviation σ are calculated using the following formulas: ; ; Where, x iLet be the predicted value for the i-th sample, and n be the number of samples. For microstructure parameters exhibiting skewed or multimodal characteristics, a Gaussian mixture model (GMM) or kernel density estimation method is used to nonparametrically fit their probability density functions. The set of parameters for all fitted probability density functions is then structurally integrated to form a confidence interval vector containing descriptions of the uncertainties of various microstructure parameters. For the confidence interval calculation, based on a confidence level α (e.g., 0.95), the upper and lower bounds are determined using the following formula: ; ; x lower and x upper Z represents the lower and upper bounds of the confidence interval, respectively. α This represents the critical value of the standard normal distribution corresponding to the confidence level. The confidence interval calculation is performed on all nodes and attributes, and the results are encapsulated as a confidence interval vector, enabling a quantitative expression of the predicted fluctuation range of the data-driven model under complex stress conditions. Through parallel distribution estimation and probability density function fitting, the set of predicted microstructure parameters output from the previous step is transformed into a confidence interval vector characterizing the model's uncertainty, achieving a comprehensive assessment of the uncertainty in the predicted insulation performance degradation trend.

[0078] S4.5: Perform structured encapsulation processing based on the set of predicted microstructure parameters and the confidence interval vector to form a complete prediction result data package containing a description of deterministic evolution trend and uncertainty boundary, which serves as the trigger input source for the physical constraint reweighting mechanism.

[0079] Step S5: Based on the predicted values ​​of the microstructure parameters for the next stage, a deviation comparison is performed with the calculation results of the preset physical aging model. When the deviation exceeds a preset threshold, a physical constraint reweighting mechanism is triggered to generate a corrected microstructure evolution path that includes dynamically adjusted edge weights and node update rules. The physical aging model follows the Fick diffusion-controlled interface hydrolysis and Arrhenius-type thermally induced crosslinking fracture laws. Specifically, it includes: S5.1: Based on the predicted set of microstructure parameters for the next stage, the preset Fick diffusion-controlled interface hydrolysis physical equation and Arrhenius-type thermally induced crosslinking fracture physical equation are called for parallel solution to generate a physical mechanism benchmark vector containing the theoretical phase interface area evolution and the theoretical polar group density gradient change.

[0080] The set of predicted microstructure parameters for the next stage, output by the forward propagation of a lightweight graph neural network, is used as input. Core indicators to be verified include physical properties such as interfacial area and polar group density gradient. A pre-defined Fick diffusion-controlled interfacial hydrolysis physical equation is invoked to theoretically model the evolution of the interfacial area during the hydrolysis process, using the following formula: ; Among them, A t Let be the theoretical phase interface area at time t during the aging stage, A0 be the initial phase interface area, D be the diffusion coefficient, C0 be the initial hydrolysis reaction concentration, L0 be the interface thickness, and t be the time step. The parameters in the above formula are configured based on material experimental data and applied to the current aging stage.

[0081] The Arrhenius-type thermally induced crosslinking fracture physics equation was called in parallel to theoretically derive the change in polar group density gradient, using the following formula: ; Where, ρ t Let ρt be the theoretical polar group density gradient at time t during the aging stage, ρ0 be the initial polar group density gradient, and Et be the theoretical polar group density gradient at time t. a Here, R is the activation energy for thermally induced crosslinking fracture, R is the gas constant, and T0 is the ambient temperature. After assigning parameter values ​​to the above formula, the theoretical values ​​of all node properties are output within a parallel computing framework.

[0082] The calculation results of Fick diffusion and Arrhenius thermal induced model are encapsulated into physical mechanism benchmark vectors, which are then mapped one-to-one with the data-driven prediction value set according to attribute categories. Through parallel solution and structured encapsulation of the above physical equations, intrinsic physical logic modeling of the microstructure evolution trend in the next stage is achieved, providing a highly reliable reference standard for subsequent deviation comparison and physical constraint reweighting mechanism.

[0083] Through the above chain processing method, the data-driven prediction results of the previous step are transformed into a physical mechanism benchmark vector containing the theoretical phase interface area evolution and the theoretical polar group density gradient change, thus realizing the endogenous verification basis of the physical mechanism of the microstructure evolution path.

[0084] S5.2: Perform element-wise difference operation and normalization on the set of predicted microstructure parameters for the next stage and the physical mechanism reference vector to calculate the deviation scalar that characterizes the deviation of the data-driven prediction from the physical intrinsic law, and determine whether the deviation scalar exceeds the preset confidence threshold.

[0085] S5.3: When the deviation scalar exceeds a preset confidence threshold, a physical constraint reweighted mask matrix is ​​constructed based on the mapping relationship between the deviation direction and the physical aging rate constant, which includes the interface coupling strength correction coefficient and the node state update damping factor, in order to quantify the degree of physical violation at each topology connection.

[0086] The physical aging rate constant refers to the key rate parameter obtained analytically or by fitting from the mathematical expression of the physical aging model (such as a chemical reaction kinetic model based on the Arrhenius equation, or a differential equation model based on diffusion-reaction coupling).

[0087] The data stream of the physical constraint reweighted mask matrix is ​​constructed by using the element-wise difference results of the predicted microstructure parameters set for the next stage and the physical mechanism baseline vector, along with the normalized deviation scalar, as input. Based on the judgment result that the deviation scalar exceeds the confidence threshold, the deviation direction information is extracted, and combined with the physical aging rate constant, a mapping relationship is established to quantitatively analyze the degree of physical violation at each node and edge connection. The interface coupling strength correction coefficient is calculated using the following formula: ; Where C is the interface coupling strength correction coefficient, α is the physical constraint weighting factor, ΔP is the prediction deviation direction, and k is the physical aging rate constant. The damping factor for node state updates is calculated using the following formula: ; Where D is the node state update damping factor, and β is the node damping weight parameter. The above correction coefficients and damping factors are mapped to the edge weight matrix and node feature aggregation rules in the dynamic heterogeneous graph structure, respectively, and a mask generation operation is performed at each topological connection. The following structured mask matrix generation method is adopted: ; Among them, M ij This represents the elements of the physical constraint reweighting mask matrix between nodes i and j. The mask generation operation is performed on all edges and nodes in batches, transforming the physical violation degree at each topological connection into a quantifiable set of correction coefficients. The graph structure parameters are dynamically adjusted by multiplying the mask matrix with the original edge weights and node feature aggregation rules point-by-point. Through the chain-like derivation process described above, the deviation scalar and its direction information output from the previous step are transformed into a physical constraint reweighting mask matrix containing interface coupling strength correction coefficients and node state update damping factors, achieving endogenous verification of the physical mechanism of the microstructure evolution path and quantitative expression of the violation degree.

[0088] S5.4: Using the physical constraint reweighting mask matrix, perform point-by-point multiplication correction operations on the initial edge weight matrix and node feature aggregation rules in the graph neural network architecture to generate a set of physical perception graph structure parameters containing dynamically adjusted edge weight values ​​and corrected node update rules.

[0089] The initial edge weight matrix and node feature aggregation rules are the core computable parameters of the dynamic heterogeneous graph data model. The element values ​​of the initial edge weight matrix are directly derived from the initial quantization values ​​assigned when abstracting the interface coupling strength into edges in step S3. These values ​​are calculated based on the material properties or image features of adjacent phase regions, constituting the initial connection strength of the graph structure. The node feature aggregation rules are defined by the graph neural network backbone architecture constructed in step S4. Their rule parameters (such as the weight matrix in the weighted summation) are optimized through backpropagation during the forward propagation training of the neural network, determining how nodes in the graph aggregate their neighbor information to update their own state.

[0090] The physical constraint reweighted mask matrix is ​​input in a structured manner and serves as the basis for correcting the initial edge weight matrix and node feature aggregation rules in the graph neural network architecture. A pointwise multiplication method is used to perform element-wise multiplication between the physical constraint reweighted mask matrix and the original edge weight matrix, generating a dynamically adjusted set of edge weight values. For the aggregation operator parameters in the node feature aggregation rules, based on the numerical distribution of the damping factor update in the node state of the mask matrix, parameter weighting correction is performed to adjust the damping of aggregation rules for nodes with high physical violation levels. The edge weight correction process is mathematically expressed using the following formula: ; Among them, A i,j M represents the original edge weights. i,j For the elements of the physical constraint reweighted mask matrix, A i,j ′ represents the weights after correction. The following damping adjustment formula is used for the node feature aggregation rule: ; Among them, w K The damping adjustment coefficient of the K-th hop neighbor is given by the original aggregation operator parameters, which are determined by DampMask. K Given, after multiplication correction, we get w. K ′ represents the modified aggregation rule parameters. The dynamically adjusted set of edge weight values ​​and the modified set of node feature aggregation rule parameters are structurally integrated to output a graph structure parameter set containing physical perception attributes, achieving a high degree of consistency between the microstructure evolution path and the intrinsic degradation logic of environmentally friendly insulating composite materials.

[0091] By using a point-by-point multiplication correction method, the physical constraint reweighting mask matrix generated in the previous step is applied to the initial edge weight matrix and node feature aggregation rules in the graph neural network architecture. This effectively achieves dynamic adjustment of edge weight values ​​and node update rules, outputting a graph structure parameter set with physical mechanism perception capabilities, providing a unique and legitimate input for subsequent microstructure evolution path backpropagation reconstruction.

[0092] S5.5: Based on the physical perception graph structure parameter set, perform backpropagation reconstruction processing on the dynamic heterogeneous graph data model, and output the corrected microstructure evolution path that eliminates physical logic conflicts and conforms to the intrinsic degradation logic of environmentally friendly insulating composite materials, as the only valid input for subsequent performance mapping decoding.

[0093] Using the physical perception graph structure parameter set as input, the node set, edge weight matrix, and attribute feature tensor of the dynamic heterogeneous graph data model are loaded. The objective function of the backpropagation reconstruction process is set as eliminating physical logic conflicts and correcting the microstructure evolution path. A pointwise multiplication correction operation is performed on the initial edge weight matrix, mapping the interface coupling strength correction coefficient in the physical constraint reweighting mask matrix to the original edge weight values, forming a dynamically adjusted edge weight matrix. The node state update damping factor is applied to the node feature aggregation rule, and the state transition formula for each node is parameterized using the following recursive formula: ; Where λ is the node state update damping factor, h t+1 h is the node state vector at the current time. t-1 This is the node state vector from the previous time step.

[0094] The modified edge weight matrix and node feature aggregation rules are then used for topological assembly. All nodes, edges, and attributes are recombined according to the physical perception parameter set to generate a dynamic heterogeneous graph structure that conforms to the intrinsic degradation logic of environmentally friendly insulating composite materials. A chain-like backpropagation operation is performed on the newly assembled heterogeneous graph structure, employing a gradient descent optimization method with the mean square error between the physical mechanism baseline vector and the predicted microstructure parameter set as the loss function. ; x i Let x be the predicted value of the i-th microstructure parameter. des,i Let L be the baseline value for the i-th physical mechanism, N be the number of samples, and L be the number of samples. MSE This represents the mean squared error loss function value.

[0095] The gradient of the loss function is calculated, and iterative optimization is performed on the corrected edge weights and node state parameters to gradually converge the evolution path of the dynamic heterogeneous graph to the intrinsic degradation logic of the environmentally friendly insulating composite material. A consistency check is performed on the snapshot of the finally converged heterogeneous graph model, including topological connectivity, attribute dimension regularity, and physical constraint compliance. The corrected microstructure evolution path obtained through backpropagation reconstruction is used as the output and designated as the sole valid input to the subsequent performance mapping decoder.

[0096] By using the above-mentioned chain-like backpropagation reconstruction method, the physical perception graph structure parameter set is effectively applied to the dynamic heterogeneous graph data model, thereby eliminating physical logic conflicts in data-driven prediction and ensuring that the microstructure evolution path strictly conforms to the intrinsic laws of material aging, providing a highly reliable input basis for accurate decoding of insulation performance decay rate.

[0097] Step S6: Using the modified microstructure evolution path as feature input, a nonlinear mapping transformation is performed through a microstructure-performance correlation decoder to calculate the insulation performance attenuation rate feature vector characterizing the intrinsic degradation logic of the environmentally friendly insulating composite material. Specifically, this includes: S6.1: Based on the node attribute matrix and edge weight adjacency list contained in the modified microstructure evolution path, graph convolution feature aggregation processing is performed on the dynamic heterogeneous graph data model to generate a high-dimensional microstructure topology embedding vector that integrates local phase region chemical geometry features and interface coupling strength.

[0098] Based on the node attribute matrix and edge-weighted adjacency list contained in the corrected microstructure evolution path, a dynamic heterogeneous graph data model is selected as the processing object. Chemical features (such as polar group density gradients) and geometric features (such as microcrack fractal dimension) in the node attribute matrix are normalized to ensure that all features participate in subsequent aggregation operations at the same scale. Using the interface coupling strength information reflected in the edge-weighted adjacency list, the physical coupling relationship between nodes is mapped to a weighted adjacency matrix, which serves as the propagation coefficient in the graph convolution operation. A graph convolution aggregation operator is used to perform multi-hop neighbor feature aggregation operations on each local phase region node. The specific calculation method is as follows: ; Among them, h v Let w be the convolutional embedding vector of node v, N(v) be the set of neighbors of node v, and w be the embedding vector of node v. v,u d represents the edge weight (interface coupling strength). v x is the degree normalization factor for node v. uLet be the attribute feature vector of neighboring node u. After convolutional aggregation of all nodes, the results are concatenated to form a high-dimensional microstructure topological embedding vector set. Batch normalization is performed on this set to eliminate potential data offsets or scale imbalances between different phase intervals. Through the above chain derivation, the corrected microstructure evolution path is transformed into a high-dimensional microstructure topological embedding vector that integrates local phase region chemical geometry features and interface coupling strength, achieving an ordered transformation from microstructure state to the input feature space of the performance mapping decoder.

[0099] S6.2: Based on the attention weight matrix, perform a weighted aggregation operation on the high-dimensional microstructure topological embedding vector to extract a multi-scale degradation key feature tensor that reflects the spatiotemporal evolution of the fractal dimension and polar group density gradient of the microcrack. The attention weight matrix is ​​calculated using the following formula: ; Where Q is the query matrix, K is the key matrix, and d k As the key vector dimension, the query matrix Q and the key matrix K are obtained by performing different linear projection transformations on the high-dimensional microstructure topology embedding vector, and the corresponding projection weight matrices are learned during model training.

[0100] Based on the aforementioned attention weight matrix A, a weighted aggregation operation is performed on the high-dimensional microstructure topological embedding vector to extract a multi-scale degradation key feature tensor that reflects the spatiotemporal evolution of the fractal dimension of microcracks and the density gradient of polar groups. The outputs of all attention heads are concatenated or averaged to achieve a fusion expression of multi-scale spatial dependencies. The fused multi-scale degradation key feature tensor is used as input to the subsequent performance correlation decoder to achieve a global characterization of the material's microscopic state evolution to macroscopic performance degradation. Through a multi-head self-attention mechanism, the high-dimensional microstructure topological embedding vector from the previous step is transformed into a multi-scale degradation key feature tensor with spatiotemporal evolution and global dependency expression capabilities, enabling accurate extraction of the core driving factors of the insulation performance degradation process.

[0101] For example, in the aging scenario of environmentally friendly insulated composite material cable supports, the high-dimensional microstructure topology embedding vector is set to 128 dimensions. Each node represents a local phase region, and the node attributes include 10 physical indicators such as polar group density gradient, microcrack fractal dimension, and pore connectivity. An 8-head self-attention mechanism is adopted, with each head generating query, key, and value vectors through a 128×16 linear projection matrix and calculating attention weights accordingly. The polar group density gradient exhibits significant spatial heterogeneity among different nodes, and the microcrack fractal dimension shows a non-linear growth trend over time. In practical applications, the attention level between each node and other nodes is calculated using the above formula, and all node attributes are weighted and aggregated to finally output a multi-scale degradation key feature tensor with a length of 128×8=1024 dimensions. On the test set, this feature tensor, when used for training the subsequent insulation performance degradation rate prediction model, significantly improves the model's ability to consistently distinguish between material aging paths and performance degradation trends under complex environments compared to traditional single-scale convolution methods, thus achieving accurate performance prediction driven by the intrinsic degradation logic of environmentally friendly insulating composite materials.

[0102] S6.3: Construct a predefined microstructure-performance correlation decoder architecture based on the multi-scale degradation key feature tensor, and use a deep fully connected neural network layer to perform nonlinear function fitting operations to establish an implicit mapping function model from the microscopic topological state space to the macroscopic insulation resistivity change rate.

[0103] Using a multi-scale degradation key feature tensor as input, a predefined microstructure-performance correlation decoder architecture is employed to construct the mapping model. Input layer parameters are initialized on the high-dimensional microstructure topological embedding vector, with initial values ​​for the weight matrix set based on the statistical distribution of node attributes and edge weights. The multi-scale degradation key feature tensor is input to the decoder backbone network, sequentially passing through several deep fully connected neural network layers for feature transformation. Each layer employs a non-linear activation function such as ReLU to enhance the model's expressive power. In this embodiment, a deep fully connected neural network with 3-5 hidden layers is preferred. For each hidden layer, feature normalization and regularization are performed to suppress overfitting and improve generalization ability. The mean squared error loss function is used as the training objective, and the predicted results of the output layer are compared with the measured macroscopic insulation resistivity change rate under supervision. The specific loss function expression is as follows: ; Where y is the predicted rate of change of insulation resistivity output by the model, y desHere, represents the measured label value, and n represents the number of samples. Based on backpropagation, the weight matrix of the fully connected layer is updated using gradients, and the Adam optimizer is used to dynamically adjust the learning rate, improving convergence speed and stability. Through the above processing method, multi-scale microstructure features are mapped to the macroscopic insulation performance decay rate, realizing implicit nonlinear function modeling from the microscopic topological state space to macroscopic performance indicators.

[0104] S6.4: Apply the implicit mapping function model to the feature data stream of the current aging stage, and perform a forward inference calculation process to solve for the characteristic vector of insulation performance decay rate that quantifies the intrinsic degradation logic of environmentally friendly insulating composite materials under electro-thermal-humidity-mechanical composite stress.

[0105] Using a multi-scale degradation key feature tensor as input, a high-dimensional microstructure topological embedding vector set for the current aging stage, along with the node attribute matrix, edge weight adjacency list, and global dependency features of the corrected microstructure evolution path, are loaded. The multi-scale degradation key feature tensor is then input into a predefined microstructure-performance correlation decoder architecture. Through parameter initialization, the weight matrices and bias terms of each layer of the fully connected neural network are set, and normalization is performed based on the statistical distribution of node attributes. The input feature stream is sequentially propagated through deep fully connected neural network layers, with each layer employing a non-linear activation function (such as ReLU) to enhance the model's ability to map complex microscopic state spaces to macroscopic performance indicators. Batch standardization and regularization are performed on the hidden layer outputs to suppress overfitting and improve the model's generalization ability. A mean squared error loss function is used to supervise the comparison between the output layer's predicted values ​​and the measured rate of change of insulation resistivity. The specific loss function is expressed as follows: ; Where y is the predicted insulation performance degradation rate output by the model, y des is the measured label value, and n is the number of samples.

[0106] Chained forward inference is performed on all neural network layers to progressively map the multi-scale degradation key feature tensor to the implicit insulation performance decay rate space, achieving a nonlinear transformation from microscopic topological states to macroscopic performance indicators. The final output insulation performance decay rate feature vector undergoes post-processing, including normalization, outlier removal, and confidence interval estimation, to ensure the prediction results possess physical rationality and statistical stability. Through the aforementioned chained derivation process, the multi-scale degradation key feature tensor obtained in the previous step is transformed into an insulation performance decay rate feature vector characterizing the intrinsic degradation logic of environmentally friendly insulating composite materials under electro-thermal-humidity-mechanical composite stress, achieving accurate decoding of macroscopic performance driven by microstructure states.

[0107] S6.5: Based on the insulation performance decay rate feature vector and combined with time integration processing, perform cumulative effect deduction processing to generate dielectric loss factor trend curves and insulation resistivity prediction sequences for characterizing the remaining life and safety status of environmentally friendly insulated composite material cable supports.

[0108] Based on the insulation performance decay rate feature vector and the timestamp index of the current aging stage, the input data is set as the set of insulation performance decay rate feature vectors output by a deep fully connected neural network decoder after the modified microstructure evolution path. Cumulative effect deduction processing is performed on this feature vector, and time integration processing is used to perform a continuous integral transformation on the insulation performance decay rate, generating a prediction sequence reflecting the trend of insulation resistivity change throughout the material's entire lifespan. Specifically, the following integral formula is used: ; Among them, R t R0 is the predicted insulation resistivity at time t during the aging stage, and S0 is the initial insulation resistivity. t Let be the characteristic vector of insulation performance decay rate, and [t0,t] be the integration interval.

[0109] The above integral results are numerically discretized to transform the continuous integral curve into a set of phased sampling points for subsequent trend analysis. Combining the frequency domain response conversion mechanism in dielectric relaxation theory, the real and imaginary parts of the complex dielectric constant are decoupled and calculated for the insulation performance decay rate eigenvector, using the following formula: ; ; Where ε′(t) is the predicted real part of the dielectric constant at time t, ε′′(t) is the predicted dielectric loss factor, S(t) is the characteristic vector of insulation performance decay rate, and ω0 is the main frequency.

[0110] The aforementioned set of phased sampling points for dielectric loss factor and the predicted insulation resistivity sequence are smoothed using cubic spline interpolation to generate a continuous and smooth trend curve for dielectric loss factor and a time series prediction result for insulation resistivity. Based on the failure threshold parameters defined in the power equipment life assessment standard, threshold determination and remaining life quantification are performed on the predicted curve, extracting the breakdown critical moment and calculating the remaining life value. Combining this with a preset safe operating range, the remaining life assessment result is transformed into a safety status assessment report through multi-level risk level discrimination rule matching.

[0111] Through the above chain-like derivation process, the insulation performance decay rate feature vector output in the previous step is transformed into a dielectric loss factor trend curve and insulation resistivity prediction sequence that characterize the remaining life and safety status of environmentally friendly insulated composite cable brackets. This achieves a high-precision mapping from microstructure evolution path to macroscopic performance indicators, providing a foundation for subsequent adaptive updates of the model driven by field monitoring data.

[0112] 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.

[0113] 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.

[0114] 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 performance prediction method for cable supports based on environmentally friendly insulating composite materials, characterized in that, include: S1: For the aging scenario of environmentally friendly insulated composite material cable brackets under composite stress environment, the accelerated aging process is collected synchronously to obtain the original multimodal characterization dataset. S2: Based on the original multimodal characterization dataset, perform image segmentation and topological analysis to extract phase interface area, pore connectivity, microcrack fractal dimension and polar group density gradient, and generate a microstructure evolution map library. S3: Based on the topological connections in the microstructure evolution map library, local phase regions are abstracted into nodes, interface coupling strength is abstracted into edges, and chemical and geometric features are abstracted into attributes to construct a dynamic heterogeneous graph data model with the current aging stage graph data model snapshot as input. S4: Input the dynamic heterogeneous graph data model into the graph neural network architecture to perform forward propagation calculation and output the set of predicted microstructure parameters for the next stage; S5: Based on the predicted set of microstructure parameters for the next stage, the deviation is compared with the calculation results of the preset physical aging model. When the deviation exceeds the preset threshold, the physical constraint reweighting mechanism is triggered to generate the corrected microstructure evolution path. S6: Using the modified microstructure evolution path as feature input, a nonlinear mapping transformation is performed through the microstructure-performance correlation decoder to calculate the insulation performance attenuation rate feature vector characterizing the intrinsic degradation logic of the environmentally friendly insulating composite material.

2. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 1, characterized in that, The original multimodal characterization dataset contains three-dimensional microstructure sequence images and local chemical bond evolution spectral data.

3. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 2, characterized in that, The accelerated aging process is simultaneously acquired, including: using high-resolution X-ray tomography to obtain the three-dimensional microstructure sequence images of the environmentally friendly insulated composite material cable bracket during the aging process; and using in-situ Raman spectroscopy to obtain the local chemical bond evolution spectral data of the environmentally friendly insulated composite material cable bracket during the aging process.

4. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 1, characterized in that, Step S3 specifically includes: The three-dimensional spatial coordinate data in the microstructure evolution map library are subjected to region clustering to identify and delineate independent local phase region entities and generate a set of local phase region nodes. Based on the spatial adjacency relationship of the local phase region node set, the interface coupling strength is calculated, and the connection weight between nodes is determined by the phase interface area and the porosity connectivity, thereby generating a heterogeneous graph edge set. Feature vector encoding is performed on the fractal dimension of the microcrack and the density gradient of polar groups to map the chemical and geometric features into high-dimensional attribute vectors, generating a node attribute feature matrix attached to the node set of the local phase region. Based on the timestamp index of the current aging stage, the local phase region node set, heterogeneous graph edge set, and node attribute feature matrix are topologically assembled to generate a snapshot of the dynamic heterogeneous graph data model. Based on the snapshot of the dynamic heterogeneous graph data model, the graph structure standardization verification is performed to check the consistency of node connectivity and attribute dimensions, and the dynamic heterogeneous graph data model is output.

5. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 4, characterized in that, Step S4 specifically includes: Perform linear projection transformation on the node attribute feature matrix contained in the dynamic heterogeneous graph data model to generate an initial node state representation set; Based on the topological adjacency relationship defined by the initial node state representation set and the heterogeneous graph edge set, a multi-hop neighbor information fusion operation is performed using the graph convolution aggregation operator to generate an enhanced node feature tensor. Temporal state deduction calculations are performed on the enhanced node feature tensor to generate the set of predicted values ​​for the microstructure parameters; The predicted set of microstructure parameters is used to perform probability density function fitting operations through a parallel distribution estimation head to generate a confidence interval vector. Based on the predicted set of microstructure parameters and the confidence interval vector, a structured encapsulation process is performed to form a complete prediction result data packet.

6. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 5, characterized in that, The complete prediction result data package contains deterministic evolution trends and descriptions of uncertainty boundaries.

7. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 5, characterized in that, The revised microstructure evolution path includes dynamically adjusted edge weights and node update rules.

8. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 5, characterized in that, Step S5 specifically includes: Based on the predicted set of microstructure parameters for the next stage, the preset Fick diffusion-controlled interface hydrolysis physical equation and Arrhenius-type thermally induced crosslinking fracture physical equation are called for parallel solution to generate a physical mechanism benchmark vector. The next stage microstructure parameter prediction set and the physical mechanism reference vector are subjected to element-wise difference operation and normalization to calculate the deviation scalar, and it is determined whether the deviation scalar exceeds the preset confidence threshold. When the deviation scalar exceeds a preset confidence threshold, a physical constraint reweighted mask matrix is ​​constructed based on the mapping relationship between the deviation direction and the physical aging rate constant. The physical constraint reweighting mask matrix is ​​used to perform point-by-point multiplication correction operations on the initial edge weight matrix and node feature aggregation rules in the graph neural network architecture to generate a set of physical perception graph structure parameters. Based on the set of physical perception graph structural parameters, the dynamic heterogeneous graph data model is subjected to backpropagation reconstruction processing, and the corrected microstructure evolution path that eliminates physical logic conflicts and conforms to the intrinsic degradation logic of environmentally friendly insulating composite materials is output.

9. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 1, characterized in that, The physical mechanism reference vector includes the theoretical phase interface area evolution and the theoretical polar group density gradient change.

10. The performance prediction method for cable supports based on environmentally friendly insulating composite materials according to claim 1, characterized in that, The physical constraint reweighted mask matrix includes interface coupling strength correction coefficients and node state update damping factors.

Images