Cable defect data enhancement method and system based on physical consistency constraint

Abstract

The invention belongs to the field of cable defect data enhancement, and relates to a cable defect data enhancement method and system based on physical consistency constraint, and the system outputs a balanced training set and a test set through a data preprocessing module; the generative model construction module is used for embedding a double-field potential function in a potential space based on a conditional variation self-encoding network encoding observable vector and a conditional label, constructing a Hamiltonian energy function fusing the constraint of the double-field potential function, introducing an adaptive kinetic energy item, and unifying a PCDH-VAE generative model training target as a Hamiltonian energy minimization problem; a dynamic situation trap mechanism is introduced into the sample generation module, adaptive quality Hamiltonian Monte Carlo sampling is adopted, and a defect enhancement sample meeting multi-physics field constraint is generated through Hamiltonian dynamic evolution and Metropolis correction; and the sample evaluation module evaluates the generated defect enhancement sample set. The problem of scarcity of cable defect data samples can be relieved.

Description

Technical Field

[0001] This invention relates to the field of cable defect data augmentation, and more specifically to a method and system for cable defect data augmentation based on physical consistency constraints. Background Technology

[0002] High-voltage cables are a critical component of power systems, and accurate identification of early-stage defects in high-voltage cables is essential for ensuring the safe and stable operation of the power grid. Currently, data-driven defect identification methods rely on a large number of labeled defect samples. However, in actual operation, defect samples are severely scarce, leading to insufficient model training and limited generalization ability.

[0003] In the prior art, a method and system for detecting cable defects in cable trenches based on machine vision are disclosed. The method enhances cable image data by using deep convolutional generative adversarial networks to improve sample diversity and model generalization ability. A multidimensional weak attention residual network is designed, which combines a weak signal attention module and a multi-scale perception module to enhance the ability to capture small defects and weak features in cable trenches, thereby improving the accuracy and robustness of visual detection.

[0004] The above-mentioned technical solutions have made progress in improving the diversity of image data and the ability to detect details, but the following technical problems still exist due to the failure to integrate physical laws: the physical reliability of the generated data under the condition of scarce samples, insufficient maintenance of the coupling relationship of multiple physical fields, poor training stability, and disconnection from the actual physical state of the cable.

[0005] In view of this, it is very necessary to provide a cable defect data enhancement method and system based on physical consistency constraints to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems in cable defect identification under the condition of sample scarcity, such as unreliable physical data, insufficient maintenance of multi-physics coupling relationship, poor training stability, and decoupling from the actual physical state of the cable. The invention provides a cable defect data enhancement method and system based on physical consistency constraints to solve the technical problems existing in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for enhancing cable defect data based on physical consistency constraints, comprising the following steps: Step S1: Obtain the raw data of cable multi-physics monitoring, perform data preprocessing, divide it into training set and test set according to the ratio, perform inverse frequency weighted sampling on the training set to obtain a balanced training set with class balance. Step S2: Construct the PCDH-VAE generative model, using the conditional variational autoencoder network as a sub-model, and encode the observable vectors and conditional labels; embed a dual potential function in the latent space of the conditional variational autoencoder network, construct the Hamiltonian energy function, unify the model optimization objective into the Hamiltonian energy minimization problem, introduce an adaptive kinetic energy term, and form a complete energy modeling framework. Step S3: Input the balanced training set into the PCDH-VAE generative model, introduce a dynamic potential well mechanism, and adaptively adjust the physical potential energy term and statistical potential energy term through time-varying tolerance threshold and stiffness coefficient to form a high-stiffness dynamic potential energy surface. On the high-stiffness dynamic potential energy surface, an adaptive mass Hamiltonian Monte Carlo sampling method is used to achieve stable sampling of the latent space through coupling of the mass matrix and local stiffness. Hamiltonian dynamic evolution and Metropolis correction are then performed to obtain the corrected latent variables. ; latent variables Input the sample to the decoder to generate defect-enhanced samples that satisfy multiphysics constraints; Step S4: Evaluate the quality of the defect enhancement samples, merge the defect enhancement samples with the training set to construct an expanded training set, train the downstream defect diagnosis model, and verify the enhancement effect through classification performance indicators.

[0008] Secondly, the present invention also provides a cable defect data enhancement system based on physical consistency constraints, comprising: The data preprocessing module is used to preprocess the raw cable monitoring data, divide the dataset, and balance it, outputting a balanced training set and a test set that maintains the original distribution. The generative model building module is used to build PCDH-VAE generative models. It encodes observable vectors and conditional labels based on conditional variational autoencoder networks, embeds dual potential functions in the latent space, constructs Hamiltonian energy functions that integrate dual potential function constraints, introduces adaptive kinetic energy terms, and unifies the training objective of PCDH-VAE generative models into the Hamiltonian energy minimization problem. The sample generation module is used to introduce a dynamic potential well mechanism, adaptively adjust the physical potential energy and statistical potential energy intensity to form a high-stiffness dynamic potential energy surface, and adopt adaptive mass Hamiltonian Monte Carlo sampling. Through Hamiltonian dynamic evolution and Metropolis correction, defect-enhanced samples that meet multi-physics constraints are generated. The sample evaluation module is used to evaluate the generated defect enhancement sample set, merge the defect enhancement samples with the training set, construct a dataset for downstream tasks, train the downstream defect identification model, and perform high-precision cable defect diagnosis.

[0009] The beneficial effects of this invention are as follows: This invention achieves class balance in the training set by performing inverse frequency-weighted sampling on multi-physics monitoring data of cables, effectively alleviating the problem of scarce cable defect data samples and providing a reliable data foundation for cable defect data enhancement.

[0010] This invention constructs a dual-field potential function, enabling the generated enhanced samples to conform to statistical characteristics while maintaining physical consistency, thus avoiding the problem of physically unreliable data generated by traditional methods.

[0011] This invention introduces a dynamic potential well mechanism to adaptively adjust the strength of physical and statistical constraints, effectively maintaining the coupling relationship between multiple physics fields and enhancing the physical rationality and consistency of the generated samples.

[0012] This invention employs the adaptive quality Hamiltonian Monte Carlo method for latent variable sampling and uses Metropolis correction to improve the stability and sampling reliability of the training process, thereby enabling the generation of defective samples to have high fidelity and physical consistency.

[0013] This invention evaluates the quality of generated samples by using statistical fidelity and physical violation rate, and then uses them in conjunction with the original training set to improve the classification performance of downstream defect diagnosis models. This effectively solves the problem of the disconnect between training data and the actual physical state of cables in traditional methods.

[0014] This invention constructs a PCDH-VAE generative model that integrates physical and statistical constraints, and combines a dynamic potential well mechanism and an adaptive Hamiltonian Monte Carlo sampling method. This enables the generation of physically consistent, well-preserved multi-physics coupling relationships, stable training, and high-fidelity cable defect enhancement samples under conditions of sample scarcity. This effectively improves the performance of downstream defect diagnosis models and solves the technical problems in existing technologies, such as unreliable physical data, insufficient preservation of multi-physics coupling relationships, unstable training, and disconnection from the actual physical state of cables.

[0015] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This is a flowchart of a cable defect data augmentation method based on physical consistency constraints; Figure 2 This is a diagram of the inverse stiffness scheduling strategy; Figure 3 This is a block diagram of a cable defect data augmentation system based on physical consistency constraints. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0019] Example 1: like Figure 1 As shown in the figure, this embodiment provides a cable defect data augmentation method based on physical consistency constraints, which includes the following steps: Step S1: Obtain the raw data of cable multi-physics monitoring, perform data preprocessing, divide it into training set and test set according to the ratio, perform inverse frequency weighted sampling on the training set to obtain a balanced training set with class balance. In step S1: In this embodiment, the data comes from the online status monitoring system and maintenance records of a provincial power company. Online monitoring records of the 110kV cross-linked polyethylene cable are obtained to construct the original multi-physics monitoring data for the cable. The original multi-physics monitoring data is simultaneously collected by on-site temperature sensors, current acquisition devices, online insulation monitoring equipment, and partial discharge detection devices, and the operating status is marked in conjunction with maintenance records. The original multi-physics monitoring data includes core temperature... Sheath temperature Leakage current Grounding circulating current There are a total of 7 observable variables, including insulation resistance R, dielectric loss tangent tanδ, and partial discharge quantity Q.

[0020] The raw data from cable multiphysics monitoring are preprocessed to generate cable multiphysics monitoring samples: Based on the long-term operating experience range of the cable and the steady-state allowable range of IEC 60287, boundary screening was performed on temperature, current and insulation parameters to eliminate outliers that obviously violate physical laws; missing records were supplemented, and data continuity was restored by prioritizing the use of statistical mean of the same category or neighborhood interpolation methods; the raw data of cable multi-physics monitoring were standardized based on the statistical quantities of training data to map the seven observable variables to a unified scale space; the labels were reviewed in conjunction with the maintenance report to ensure that the category labeling was consistent with the actual operating status, and seven observable vectors x were generated.

[0021] The processed cable multi-physics monitoring samples include four operating states: 742 samples of normal operation, 92 samples of main insulation abnormalities, 107 samples of joint overheating, and 102 samples of sheath damage. These four operating states are then categorized into four classes. Due to the significant long-tail distribution of the categories, with a ratio of approximately 8:1 between normal and defective classes, a stratified random sampling method was used to divide the cable multi-physics monitoring samples into training and testing sets at a 7:3 ratio, ensuring that the proportions of each category remain consistent between the training and testing sets.

[0022] To mitigate the bias caused by class imbalance in parameter updates, an inverse frequency-weighted sampling mechanism is introduced: a weight coefficient is calculated based on the number of samples of each class in the training set, ensuring that the probability of a sample being selected from each class is inversely proportional to its class frequency; during batch data construction, random sampling is performed according to this weight, giving each class an approximately balanced opportunity to participate in the training process, resulting in a balanced training set with a class distribution approaching 1:1:1:1. The test data maintains its original proportions and is used for generalization evaluation in real-world scenarios.

[0023] Through the above operations, we obtain a balanced training set after preprocessing and inverse frequency-weighted sampling, and a test set that maintains the true distribution, providing a stable and reliable data foundation for subsequent operations.

[0024] Step S2: Construct the PCDH-VAE generative model, using the conditional variational autoencoder network as a sub-model, and encode the observable vectors and conditional labels; embed a dual potential function in the latent space of the conditional variational autoencoder network, construct the Hamiltonian energy function, unify the model optimization objective into the Hamiltonian energy minimization problem, introduce an adaptive kinetic energy term, and form a complete energy modeling framework. Step S2 specifically includes: Step S21: Construct a conditional variational autoencoder network; Using a conditional variational autoencoder network as a sub-model, the conditional variational autoencoder network includes an encoder and a decoder. The encoder is used to map the observable vectors and conditional labels to the latent distribution parameter space, and the decoder is used to map the latent variables back to the data space, thereby realizing the reconstruction and generation of the original monitoring vectors.

[0025] The observable vector x and the defect category label c are jointly encoded to form a conditional input vector, which is then fed into the encoder of a conditional variational autoencoder network. The encoder outputs the parameters of the latent distribution, including the mean vector and variance vector, through multiple layers of nonlinear mapping, thus establishing a conditional approximate posterior distribution. Where z is a latent variable, q is an approximate posterior distribution, and φ is the encoder parameter.

[0026] By reparameterizing the latent variable z to make it differentiable, the random sampling process can participate in gradient backpropagation, continuously updating the latent distribution parameters during training, and realizing the learning of the latent structure of multi-physics monitoring data.

[0027] The sampled latent variable z and the defect category label c are jointly input into the decoder to establish a generation correspondence between the latent space and the data space, so that different operating states form a distinctive distribution structure in the latent space.

[0028] Step S22: Construct the dual potential function; In the latent space of the conditional variational autoencoder network, a dual-field potential function consisting of a physical potential term and a statistical potential term is introduced. The physical potential term is derived from the constraint expression constructed by the thermodynamic constitutive relation and dielectric constitutive relation of the steady-state high-voltage cable, and is used to characterize the degree of deviation of the generated samples corresponding to the latent variables from the physical laws. The statistical potential term is derived from the lower bound of negative evidence and the moment-based statistical feature constraint, and is used to characterize the degree of approximation of the latent distribution to the real data distribution and the ability to preserve defect features.

[0029] The physical potential energy construction term includes a thermal potential construction sub-stage, an electrical residual construction sub-stage, and a total physical potential energy construction sub-stage, wherein: The thermal potential construction sub-stage is as follows: Under thermal equilibrium, radial heat conduction and surface convection heat dissipation are governed by Fourier's law and Newton's law of cooling. This is achieved through core temperature... Sheath temperature and ambient temperature Calculate the thermal residual and express it as thermal potential. :

[0030] in , and These represent the core temperature, sheath temperature, and ambient temperature, respectively. Parameters Indicates thermal conductivity, The thickness represents the equivalent thermal resistance, and h represents the convective heat transfer coefficient.

[0031] The residual dielectric construction sub-stage is characterized by: dielectric loss being characterized by leakage current. Insulation resistance R and dielectric loss tangent The inherent coupling between them, δ is the dielectric loss angle of the cable insulation material, and the residual current is defined as: Where U represents the operating voltage.

[0032] The sub-stage of constructing the total physical potential energy is as follows: the thermal potential and electrical residue are combined by weighted sum to obtain the total physical potential energy, and the PCDH-VAE generation model is constrained to satisfy the multiphysics laws.

[0033] The process of constructing the statistical potential term is as follows: Considering the differences between the training and generation phases of the cable condition monitoring system, where real samples are not directly observable under the target defect conditions, a statistical potential is constructed. The training phase refers to the PCDH-VAE generative model optimizing its parameters on real observation data from the cable condition monitoring system, while simultaneously learning the data distribution and physical constraints. The generation phase refers to the process after the PCDH-VAE generative model has been trained, where defect-enhanced samples are generated from the latent space, at which point real observations are unavailable.

[0034] During the training phase, the negative evidence lower bound (negative ELBO) is used as the statistical potential to calculate the KL divergence between the approximate posterior distribution and the prior distribution, which measures the degree of deviation from the latent distribution. The negative evidence lower bound is: ,in, It is the KL divergence, used to measure the distance between two distributions; It is an approximate posterior distribution. It is a prior distribution; It's a decoder. It is a balancing factor.

[0035] During the generation phase, the unavailability of real observations causes the reconstruction loss to fail. To guide the latent variables along the target defect condition c... target Evolution, introducing moment-based statistical potentials, preserves defect-specific statistical properties by using the mean, standard deviation, and shape factor of each dimension of the input data. Specifically: ; Where D is the dimension of the input data. It is the value of the sample in the j-th dimension. , and Let represent the mean, standard deviation, and shape factor of the j-th feature under the target defect condition c, respectively. The statistical potential term can preserve the characteristic distribution of defect categories in the absence of direct observation, providing statistical guidance for potential space sampling.

[0036] Step S23: Construct the Hamiltonian energy function; When constructing the Hamiltonian energy function, a momentum variable p of the same dimension is introduced on the basis of the potential variable z to form an extended state variable (z,p), and the total Hamiltonian is expressed as the sum of the potential energy term and the kinetic energy term, thereby constructing a complete potential space dynamic system.

[0037] Meanwhile, an adaptive kinetic energy term is introduced. By setting the momentum variable p to follow a zero-mean Gaussian distribution and combining it with the mass matrix to scale different latent dimensions, the latent variables have good numerical stability and ergodicity, thereby improving the efficiency and stability of the Hamiltonian sampling process.

[0038] Specifically, the physical potential energy term, statistical potential energy term, and adaptive kinetic energy term in the dual-field potential function are weighted and combined to construct the total Hamiltonian energy function, yielding the total Hamiltonian H. Within the Hamiltonian energy function framework, to construct a stable dynamic evolution process within the potential space, an auxiliary momentum variable p is introduced based on the latent variable z, and the total Hamiltonian is constructed as the sum of the total physical potential energy term and the adaptive kinetic energy term, i.e. ; Among them, the total physical potential energy term Composed of both physical potential energy and statistical potential energy, it can be expressed as: ; in, This represents the physical potential energy term constructed from the steady-state thermal constraint and dielectric coupling relationship, used to measure the degree of deviation of the defect-enhanced sample from the steady-state thermal constraint and dielectric coupling relationship; The statistical potential term is used to characterize reconstruction error, measure the difference between the latent distribution and the data distribution, and the degree of preservation of defect feature distribution; the adaptive kinetic energy term is also included. Used to enhance the stability of the potential space sampling process.

[0039] The adaptive kinetic energy term is defined by a momentum variable p with the same dimension as the latent variable z in the latent space, and it is assumed to follow a zero-mean Gaussian distribution: .

[0040] Step S24: Construct the optimization objective of the PCDH-VAE generative model; The optimization objective of the PCDH-VAE generative model is unified as the Hamiltonian energy minimization problem, which enables the encoder and decoder parameters to be updated collaboratively under the combined effect of statistical consistency constraints and physical consistency constraints. During training, the energy gradient guides the latent variables to converge toward a feasible manifold that satisfies multiple physical constraints.

[0041] By minimizing the total Hamiltonian energy, the encoder and decoder parameters are simultaneously optimized under the dual constraints of statistical and physical consistency. Furthermore, a parameter tuning mechanism that varies with the training phase allows the statistical and physical potentials to exhibit different degrees of dominance at different stages, achieving a smooth transition from data-driven representation learning to physically consistent generation. Ultimately, a latent distribution structure constrained by multiple physical constraints is obtained, laying the foundation for subsequent stable sampling and high-fidelity sample generation.

[0042] Step S25: Construct the PCDH-VAE generative model; By unifying and integrating the encoder, decoder, dual potential function, and Hamiltonian energy function, the PCDH-VAE generative model forms an energy distribution structure in the latent space that is constrained by the dual potential function. Under this energy distribution structure, the sampling, evolution, and mapping processes of latent variables are all guided by the unified energy landscape. This enables the PCDH-VAE generative model to not only possess conditional generation capabilities and latent space representation capabilities, but also to achieve adaptive adjustment of the latent space structure under multi-physics constraints, thereby forming a constrained latent distribution manifold.

[0043] The constructed PCDH-VAE generative model can generate defect-enhanced samples in subsequent sampling processes, achieving a balance between physical consistency and data diversity.

[0044] In this model, the latent space dimension of the PCDH-VAE generative model is set to 16. The encoder adopts a two-layer network structure with 64 and 32 hidden units, respectively, while the decoder adopts a symmetrical two-layer network structure with 32 and 64 hidden units, respectively. Through these additions, step S2 not only completes the construction of the conditional variational autoencoder network structure but also achieves unified modeling of physical and statistical constraints at the latent space level. Furthermore, it combines this with an adaptive kinetic energy mechanism to form a complete energy modeling framework, providing a consistent theoretical basis and structural support for dynamic potential energy regulation and stable sampling in subsequent steps.

[0045] Step S3: Input the balanced training set into the PCDH-VAE generative model, introduce a dynamic potential well mechanism, and adaptively adjust the physical potential energy term and statistical potential energy term through time-varying tolerance threshold and stiffness coefficient to form a high-stiffness dynamic potential energy surface. On the high-stiffness dynamic potential energy surface, an adaptive mass Hamiltonian Monte Carlo sampling method is used to achieve stable sampling of the latent space through coupling of the mass matrix and local stiffness. Hamiltonian dynamic evolution and Metropolis correction are then performed to obtain the corrected latent variables. ; latent variables Input the sample to the decoder to generate defect-enhanced samples that satisfy multiphysics constraints; In step S3: During the training of the PCDH-VAE generative model, the batch size is set to 32, the number of training rounds is 1000, and the initial learning rate is set to 0.001. To simultaneously satisfy multiphysics constraints and statistical distribution constraints during latent space sampling, the physical stiffness range is set to 5 to 60, the statistical stiffness range is set to 40 to 50, and a threshold smoothing factor of 0.95 is used to ensure the stability of dynamic potential energy convergence. In the adaptive mass Hamiltonian Monte Carlo sampling stage, the inertia scaling factor coefficient is set to 0.8, the symplectic transform step size is 0.005, and the stride step size is 20, used to achieve efficient exploration of latent variables on the high-stiffness dynamic potential energy surface and numerically stable dynamic evolution.

[0046] Step S3 specifically includes: Step S31: Introduce a dual-field dynamic potential well mechanism; To address the gradient conflict caused by static physical constraints between data fitting and physical compliance, and to avoid convergence instability or mode collapse during training, a dual-field dynamic potential well mechanism is introduced.

[0047] The dual-field dynamic potential well mechanism uses the training period t as the evolution variable and transforms static physical and statistical constraints into a dynamic energy landscape that changes over time through homotopy transformation, thereby adaptively adjusting the strength of physical and statistical constraints.

[0048] Step S31 specifically includes: Step S311, the time-varying dynamic stage of potential parameters; To realize the time evolution of the potential energy surface, two adaptive control parameters are defined: dynamic tolerance threshold W(t) and dynamic stiffness coefficient k(t). W(t) is used to control the compliance range of the potential energy deviation, and k(t) is used to adjust the strength of the potential energy constraint.

[0049] To ensure that the contraction process of the potential energy surface remains consistent with the model's learning dynamics, an exponential moving average (EMA) is used to track the residual statistics of the physical potential energy and the statistical potential energy. The calculation form is as follows: .

[0050] Where t represents the training epoch, B represents the batch size, and i represents the i-th sample. It is a smoothing factor; This indicates whether j originates from physical potential energy or statistical potential energy; phys represents physical potential energy, and stat represents statistical potential energy.

[0051] Adopting an inverse stiffness scheduling strategy, such as Figure 2 As shown, for physical stiffness k phys and statistical stiffness k stat Dynamic adjustments are made: In the early stages of training, the physical stiffness... k phys Gradually increase the physical stiffness to suppress gradient explosion in the early stages and avoid premature physical constraints; in the later stages of training, increase the physical stiffness. k phys Maintain strict physical compliance, statistical stiffness k stat By appropriately relaxing the constraints, the flexibility of the potential space is enhanced, enabling the model to perform reasonable extrapolation under physical correction. Through the above-mentioned inverse evolution strategy, a dynamic balance is achieved between gradually strengthening physical constraints and moderately softening statistical constraints.

[0052] Step S312: Construct a dynamic potential energy surface; Based on the dynamic tolerance threshold W(t), the dynamic stiffness coefficient k(t), and the inverse stiffness scheduling strategy, a high-stiffness dynamic potential energy surface in a quadratic piecewise form is constructed. The total dynamic potential energy function... The calculation formula is:

[0053] in, The current potential energy loss is represented by the "+" sign, which indicates the physical hardening process. The symbol "" represents the controlled relaxation process of the statistical boundary.

[0054] Through the above construction method, dynamic physical potential V phys,dyn As training progresses, the acceptable compliance region is gradually tightened, allowing the defect-enhanced samples to gradually approach the physically feasible manifold; dynamic statistical potential V stat,dyn The initial rigid distribution constraints were transformed into more flexible statistical boundaries, which improved the diversity of defective samples while ensuring data quality.

[0055] Step S313, the co-evolution stage of the joint potential field; Based on the high-stiffness dynamic potential energy surface in the form of a second-order piecewise segment, the dynamic physical potential... V phys,dyn and dynamic statistical potential V stat,dyn The evolution follows an "anchor-transition-correction" process: In the early stages (t≤T), dynamic statistical potential dominates, stabilizing the potential topological structure and preventing premature introduction of strong physical constraints that could lead to mode collapse. During the transition phase (t≈0.6T), the sampled particles are simultaneously subjected to the competitive gradient effects of the dynamic physical potential and the dynamic statistical potential, thereby achieving coordinated regulation of the dual potentials. In the final stage (t→T), physical stiffness gradually becomes dominant, constraining the defect-enhanced sample to a physically acceptable manifold region, while the softened dynamic statistical potential allows for limited correction adjustments.

[0056] Through three stages of evolution, a fully trained PCDH-VAE model and its corresponding high-stiffness dynamic potential energy surface are obtained. In the environment of small sample data augmentation, physical compliance is guaranteed and the diversity of defect augmentation samples is improved, providing a stable and high-quality data source for downstream defect identification.

[0057] In step S32, the adaptive mass Hamiltonian Monte Carlo sampling stage, due to the combined effect of strong physical constraints and dynamic potential well mechanism, the high-stiffness dynamic potential energy surface exhibits high curvature and high stiffness characteristics in some directions. Conventional sampling methods are prone to numerical instability or step size limitations in deep potential wells.

[0058] To ensure stable and efficient in-depth exploration of the latent space under strong constraints, an adaptive mass Hamiltonian Monte Carlo sampling strategy is introduced in the neighborhood of the defect-enhanced sample distribution on the high-stiffness dynamic potential energy surface for latent space sampling. Geometric inertia matching is achieved through coupling the mass matrix with local stiffness, and Hamiltonian dynamic evolution and Metropolis correction are performed in the extended phase space to obtain the latent variables. latent variables Input the decoder of the PCDH-VAE generative model and output physically consistent and more diverse defect-enhanced samples.

[0059] Step S32 specifically includes: Step S321: Construct the adaptive quality matrix; Due to the stiffness mismatch between the physical potential energy term and the statistical potential energy term, the standard Hamiltonian Monte Carlo (HMC) method is prone to numerical oscillations in high-stiffness dynamic potential energy surfaces.

[0060] Therefore, a diagonally time-varying adaptive mass matrix M(t) proportional to the local potential stiffness is constructed, and a geometric inertia matching strategy is adopted to enable the particle to obtain greater inertia in the high-stiffness direction. The adaptive mass matrix is ​​defined as: .

[0061] Where diag represents the diagonal matrix form, η is the inertia scaling factor, and ε is a small constant to prevent numerical singularities. The geometric inertia matching strategy can increase particle inertia in the high-stiffness direction, smooth steep energy gradients, and achieve stable skip integration without setting an excessively small integration step size.

[0062] Step S322, Extended phase space and Hamiltonian dynamics evolution stage; Introducing auxiliary momentum variables p ~N(0, M ( t This expands the original d-dimensional latent space into a 2d-dimensional phase space. The Hamiltonian obtained from this is defined as:

[0063] According to the Hamiltonian equation, it evolves as follows:

[0064] in, For the simulation time step in HMC sampling, This represents gradient operation. The total dynamic potential energy. The symplectic structure of Hamiltonian dynamics maintains the phase space volume invariant and ensures that the resulting Markov chain satisfies detailed equilibrium conditions, enabling the sampling process to expand the structure at a deeper level around the defect-enhanced sample distribution.

[0065] Step S323, Discretization and Metropolis Correction Stage; The continuous dynamical system is discretized using a time-reversible leapfrog integrator. After L integrations with a step size δ, candidate states are obtained, and the acceptable candidate states are calculated according to the Metropolis criterion. ; in, Let be the probability of accepting the sample, and exp be the exponential operation. Original Hamiltonian energy, This is the updated Hamiltonian energy. The correction process eliminates numerical bias caused by discretization, ensuring that the sampling results converge to the accurate target distribution while strictly adhering to physical and statistical distribution constraints.

[0066] Step S324: Generate defect enhancement samples; After completing adaptive quality Hamiltonian Monte Carlo sampling, latent variables that satisfy the target distribution and strictly adhere to physical and statistical distribution constraints are obtained. Then, latent variables Input the trained PCDH-VAE generative model. Guided by a high-stiffness dynamic potential surface formed by dual potential function constraints, the decoder converts the latent variables... Mapping back to the data space generates physically consistent and more diverse defect enhancement samples.

[0067] Because the sampling process has completed physical consistency screening and statistical distribution correction on the high-stiffness dynamic potential energy surface, the defect enhancement samples output by the decoder not only maintain statistical distribution characteristics consistent with the real defect category, but also have thermal-electric coupling consistency under multi-physics field constraints, thus providing diversified and highly reliable data support for the subsequent defect identification model.

[0068] Step S4: Evaluate the quality of the defect enhancement samples, merge the defect enhancement samples with the training set to construct an expanded training set, train the downstream defect diagnosis model, and verify the enhancement effect through classification performance indicators.

[0069] In step S4: a comprehensive evaluation index system covering statistical distribution consistency, physical compliance and diagnostic effectiveness is constructed; the statistical fidelity and physical violation rate of the defect enhancement samples are calculated; a systematic quality evaluation of the defect enhancement samples is conducted; and the improvement effect of the defect enhancement samples on the downstream defect identification performance is examined.

[0070] The comprehensive evaluation system includes three dimensions: Statistical fidelity dimension: The distributional similarity between the enhanced defect samples and the real samples from the cable condition monitoring system is quantified using Fréchet Inception distance (FID), Wasserstein distance (WD), and maximum mean difference (MMD). To avoid bias in the overall results due to prior distribution imbalance, the distances between the generated sub-distributions and the real sub-distributions are calculated within each defect category, and then the arithmetic mean of the results for each category is taken to obtain a statistical consistency measure under class equilibrium conditions.

[0071] Physical Consistency Dimension: This dimension assesses the degree to which defect-enhanced samples meet engineering constraints through the Physical Violation Rate (PVR). This metric is defined as the percentage of defect-enhanced samples that violate the physical threshold specified in IEC 60287. = 24 W / m², The proportion of samples with a physical violation rate (e.g., 0.5kV) is used to characterize the compliance of defect-enhanced samples under steady-state thermal and dielectric constraints. The lower the physical violation rate, the more the defect-enhanced samples conform to the multiphysics coupling law, and the higher their engineering feasibility.

[0072] Diagnostic utility dimension: The enhanced defect samples are merged with the training set data to form an expanded defect recognition training set, which is used to train the downstream defect diagnosis model. A macro-level F1 score is used to evaluate the classification performance of each defect category, especially sparse defect categories. By comparing the changes in classification metrics before and after enhancement, the effectiveness of the enhanced defect samples in improving diagnostic performance is verified.

[0073] The above three-dimensional evaluation framework enables a unified measurement of the quality, physical effectiveness, and actual diagnostic contribution of defect-enhanced samples, thereby forming a complete verification loop for engineering applications.

[0074] Example 2: like Figure 3 As shown in the figure, this embodiment provides a cable defect data enhancement system based on physical consistency constraints, including: Data preprocessing module 1 is used to preprocess the raw multi-physics monitoring data of cables from the cable condition monitoring system to generate multi-physics monitoring samples. Preprocessing includes threshold judgment of temperature, current, and insulation parameters to remove abnormal records exceeding the physical allowable range; imputation of missing data using statistical means of the same category or neighborhood interpolation; standardization of seven observable variables based on training data statistics to map each variable to a unified scale space; and verification and correction of operating status labels based on maintenance reports. The multi-physics monitoring samples are divided into training and test sets in a 7:3 ratio. To address the imbalance in the class distribution of the training set, an inverse frequency weighting coefficient is calculated based on the number of samples in each category, making the sample selection probability inversely proportional to the class frequency. During batch construction, samples are randomly selected according to the weights to form a balanced training set with a class distribution approaching 1:1:1:1. Data preprocessing module 1 ensures the integrity and balance of the input data, provides a reliable data foundation, reduces the impact of class bias on model performance, and effectively alleviates the class imbalance problem.

[0075] Module 2, the generative model building module, is used to construct the PCDH-VAE generative model. It employs a conditional variational autoencoder network as a sub-model, setting up encoder and decoder structures. It jointly encodes the observable vectors and conditional labels as input, outputting latent distribution parameters in the latent space and obtaining latent variables through reparameterization. A dual-field potential function, consisting of physical and statistical potential terms, is embedded in the latent space. The physical potential term is constructed from thermodynamic and dielectric constitutive relations, while the statistical potential term is constructed from negative evidence lower bounds and moment-based statistical characteristic constraints. A momentum variable of the same dimension is introduced into the latent variables, constructing a Hamiltonian energy function containing potential and kinetic energy terms. The momentum variable is set to follow a zero-mean Gaussian distribution, and the mass matrix is ​​used to scale each latent dimension. The physical, statistical, and kinetic energy terms are weighted and combined to form the total Hamiltonian energy function. Minimizing the total Hamiltonian energy is used as the model optimization objective, jointly updating the encoder and decoder parameters. By introducing physical and statistical potential terms into the latent space, and unifying them within the Hamiltonian energy modeling framework, the PCDH-VAE generative model can simultaneously satisfy physical constraints and data distribution constraints during training, thereby improving the physical consistency and distributional expressiveness of the latent representation.

[0076] Sample generation module 3 inputs the balanced training set into the PCDH-VAE generation model, introduces a dual-field dynamic potential well mechanism with the training period as the evolution variable, and transforms the physical potential energy term and statistical potential energy term into a time-varying dynamic energy landscape through homotopy transformation; sets a dynamic tolerance threshold W(t) and a dynamic stiffness coefficient k(t), and tracks the physical potential energy residual and statistical potential energy residual based on exponential moving average, and dynamically adjusts the physical stiffness and statistical stiffness in combination with an inverse stiffness scheduling strategy; constructs a high-stiffness dynamic potential energy surface in a quadratic piecewise form based on the dynamic tolerance threshold, dynamic stiffness coefficient, and stiffness scheduling results, and performs training... The process involves the co-evolution of dynamic physical potential and dynamic statistical potential in a "anchor-transition-correction" phase. An adaptive mass Hamiltonian Monte Carlo sampling method is constructed on the high-stiffness dynamic potential energy surface. A mass matrix proportional to the local potential stiffness is set, and auxiliary momentum variables are introduced to extend the phase space. Dynamic evolution is performed according to the Hamiltonian equation, and discretization is performed using a leapfrog integrator. Correction is then performed using the Metropolis criterion to obtain latent variables that satisfy the target distribution. These latent variables are input into the decoder and mapped to the data space under the constraints of the dynamic potential energy surface, generating defect enhancement samples that satisfy multiphysics constraints. Through the synergistic effect of the dynamic potential well mechanism and the adaptive mass Hamiltonian Monte Carlo sampling method, stable sampling and deep exploration are achieved under the high-stiffness dynamic potential energy surface, generating high-quality defect enhancement samples that combine physical compliance and diversity.

[0077] Sample evaluation module 4 is used to construct a three-dimensional comprehensive evaluation index system covering statistical fidelity, physical consistency, and diagnostic effectiveness, and to systematically evaluate the quality of the generated defect enhancement samples. The qualified defect enhancement samples are merged with the original training set to construct an expanded defect recognition training set for training downstream defect diagnosis models. The training results are quantitatively verified using macroscopic F1 scores, enabling the evaluation and feedback of the performance improvement of cable defect diagnosis by the defect enhancement samples. High-quality defect enhancement samples are selected through the three-dimensional comprehensive evaluation system and combined with the training set for downstream model training, improving the classification performance of the defect diagnosis model under sparse categories and multiple physical constraints, and enabling the enhancement samples to effectively support practical engineering diagnosis.

[0078] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for enhancing cable defect data based on physical consistency constraints, characterized in that, Includes the following steps: Step S1: Obtain the raw data of cable multi-physics monitoring, perform data preprocessing, divide it into training set and test set according to the ratio, perform inverse frequency weighted sampling on the training set to obtain a balanced training set with class balance. Step S2: Construct a PCDH-VAE generative model, using a conditional variational autoencoder network as a sub-model, to encode the observable vectors and conditional labels. By embedding a dual potential function into the latent space of a conditional variational autoencoder network, a Hamiltonian energy function is constructed, and the model optimization objective is unified into a Hamiltonian energy minimization problem. An adaptive kinetic energy term is introduced to form a complete energy modeling framework. Step S3: Input the balanced training set into the PCDH-VAE generative model, introduce a dynamic potential well mechanism, and adaptively adjust the physical potential energy term and statistical potential energy term through time-varying tolerance threshold and stiffness coefficient to form a high-stiffness dynamic potential energy surface. On the high-stiffness dynamic potential energy surface, an adaptive mass Hamiltonian Monte Carlo sampling method is used to achieve stable sampling of the latent space through coupling of the mass matrix and local stiffness. Hamiltonian dynamic evolution and Metropolis correction are then performed to obtain the corrected latent variables. ; latent variables Input the sample to the decoder to generate defect-enhanced samples that satisfy multiphysics constraints; Step S4: Evaluate the quality of the defect enhancement samples, merge the defect enhancement samples with the training set to construct an expanded training set, train the downstream defect diagnosis model, and verify the enhancement effect through classification performance indicators.

2. The cable defect data augmentation method based on physical consistency constraints according to claim 1, characterized in that, In step S1: the raw data for cable multiphysics monitoring includes core temperature. Sheath temperature Leakage current Grounding circulating current There are a total of 7 observable variables, including insulation resistance R, dielectric loss tangent tanδ, and partial discharge quantity Q; The raw data of cable multi-physics field monitoring is preprocessed to generate cable multi-physics field monitoring samples. The cable multi-physics field monitoring samples include four operating states: normal operation, main insulation abnormality, joint overheating and sheath damage, seven observable variables, seven observable vectors x are generated, and four categories are generated according to the four operating states. When dividing the data, a stratified random sampling method was adopted, and the cable multi-physics monitoring samples were divided into training set and test set in a 7:3 ratio. An inverse frequency weighted sampling mechanism was introduced to generate a balanced training set with a category distribution that tends to be 1:1:1:

1.

3. The cable defect data augmentation method based on physical consistency constraints according to claim 1, characterized in that, Step S2 includes: Step S21: Construct a conditional variational autoencoder network; The observable vector x and the defect category label c are jointly encoded to form a conditional input vector, which is then fed into the encoder of a conditional variational autoencoder network. The encoder outputs the parameters of the latent distribution through multiple layers of nonlinear mapping, establishing a conditional approximate posterior distribution. Where z is a latent variable, q is an approximate posterior distribution, and φ is an encoder parameter; the latent variable z is differentiable through reparameterization, and the sampled latent variable z is jointly input into the decoder along with the defect category label c; Step S22: Construct the dual potential function; In the latent space of the conditional variational autoencoder network, a dual-field potential function consisting of physical potential and statistical potential is introduced to construct the physical potential term, which includes a thermal potential construction sub-stage, an electric residual construction sub-stage, and a total physical potential construction sub-stage. The process of constructing the statistical potential term is as follows: During the training phase, the lower bound of negative evidence is used as the statistical potential, and the KL divergence between the approximate posterior distribution and the prior distribution is calculated. During the generation phase, a moment-based statistical potential is introduced to preserve the statistical properties of the defects by using the mean, standard deviation, and shape factor of each dimension of the input data. ; Where D is the dimension of the input data. It is the value of the sample in the j-th dimension. , and Let represent the mean, standard deviation, and shape factor of the j-th feature under the target defect condition c, respectively. Step S23: Construct the Hamiltonian energy function; Based on the latent variable z, a momentum variable p of the same dimension is introduced to form an extended state variable (z,p). The total energy is expressed as the sum of the potential energy term and the kinetic energy term. An adaptive kinetic energy term is introduced. By setting the momentum variable p to follow a zero-mean Gaussian distribution, the scale of different latent dimensions is adjusted in combination with the mass matrix. Step S24: Construct the optimization objective of the PCDH-VAE generative model; The optimization objective of the PCDH-VAE generative model is unified as the Hamiltonian energy minimization problem. Step S25: Construct the PCDH-VAE generative model; By unifying and integrating the encoder, decoder, dual potential function, and Hamiltonian energy function, the PCDH-VAE generative model forms an energy distribution structure in the latent space that is constrained by the dual potential function. The latent space dimension of the PCDH-VAE generative model is set to 16. The encoder adopts a two-layer network structure with 64 and 32 hidden units respectively, and the decoder adopts a symmetrical two-layer network structure with 32 and 64 hidden units respectively.

4. The cable defect data augmentation method based on physical consistency constraints according to claim 3, characterized in that, The aforementioned thermal potential construction sub-stage is: through core temperature... Sheath temperature and ambient temperature Calculate the thermal residual and express it as thermal potential. : in , and These represent the core temperature, sheath temperature, and ambient temperature, respectively; parameters Indicates thermal conductivity, The thickness represents the equivalent thermal resistance, and h represents the convective heat transfer coefficient. The residual dielectric construction sub-stage is characterized by: dielectric loss being characterized by leakage current. Insulation resistance R and dielectric loss tangent The inherent coupling between them, δ is the dielectric loss angle of the cable insulation material, and the residual current is defined as: Where U represents the operating voltage; The sub-stage for constructing the total physical potential energy is: combining the thermal potential and electrical residue by weighted sum to obtain the total physical potential energy.

5. The cable defect data augmentation method based on physical consistency constraints according to claim 4, characterized in that, The lower bound for negative evidence is: ,in, It is the KL divergence, used to measure the distance between two distributions; It is an approximate posterior distribution. It is a prior distribution; It's a decoder. It is a balancing factor; Step S23 includes: By weighting and combining the physical potential energy term, statistical potential energy term, and adaptive kinetic energy term in the dual-field potential function, a total Hamiltonian energy function is constructed, yielding the total Hamiltonian H. Within the Hamiltonian energy function framework, an auxiliary momentum variable p is introduced based on the latent variable z, constructing the total Hamiltonian as the sum of the total physical potential energy term and the adaptive kinetic energy term. ; Among them, the total physical potential energy term Composed of both physical potential energy and statistical potential energy, it can be expressed as: ; in, Represents the physical potential energy term; The statistical potential energy term and the adaptive kinetic energy term are defined by a momentum variable p with the same dimension as the latent variable z in the latent space, and it is assumed to follow a zero-mean Gaussian distribution: .

6. The cable defect data augmentation method based on physical consistency constraints according to claim 5, characterized in that, Step S3 includes: Step S31: Introduce a dual-field dynamic potential well mechanism; The dual-field dynamic potential well mechanism uses the training period t as the evolution variable and transforms static physical constraints and statistical constraints into a dynamic energy landscape that changes over time through homotopy transformation, thereby adaptively adjusting the strength of physical and statistical constraints. Step S32, Adaptive Quality Hamiltonian Monte Carlo Sampling Stage; An adaptive mass Hamiltonian Monte Carlo sampling strategy is introduced for latent space sampling. Geometric inertia matching is achieved through coupling the mass matrix with local stiffness. Hamiltonian dynamics evolution and Metropolis correction are then performed in the extended phase space to obtain the latent variables. latent variables Input the decoder of the PCDH-VAE generative model and output defect-enhanced samples.

7. The cable defect data augmentation method based on physical consistency constraints according to claim 6, characterized in that, Step S31 specifically includes: Step S311, the time-varying dynamic stage of potential parameters; Two adaptive control parameters are defined: dynamic tolerance threshold W(t) and dynamic stiffness coefficient k(t), where W(t) is used to control the compliance range of the potential energy deviation, and k(t) is used to adjust the strength of the potential energy constraint. The residual statistics of physical potential energy and statistical potential energy are tracked using exponential moving average, and the calculation form is as follows: ; Where t represents the training epoch, B represents the batch size, and i represents the i-th sample. It is a smoothing factor; This indicates whether j comes from physical potential energy or statistical potential energy; phys represents physical potential energy, and stat represents statistical potential energy. An inverse stiffness scheduling strategy is adopted to adjust the physical stiffness. k phys and statistical stiffness k stat Dynamic adjustments are made: In the early stages of training, the physical stiffness... k phys Gradually increase the physical stiffness to suppress gradient explosion in the early stages and avoid premature physical constraints; in the later stages of training, increase the physical stiffness. k phys Maintain strict physical compliance, statistical stiffness k stat Appropriate relaxation; Step S312: Construct a dynamic potential energy surface; Based on the dynamic tolerance threshold W(t), the dynamic stiffness coefficient k(t), and the inverse stiffness scheduling strategy, a high-stiffness dynamic potential energy surface in the form of a quadratic piecewise segment is constructed, and the total dynamic potential energy function is... The calculation formula is: in, The current potential energy loss is represented by the "+" sign, indicating a physical hardening process. The symbol "" represents the controlled relaxation process of the statistical boundary; Step S313, the co-evolution stage of the joint potential field; Based on the high-stiffness dynamic potential energy surface in the form of a second-order piecewise segment, the dynamic physical potential... and dynamic statistical potential It follows an evolutionary process of "anchor-transition-correction".

8. The cable defect data augmentation method based on physical consistency constraints according to claim 7, characterized in that, Step S32 specifically includes: Step S321: Construct the adaptive quality matrix; A diagonally time-varying adaptive mass matrix M(t) proportional to the local potential stiffness is constructed. A geometric inertia matching strategy is employed. The adaptive mass matrix is ​​defined as follows: Where, diag represents the diagonal matrix form, η is the inertia scaling factor, and ε is a small constant to prevent numerical singularity; Step S322, Extended phase space and Hamiltonian dynamics evolution stage; Introducing auxiliary momentum variables p ~N(0, M ( t Expanding the original d-dimensional latent space into a 2d-dimensional phase space, the Hamiltonian is defined as follows: According to the Hamiltonian equation, it evolves as follows: in, For the simulation time step in HMC sampling, This represents gradient operation. V total,dyn The total dynamic potential energy; Step S323, Discretization and Metropolis Correction Stage; Discretization is performed using a time-reversible leapfrog integrator. After L integrations with a step size δ, candidate states are obtained. The accepted candidate states are then calculated according to the Metropolis criterion. ; in, Let be the probability of accepting the sample, and exp be the exponential operation. Original Hamiltonian energy, For the updated Hamiltonian energy; Step S324: Generate defect enhancement samples; Perform adaptive quality Hamiltonian Monte Carlo sampling to obtain latent variables. Then, latent variables The trained PCDH-VAE generative model is input, and the decoder of the PCDH-VAE generative model, guided by the high-stiffness dynamic potential energy surface formed by the dual potential function constraints, converts the latent variables... Map back to the data space to generate defect-enhanced samples.

9. A cable defect data enhancement system based on physical consistency constraints, characterized in that, include: Data preprocessing module (1), generative model building module (2), sample generation module (3) and sample evaluation module (4); The data preprocessing module (1) is used to preprocess the original cable monitoring data, divide the dataset and balance it, and output a balanced training set and a test set that maintains the original distribution. The generative model building module (2) is used to build a PCDH-VAE generative model. It encodes observable vectors and conditional labels based on a conditional variational autoencoder network, embeds a dual potential function in the latent space, constructs a Hamiltonian energy function that integrates dual potential function constraints, introduces an adaptive kinetic energy term, and unifies the training objective of the PCDH-VAE generative model into a Hamiltonian energy minimization problem. The sample generation module (3) introduces a dynamic potential well mechanism to adaptively adjust the physical potential energy and statistical potential energy intensity to form a high-stiffness dynamic potential energy surface. It adopts adaptive mass Hamiltonian Monte Carlo sampling and generates defect-enhanced samples that meet multi-physics constraints through Hamiltonian dynamic evolution and Metropolis correction. The sample evaluation module (4) is used to evaluate the generated defect enhancement samples, merge the defect enhancement samples with the training set, construct a dataset for downstream tasks, train the downstream defect identification model, and perform high-precision cable defect diagnosis.

10. A cable defect data enhancement system based on physical consistency constraints according to claim 9, characterized in that, The data preprocessing module (1) is used to preprocess the original data of cable multi-physical field monitoring to generate cable multi-physical field monitoring samples. The cable multi-physical field monitoring samples are divided into training set and test set according to a 7:3 ratio. In order to address the problem of unbalanced distribution of training set categories, the inverse frequency weight coefficient is calculated based on the number of samples in each category, so that the sample selection probability is inversely proportional to the category frequency. During the batch construction process, random sampling is performed according to the weight to form a balanced training set with a category distribution that tends to be 1:1:1:

1. The generative model building module (2) is used to build the PCDH-VAE generative model. It adopts a conditional variational autoencoder network as a sub-model, performs joint input encoding on the observable vector and conditional label, and obtains latent variables in the latent space. It embeds a dual potential function in the latent space, introduces momentum variables of the same dimension, and constructs a Hamiltonian energy function. It adjusts the scale of each latent dimension through the mass matrix. It combines the physical potential energy term, the statistical potential energy term and the kinetic energy term in a weighted combination to form the total Hamiltonian energy function. It takes the minimization of the total Hamiltonian energy as the model optimization objective, and jointly updates the encoder parameters and decoder parameters. By introducing the physical potential energy term and the statistical potential energy term in the latent space, it is unified in the Hamiltonian energy modeling framework, so that the PCDH-VAE generative model can simultaneously satisfy physical constraints and data distribution constraints during the training process. The sample generation module (3) introduces a dual-field dynamic potential well mechanism with the training period as the evolution variable, sets a dynamic tolerance threshold W(t) and a dynamic stiffness coefficient k(t), and dynamically adjusts the physical stiffness and statistical stiffness in combination with an inverse stiffness scheduling strategy. A high-stiffness dynamic potential energy surface is constructed based on dynamic tolerance threshold, dynamic stiffness coefficient, and stiffness scheduling results. An adaptive mass Hamiltonian Monte Carlo sampling method is then built on the high-stiffness dynamic potential energy surface. A mass matrix is ​​set, and auxiliary momentum variables are introduced to extend to the phase space. Dynamic evolution is performed according to the Hamiltonian equation, and discretization calculation is carried out. The results are corrected by the Metropolis criterion to obtain latent variables that satisfy the target distribution. The latent variables are input into the decoder and mapped to the data space under the constraints of the dynamic potential energy surface to generate defect enhancement samples that satisfy multiphysics constraints. The sample evaluation module (4) is used to construct a three-dimensional comprehensive evaluation index system covering statistical fidelity, physical consistency and diagnostic effectiveness, to conduct a systematic quality evaluation of the generated defect enhancement samples, to merge the qualified defect enhancement samples with the original training set, to construct an expanded defect recognition training set, to train the downstream defect diagnosis model, and to quantitatively verify the training results through macro F1 scores.

Images