CN-121997053-A - Cable defect data enhancement method and system based on physical consistency constraint

CN121997053ACN 121997053 ACN121997053 ACN 121997053ACN-121997053-A

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, wherein the system outputs a balanced training set and a test set through a data preprocessing module; the system comprises a generation model construction module, a sample generation module, a sample evaluation module and a sample collection module, wherein the generation model construction module is used for encoding an observable vector and a condition label based on a condition variation self-encoding network, embedding a double-field potential function in a potential space, constructing a Hamiltonian energy function fused with double-potential-field-function constraint, introducing an adaptive kinetic energy term, unifying PCDH-VAE generation model training targets into Hamiltonian energy minimization problem, introducing a dynamic potential well mechanism, adopting adaptive quality Hamiltonian Monte Carlo sampling, generating a defect enhancement sample meeting multiple physical-field constraint through Hamiltonian dynamic evolution and Metropolis correction, and evaluating the generated defect enhancement sample set by the sample evaluation module. The invention can relieve the problem of scarcity of cable defect data samples.

Inventors

LI QINGQUAN
LIU ZHANQING
MENG FANBO
MA NUO
ZHAO SHIBO
WU WANHAO
Dong Yetong
YANG GUANGDI

Assignees

山东大学

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (10)

1. The cable defect data enhancement method based on physical consistency constraint is characterized by comprising the following steps of: step S1, acquiring cable multi-physical field monitoring original data, dividing the data into a training set and a testing set according to a proportion after data preprocessing, and carrying out inverse frequency weighted sampling on the training set to obtain a balanced training set with balanced categories; step S2, constructing a PCDH-VAE generation model, taking a conditional variation self-coding network as a sub-model, and coding an observable vector and a conditional label, embedding a double-field potential function into a potential space of the conditional variation self-coding network, constructing a Hamiltonian energy function, unifying a model optimization target as a Hamiltonian energy minimization problem, and introducing an adaptive kinetic energy term to form a complete energy modeling framework; S3, inputting a balanced training set into a PCDH-VAE generation model, introducing a dynamic potential well mechanism, adaptively adjusting a physical potential energy item and a statistical potential energy item through a time-varying tolerance threshold value and a rigidity coefficient to form a high rigidity dynamic potential energy surface, adopting a self-adaptive mass Hamiltonian Monte Carlo sampling method on the high rigidity dynamic potential energy surface, realizing potential space stable sampling through coupling of a mass matrix and local rigidity, executing Hamiltonian dynamics evolution and Metropolis correction, and obtaining corrected potential variables ; Potential variables Inputting the decoder to generate a defect enhancement sample meeting the multi-physical field constraint; And S4, evaluating the quality of the defect enhancement sample, combining the defect enhancement sample with a training set, constructing an extended training set, training a downstream defect diagnosis model, and verifying the enhancement effect through the classification performance index.
2. The method for enhancing cable defect data based on physical consistency constraints as claimed in claim 1, wherein in said step S1, the cable multi-physical-field monitoring raw data comprises core temperature Sheath temperature Leakage current Ground loop current The insulation resistance R, the dielectric loss tangent tan delta and the partial discharge quantity Q are 7 observable variables in total; Preprocessing original data of cable multi-physical field monitoring to generate a cable multi-physical field monitoring sample, wherein the cable multi-physical field monitoring sample comprises four operating states including normal operation, main insulation abnormality, joint overheating and sheath damage, 7 observable variables are generated, 7 observable vectors x are generated, and four categories are generated according to the four operating states; And when data is divided, a layering random extraction mode is adopted, a cable multi-physical field monitoring sample is divided into a training set and a testing set according to a ratio of 7:3, an inverse frequency weighted sampling mechanism is introduced, and a balanced training set with category distribution tending to 1:1:1:1 is generated.
3. The method for enhancing cable defect data based on physical consistency constraints as claimed in claim 1, wherein said step S2 comprises: s21, constructing a condition variation self-coding network; The observable vector x and the defect type label c are combined and encoded to form a conditional input vector, the conditional input vector is sent to an encoder of a conditional variation self-encoding network, the encoder outputs parameters of potential distribution through multi-layer nonlinear mapping, and a conditional approximate posterior distribution is established Wherein z is a latent variable, q is an approximate posterior distribution, and φ is an encoder parameter; the potential variable z can be micro-sampled through re-parameterization, and the sampled potential variable z and the defect type label c are input into a decoder in a combined way; S22, constructing a double-field potential function; Introducing a double-field potential function formed by physical potential energy items and statistical potential energy items into a potential space of a condition variation self-coding network, wherein the physical potential energy item construction comprises a thermal potential construction sub-stage, an electrical residual construction sub-stage and a total physical potential energy construction sub-stage; the process of constructing the statistical potential energy item is that in the training stage, the negative evidence lower bound is adopted as the statistical potential, and KL divergence between the approximate posterior distribution and the prior distribution is calculated; In the generation stage, a statistical potential based on moment is introduced, and the statistical properties of the defects are maintained through the mean value, standard deviation and shape factor of each dimension of input data: ; where D is the dimension of the input data, Is the value of the sample in the j-th dimension, , And Mean, standard deviation and shape factor of the jth feature under the target defect condition c are respectively represented; s23, constructing a Hamiltonian energy function; Introducing a momentum variable p with the same dimension on the basis of a potential variable z to form an expansion state variable (z, p), expressing total energy as the sum of potential energy items and kinetic energy items, introducing a self-adaptive kinetic energy item, and performing scale adjustment on different potential dimensions by setting the momentum variable p to obey zero-mean Gaussian distribution and combining a quality matrix; S24, constructing a PCDH-VAE generation model optimization target; unifying PCDH-VAE generation model optimization targets into Hamiltonian energy minimization problems; s25, constructing a PCDH-VAE generation model; Uniformly integrating an encoder, a decoder, a double-field potential function and a Hamiltonian energy function, so that the PCDH-VAE generation model forms an energy distribution structure under the constraint and combined action of the double-field potential function in a potential space; The potential space dimension of the PCDH-VAE generation model is set to be 16, the encoder adopts a two-layer network structure, the number of hidden units is respectively 64 and 32, the decoder adopts a symmetrical two-layer network structure, and the number of hidden units is respectively 32 and 64.
4. A method for enhancing cable defect data based on physical uniformity constraints according to claim 3, wherein said thermal potential building sub-stage comprises passing core temperature Sheath temperature And ambient temperature Calculating a thermal residual and expressing the thermal residual as a thermal potential : Wherein the method comprises the steps of 、 And Respectively representing core temperature, sheath temperature and ambient temperature, parameters Represents the thermal conductivity of the material, The equivalent thermal resistance thickness is represented, and h represents the convective heat transfer coefficient; The electrical residual construction sub-stage is that dielectric loss is characterized by leakage current Insulation resistance R and dielectric loss tangent Intrinsic coupling between, δ is the dielectric loss angle of the cable insulation, and electrical residue is defined as: wherein U represents an operating voltage; the total physical potential energy construction sub-stage is to combine the thermal potential and the electrical residual according to the weighted sum to obtain the total physical potential energy.
5. The method for enhancing cable defect data based on physical consistency constraints of claim 4, wherein the negative evidence lower bound is: , wherein, KL divergence, which is used to measure the distance between two distributions; is an approximate posterior distribution of the data, Is a priori distribution; Is a decoder and is provided with a signal processing unit, Is a balance factor; step S23 includes: The physical potential energy items, the statistical potential energy items and the self-adaptive kinetic energy items in the double-field potential function are combined in a weighting manner, a total Hamiltonian energy function is constructed, the total Hamiltonian quantity H is obtained, an auxiliary momentum variable p is introduced on the basis of a potential variable z in a Hamiltonian energy function frame, the total Hamiltonian quantity is constructed to be the sum of the total physical potential energy items and the self-adaptive kinetic energy items, ; Wherein, the total physical potential energy item The physical potential energy item and the statistical potential energy item are combined, and can be expressed as: ; Wherein, the Representing a physical potential energy item; representing a statistical potential energy term, the adaptive potential energy term is formed by defining a momentum variable p with the same dimension as a potential variable z in a potential space and setting the momentum variable p to follow zero-mean Gaussian distribution: 。
6. The method for enhancing cable defect data based on physical consistency constraints as claimed in claim 5, wherein said step S3 comprises: step S31, introducing a double-field dynamic potential well mechanism; The double-field dynamic potential well mechanism takes a training period t as an evolution variable, converts static physical constraint and statistical constraint into dynamic energy landscape which changes along with time through homotopy transformation, and carries out self-adaptive adjustment on the physical constraint and statistical constraint intensity; Step S32, a self-adaptive quality Hamiltonian Monte Carlo sampling stage; Introducing a self-adaptive mass Hamiltonian Monte Carlo sampling strategy to perform potential space sampling, realizing geometric inertia matching through coupling of a mass matrix and local rigidity, and performing Hamiltonian dynamics evolution and Metropolis correction in an expanded phase space to obtain potential variables Latent variable The decoder of the PCDH-VAE generating model is input and the defect enhancement sample is output.
7. The method for enhancing cable defect data based on physical consistency constraint according to claim 6, wherein said step S31 specifically comprises: Step S311, a time-varying dynamic phase of the potential parameter; Defining two self-adaptive control parameters, namely a dynamic tolerance threshold W (t) and a dynamic stiffness coefficient k (t), wherein W (t) is used for controlling a compliant range of potential energy allowable deviation, and k (t) is used for adjusting the strength of potential energy constraint; Tracking the residual statistic of the physical potential energy and the residual statistic of the statistical potential energy by adopting an exponential moving average, wherein the calculation form is as follows: ; wherein t represents the training round, B represents the batch size, i represents the ith sample, Is a smoothing factor; whether j is from physical potential or statistical potential, phys is physical potential, and stat is statistical potential; The inverse stiffness scheduling strategy is adopted to dynamically adjust the physical stiffness k phys and the statistical stiffness k stat , wherein the physical stiffness k phys is gradually increased in the early stage of training, the gradient explosion in the early stage is restrained, and the premature physical constraint is avoided; Step S312, constructing a dynamic potential energy surface; based on a dynamic tolerance threshold W (t), a dynamic stiffness coefficient k (t) and a reverse stiffness scheduling strategy, constructing a high-stiffness dynamic potential energy surface in a secondary sectional form, and a total dynamic potential energy function The calculation formula is as follows: Wherein, the For the current potential energy loss, "+" sign indicates the physical hardening process, " The symbol represents a controlled relaxation process of the statistical boundary; step S313, the co-evolution stage of the combined potential field; on the basis of the high-rigidity dynamic potential energy surface in the secondary sectional form, dynamic physical potential Dynamic statistics of potential The evolution process of "anchor-transition-correction" was followed.
8. The method for enhancing cable defect data based on physical consistency constraints as claimed in claim 7, wherein said step S32 specifically comprises: S321, constructing a self-adaptive quality matrix; Constructing a diagonal time-varying adaptive mass matrix M (t) proportional to the local potential stiffness, and adopting a geometric inertia matching strategy, wherein the adaptive mass matrix is defined as: wherein diag represents a diagonal matrix form, eta is an inertial scale factor, and epsilon is a tiny constant for preventing numerical singularities; Step S322, expanding a phase space and a Hamiltonian dynamics evolution stage; Introducing auxiliary momentum variables p-N (0, M (t)), expanding an original d-dimensional potential space into a 2 d-dimensional phase space, and defining the obtained Hamiltonian amount as follows: The Hamiltonian equation is developed as follows: Wherein, the For the analog time step in the HMC sampling, Representing gradient operation, V total,dyn is total dynamic potential energy; step S323, discretizing and Metropolis correction stage; Discretizing by using a time-reversible frog-leaping integrator, performing L times of integration by a step delta to obtain candidate states, and calculating an acceptance candidate state according to a Metropolis criterion: ; Wherein, the For the sample reception probability, exp is an exponential operation, Is the original Hamiltonian energy, and the energy is the original Hamiltonian energy, The updated Hamiltonian energy; step S324, generating a defect enhancement sample; completing self-adaptive quality Hamiltonian Monte Carlo sampling to obtain potential variables Thereafter, the latent variable Inputting the trained PCDH-VAE generation model, and leading a decoder of the PCDH-VAE generation model to guide potential variables under the high-rigidity dynamic potential energy surface formed by double-field potential function constraint Mapping back to the data space generates defect enhanced samples.
9. The cable defect data enhancement system based on physical consistency constraint is characterized by comprising a data preprocessing module (1), a generation model construction module (2), a sample generation module (3) and a sample evaluation module (4); The data preprocessing module (1) is used for preprocessing, dividing a data set and balancing the original cable monitoring data, and outputting a balanced training set and a test set keeping original distribution; The generating model construction module (2) is used for constructing a PCDH-VAE generating model, encoding an observable vector and a condition label based on a condition variation self-coding network, embedding a double-field potential function in a potential space, constructing a Hamiltonian energy function fused with constraint of the double-potential function, introducing a self-adaptive kinetic energy item, and unifying a PCDH-VAE generating model training target into a Hamiltonian energy minimization problem; The sample generation module (3) introduces a dynamic potential well mechanism, adaptively adjusts physical potential energy and statistical potential energy intensity to form a high-rigidity dynamic potential energy surface, adopts adaptive mass Hamiltonian Monte Carlo sampling, and generates a defect enhancement sample meeting multiple physical field constraints through Hamiltonian dynamic evolution and Metropolis correction; The sample evaluation module (4) is used for evaluating the generated defect enhancement sample, combining the defect enhancement sample with the training set, constructing a data set for a downstream task, training a downstream defect recognition model and performing high-precision cable defect diagnosis.
10. The system for enhancing cable defect data based on physical consistency constraints as recited in claim 9, The data preprocessing module (1) is used for preprocessing the original data of the cable multi-physical field monitoring to generate cable multi-physical field monitoring samples, dividing the cable multi-physical field monitoring samples into a training set and a testing set according to a ratio of 7:3, aiming at the problem of unbalanced distribution of the category of the training set, calculating an inverse frequency weight coefficient according to the number of samples of each category to enable the sample selection probability to be inversely proportional to the category frequency, and randomly extracting according to the weight in the batch construction process to form a balanced training set with the category distribution tending to 1:1:1:1; The generation model construction module (2) is used for constructing a PCDH-VAE generation model, adopting a condition variation self-coding network as a sub-model, carrying out joint input coding on an observable vector and a condition label to obtain a potential variable in a potential space, embedding a double-field potential function in the potential space, introducing a homodimensional momentum variable, constructing a Hamiltonian energy function, carrying out scale adjustment on each potential dimension through a quality matrix, carrying out weighted combination on a physical potential energy item, a statistical potential energy item and a kinetic energy item to form a total Hamiltonian energy function, carrying out joint update on encoder parameters and decoder parameters by taking total Hamiltonian energy minimization as a model optimization target, and unifying the physical potential energy item and the statistical potential energy item in the potential space to the Hamiltonian energy modeling frame so that the PCDH-VAE generation model synchronously meets physical constraint and data distribution constraint in the training process; The sample generation module (3) introduces a double-field dynamic potential well mechanism taking a training period as an evolution variable, sets a dynamic tolerance threshold W (t) and a dynamic stiffness coefficient k (t), and dynamically adjusts physical stiffness and statistical stiffness by combining an inverse stiffness scheduling strategy; constructing a high-rigidity dynamic potential energy surface based on a dynamic tolerance threshold, a dynamic rigidity coefficient and a rigidity scheduling result, constructing a self-adaptive mass Hamiltonian Monte Carlo sampling on the high-rigidity dynamic potential energy surface, setting a mass matrix, introducing an auxiliary momentum variable to expand to a phase space, executing dynamic evolution according to a Hamiltonian equation, performing discretization calculation, correcting by a Metropolis criterion to obtain potential variables meeting target distribution, inputting the potential variables into a decoder, mapping the potential variables to a data space under the constraint of the dynamic potential energy surface, and generating a defect enhancement sample meeting the constraint of multiple physical fields; The sample evaluation module (4) is used for constructing a three-dimensional comprehensive evaluation index system covering statistical fidelity, physical consistency and diagnosis effectiveness, performing systematic quality evaluation on the generated defect enhancement samples, combining the defect enhancement samples qualified in evaluation with the original training set, constructing an expanded defect identification training set, and performing quantitative verification on training results through macroscopic F1 scores, wherein the training set is used for training a downstream defect diagnosis model.

Description

Cable defect data enhancement method and system based on physical consistency constraint Technical Field The invention relates to the field of cable defect data enhancement, in particular to a method and a system for enhancing cable defect data based on physical consistency constraint. Background The high-voltage cable is a key component of the power system, and accurate identification of early defects of the high-voltage cable is important to ensuring safe and stable operation of a power grid. At present, a defect identification method based on data driving depends on a large number of defect samples with labels, however, the defect samples are seriously scarcity in actual operation, so that model training is insufficient and generalization capability is limited. In the prior art, a method and a system for detecting cable defects in a cable pit based on machine vision are disclosed, wherein the method and the system are used for generating network-reinforced cable image data through deep convolution, improving sample diversity and model generalization capability, designing a multi-dimensional weak attention residual error network, combining a weak signal attention module and a multi-scale perception module, and strengthening the capturing capability of small defects and weak features in the cable pit, thereby improving the accuracy and the robustness of visual detection. The technical scheme has improved the diversity of image data and the capability of detecting details, but still has the following technical problems caused by the fact that the physical rule is not fused, the physical untrustworthy of generating data under the condition of scarce samples, the multi-physical field coupling relation is not enough, the training stability is poor and the cable is disconnected from the actual physical state. In view of the foregoing, it is desirable to provide a method and system for enhancing cable defect data based on physical consistency constraints to address the above-mentioned drawbacks of the prior art. Disclosure of Invention The invention aims to solve the technical problems of unreliable generated data physics, insufficient maintenance of multi-physics field coupling relation, poor training stability and dislocation with the actual physical state of a cable under the condition of sparse samples in the prior art of cable defect identification, and provides a cable defect data enhancement method and system based on physical consistency constraint, so as to solve the technical problems in the prior art. In order to achieve the above purpose, the present invention provides the following technical solutions: in a first aspect, the present invention provides a method for enhancing cable defect data based on physical consistency constraints, comprising the steps of: step S1, acquiring cable multi-physical field monitoring original data, dividing the data into a training set and a testing set according to a proportion after data preprocessing, and carrying out inverse frequency weighted sampling on the training set to obtain a balanced training set with balanced categories; step S2, constructing a PCDH-VAE generation model, taking a conditional variation self-coding network as a sub-model, and coding an observable vector and a conditional label, embedding a double-field potential function into a potential space of the conditional variation self-coding network, constructing a Hamiltonian energy function, unifying a model optimization target as a Hamiltonian energy minimization problem, and introducing an adaptive kinetic energy term to form a complete energy modeling framework; S3, inputting a balanced training set into a PCDH-VAE generation model, introducing a dynamic potential well mechanism, adaptively adjusting a physical potential energy item and a statistical potential energy item through a time-varying tolerance threshold value and a rigidity coefficient to form a high rigidity dynamic potential energy surface, adopting a self-adaptive mass Hamiltonian Monte Carlo sampling method on the high rigidity dynamic potential energy surface, realizing potential space stable sampling through coupling of a mass matrix and local rigidity, executing Hamiltonian dynamics evolution and Metropolis correction, and obtaining corrected potential variables ; Potential variablesInputting the decoder to generate a defect enhancement sample meeting the multi-physical field constraint; And S4, evaluating the quality of the defect enhancement sample, combining the defect enhancement sample with a training set, constructing an extended training set, training a downstream defect diagnosis model, and verifying the enhancement effect through the classification performance index. In a second aspect, the present invention also provides a system for enhancing cable defect data based on physical consistency constraints, comprising: The data preprocessing module is used for preprocessing, dividing a data set and balancing the origina