Ultrahigh-voltage cable insulation detection method and ultrahigh-voltage cable maintenance method

Abstract

The invention discloses an ultrahigh-voltage cable insulation detection method and an ultrahigh-voltage cable maintenance method, and relates to the technical field of crosslinked polyethylene cable manufacturing, and the method comprises the steps: carrying out the in-situ and non-contact terahertz time-domain spectrum three-dimensional full-field scanning of a produced cable, obtaining the time-domain waveform of each space point of a cable insulation layer in the cable, and carrying out the in-situ and non-contact terahertz time-domain spectrum three-dimensional full-field scanning; according to the method, the insulation microstructure is comprehensively perceived through non-destructive terahertz scanning, the manufacturing data and the molecular dynamics mechanism are combined, the crossing from microdefect positioning to long-term performance prediction is realized by using the mixed AI model, and the method has the advantages that the method is simple and convenient, and the cost is low. Therefore, after the production of the cable is completed, in-situ and quantitative evaluation and life prediction can be carried out on the insulation reliability of the cable, a closed loop for guiding process optimization is formed, and the fundamental transformation from passive inspection to active predictive quality control is finally realized.

Description

Technical Field

[0001] This invention relates to the field of cross-linked polyethylene cable manufacturing technology, specifically to methods for testing insulation of ultra-high voltage cables and methods for maintaining ultra-high voltage cables. Background Technology

[0002] Cross-linked polyethylene (XLPE) cables have become core equipment in the field of ultra-high voltage power transmission due to their excellent electrical and mechanical properties. The long-term operational reliability of the cable fundamentally depends on the quality of its insulation layer; therefore, strict quality testing of the insulation layer is essential after production.

[0003] Currently, the insulation testing methods commonly used in the industry are mainly based on electrical performance testing, such as partial discharge detection, power frequency withstand voltage testing, or ultra-low frequency dielectric loss testing. These methods assess the electrical strength of insulation or detect macroscopic defects by applying high voltage, playing an important role in engineering. However, these traditional methods have a fundamental limitation: they are essentially "post-hoc," pass / fail tests based on "destructive stress." Their detection signals reflect the instantaneous performance of insulation under the test voltage, or the response to macroscopic defects (such as air gaps and impurities) that have developed to a certain scale. They cannot non-destructively and directly perceive the microstructural state that determines the long-term aging performance of insulation (such as the uniformity of molecular chain crosslinking, micron-level micropore distribution, early impurity dispersion, etc.), let alone assess the evolution trend of these microstructures under the combined effects of electric and thermal fields over decades.

[0004] Therefore, a prominent technical problem facing existing technologies is how to detect existing microscopic insulation defects after the cable is manufactured without damage and with comprehensive scanning, and how to quantitatively assess and predict its long-term reliability (or remaining electrical life) under ultra-high voltage operating conditions. Summary of the Invention

[0005] To overcome the shortcomings mentioned above, this invention aims to provide a technical solution that can solve the aforementioned problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The insulation testing method for ultra-high voltage cables includes the following steps: S100: Performs in-situ, non-contact terahertz time-domain spectral three-dimensional full-field scanning on the finished cable to obtain the time-domain waveforms of each spatial point in the cable insulation layer and calculates and generates a four-dimensional database containing spatial coordinates and corresponding dielectric spectra. S200: Synchronously collect key process parameters of the entire cable production process, and digitally associate the key process parameters with the corresponding cable segment spatial coordinates in S100 to generate associated process parameters. S300: Based on molecular dynamics simulation, a mechanistic knowledge base is constructed to understand the microscopic defect structure of cross-linked polyethylene under different process deviations and its evolution under electro-thermal stress, and key mechanistic characteristic parameters are extracted. S400: Building Mechanism-Data Hybrid Driven Artificial Intelligence Models; The dielectric spectrum of S100, the associated process parameters of S200, and the key mechanism feature parameters matched from S300 are used as mixed inputs, and the micro-defect type, severity level, and long-term performance prediction index are used as outputs to train the mechanism-data hybrid driven artificial intelligence model. S500: Input the dielectric spectrum and associated process parameters of the cable under test into the trained mechanism-data hybrid driven artificial intelligence model to obtain the three-dimensional defect distribution and long-term performance prediction results of the cable insulation layer, and generate a visual quality assessment report accordingly. S600: Based on the correlation analysis between the three-dimensional defect distribution and related process parameters in the visualized quality assessment report, generate production process optimization feedback suggestions.

[0007] As a further aspect of the present invention: step S100 specifically includes: S110: Pass the completed cable axially through a ring-shaped terahertz time-domain spectroscopy scanning device at a constant speed. S120: Control the transmitting and receiving units of the terahertz time-domain spectral scanning device to rotate at a constant speed along the circumference of the cable and move synchronously along the axial direction of the cable, so that the terahertz pulse beam scans the cable insulation layer point by point in a spiral trajectory until it covers the entire cable. S130: At each scanning point, record the time-domain waveform of the transmitted or reflected terahertz pulse and separate the main pulse signal characterizing the bulk properties of the insulating layer. S140: Perform Fourier transform on the main pulse signal of each scanning point to obtain the frequency domain spectrum of the corresponding scanning point, and calculate the real part ε'(ω) and imaginary part ε''(ω) of the complex dielectric constant of the corresponding scanning point at different frequencies to form the dielectric spectrum of the corresponding scanning point; S150: The dielectric spectrum of each scan point is associated with and stored with spatial coordinates, including axial position, radial depth and circumferential angle, thereby constructing the four-dimensional database.

[0008] As a further aspect of the present invention: step S200 specifically includes: S210: For the cable segment to be evaluated, from the start to the end of its production, a set of predefined key process parameters is collected in real time from the production control system. S220: Arrange and integrate the set of key process parameters in chronological order of production time, and associate them with the unique production identifier of the cable segment to generate a structured digital model of process parameters. S230: Based on the historical traction speed record of the cable segment, calculate the axial position of the insulation layer at each point in time on the production line. S240: Based on the correspondence calculated in S230, establish a mapping relationship between the parameter set at each time point in the digital model of the process parameters and all scanning points at the corresponding axial position (Z) in the four-dimensional database constructed in S150, complete the digital association, and generate the associated process parameters.

[0009] As a further aspect of the present invention: step S300 specifically includes: S310: Establish an initial atomic-level molecular dynamics model of cross-linked polyethylene, using COMPASS or a similar force field, and set the density, degree of cross-linking, and chain length distribution consistent with the cable insulation material. S320: In the initial model, different preset process defects are introduced by changing the simulation conditions to construct multiple micro-defect structure models; The preset process defects include: insufficient local cross-linking degree, embedding of specific impurity atoms or clusters, and crystallinity gradient distribution caused by non-uniform temperature field. S330: For each of the aforementioned micro-defect structure models, apply constant electrical and thermal stresses corresponding to the actual operating conditions of the cable in the simulation environment, perform long-term molecular dynamics simulation, and record the evolution process of the microstructure of the micro-defect structure model over simulation time. S340: Extract key mechanism parameters characterizing defect dynamics from the evolution process. These key mechanism parameters include: the size and distribution rate of free volume pores, the anisotropic parameters of molecular chain segment mobility, and the evolution trend of trap charge density and energy level distribution. S350: The initial conditions of each simulation case, the corresponding microscopic defect structure model, and the key mechanism feature parameters evolved from it are associated and stored to construct the mechanism knowledge base.

[0010] As a further aspect of the present invention: step S400 specifically includes: S410: Construct a mechanism-data hybrid artificial intelligence model with multimodal input branches, wherein the mechanism-data hybrid artificial intelligence model includes: The first input branch is used to receive normalized dielectric spectrum data from a specific spatial point of S100; The second input branch is used to receive a vectorized set of complete critical process parameters associated with the spatial point from S200; The third input branch is used to receive key mechanism feature parameters from S300 that best match the set of key process parameters, and the key mechanism feature parameters are embedded as physical constraint terms. S420: Prepare the training dataset, which uses the S100, S200 and S300 data of historically produced cables as input features, and the actual defect types, locations and performance degradation data of the corresponding cable segments obtained in accelerated aging tests or long-term operation as supervision labels. S430: Train the mechanism-data hybrid driven artificial intelligence model. Its total loss function consists of three parts: the main task loss for defect classification and localization, the regression task loss for long-term performance prediction, and a mechanism consistency regularization loss that specifically penalizes the prediction results of the mechanism-data hybrid driven artificial intelligence model for violating the physical laws revealed by the key mechanism feature parameters in S300. S440: After training is completed, the parameters of the mechanism-data hybrid-driven artificial intelligence model are solidified, so that the mechanism-data hybrid-driven artificial intelligence model has the following functions: based on the input multimodal data, it synchronously outputs the micro-defect classification and level corresponding to the current spatial point, as well as the performance degradation curve and remaining electrical lifetime prediction value under the preset operating conditions at that point.

[0011] As a further aspect of the present invention: step S500 specifically includes: S510: The four-dimensional database data of the cable under test obtained in S100, and the associated process parameters obtained and spatiotemporally correlated in S200, are synchronously input into the mechanism-data hybrid driven artificial intelligence model trained and solidified in S400. S520: The mechanism-data hybrid driven artificial intelligence model traverses and processes all scan points in spatial coordinate order, and outputs the micro-defect classification and grade label of each scan point in parallel, as well as the corresponding predicted remaining lifetime value array. The predicted remaining lifetime value array represents the failure probability under different operating years. S530: Based on the micro-defect classification and grade labeling of all scanning points, a three-dimensional interpolation algorithm is used to generate a three-dimensional defect type distribution cloud map and defect grade distribution cloud map of the cable insulation layer; S540: Based on the array of predicted remaining lifetime values ​​of all scan points, extract the failure probability under the same service life threshold, generate a three-dimensional heat map of the failure probability of the cable insulation layer at different service time points; and calculate the minimum predicted remaining lifetime value of the entire cable. S550: Combining a three-dimensional defect distribution cloud map and a failure probability heat map, the cable is comprehensively rated according to a preset quality grading rule. The quality grading rule is based at least on the overall defect density and the minimum predicted remaining life value, and divides the cable into at least four levels: superior, qualified, reworkable and unqualified. S560: Automatically generates a visualized quality assessment report that includes the aforementioned 3D distribution cloud map, heat map, quality grade, list of specific defect location coordinates, and maintenance recommendations.

[0012] As a further aspect of the present invention: step S600 specifically includes: S610: Analyze the visual quality assessment report generated by S500 and extract the spatial location information, type, and level of all defects marked therein; S620: Based on the mapping relationship established in S200, map each defect location back to its corresponding key process parameter that takes effect when producing the insulation layer at that location; S630: Using statistical correlation analysis or machine learning attribution algorithms, compare the set of key process parameters for all defect points with the set of key process parameters for defect-free or excellent points to identify at least one key process parameter and its deviation direction that is significantly related to a specific defect type and / or low predicted lifetime. S640: Based on the identification results of S630, for each key process parameter, combined with its physical allowable range and the corresponding defect evolution knowledge in the S300 mechanistic knowledge base, a list of process optimization suggestions is generated, which includes the specific parameter adjustment amount, adjustment priority and expected quality improvement after adjustment. S650: The process optimization suggestion list is pushed to the corresponding execution unit of the production control system in real time in the form of structured instructions to guide the closed-loop control of the production process of the next batch or the current online cables.

[0013] As a further aspect of the present invention: in step S410, the key mechanistic feature parameters received by the third input branch are structurally fused into the forward propagation process of the hybrid neural network model through at least one of the following methods: Method 1: The key mechanism feature parameters are input into a dedicated auxiliary decoder network, which attempts to reconstruct these key mechanism feature parameters from the hidden features of the intermediate layer of the mechanism-data hybrid driven artificial intelligence model, thereby forcing the main network to learn feature representations consistent with physical laws; Method 2: The key mechanism feature parameters are used to construct a prior physical rule sub-network. The output of the physical rule sub-network is weighted and fused or gated with the intermediate features of the main network to selectively strengthen signal paths that conform to physical laws at different stages of inference. Method 3: The key mechanism feature parameters are used as a set of ideal feature vectors to initialize or apply bias to some weight matrices of the core layer of the mechanism-data hybrid driven artificial intelligence model, so that the mechanism-data hybrid driven artificial intelligence model has a reasoning tendency that conforms to physical laws from the beginning of training.

[0014] As a further aspect of the present invention: in step S410, the feature vector Fd obtained after encoding the data from the first and second input branches, and the key mechanism feature parameter vector Fm from the third input branch, are fused through a mechanism-guided fusion module to generate a hybrid feature vector Fhybrid; The mechanism-guided fusion module performs the following operations: S411: Input the key mechanism feature parameter vector Fm into a parameterized network to generate a set of attention weight vectors Wa or a feature transformation matrix Mm; S412: Using Wa or Mm, the feature vector Fd is weighted, filtered, or projected to enhance the feature dimensions related to the current physical mechanism in the output feature Fhybrid. Furthermore, the total loss function Ltotal in step S430 is specifically as follows: Ltotal=α(λLcls+(1 λ)Lreg_life)+βLmech+γLreg Where: Lcls is the cross-entropy loss for the defect classification task; Lreg_life is the smoothing L1 loss for the lifetime prediction task; Lmech is the mechanism consistency regularization loss, the calculation of which depends on the deviation between the intermediate or final result predicted by the model and the physical relationship implied by the mechanism feature parameter Fm; Lreg is the weight regularization loss; α, β, γ, λ are preset positive weight coefficients.

[0015] A method for maintaining an ultra-high voltage cable involves first inspecting the current cable segment using the ultra-high voltage cable insulation testing method described in any one of claims 1 to 9, and then maintaining the cable segment.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention comprehensively perceives the microstructure of insulation through non-destructive terahertz scanning, and combines manufacturing data with molecular dynamics mechanisms. It utilizes a hybrid AI model to achieve a leap from locating microscopic defects to predicting long-term performance. This allows for in-situ, quantitative assessment of insulation reliability and life prediction after cable production is completed, forming a closed loop that guides process optimization. Ultimately, it achieves a fundamental shift from passive inspection to proactive predictive quality control. Attached Figure Description

[0017] Figure 1 This is a flowchart of steps S100-S600 in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 The insulation testing method for ultra-high voltage cables includes the following steps: S100: Performs in-situ, non-contact terahertz time-domain spectral three-dimensional full-field scanning on the finished cable to obtain the time-domain waveforms of each spatial point in the cable insulation layer and calculates and generates a four-dimensional database containing spatial coordinates and corresponding dielectric spectra. S200: Synchronously collect key process parameters of the entire cable production process, and digitally associate the key process parameters with the corresponding cable segment spatial coordinates in S100 to generate associated process parameters. S300: Based on molecular dynamics simulation, a mechanistic knowledge base is constructed to understand the microscopic defect structure of cross-linked polyethylene under different process deviations and its evolution under electro-thermal stress, and key mechanistic characteristic parameters are extracted. S400: Building Mechanism-Data Hybrid Driven Artificial Intelligence Models; The dielectric spectrum of S100, the associated process parameters of S200, and the key mechanism feature parameters matched from S300 are used as mixed inputs, and the micro-defect type, severity level, and long-term performance prediction index are used as outputs to train the mechanism-data hybrid driven artificial intelligence model. S500: Input the dielectric spectrum and associated process parameters of the cable under test into the trained mechanism-data hybrid driven artificial intelligence model to obtain the three-dimensional defect distribution and long-term performance prediction results of the cable insulation layer, and generate a visual quality assessment report accordingly. S600: Based on the correlation analysis between the three-dimensional defect distribution and related process parameters in the visualized quality assessment report, generate production process optimization feedback suggestions; As described in the published patent CN114739946A, "A Terahertz Imaging Device and Method for Non-destructive Testing of Cables", terahertz imaging is a common choice in cable testing-related technologies. The non-contact terahertz time-domain spectroscopy three-dimensional full-field scanning method adopted in the S100 step enables non-destructive, full-domain, and quantitative characterization of the micro-dielectric properties of cross-linked polyethylene insulation layer. This overcomes the limitations of traditional electrical testing methods, which are highly destructive and rely on sampling. It can keenly detect early micro-defects (such as micropores, impurity agglomeration, and uneven cross-linking) that have not yet caused obvious electrical signal anomalies. Secondly, through the S200 and S300 steps, the insulation performance is creatively integrated with the defect evolution mechanism revealed by the whole process parameters and molecular dynamics simulation, which makes the detection go from simple state judgment to a deep understanding of the causal chain of "manufacturing process-microstructure-long-term performance". Furthermore, by leveraging the mechanism-data hybrid AI model built through the S400 steps, it was the first time that individualized and quantitative predictions of the long-term aging behavior and remaining electrical life of cables under specific operating stresses could be made when the cables were laid down, transforming quality control from "passive post-event inspection" to "proactive pre-event prediction". Finally, through the S500 and S600 steps, a complete closed loop integrating "intelligent online assessment, visual grading, and precise process feedback" is formed. This not only guides the precise handling of current products (such as graded use) but also continuously optimizes production processes, fundamentally improving the long-term reliability of batch products, thereby significantly reducing the total life cycle operation and maintenance costs while ensuring grid security. This invention comprehensively perceives the microstructure of insulation through non-destructive terahertz scanning, and combines manufacturing data with molecular dynamics mechanisms. It utilizes a hybrid AI model to achieve a leap from locating microscopic defects to predicting long-term performance. This allows for in-situ, quantitative assessment of insulation reliability and life prediction after cable production is completed, forming a closed loop that guides process optimization. Ultimately, it achieves a fundamental shift from passive inspection to proactive predictive quality control.

[0020] In this embodiment of the invention, step S100 specifically includes: S110: Pass the completed cable axially through a ring-shaped terahertz time-domain spectroscopy scanning device at a constant speed. S120: Control the transmitting and receiving units of the terahertz time-domain spectral scanning device to rotate at a constant speed along the circumference of the cable and move synchronously along the axial direction of the cable, so that the terahertz pulse beam scans the cable insulation layer point by point in a spiral trajectory until it covers the entire cable. S130: At each scanning point, record the time-domain waveform of the transmitted or reflected terahertz pulse and separate the main pulse signal characterizing the bulk properties of the insulating layer. S140: Perform Fourier transform on the main pulse signal of each scanning point to obtain the frequency domain spectrum of the corresponding scanning point, and calculate the real part ε'(ω) and imaginary part ε''(ω) of the complex dielectric constant of the corresponding scanning point at different frequencies to form the dielectric spectrum of the corresponding scanning point; S150: The dielectric spectrum of each scan point is associated with and stored with spatial coordinates, including axial position, radial depth and circumferential angle, thereby constructing the four-dimensional database; By limiting the "spiral trajectory scan" (S120), the terahertz pulse beam is ensured to conduct continuous and uninterrupted point-by-point detection of the cylindrical cable insulation layer in space, eliminating detection blind spots and achieving "full field" coverage. By clearly defining the need to "separate the main pulse signal that characterizes the properties of the insulation layer" (S130), interference signals generated by reflections from the conductor surface are effectively filtered out, thereby significantly improving the signal-to-noise ratio and intrinsic reliability of the detection data, making the measurement of the properties of the insulation material itself purer and more accurate. Furthermore, by specifying the use of "Fourier transform to obtain the frequency domain spectrum and calculate the complex dielectric constant" (S140), not only is the dielectric response of the material obtained over a wide frequency band, but its spectral characteristics are also extremely sensitive to changes in microstructure, providing rich feature dimensions for distinguishing different types of defects (such as polar impurities and micropores). Finally, by strictly correlating the dielectric spectrum with the three-dimensional coordinates of "axial position, radial depth and circumferential angle" (S150), a unique spatial-spectral four-dimensional database was constructed. This database can not only locate the geometric position of defects with extremely high spatial resolution, but also reveal the changes in the intrinsic state of the material at the defect. This provides standardized input data with both spatial accuracy and physical depth for the subsequent hybrid AI model, which is the premise and foundation for the realization of the entire predictive evaluation scheme. In summary, spiral trajectory scanning achieved comprehensive data acquisition of the cable insulation layer without blind spots, ensuring the integrity of the inspection. By separating and analyzing the terahertz main pulse signal characterizing the insulation's intrinsic properties, interference such as conductor interface reflections was effectively eliminated, significantly improving the signal-to-noise ratio and measurement accuracy. Furthermore, Fourier transform was used to convert the time-domain waveform into a frequency-domain dielectric spectrum and accurately correlate it with three-dimensional spatial coordinates, thereby constructing a high-quality four-dimensional database that can simultaneously reflect the spatial distribution and frequency dependence characteristics of the dielectric properties of the insulation material. This lays an irreplaceable quantitative data foundation for subsequent accurate defect identification and reliable lifetime prediction.

[0021] In this embodiment of the invention, step S200 specifically includes: S210: For the cable segment to be evaluated, from the start to the end of its production, a set of predefined key process parameters is collected in real time from the production control system. S220: Arrange and integrate the set of key process parameters in chronological order of production time, and associate them with the unique production identifier of the cable segment to generate a structured digital model of process parameters. S230: Based on the historical traction speed record of the cable segment, calculate the axial position of the insulation layer at each point in time on the production line. S240: Based on the correspondence calculated in S230, establish a mapping relationship between the parameter set at each time point in the digital model of the process parameters and all scanning points at the corresponding axial position (Z) in the four-dimensional database constructed in S150, complete the digital association, and generate the associated process parameters. By constructing a complete data collection, modeling, and mapping system, significant multi-level effects were achieved. First, by synchronously collecting the key process parameter set (S210) of the entire process from conductor preheating to cooling and forming in real time and constructing a structured digital model of process parameters (S220), a unique and tamper-proof "digital twin manufacturing history" is generated for each section of cable, which completely solves the problem of disconnection between process data and physical products and the difficulty of traceability in the relevant production process; Secondly, an innovative time-space conversion algorithm (S230) based on traction speed history was introduced to accurately convert the time series process parameters in the continuous production process into a function of the three-dimensional spatial axial position of the cable insulation layer, thus overcoming the technical challenge of matching process conditions with product position in dynamic production lines. Finally, through the established parameter-space mapping relationship (S240), any microscopic dielectric property anomaly detected in S100 can be uniquely locked to the specific production time and corresponding complete set of process conditions that caused the anomaly. This precise causal correlation capability not only provides high-quality, highly interpretable training data for subsequent AI models to learn the complex relationship between "process-structure-performance", but more importantly, it provides a direct engineering path for achieving precise root cause diagnosis and process optimization, thereby shifting quality control from back-end screening of results to front-end precise regulation of the manufacturing process. In summary, by synchronously collecting key process parameters in real time and constructing a structured digital model (S210-S220), the integrity and traceability of the "manufacturing fingerprint" of each cable segment were ensured. Furthermore, through spatiotemporal mapping technology based on traction speed (S230-S240), the correspondence between process parameters and the three-dimensional spatial location of insulation dielectric properties was accurately established. For the first time, a quantitative correlation from macroscopic process conditions to microscopic insulation performance was achieved, providing a precise data chain for subsequent analysis of the process root causes of defects. Ultimately, this refined correlation makes quality assessment no longer an isolated result judgment, but can directly trace back to the deviation of specific processes and parameters, thus laying a key foundation for achieving data-driven closed-loop optimization of production processes.

[0022] In this embodiment of the invention, step S300 specifically includes: S310: Establish an initial atomic-level molecular dynamics model of cross-linked polyethylene, using COMPASS or a similar force field, and set the density, degree of cross-linking, and chain length distribution consistent with the cable insulation material. S320: In the initial model, different preset process defects are introduced by changing the simulation conditions to construct multiple micro-defect structure models; The preset process defects include: insufficient local cross-linking degree, embedding of specific impurity atoms or clusters, and crystallinity gradient distribution caused by non-uniform temperature field. S330: For each of the aforementioned micro-defect structure models, apply constant electrical and thermal stresses corresponding to the actual operating conditions of the cable in the simulation environment, perform long-term molecular dynamics simulation, and record the evolution process of the microstructure of the micro-defect structure model over simulation time. S340: Extract key mechanism parameters characterizing defect dynamics from the evolution process. These key mechanism parameters include: the size and distribution rate of free volume pores, the anisotropic parameters of molecular chain segment mobility, and the evolution trend of trap charge density and energy level distribution. S350: The initial conditions of each simulation case, the corresponding microscopic defect structure model, and the key mechanism feature parameters evolved therefrom are associated and stored to construct the mechanism knowledge base; First, a high-fidelity virtual material experimental system was constructed: by adopting a widely validated dedicated force field (S310) such as COMPASS, and setting key parameters such as density and cross-linking degree that correspond to real ultra-high voltage cable insulation materials, it was ensured that the initial simulation model could highly reproduce the atomic and molecular structure of actual cross-linked polyethylene. On this basis, by actively introducing structural deviations (S320) that strictly correspond to real process defects such as "insufficient local cross-linking degree", "specific impurity embedding" and "non-uniform crystallization distribution", a series of accurate "cause-symptom" control models were constructed. This is equivalent to establishing standard digital specimens at the atomic scale for each "material internal injury" that may be caused by process fluctuations, so that subsequent analysis has clear direction and physical basis. Secondly, it enables the "time-accelerated" evolution observation of microscopic defects under service stress: by applying constant electrical and thermal stress (S330) equivalent to actual operation in the simulation environment and performing long-term molecular dynamics simulations on these defect models, this method can "watch" the material aging process that usually takes decades to occur in reality in the computer with extremely high time resolution. This breaks through the limitations of long experimental cycles and the inability to directly observe microscopic processes in the real world, and can reveal in advance how different initial defects (such as micropores and impurities) gradually evolve into microscopic failure mechanisms such as free volume channels or deep trap energy levels that lead to breakdown. Then, core physical descriptors bridging the microscopic and macroscopic worlds were extracted: from the above evolutionary process, key mechanistic characteristic parameters such as "the rate of change of free volume pores," "the anisotropy parameters of molecular chain segment mobility," and "the evolution trend of trap charge density and energy level distribution" were systematically extracted (S340). These parameters are not arbitrary data, but microscopic physical quantities that directly determine the macroscopic dielectric properties (such as dielectric constant, loss, and conductivity) and aging behavior of materials. For example, the expansion of free volume is directly related to local electric field distortion and a decrease in breakdown strength; the evolution of trap characteristics is directly related to space charge accumulation and dielectric loss. Therefore, these characteristic parameters constitute a quantitative bridge connecting atomic-scale structural variations and macroscopic measurable properties (and terahertz response). Finally, a mechanistic knowledge base supporting reliable predictions by the AI ​​model was formed: the final constructed mechanistic knowledge base (S350) stores complete causal chain cases of "specific process conditions → initial microstructure → long-term evolution characteristics". The core value of this knowledge base is that when it provides input for the subsequent hybrid AI model (S400), it does not provide raw, massive atomic motion trajectory data, but rather provides high-level knowledge that has been "refined" and "interpreted" by physical laws - that is, "if there is a certain process deviation, how will its microstructure be, and what physical law is most likely to deteriorate under future stress"; this makes the training of the AI ​​model not only dependent on limited, labeled historical data, but also constrained and guided by universal physical laws; In summary, by constructing a virtual experimental environment with a clear mechanism at the atomic scale, an indispensable physical core and prior knowledge support are provided for the entire predictive evaluation scheme. First, by establishing an atomic-level model based on dedicated force fields such as COMPASS and setting parameters consistent with real materials (S310), it is ensured that the starting point of the simulation is physically realistic and engineering-related, rather than theoretical assumptions. Furthermore, by pre-setting microscopic defects directly corresponding to process deviations (such as local low cross-linking degree, impurity embedding, etc., S320) and applying electro-thermal stress equivalent to actual operating conditions for long-term simulation (S330), the dynamic evolution process of various initial defects under decades of service stress is "pre-enacted" in a computer for the first time. This surpasses the time and observation scale limitations of any real-world accelerated aging experiment. More importantly... The key mechanistic parameters (S340) extracted from this evolutionary process, such as the free volume change rate, chain segment mobility, and trap characteristics, are a set of quantitative physical descriptors connecting microstructural variations with macroscopic dielectric property degradation (which can be perceived by terahertz spectrum) and eventual insulation failure. Finally, the mechanistic knowledge base (S350) composed of these cases, as high-order knowledge "distilled" from physical laws, is embedded in the subsequent AI model (S400), forcing its prediction results to conform to the basic laws of materials science. This fundamentally solves the problems of unreliability of the "black box" and high extrapolation risk in the prediction of long-term invisible behavior by purely data-driven models. This gives the lifetime prediction of this scheme solid scientific explanatory power and extremely high reliability. This is the core technical guarantee for moving from "statistical correlation" to "mechanistic prediction".

[0023] In this embodiment of the invention, step S400 specifically includes: S410: Construct a mechanism-data hybrid artificial intelligence model with multimodal input branches, wherein the mechanism-data hybrid artificial intelligence model includes: The first input branch is used to receive normalized dielectric spectrum data from a specific spatial point of S100; The second input branch is used to receive a vectorized set of complete critical process parameters associated with the spatial point from S200; The third input branch is used to receive key mechanism feature parameters from S300 that best match the set of key process parameters, and the key mechanism feature parameters are embedded as physical constraint terms. S420: Prepare the training dataset, which uses the S100, S200 and S300 data of historically produced cables as input features, and the actual defect types, locations and performance degradation data of the corresponding cable segments obtained in accelerated aging tests or long-term operation as supervision labels. S430: Train the mechanism-data hybrid driven artificial intelligence model. Its total loss function consists of three parts: the main task loss for defect classification and localization, the regression task loss for long-term performance prediction, and a mechanism consistency regularization loss that specifically penalizes the prediction results of the mechanism-data hybrid driven artificial intelligence model for violating the physical laws revealed by the key mechanism feature parameters in S300. S440: After training is completed, solidify the parameters of the mechanism-data hybrid driven artificial intelligence model so that the mechanism-data hybrid driven artificial intelligence model has the following functions: based on the input multimodal data, synchronously output the micro-defect classification and level corresponding to the current spatial point, as well as the performance degradation curve and remaining electric lifetime prediction value under the preset operating conditions at that point; By finely defining the construction steps of mechanism-data hybrid driven artificial intelligence models, the value of this architecture lies not only in improving prediction accuracy, but also in ensuring the physical credibility and engineering applicability of the prediction results. Firstly, information complementarity and synergistic effect are achieved through a multimodal fusion architecture; by designing three independent input branches (S410), the model systematically integrates key information of different natures and from different sources: The first branch (dielectric spectrum) provides a "snapshot of the current health status" of the cable insulation, which is the most direct physical response signal of microscopic defects; the second branch (key process parameters throughout the entire process) provides a "historical record of causes" leading to the current health status, establishing a causal chain from the source of manufacturing to the product status; the third branch (key mechanism characteristic parameters) provides a "future evolution script" based on physical laws, indicating the possible development direction of a specific defect structure under stress. This architecture forces the model to consider the current situation, trace the root cause, and predict the trend simultaneously when making decisions, realizing a deep integration of information in the time dimension and causal logic, and overcoming the misjudgment or omission that is easily caused when making judgments based on a single data source (such as relying solely on detection signals). Secondly, the mechanistic consistency regularization loss ensures the physical reliability and extrapolation robustness of the predictions, which is the core innovation that distinguishes this model from all purely data-driven methods (S430). This loss function, in addition to the standard data fitting objective (main task loss), adds a strong constraint: the model's internal inference process and its intermediate or final predictions (e.g., predicting the free volume fraction at a future point) must be as consistent as possible with the physically verified evolutionary laws provided by molecular dynamics simulations (S300). Effect 1: Injecting physical priors overcomes data sparsity. Long-term aging data of cables, especially extreme case data leading to eventual failure, is extremely rare in reality. Pure data models are prone to failure in such "long-tail problems." Mechanism regularization essentially provides the model with... The massive amount of "virtual experimental data" and universal physical rules enable it to make scientifically sound predictions even in the absence of real failure data. Secondly, it enhances the interpretability and reliability of the model. This constraint forces the model's learning process to go beyond simply finding statistical correlations; it requires understanding and approximating the underlying physical mechanisms. This makes the model's predictions no longer "black box" outputs, but rather inferences based on known physical laws, greatly increasing engineers' confidence in the predictions—a prerequisite for engineering applications. Thirdly, it ensures the robustness of long-term predictions. For lifespan predictions spanning decades, extrapolation based purely on historical data carries significant uncertainty. Mechanism regularization anchors the predictions to relatively stable physical laws, significantly reducing extrapolation risks and making long-term predictions unprecedentedly robust. Finally, the integrated model output revolutionizes assessment efficiency and decision support. The trained model (S440) has the function of "synchronous output" of defect diagnosis and life prediction. This means that for any point on the insulation layer, a single forward inference can simultaneously obtain: immediate diagnostic conclusion: what the defect is (classification) and how serious it is (level); long-term risk prediction: performance degradation curve and remaining electrical life under preset operating conditions at that point. This "one-stop" output integrates the incomplete information that traditionally requires multiple independent tests and multiple analyses to piece together into a unified and coherent assessment report. It enables the production line to grasp the quality status and potential risks of each cable segment in real time and comprehensively, providing a unique, efficient and reliable decision-making basis for subsequent accurate classification based on predictive life (S550), differentiated commissioning strategy formulation, and proactive maintenance arrangements. In summary, through innovative architecture design and training mechanisms, this scheme elevates its predictive capabilities to a new level that combines high accuracy and high reliability. Firstly, by establishing a multimodal input branch (S410), it structurally integrates the dielectric spectrum reflecting the "current state," process parameters revealing the "manufacturing root," and mechanistic features predicting "evolutionary laws," enabling the model to simultaneously reason based on data correlation and physical mechanisms, thus solving the problem of the one-sidedness of information from a single data source. Secondly, its unique mechanism consistency regularization loss function (S430) forces the model's internal representation to align with the material degradation physical laws revealed by molecular dynamics simulations (such as free volume) during training. The growth trend is consistent with the principle of using first principles as a "mentor" to constrain and guide the learning process of AI. This fundamentally overcomes the "black box" extrapolation risk and unreliable results caused by data sparsity when pure data-driven models predict long-term, rare failure modes. Finally, the trained model (S440) can synchronously output accurate defect diagnosis and quantitative long-term performance curves for any scan point, realizing integrated intelligent evaluation from "detecting the present" to "predicting the future". It provides a unique and reliable core criterion for the final realization of cable classification and decision-making based on predictive life. It is the intelligent hub that distinguishes this solution from all traditional detection methods and has outstanding creativity.

[0024] In this embodiment of the invention, step S500 specifically includes: S510: The four-dimensional database data of the cable under test obtained in S100, and the associated process parameters obtained and spatiotemporally correlated in S200, are synchronously input into the mechanism-data hybrid driven artificial intelligence model trained and solidified in S400. S520: The mechanism-data hybrid driven artificial intelligence model traverses and processes all scan points in spatial coordinate order, and outputs the micro-defect classification and grade label of each scan point in parallel, as well as the corresponding predicted remaining lifetime value array. The predicted remaining lifetime value array represents the failure probability under different operating years. S530: Based on the micro-defect classification and grade labeling of all scanning points, a three-dimensional interpolation algorithm is used to generate a three-dimensional defect type distribution cloud map and defect grade distribution cloud map of the cable insulation layer; S540: Based on the array of predicted remaining lifetime values ​​of all scan points, extract the failure probability under the same service life threshold, generate a three-dimensional heat map of the failure probability of the cable insulation layer at different service time points; and calculate the minimum predicted remaining lifetime value of the entire cable. S550: Combining a three-dimensional defect distribution cloud map and a failure probability heat map, the cable is comprehensively rated according to a preset quality grading rule. The quality grading rule is based at least on the overall defect density and the minimum predicted remaining life value, and divides the cable into at least four levels: superior, qualified, reworkable and unqualified. S560: Automatically generate a visual quality assessment report that includes the aforementioned three-dimensional distribution cloud map, heat map, quality grade, list of specific defect location coordinates, and maintenance recommendations; The complex data generated in the preceding steps is transformed into highly operable intelligent outputs that can directly drive production decisions, specifically in the following key aspects: First, a panoramic and visual reconstruction from "point data" to "field information" was achieved. This step, through a precisely defined algorithm process, integrates the discrete prediction results of AI models for hundreds of thousands or even millions of independent spatial points into a global and intuitive visual presentation. The 3D defect distribution cloud map (S530) uses a 3D interpolation algorithm to classify and label point-like defects, generating a continuous color cloud map covering the entire volume of the cable insulation layer. This allows engineers to instantly identify defect clusters and distribution patterns (e.g., axial strip defects may indicate extrusion problems, while circumferential ring defects may indicate uneven temperature distribution in the cross-linking tube), enabling comprehensive insulation analysis. The macroscopic understanding of spatial heterogeneity in quality provides a global perspective that traditional detection methods based on single-point or local sampling cannot offer. The time-varying failure probability heatmap (S540) directly visualizes the predictive capabilities of this solution. By extracting the failure probability of all scanned points at the same future time point (e.g., 10 or 20 years after commissioning) and generating a 3D heatmap of that time point, the risk evolution process of cable insulation can be dynamically displayed. Management can intuitively see how high-risk areas expand or connect from scattered points as service time progresses, thus providing unprecedented forward-looking data for preventative maintenance and upgrade plans of the power grid. Secondly, a new intelligent and quantitative grading standard based on predictive indicators has been established. The quality grading rules defined in S550 serve as a core bridge connecting technical forecasting and business decisions. The scientific nature of the core indicators: the rules abandon the traditional pass / fail electrical performance thresholds, instead using the minimum predicted remaining life value (reflecting the expected service life of the weakest point) and overall defect density (reflecting the uniformity of insulation and overall health) as grading criteria. These two indicators are directly related to the long-term reliability and safe operating cycle of cables, making their decision-making significance more profound. The grading strategy is refined and maximizes value: cables are divided into multiple grades such as "superior grade, qualified grade, repair grade, and non-qualified grade," achieving refined product management and maximizing value. For example, "superior grade" cables can be prioritized for important lines or harsh environments; "qualified grade" cables are used for general lines; "repair grade" cables can be downgraded through partial repair (such as targeted treatment after location), avoiding the huge waste caused by overall scrapping; and "non-qualified grade" cables are resolutely rejected. This "graded disposal" strategy based on predicted lifespan generates significant economic benefits compared to the traditional "binary judgment" (qualified / scrapped). Finally, a structured and executable report driving closed-loop action is generated. The inspection report generated by S560 is far more than a simple data list; it is a comprehensive decision support document integrating diagnosis, prediction, and recommendations. The content is structured and traceable: the report not only includes cloud maps and heat maps but also a list of specific defect location coordinates. This provides a precise "navigation map" for subsequent possible re-inspections, location repairs, or sampling dissection verifications, achieving full-process data traceability. The decision support is direct: the "quality level" in the report directly corresponds to warehouse management, shipping, or installation instructions; while the "maintenance recommendations" provide the recipient (such as the operating unit) with personalized maintenance tips for that section of cable (such as "avoid long-term overload operation" and "focus on the axial section from X meters to Y meters"). This allows cutting-edge inspection technology to seamlessly connect and directly empower downstream production, logistics, and maintenance processes. In summary, through detailed evaluation and grading steps, the complex prediction results of the hybrid AI model are transformed into intuitive and quantitative tools that can be directly used for production decisions, constructing a complete intelligent decision-making closed loop: First, through parallel processing and three-dimensional interpolation algorithms, the massive point data output by the model is efficiently and accurately synthesized into a panoramic three-dimensional defect distribution cloud map and a time-varying failure probability heat map, making hidden micro-defects and future risks "visualized," greatly improving the intuitiveness and depth of quality assessment; Second, an innovative multi-rule dynamic grading mechanism based on the minimum predicted remaining lifetime value and overall defect density is introduced, realizing an intelligent mapping from continuous predicted values ​​to discrete disposal instructions (excellent, acceptable, returned, scrapped), enabling quality control decisions to leap from experience thresholds to predictive quantitative indicators; Finally, the automatically generated structured inspection report integrates all key information and action recommendations, seamlessly connecting cutting-edge inspection and prediction technologies to existing production management and quality control processes, thereby not only achieving accurate "physical examination" and "prognosis" of individual cables, but also significantly improving the automation, scientification, and forward-looking level of production line quality management.

[0025] In this embodiment of the invention, step S600 specifically includes: S610: Analyze the visual quality assessment report generated by S500 and extract the spatial location information, type, and level of all defects marked therein; S620: Based on the mapping relationship established in S200, map each defect location back to its corresponding key process parameter that takes effect when producing the insulation layer at that location; S630: Using statistical correlation analysis or machine learning attribution algorithms, compare the set of key process parameters for all defect points with the set of key process parameters for defect-free or excellent points to identify at least one key process parameter and its deviation direction that is significantly related to a specific defect type and / or low predicted lifetime. S640: Based on the identification results of S630, for each key process parameter, combined with its physical allowable range and the corresponding defect evolution knowledge in the S300 mechanistic knowledge base, a list of process optimization suggestions is generated, which includes the specific parameter adjustment amount, adjustment priority and expected quality improvement after adjustment. S650: The process optimization suggestion list is pushed to the corresponding execution unit of the production control system in real time in the form of structured instructions to guide the closed-loop control of the production process of the next batch or the current online cable. First, it achieves precise reverse tracing from macroscopic quality defects to microscopic process parameters. Through the precise operation of S610 and S620, this solution establishes a reverse mapping mechanism. When any "defect point" is identified in S500, the system can immediately trace back and lock onto the set of all process parameters used in manufacturing that specific insulation point on the production line through the established "spatiotemporal mapping relationship" (derived from S200 and S240). This solves the fundamental pain point of traditional quality control of "knowing what (there is a defect) but not knowing why (why there is a defect)", accurately locating quality problems from the end of the product to specific equipment, workstations, and even time points in the production process, providing a unique "target" for precise intervention. Secondly, it has achieved a leap from human experience-based judgment to data-driven attribution-based scientific decision-making. Based on precise tracing, the S630 uses statistical correlation analysis or machine learning attribution algorithms (such as SHAP value analysis) to systematically compare a large number of process parameter sets of "defect points" and "normal points". This process can automatically and objectively identify key process parameters that are statistically significantly correlated with specific defect types (such as micropores) or low life prediction values ​​(for example, identifying "temperature fluctuation in the third zone of the cross-linking tube" as the main cause of insufficient local cross-linking degree) and clarify its deviation direction (such as "low temperature"). This completely replaces the vague diagnosis that relies on the experience of experienced workers, and makes the root cause analysis based on multivariate and big data analysis. The conclusions are more scientific and comprehensive, and can discover complex interactive effects that are difficult for the human brain to perceive. Then, a quantitative parameter tuning scheme that combines physical feasibility and optimization expectation is generated. Step S640 deeply integrates data-driven attribution results with domain knowledge. It not only proposes adjustment suggestions based on analysis results, but more importantly, it combines the physical allowable range of the process parameter (equipment limits, material properties) with the knowledge in the S300 mechanism knowledge base about how the parameter affects the microstructure evolution. The resulting suggestion list not only includes the specific adjustment amount of "increasing the temperature by X degrees Celsius", but may also include "priority ranking" (addressing the parameter with the greatest impact first) and "expected quality improvement" (expected to increase the minimum remaining lifetime by Y%). This makes the optimization suggestion no longer a simple alarm, but a process improvement scheme that has been "virtually verified" and can be quantified and expected, which greatly improves the adoption rate and implementation success rate of the suggestion. Finally, the most transformative effect of the "perception-decision-execution" real-time production closed loop and continuous evolution capability is reflected in the S650. Optimization suggestions no longer end with paper reports, but are fed back in real time to the corresponding execution units (such as heaters and extruder speed controllers) of the production control system (such as DCS or MES) in the form of structured, machine-readable instructions. This means: Real-time correction: For subsequent cables still on the production line, the system can immediately fine-tune the process to prevent defects from continuing to occur; Closed-loop control: The production system transforms from an open-loop, fixed operating mode to an intelligent closed-loop system that can adaptively adjust based on real-time product quality feedback; Knowledge accumulation and iteration: Each adjustment and its effect (verified through the detection of subsequent cables) will generate new data, feeding back into the optimization algorithm and AI model, enabling the system to have the ability to continuously learn and self-optimize the process. In summary, the detection and evaluation results are transformed into direct productivity: Through precise reverse mapping of S610-S620, the spatial location of insulation defects is uniquely mapped to the set of process parameters used during production, enabling precise tracing of quality issues. Furthermore, with the statistical or machine learning attribution analysis of S630, key process parameters and their deviation directions leading to specific defects are automatically identified from massive amounts of data, transforming root cause analysis from vague inferences based on experience to data-driven scientific diagnosis. On this basis, S640 generates a quantitative optimization suggestion list combining physical allowable ranges and mechanistic knowledge, ensuring that adjustment plans are both targeted and feasible. Finally, S650 feeds the suggestions back to the production control system in real time as structured instructions, directly driving online adjustments to process parameters. This not only corrects problems in the current batch but also promotes continuous iteration and optimization of the production process, forming a self-evolving cycle of "detection-prediction-optimization." This systematically improves the insulation quality and long-term reliability of ultra-high voltage cables from the root cause and significantly reduces batch quality risks and production costs caused by process fluctuations.

[0026] In this embodiment of the invention, in step S410, the key mechanistic feature parameters received by the third input branch are structurally fused into the forward propagation process of the hybrid neural network model through at least one of the following methods: Method 1: The key mechanism feature parameters are input into a dedicated auxiliary decoder network, which attempts to reconstruct these key mechanism feature parameters from the hidden features of the intermediate layer of the mechanism-data hybrid driven artificial intelligence model, thereby forcing the main network to learn feature representations consistent with physical laws; Method 2: The key mechanism feature parameters are used to construct a prior physical rule sub-network. The output of the physical rule sub-network is weighted and fused or gated with the intermediate features of the main network to selectively strengthen signal paths that conform to physical laws at different stages of inference. Method 3: The key mechanism feature parameters are used as a set of ideal feature vectors to initialize or apply bias to some weight matrices of the core layer of the mechanism-data hybrid driven artificial intelligence model, so that the mechanism-data hybrid driven artificial intelligence model has a reasoning tendency that conforms to physical laws from the beginning of training. By specifically defining the structured fusion of key mechanistic feature parameters in neural networks, multi-layered and substantially progressive effects are achieved, transforming physical laws from external constraints into the model's intrinsic reasoning framework, realizing a paradigm shift from "data fitting" to "physics-guided reasoning." Method 1 constructs a deep forced alignment mechanism between physical laws and data representation (for "Method 1"). Method 1 requires the model to actively reconstruct key mechanistic feature parameters (such as free volume change rate) from its internal hidden features through a dedicated auxiliary decoder. This generates a strong internal learning pressure, forcing the backbone of the neural network to learn intermediate feature representations that can simultaneously interpret the observed data (dielectric spectrum) and derive the physical laws (mechanistic parameters). This is equivalent to establishing a "physical correctness" verification channel within the model, significantly improving the interpretability of model features. These intermediate features are no longer black boxes but are associated with explicit physical quantities (such as molecular chain mobility and trap density). When the model faces new samples, even if its final output (such as lifetime) is new, the physical meaning corresponding to the intermediate features supporting the output is clear and consistent with the laws, thereby fundamentally enhancing the credibility of the prediction results and engineers' trust in the model. Method 2 implements dynamic inference path control based on physical rules (for "Method 2"). Method 2 constructs an independent physical rule sub-network, which receives mechanistic parameters and generates a control signal (such as a weight or gating signal) that is fused with the features of the main network. This effectively introduces conditional computation capabilities into the model. For different types of defects (such as thermal aging-dominated defects and electrical aging-dominated defects), the physical rule sub-network can dynamically adjust the importance of different feature channels in the main network, selectively activating or suppressing certain signal pathways, enabling the model to apply different physical rules for inference "according to local conditions." This greatly improves the model's situational adaptability and generalization ability. For defect combinations or new failure modes that are not fully present in the training data but can be identified by the physical rule sub-network, the model can provide more reasonable predictions by activating the corresponding physical rule pathways, rather than producing absurd extrapolation results. This solves the core challenge of AI dealing with "unknown unknowns" risks in industrial scenarios. Method 3 ensures that the model has physically reasonable prior preferences from the initialization stage (for Method 3). Method 3 uses mechanistic parameters as ideal feature vectors to physically guide the initialization or apply biases to the model's core weights. This ensures that the model's parameter space is already in a physically reasonable region before gradient descent training begins. Compared to completely random initialization, this is equivalent to providing the model with the correct "starting direction." First, it significantly accelerates model convergence and reduces dependence on massive amounts of labeled data. Second, and more importantly, it reduces the risk of the model converging to a purely mathematically optimal but physically absurd local optimum. In cases of limited data or high noise, this physically guided initialization plays a crucial "stabilizer" role, ensuring the robustness of the training process and the fundamental reliability of the final model.

[0027] In this embodiment of the invention, in step S410, the feature vector Fd obtained after encoding the data from the first and second input branches, and the key mechanism feature parameter vector Fm from the third input branch, are fused through a mechanism-guided fusion module to generate a hybrid feature vector F. hybrid ; The mechanism-guided fusion module performs the following operations: S411: Transfer the key mechanism feature parameter vector F m Input a parameterized network and generate a set of attention weight vectors W a Or a characteristic transformation matrix M m ; S412: Utilizing the W a Or M m For the feature vector F d Perform weighted filtering or projection transformation to make the output feature F hybrid The feature dimensions related to the current physical mechanisms are enhanced; Furthermore, the total loss function L in step S430 total Specifically: L total =α(λL cls +(1 λ)L reg_life )+β Lmech +γL reg Where: L cls Cross-entropy loss for defect classification tasks; L reg_life Smoothing L1 loss for lifetime prediction tasks; L mech The mechanism consistency regularization loss is calculated based on the deviation between the intermediate or final results predicted by the model and the physical relationship implied by the mechanism feature parameter Fm; L regα, β, γ, λ are the weighted regularization loss; α, β, γ, λ are the preset positive weight coefficients; First, the defined mechanism-guided fusion modules (S411-S412) transform physical mechanism features from static inputs into dynamic controllers, actively generating attention weights or transformation matrices to filter and reshape data features. This forces the alignment of data representation with physical laws along the critical path of model inference, ensuring the physical interpretability of the model's internal logic. Second, the defined total loss function L... total A specific weighted combination, especially the mechanism consistency loss term L. mech The quantitative constraints on the deviation between the prediction results and the physical relationship establish a multi-objective, mechanism-oriented optimization blueprint for the model training process, enabling it to simultaneously pursue classification accuracy, prediction accuracy, and physical reliability.

[0028] A method for maintaining an ultra-high voltage cable involves first inspecting the current cable segment using the ultra-high voltage cable insulation testing method described in any one of claims 1 to 9, and then maintaining the cable segment.

[0029] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for testing the insulation of ultra-high voltage cables, characterized in that, Includes the following steps: S100: Performs in-situ, non-contact terahertz time-domain spectral three-dimensional full-field scanning on the finished cable to obtain the time-domain waveforms of each spatial point in the cable insulation layer and calculates and generates a four-dimensional database containing spatial coordinates and corresponding dielectric spectra. S2 00: Synchronously collect key process parameters of the entire cable production process, and digitally associate the key process parameters with the corresponding cable segment spatial coordinates in S100 to generate associated process parameters; S300: Based on molecular dynamics simulation, a mechanistic knowledge base is constructed to understand the microscopic defect structure of cross-linked polyethylene under different process deviations and its evolution under electro-thermal stress, and key mechanistic characteristic parameters are extracted. S400: Building Mechanism-Data Hybrid Driven Artificial Intelligence Models; The dielectric spectrum of S100, the associated process parameters of S200, and the key mechanism feature parameters matched from S300 are used as mixed inputs, and the micro-defect type, severity level, and long-term performance prediction index are used as outputs to train the mechanism-data hybrid driven artificial intelligence model. S500: Input the dielectric spectrum and associated process parameters of the cable under test into the trained mechanism-data hybrid driven artificial intelligence model to obtain the three-dimensional defect distribution and long-term performance prediction results of the cable insulation layer, and generate a visual quality assessment report accordingly. S6 00: Based on the correlation analysis between the three-dimensional defect distribution and related process parameters in the visualized quality assessment report, generate production process optimization feedback suggestions.

2. The method for detecting the insulation of ultra-high voltage cables according to claim 1, characterized in that, Step S100 specifically includes: S110: Pass the completed cable axially through a ring-shaped terahertz time-domain spectroscopy scanning device at a constant speed. S120: Control the transmitting and receiving units of the terahertz time-domain spectral scanning device to rotate at a constant speed along the circumference of the cable and move synchronously along the axial direction of the cable, so that the terahertz pulse beam scans the cable insulation layer point by point in a spiral trajectory until it covers the entire cable. S130: At each scanning point, record the time-domain waveform of the transmitted or reflected terahertz pulse and separate the main pulse signal characterizing the bulk properties of the insulating layer. S140: Perform Fourier transform on the main pulse signal of each scanning point to obtain the frequency domain spectrum of the corresponding scanning point, and calculate the real part ε'(ω) and imaginary part ε''(ω) of the complex dielectric constant of the corresponding scanning point at different frequencies to form the dielectric spectrum of the corresponding scanning point; S150: The dielectric spectrum of each scan point is associated with and stored with spatial coordinates, including axial position, radial depth and circumferential angle, thereby constructing the four-dimensional database.

3. The method for testing the insulation of ultra-high voltage cables according to claim 2, characterized in that, Step S200 specifically includes: S210: For the cable segment to be evaluated, from the start to the end of its production, a set of predefined key process parameters is collected in real time from the production control system. S220: Arrange and integrate the set of key process parameters in chronological order of production time, and associate them with the unique production identifier of the cable segment to generate a structured digital model of process parameters. S230: Based on the historical traction speed record of the cable segment, calculate the axial position of the insulation layer at each point in time on the production line. S240: Based on the correspondence calculated in S230, establish a mapping relationship between the parameter set at each time point in the digital model of the process parameters and all scanning points at the corresponding axial positions in the four-dimensional database constructed in S150, complete the digital association, and generate associated process parameters.

4. The method for detecting the insulation of ultra-high voltage cables according to claim 3, characterized in that, Step S300 specifically includes: S310: Establish an initial atomic-level molecular dynamics model of cross-linked polyethylene, using COMPASS or a similar force field, and set the density, degree of cross-linking, and chain length distribution consistent with the cable insulation material. S320: In the initial model, different preset process defects are introduced by changing the simulation conditions to construct multiple micro-defect structure models; The preset process defects include: insufficient local cross-linking degree, embedding of specific impurity atoms or clusters, and crystallinity gradient distribution caused by non-uniform temperature field. S330: For each of the aforementioned micro-defect structure models, apply constant electrical and thermal stresses corresponding to the actual operating conditions of the cable in the simulation environment, perform long-term molecular dynamics simulation, and record the evolution process of the microstructure of the micro-defect structure model over simulation time. S340: Extract key mechanism parameters characterizing defect dynamics from the evolution process. These key mechanism parameters include: the size and distribution rate of free volume pores, the anisotropic parameters of molecular chain segment mobility, and the evolution trend of trap charge density and energy level distribution. S350: The initial conditions of each simulation case, the corresponding microscopic defect structure model, and the key mechanism feature parameters evolved from it are associated and stored to construct the mechanism knowledge base.

5. The method for detecting the insulation of ultra-high voltage cables according to claim 4, characterized in that, Step S400 specifically includes: S410: Construct a mechanism-data hybrid artificial intelligence model with multimodal input branches, wherein the mechanism-data hybrid artificial intelligence model includes: The first input branch is used to receive normalized dielectric spectrum data from a specific spatial point of S100; The second input branch is used to receive a vectorized set of complete critical process parameters associated with the spatial point from S200; The third input branch is used to receive key mechanism feature parameters from S300 that best match the set of key process parameters, and the key mechanism feature parameters are embedded as physical constraint terms. S420: Prepare the training dataset, which uses the S100, S200 and S300 data of historically produced cables as input features, and the actual defect types, locations and performance degradation data of the corresponding cable segments obtained in accelerated aging tests or long-term operation as supervision labels. S430: Train the mechanism-data hybrid driven artificial intelligence model. Its total loss function consists of three parts: the main task loss for defect classification and localization, the regression task loss for long-term performance prediction, and a mechanism consistency regularization loss that specifically penalizes the prediction results of the mechanism-data hybrid driven artificial intelligence model for violating the physical laws revealed by the key mechanism feature parameters in S300. S440: After training is completed, the parameters of the mechanism-data hybrid-driven artificial intelligence model are solidified, so that the mechanism-data hybrid-driven artificial intelligence model has the following functions: based on the input multimodal data, it synchronously outputs the micro-defect classification and level corresponding to the current spatial point, as well as the performance degradation curve and remaining electrical lifetime prediction value under the preset operating conditions at that point.

6. The method for testing the insulation of ultra-high voltage cables according to claim 5, characterized in that, Step S500 specifically includes: S510: The four-dimensional database data of the cable under test obtained in S100, and the associated process parameters obtained and spatiotemporally correlated in S200, are synchronously input into the mechanism-data hybrid driven artificial intelligence model trained and solidified in S400. S520: The mechanism-data hybrid driven artificial intelligence model traverses and processes all scan points in spatial coordinate order, and outputs the micro-defect classification and grade label of each scan point in parallel, as well as the corresponding predicted remaining lifetime value array. The predicted remaining lifetime value array represents the failure probability under different operating years. S530: Based on the micro-defect classification and grade labeling of all scanning points, a three-dimensional interpolation algorithm is used to generate a three-dimensional defect type distribution cloud map and defect grade distribution cloud map of the cable insulation layer; S540: Based on the array of predicted remaining lifetime values ​​of all scan points, extract the failure probability under the same service life threshold, generate a three-dimensional heat map of the failure probability of the cable insulation layer at different service time points; and calculate the minimum predicted remaining lifetime value of the entire cable. S550: Combining a three-dimensional defect distribution cloud map and a failure probability heat map, the cable is comprehensively rated according to a preset quality grading rule. The quality grading rule is based at least on the overall defect density and the minimum predicted remaining life value, and divides the cable into at least four levels: superior, qualified, reworkable and unqualified. S560: Automatically generates a visualized quality assessment report that includes the aforementioned 3D distribution cloud map, heat map, quality grade, list of specific defect location coordinates, and maintenance recommendations.

7. The method for testing the insulation of ultra-high voltage cables according to claim 6, characterized in that, Step S600 specifically includes: S610: Analyze the visual quality assessment report generated by S500 and extract the spatial location information, type, and level of all defects marked therein; S620: Based on the mapping relationship established in S200, map each defect location back to its corresponding key process parameter that takes effect when producing the insulation layer at that location; S630: Using statistical correlation analysis or machine learning attribution algorithms, compare the set of key process parameters for all defect points with the set of key process parameters for defect-free or excellent points to identify at least one key process parameter and its deviation direction that is significantly related to a specific defect type and / or low predicted lifetime. S640: Based on the identification results of S630, for each key process parameter, combined with its physical allowable range and the corresponding defect evolution knowledge in the S300 mechanistic knowledge base, a list of process optimization suggestions is generated, which includes the specific parameter adjustment amount, adjustment priority and expected quality improvement after adjustment. S650: The process optimization suggestion list is pushed to the corresponding execution unit of the production control system in real time in the form of structured instructions to guide the closed-loop control of the production process of the next batch or the current online cables.

8. The method for testing the insulation of ultra-high voltage cables according to claim 5, characterized in that, In step S410, the key mechanistic feature parameters received by the third input branch are structurally fused into the forward propagation process of the hybrid neural network model through at least one of the following methods: Method 1: The key mechanism feature parameters are input into a dedicated auxiliary decoder network, which attempts to reconstruct these key mechanism feature parameters from the hidden features of the intermediate layer of the mechanism-data hybrid driven artificial intelligence model, thereby forcing the main network to learn feature representations consistent with physical laws; Method 2: The key mechanism feature parameters are used to construct a prior physical rule sub-network. The output of the physical rule sub-network is weighted and fused or gated with the intermediate features of the main network to selectively strengthen signal paths that conform to physical laws at different stages of inference. Method 3: The key mechanism feature parameters are used as a set of ideal feature vectors to initialize or apply bias to some weight matrices of the core layer of the mechanism-data hybrid driven artificial intelligence model, so that the mechanism-data hybrid driven artificial intelligence model has a reasoning tendency that conforms to physical laws from the beginning of training.

9. The method for testing the insulation of ultra-high voltage cables according to claim 5, characterized in that, In step S410, the feature vector Fd obtained after encoding the data from the first and second input branches, and the key mechanism feature parameter vector Fm from the third input branch, are fused through a mechanism-guided fusion module to generate a hybrid feature vector F. hybrid ; The mechanism-guided fusion module performs the following operations: S411: Transfer the key mechanism feature parameter vector F m Input a parameterized network and generate a set of attention weight vectors W a Or a characteristic transformation matrix M m ; S412: Utilizing the W a Or M m For the feature vector F d Perform weighted filtering or projection transformation to make the output feature F hybrid The feature dimensions related to the current physical mechanisms are enhanced; Furthermore, the total loss function L in step S430 total Specifically: L total =α(λL cls +(1 λ)L reg_life )+b Lmech +γL reg Where: L cls Cross-entropy loss for defect classification tasks; L reg_life Smoothing L1 loss for lifetime prediction tasks; L mech The mechanism consistency regularization loss is calculated based on the deviation between the intermediate or final results predicted by the model and the physical relationship implied by the mechanism feature parameter Fm; L reg α, β, γ, λ are the weighted regularization loss; α, β, γ, λ are the preset positive weight coefficients.

10. A method for maintaining ultra-high voltage cables, characterized in that, First, the current cable segment is tested using the ultra-high voltage cable insulation testing method described in any one of claims 1 to 9, and then the cable segment is maintained.

Images