Infrared synthetic data and domain adaptive method based on digital twinning

Abstract

The invention relates to the technical field of infrared thermal image detection and digital twinning of power equipment, and provides an infrared synthetic data and domain self-adaption method based on digital twinning, which comprises the following steps of: constructing a digital twinning body of the power equipment, and aligning parameters of the digital twinning body with statistical characteristics of a real infrared sample through reverse calibration; performing multi-physical field thermal simulation to generate a temperature field; performing radiation transfer and imaging chain synthesis based on the temperature field to generate a synthesized infrared image; automatically marking the synthesized infrared image to obtain synthesized data; performing multi-target domain adaptive training by using the synthetic data as a source domain and real infrared data as a target domain to obtain an adaptive model; using the adaptive model to generate a confidence pseudo label for the target domain data, and performing gradual self-training based on the confidence pseudo label to update the adaptive model; during testing, self-adaption is carried out when testing is carried out on data without target domains, and uncertainty is output; and adding an auditing label to the output result to form a traceable evidence chain.

Description

Technical Field

[0001] This invention relates to the field of infrared thermal imaging inspection and digital twin technology for power equipment, and in particular to an infrared synthetic data and domain adaptive method based on digital twins. Background Technology

[0002] In existing technologies, intelligent diagnostics of power equipment using infrared thermal imaging mainly follows two paths: one focuses on image domain synthesis and enhancement, directly generating infrared images through generative models such as style transfer and GAN / CycleGAN. However, this approach lacks explicit physical modeling of the "temperature field—radiative transfer—camera response" link. It fails to refine key aspects such as directional emissivity under different materials and incident angles, atmospheric transmittance and path radiation, pixel field of view (IFOV) and point spread function, camera radiometric calibration, and non-uniform noise. This results in a lack of physical reversibility between radiance and surface temperature. Estimation and defect classification are unavailable; another type of supervised learning and domain adaptation method relies on real data-driven approaches, but power line inspection annotation is costly, the scenario evolves, and there are significant differences across devices. Without hard physical constraints, domain adaptation is prone to learning "style bias," and model stability decreases significantly when observation distance, viewing angle, weather conditions, or load status change. Furthermore, existing technologies generally suffer from insufficient auditability and interpretability. Most processes only output the final result, lacking intermediate physical quantities such as temperature field, radiation field decomposition, and transmittance estimation for verification. Moreover, for multi-target domain migration across different machine types, sites, and seasons, a unified global alignment strategy is often used, making it difficult to simultaneously handle target distribution drift, conditional distribution drift, and label imbalance. There is a lack of specific... The physical structured loss design for regression and defect classification tasks makes the model prone to "apparent accuracy but physical distortion" when crossing domains, making it difficult to maintain the consistency and reliability of key indicators in real inspection environments. Summary of the Invention

[0003] To address the aforementioned shortcomings, this invention aims to propose an infrared synthetic data and domain adaptation method based on digital twins. The method constructs a full-link digital twin encompassing "temperature field—radiative transfer—camera response," generating physically consistent labeled synthetic data through reverse calibration. It integrates physical constraints and multi-target domain adaptation to achieve robust transfer across devices, distances, viewing angles, and weather conditions. By self-training and adaptively testing with confident pseudo-labels, it reduces label dependency and adapts to deployment drift. Finally, it outputs auditable temperature / / Defect / uncertainty results enable traceable infrared diagnostics.

[0004] To achieve this objective, the present invention adopts the following technical solution: An infrared synthetic data and domain adaptation method based on digital twins includes the following steps: S1: Construct a digital twin of the power equipment and align the parameters of the digital twin with the statistical characteristics of the real infrared sample through reverse calibration; S2: Perform multiphysics thermal simulation based on the digital twin to generate a temperature field; S3: Based on the temperature field, perform radiative transfer and imaging chain synthesis to generate a synthetic infrared image; S4: Automatically label the synthesized infrared image to obtain synthesized data with physical tags; S5: Using the synthetic data with physical labels as the source domain and the real infrared data as the target domain, perform multi-target domain adaptive training to obtain an adaptive model; S6: Use the adaptive model to generate confidence pseudo-labels for the target domain data, and perform stepwise self-training based on the confidence pseudo-labels to update the adaptive model; S7: During testing, adaptive testing is performed on data from unseen target domains, and the uncertainty is output. S8: Attach audit tags to the output results to form a traceable chain of evidence.

[0005] Preferably, step S1 includes: Construct a digital twin including geometry, materials, operating conditions, and camera response parameters; parameter set Including material thermal parameters such as thermal conductivity ,density Specific heat capacity directional emissivity Atmospheric extinction coefficient And camera response and noise parameters; Based on the statistical characteristics of several real infrared samples Alignment Synthesis Statistical Features Optimal parameters Satisfying the relation: ; in, Represents the distribution distance metric. Indicates in the parameter The statistical characteristics of the synthetic data are as follows: Represents the statistical characteristics of a real sample; First, fix the thermal parameters and adjust the radiation and imaging chain parameters, then fine-tune the heat source and boundary coefficient until the distribution distance is reached. Less than the threshold .

[0006] Preferably, step S2 includes: Under given load and environmental boundary conditions, the heat diffusion equation and boundary heat transfer equation are solved to generate a temperature field, wherein the heat diffusion equation satisfies the following relationship: ; The boundary heat transfer equation satisfies the following relationship: ; in, Indicates the density of the material. Indicates specific heat capacity. Represents the temperature field. Indicates time, Indicates thermal conductivity. Indicates the body's heat source. Represents the spatial derivative along the normal direction. Indicates the convective heat transfer coefficient. Indicates ambient temperature. Represents the Stefan-Boltzmann constant. Indicates surface emissivity; The heat diffusion equation is solved discretically using an implicit time-progression method, and the discretized form satisfies the following relationship: ; in, Indicates the time step. Represents the heat capacity matrix. Represents the thermal conductivity and boundary term stiffness matrix. Represents the equivalent heat source vector. and These represent the temperature vectors at the current time step and the next time step, respectively. By locally increasing the contact resistance Or reduce thermal conductivity Injecting defects in a manner that generates multi-level temperature differences The sample.

[0007] Preferably, step S3 includes: Radiative transfer and imaging chain synthesis are performed based on the temperature field, wherein atmospheric transmittance Satisfying the relation: ; Path radiation Satisfying the relation: ; in, Indicates the distance of propagation Atmospheric transmittance below Indicates the extinction coefficient. Indicates path radiation. This represents the bandwidth-averaged Planck radiation. Indicates the equivalent atmospheric temperature; Radiation reaching the detector Satisfying the relation: ; Camera code value Satisfying the relation: ; Among them, total noise Satisfying the relation: ; in, This indicates the amount of radiation reaching the detector. Indicates the angle of incidence Directional emissivity below Indicates environmental reflected radiation. Indicates camera code value, Represents the quantization operator. and These represent camera gain and bias, respectively. Indicates as distance Scaled IFOV / PSF convolution kernels, This represents the convolution operation. Indicates total noise. This indicates non-uniform response noise. This indicates noise due to unevenness in the dark field. Represents equivalent temperature difference noise. This represents point-by-point multiplication; Output a synthesized code value map and a complete physical tag that match the target camera specifications.

[0008] Preferably, the step S3 includes the following: While maintaining temperature order and relative temperature difference ΔT, the appearance style of the synthesized infrared image is mapped to the real domain, where the mapping G from the synthesized domain to the real domain and the mapping F from the real domain to the synthesized domain are defined. Style bridging is achieved by minimizing the following objective function: ; Among them, physical consistency regularization term Satisfying the relation: ; in, This represents the mapping function from the synthetic domain to the real domain. This represents the mapping function from the real domain to the synthetic domain. Indicating resistance to loss, , and represent the weight coefficients of the cycle consistency loss, identity loss, and physical consistency regularization term, respectively. Describing the L1 norm, This represents a physical consistency regularization term. This represents the inversion operator from radiation to temperature. Indicates in the region Temperature order mapping function within, This represents the gradient operator along the tangent of the isothermal surface. This indicates the rank constraint evaluation region.

[0009] Preferably, step S4 includes: Pixel-level and target-level physical tags and metadata are directly derived from the digital twin, wherein the tag set Satisfying the relation: ; Metadata collection Satisfying the relation: ; in, Indicates the temperature field distribution. Indicates the temperature difference distribution. This represents a defect segmentation mask. This represents the equivalent heat source distribution map. Indicates the imaging distance. Indicates the imaging perspective; Latin hypercube sampling or stratified sampling methods were used to cover the parameter space of machine type, distance, viewing angle, operating condition, and season, and different The levels are resampled and balanced to construct a source domain training set and a target domain unlabeled set.

[0010] Preferably, step S5 includes: Using the synthetic data with physical tags as the source domain Using real infrared data as the target domain Through feature extraction network Simultaneously outputs defect segmentation and temperature difference. The regression and classification results show that the overall loss function satisfies the following relationship: ; in, This includes classification cross-entropy loss, Dice loss, and... The source domain supervised loss of the regression Huber loss, Indicates domain adversarial loss, This represents the statistical distribution matching term of the maximum mean difference. This represents a physical consistency regularization term. , and Indicates the weighting coefficient. and These represent samples from the source and target domains, respectively. This represents a feature extraction network; The domain adversarial loss Satisfying the relation: ; in, Discriminator for the domain, Represents the expectation operator. and These represent samples from the source and target domains, respectively. The statistical distribution matching item Satisfying the relation: ; in, Represents the kernel mapping function. Denotes the Hilbert space norm of the regenerating kernel. and These represent the number of samples in the source domain and the target domain, respectively. and They represent the first The source domain sample and the first Features of each target domain sample; The physical consistency regularization term Satisfying the relation: ; in, , and This represents the regularization weight coefficient. Indicates the temperature order maintenance item. Indicates the isothermal surface smoothing term. This represents the radiation field or equivalent quantity obtained through inversion. Represents the distance-dependent point spread function / instantaneous field-of-view convolution kernel; Use metadata Conditional modulation is performed using characteristic linear modulation or conditional normalization methods to characterize equipment, attitude, and weather differences.

[0011] Preferably, step S6 includes: Generate pseudo-tags in the target domain. Through the quantile threshold of uncertainty or residual To control the quality of pseudo-tags, the accept function must satisfy the following relation: ; The pseudo-labeled dataset satisfies the following relation: ; in, The pseudo-labels representing the target domain samples. This represents the uncertainty measurement function. This represents a threshold based on quantiles. Indicates an indicator function, This represents the accepted pseudo-labeled dataset; Gradually relax threshold parameters using a course-based learning approach. To expand the pseudo-label dataset And through weights The pseudo-label supervision loss is accumulated into the overall training objective, and the update relation satisfies: ; in, This represents the loss weight for pseudo-label supervision. This represents the supervised loss for false-labeled samples. This indicates the overall training objective.

[0012] Preferably, step S7 includes: During the deployment phase, when performing lightweight testing on unseen target domain data, an adaptive approach is adopted. This is achieved by freezing backbone network parameters and updating only normalized layer parameters and small output header parameters to minimize prediction entropy, optimizing the objective to satisfy the following relation: ; in, This indicates the normalization layer and small output header parameters to be updated. Indicates category The predicted probability, This represents the summation over all categories. This indicates taking the expectation of the samples in the target domain; Uncertainty is estimated using heteroscedastic regression combined with Monte Carlo Dropout or deep ensemble methods. The heteroscedastic regression loss function satisfies the following relationship: ; in, Represents the actual value. This represents the model's predicted value. Indicates the prediction variance. Represents square Norm; predicted mean and cognitive uncertainty Obtained through statistical analysis of multiple forward propagations; Output set and model version information and timestamp, among which This represents a temperature difference distribution map. This represents a defect segmentation mask. This represents a temperature distribution map. This represents the uncertainty distribution diagram. Represents metadata.

[0013] Preferably, step S8 includes: Add an audit label to the output results, the audit label Satisfying the relation: ; in, Indicates the model identifier and version number. Indicates the site identifier. Indicates the imaging distance. Indicates the imaging perspective. Indicates ambient temperature. Indicates relative humidity. Indicates wind speed. Indicates calibration marking, Represents a timestamp; When comparing across different models or seasons, physical consistency indicators are included, such as temperature order-related indicators, reprojection error, and the proportion of equivalent temperature difference noise. When the physical consistency indicators do not reach a preset threshold, a prompt for review or minor calibration is generated.

[0014] One of the above technical solutions has the following advantages or beneficial effects: This invention utilizes a digital twin modeling and synthesis workflow constructed through S1 to S4. Through reverse calibration, it aligns the twin parameters with the real statistical distribution. Based on explicit modeling of multiphysics thermal simulation and radiative transfer imaging chains, it incorporates physical factors such as material emissivity, atmospheric transmittance, and camera response noise into the synthesis process, directly outputting a temperature field... High-fidelity data from physical tags such as defect masks fundamentally solves the problem of irreversible radiation-temperature mapping in traditional methods. The unquantifiable challenges enable synthetic data to possess interpretability that strictly aligns with physical laws. Furthermore, multi-target domain adaptive training uses this synthetic data as the source domain, fusing adversarial alignment, statistical distribution matching, and physical consistency regularization (structured constraints such as temperature order preservation and isothermal surface smoothing) to force the model to learn the physical essence rather than superficial style when migrating across device models, imaging distances, observation angles, and meteorological conditions. This effectively mitigates performance fluctuations caused by domain shifts. Building upon this, a confidence pseudo-label screening and curriculum-based self-training mechanism continuously expand reliable supervision signals in weakly labeled target domains. Combined with adaptive dynamic model adjustments during testing to adapt to deployment environment drift, the model simultaneously outputs heteroscedasticity uncertainty to quantify prediction reliability, significantly reducing reliance on manual annotation and improving field adaptability. Finally, by attaching audit tags such as model version, operating condition metadata, and physical consistency indicators, a traceable chain of evidence is constructed for each diagnostic conclusion, enabling maintenance personnel to verify the rationality of intermediate physical quantities such as temperature reconstruction and radiometric decomposition. This addresses the shortcomings of insufficient auditability in existing technologies, achieving a physically reversible, cross-domain robust, and engineering-usable closed-loop infrared diagnostic process with low annotation costs. Attached Figure Description

[0015] 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.

[0016] Figure 1 This is a flowchart of the infrared synthetic data and domain adaptation method based on digital twin provided in an embodiment of the present invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] In this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0019] An infrared synthetic data and domain adaptation method based on digital twins, such as Figure 1 As shown, it includes the following steps: S1: Construct a digital twin of the power equipment and align the parameters of the digital twin with the statistical characteristics of the real infrared sample through reverse calibration; S2: Perform multiphysics thermal simulation based on the digital twin to generate a temperature field; S3: Based on the temperature field, perform radiative transfer and imaging chain synthesis to generate a synthetic infrared image; S4: Automatically label the synthesized infrared image to obtain synthesized data with physical tags; S5: Using the synthetic data with physical labels as the source domain and the real infrared data as the target domain, perform multi-target domain adaptive training to obtain an adaptive model; S6: Use the adaptive model to generate confidence pseudo-labels for the target domain data, and perform stepwise self-training based on the confidence pseudo-labels to update the adaptive model; S7: During testing, adaptive testing is performed on data from unseen target domains, and the uncertainty is output. S8: Attach audit tags to the output results to form a traceable chain of evidence.

[0020] It should be noted that a digital twin is a high-precision mirror model of power equipment in virtual space, integrating multi-dimensional information such as geometric structure, material properties, thermodynamic parameters, and camera imaging characteristics. Reverse calibration refers to using the statistical distribution characteristics of real infrared images as an optimization target to automatically adjust unknown parameters in the twin, ensuring that the output synthetic data closely matches the field data in macroscopic statistical quantities such as radiation histograms and gradient distributions, achieving model self-consistency without manual intervention. Multiphysics thermal simulation, under given boundary conditions such as load current, ambient temperature, and solar radiation, uses numerical calculation methods to solve for the coupled processes of heat conduction, surface convection heat transfer, and radiative heat dissipation within the equipment, generating transient or steady-state temperature distribution fields that conform to physical conservation laws. Radiative transfer and imaging chain synthesis converts the temperature field into a radiative energy signal that the detector can receive, and fully simulates atmospheric attenuation, environmental reflection, optical system blurring, and camera noise, ultimately outputting a digital code image consistent with the specifications of a real infrared camera. Physical labels are pixel-level or target-level annotations containing true temperature values, temperature difference levels, defect location masks, and equivalent heat source distributions. They are directly exported from the simulation process and require no manual annotation. Multi-target domain adaptive training uses labeled synthetic data as the source domain and unlabeled real data as the target domain. It employs multiple mechanisms such as adversarial learning, statistical distribution matching, and physical consistency constraints for joint training, enabling the model to adapt to distribution differences across multiple target domains, including cross-aircraft types, distances, and weather conditions. Confidence pseudo-labels are high-confidence predictions of the target domain data. Reliable samples are screened through uncertainty quantification. Stepwise self-training uses a course-based learning strategy, gradually incorporating pseudo-labels into the supervision, progressively improving the model's adaptability to specific patterns in the target domain. Test-time adaptation involves updating some parameters (such as statistics in the batch normalization layer) online for single or small batches of test samples during the model deployment phase. This allows the model to quickly adapt to the transient characteristics of the current data distribution. Uncertainty output is quantified through heteroscedasticity regression or multiple sampling statistics to determine the range of confidence in the prediction results. Audit tags bind metadata such as model version, deployment site, shooting parameters, weather conditions, and calibration information to diagnostic results, forming a traceable and verifiable chain of evidence to meet the compliance requirements of industrial safety audits.

[0021] Understandably, this invention constructs a complete closed loop from data generation to model adaptation to result auditing, systematically solving the problems of physical distortion and insufficient generalization ability caused by the fragmented processing of "data synthesis" and "domain adaptation" in existing infrared diagnostic technologies. Steps S1 to S3 generate synthetic data with true temperature values ​​and radiation reversibility from the bottom up through digital twins and physical simulation, avoiding the expensive manual annotation costs and annotation inconsistencies. Step S4 automatically exports physical labels, so that each pixel carries structured information such as temperature and defect level, providing a high-quality truth foundation for subsequent supervised learning. Steps S5 and S6 transfer rich physical knowledge from the source domain to the target domain through multi-target domain adaptation and confidence pseudo-label self-training, while using target domain data self-supervision to gradually internalize the unique patterns of real-world scenarios, solving the performance degradation problem across equipment and weather conditions. Step S7's test-time adaptation and uncertainty output endow the model with online evolution and risk assessment capabilities, adapting to dynamically changing inspection environments. Step S8's audit label mechanism fills the gap in AI diagnostics for industrial safety compliance, making each judgment traceable to specific physical links and parameter configurations. This process embeds physical laws into the AI ​​training objectives, achieving a paradigm shift from "data-driven" to "physical-data dual-driven," ensuring that the model possesses both cross-domain generalization capabilities and maintains accuracy in complex and ever-changing power inspection scenarios. Accuracy of quantitative estimation of physical quantities and interpretability of decision-making.

[0022] Preferably, step S1 includes: Construct a digital twin including geometry, materials, operating conditions, and camera response parameters; parameter set Including material thermal parameters such as thermal conductivity ,density Specific heat capacity directional emissivity Atmospheric extinction coefficient And camera response and noise parameters; Based on the statistical characteristics of several real infrared samples Alignment Synthesis Statistical Features Optimal parameters Satisfying the relation: ; in, Represents the distribution distance metric. Indicates in the parameter The statistical characteristics of the synthetic data are as follows: Represents the statistical characteristics of a real sample; First, fix the thermal parameters and adjust the radiation and imaging chain parameters, then fine-tune the heat source and boundary coefficient until the distribution distance is reached. Less than the threshold .

[0023] It should be noted that the parameter set It is a core component of the digital twin, among which the material's thermal parameters include thermal conductivity. (unit Density is a characterizing property of a material's ability to conduct heat. ( ) and specific heat capacity ( These three factors collectively determine the thermal inertia of a material, and together they form the heat capacity matrix in the thermal diffusion equation. With stiffness matrix Directional emissivity directly affects the dynamic response speed and steady-state distribution of the temperature field. It is the angle between the surface normal and the observation direction. The function describes the angular non-uniformity of infrared radiation intensity in a material, and its accurate modeling determines the accuracy of the surface spontaneous emission term in radiative transfer calculations. Atmospheric extinction coefficient. (unit This comprehensively reflects the attenuation effect of atmospheric molecular absorption and aerosol scattering on infrared radiation, and its value changes dynamically with meteorological conditions. Camera response and noise parameters include gain. Bias The PRNU coefficient matrix, DSNU mean, and NETD standard deviation comprehensively characterize the nonlinear transformation from radiometric to code value by the camera and the statistical characteristics of various noise sources. Distribution distance metric. The core of constructing the inverse calibration optimization objective is the maximum mean difference (MMD), which maps the distribution to the regenerating kernel Hilbert space using a kernel function to calculate the mean difference, suitable for high-dimensional feature alignment; the bulldozer distance (EMD) measures the distribution difference by solving for the optimal transmission scheme, showing significant performance in histogram matching; and the multi-scale L2 distance calculates the L2 norm at different resolutions of the image pyramid, taking into account both global statistics and local texture. Statistical features and It is a mathematical abstraction of data distribution, covering the first moment (mean), second moment (variance), higher moments (skewness, kurtosis) of radiation values, and spatial statistics (gradient distribution, entropy, hotspot connected area distribution). Its extraction process requires nonlinear transformation (such as histogram equalization) and spatial filtering (such as Gaussian pyramid decomposition) of the radiation code value map to capture multi-scale statistical characteristics.

[0024] Understandably, the domain gap in infrared imaging of power equipment stems from physical parameter mismatch and differences in camera characteristics. If the parameters of the digital twin deviate too much from reality, all subsequent domain adaptation efforts will be based on an unreliable source domain. Therefore, it is essential to clearly define the complete composition of the parameter set and align it to the statistical characteristics of real data through inverse calibration. Using a distribution distance metric to construct an unsupervised optimization objective avoids the dependence on pixel-level registration in solving the inverse problem, requiring only macroscopic statistics to drive parameter adjustments, significantly reducing the cost of calibration data acquisition. Furthermore, the alternating optimization strategy addresses the ill-conditioned coupling problem between parameters: thermal parameters... Influencing the spatiotemporal evolution of the temperature field, radiation parameters The imaging parameters that affect the absolute value of radiance ( This affects the dynamic range of the code value and noise texture. If optimized synchronously, it can easily lead to optimization oscillations and local optima. First, the thermal parameters are fixed and the radiation imaging chain is adjusted, which is equivalent to calibrating the "image generator" to make its appearance match the real camera. Then, the boundary conditions of the heat source are fine-tuned, which is equivalent to correcting the deviation of the "temperature source" on the calibrated imaging chain. This hierarchical optimization ensures that each sub-problem is more convex and converges more stably. In the end, the synthetic data is highly consistent with the real inspection data in terms of radiation physics properties and statistical appearance, laying a reliable source domain foundation for subsequent cross-domain migration.

[0025] Preferably, step S2 includes: Under given load and environmental boundary conditions, the heat diffusion equation and boundary heat transfer equation are solved to generate a temperature field, wherein the heat diffusion equation satisfies the following relationship: ; The boundary heat transfer equation satisfies the following relationship: ; in, Indicates the density of the material. Indicates specific heat capacity. Represents the temperature field. Indicates time, Indicates thermal conductivity. Indicates the body's heat source. Represents the spatial derivative along the normal direction. Indicates the convective heat transfer coefficient. Indicates ambient temperature. Represents the Stefan-Boltzmann constant. Indicates surface emissivity; The heat diffusion equation is solved discretically using an implicit time-progression method, and the discretized form satisfies the following relationship: ; in, Indicates the time step. Represents the heat capacity matrix. Represents the thermal conductivity and boundary term stiffness matrix. Represents the equivalent heat source vector. and These represent the temperature vectors at the current time step and the next time step, respectively. By locally increasing the contact resistance Or reduce thermal conductivity Injecting defects in a manner that generates multi-level temperature differences The sample.

[0026] It should be noted that the thermal diffusion equation is a parabolic partial differential equation describing the heat conduction process in a solid medium, where the material density... With specific heat capacity The product of these factors constitutes the volumetric heat capacity, which determines the time-response characteristics of the temperature field to thermal disturbances. Thermal conductivity, This captures the non-uniform heat flow caused by the anisotropic thermal conductivity of the material. The heat source density represents the power density distribution of Joule heat or dielectric losses within the solid surface. The boundary heat transfer equation characterizes the energy exchange between the solid surface and the fluid environment (air) and the radiation environment. The convective heat transfer coefficient is related to wind speed, surface roughness, and fluid properties. For the radiative heat transfer term, it follows the Stefan-Boltzmann law. For surface emissivity, constant Implicit time progression methods advance the time at the current time step. To solve for the unknown temperature field, we construct a system of algebraic equations, where C in the discretized form is based on the material density. With specific heat capacity The diagonal heat capacity matrix is ​​constructed from the product of these. To cover Spatial Discrete Laplace Operator and Boundary Terms , The sum of the stiffness matrices, For inclusion With respect to the load vector of the boundary heat flux, this scheme is unconditionally stable, allowing for the selection of a larger value. Without inducing numerical oscillations, defect injection modifies the resistance of the contact interface in the geometric model locally. Alternatively, the k-value of the material's microstructure can be used to simulate poor contact or material degradation in real equipment at the simulation level. Increasing the contact loss leads to a higher level of contact loss. Both factors reduce resistance to heat conduction, leading to the formation of localized overheating hotspots. The difference between the hot spot temperature and the ambient temperature is used to quantify the severity of defects.

[0027] It is understood that this invention aims to obtain the temperature field of the device surface undergoing real spatiotemporal evolution by solving physical conservation equations, rather than relying on empirical formulas or simplified assumptions to generate temperature maps. Since the heating of power equipment is dynamic (e.g., intraday load fluctuations, changes in solar intensity angle), implicit propagation ensures numerical stability and computational efficiency over large time steps, avoiding the limitations imposed by CFL conditions in explicit schemes. The problem of excessively small values ​​and long simulation times exists. The boundary heat transfer equations comprehensively consider both convection and radiation heat dissipation mechanisms. Especially in outdoor substation scenarios, radiation heat dissipation can account for 30%-50% of the total heat dissipation; ignoring this mechanism will severely overestimate equipment temperature. Defect injection is achieved through precise control within the digital model. or It can generate thermal defect samples ranging from minor to severe, allowing the model to learn different... The discrimination boundary of the level is determined to avoid overfitting due to the synthetic data only covering a single anomalous pattern. The temperature field output in this step is the sole driving force for subsequent radiation calculations, and its physical accuracy directly determines the quality of the synthesized image. The reliability estimation fundamentally solves the shortcomings of traditional methods, such as the lack of true temperature values ​​and the inability to verify the reversibility of radiation, and provides thermodynamic scenarios with clear physical meaning to support subsequent domain adaptation.

[0028] Preferably, step S3 includes: Radiative transfer and imaging chain synthesis are performed based on the temperature field, wherein atmospheric transmittance Satisfying the relation: ; Path radiation Satisfying the relation: ; in, Indicates the distance of propagation Atmospheric transmittance below Indicates the extinction coefficient. Indicates path radiation. This represents the bandwidth-averaged Planck radiation. Indicates the equivalent atmospheric temperature; Radiation reaching the detector Satisfying the relation: ; Camera code value Satisfying the relation: ; Among them, total noise Satisfying the relation: ; in, This indicates the amount of radiation reaching the detector. Indicates the angle of incidence Directional emissivity below Indicates environmental reflected radiation. Indicates camera code value, Represents the quantization operator. and These represent camera gain and bias, respectively. Indicates as distance Scaled IFOV / PSF convolution kernels, This represents the convolution operation. Indicates total noise. This indicates non-uniform response noise. This indicates noise due to unevenness in the dark field. Represents equivalent temperature difference noise. This represents point-by-point multiplication; Output a synthesized code value map and a complete physical tag that match the target camera specifications.

[0029] It should be noted that atmospheric transmittance describes the degree of energy attenuation of infrared radiation as it propagates through the air. Its value decreases exponentially with increasing transmission distance. The extinction coefficient comprehensively reflects the absorption and scattering capabilities of water vapor, carbon dioxide, and suspended particles in the atmosphere for a specific infrared band. Path radiation refers to the infrared radiation emitted by the atmosphere itself due to thermal motion. This energy is unrelated to equipment radiation but is received by the detector and superimposed on the signal. Its contribution is particularly significant in haze or high humidity environments. The amount of radiation reaching the detector is the sum of three contributions: first, the infrared radiation emitted by the equipment surface itself, the intensity of which depends on the surface temperature and material emission characteristics; second, the radiation reflected from the surrounding environment (such as the sky, ground, and adjacent equipment); and third, atmospheric path radiation. The camera code value is the integer value output by the infrared detector after photoelectric conversion, amplification, and analog-to-digital conversion of the received radiation energy. Gain determines the amplification factor of the radiation to the code value, while bias represents the dark level at zero input. Together, they constitute the camera's radiometric calibration curve. IFOV refers to the instantaneous field of view of the camera, which determines the spatial resolution of a single pixel. PSF is the point spread function, which describes the blurring effect of an optical system on a point light source; its convolution operation simulates sub-pixel mixing. Response inhomogeneity noise is fixed-pattern noise caused by gain differences between pixels, manifesting as fixed stripes or spots on the image. Dark-field inhomogeneity noise is the difference in dark current among pixels, constituting the spatial undulations of the image background baseline. Equivalent temperature difference noise is the thermal noise limit of the detector, manifesting as random flickering particle noise. These three together constitute the total noise of the camera. The quantization operator discretizes a continuous analog signal into a fixed-bit digital signal, completing the analog-to-digital conversion. Physical tags refer to the ground truth information such as temperature value, radiation component, and noise component corresponding to each pixel, which can be used for subsequent supervised learning and audit traceability.

[0030] Understandably, the core purpose of step S3 is to convert the physical quantity of the temperature field obtained in step S2 into a pixel-by-pixel digital code image that the infrared camera would actually capture, through a rigorous radiative transfer and imaging link, thereby completely reproducing the entire process from photon emission to electron readout in real-world inspections. Traditional methods rely on generative adversarial networks to directly output images, which cannot guarantee the reversibility between radiation values ​​and temperature, leading to severe distortion in temperature estimation. This invention uses an analytical model to accurately calculate the amount of radiation reaching the detector for each pixel, ensuring it strictly follows Planck's radiation law and Lambert's cosine law. It simulates the sub-pixel mixing effect during long-distance imaging using IFOV and PSF convolution kernels, reproducing hotspot diffusion and edge blurring phenomena. By superimposing three types of noise—PRNU, DSNU, and NETD—it generates a noise texture consistent with the radiation characteristics of a specific camera model. Pixel-level inversion based on first principles of physics ensures that each grayscale value in the synthesized image can be traced bidirectionally to surface temperature, material emissivity, atmospheric transmittance, and noise components. This gives the temperature difference estimation the same reliability as real physical measurements, providing verifiable radiometric evidence for subsequent domain adaptation. This addresses the fundamental defect in existing technologies where synthesized data is "appearing similar but physically distorted."

[0031] Preferably, the step S3 includes the following: While maintaining temperature order and relative temperature difference ΔT, the appearance style of the synthesized infrared image is mapped to the real domain, where the mapping G from the synthesized domain to the real domain and the mapping F from the real domain to the synthesized domain are defined. Style bridging is achieved by minimizing the following objective function: ; Among them, physical consistency regularization term Satisfying the relation: ; in, This represents the mapping function from the synthetic domain to the real domain. This represents the mapping function from the real domain to the synthetic domain. Indicating resistance to loss, , and represent the weight coefficients of the cycle consistency loss, identity loss, and physical consistency regularization term, respectively. Describing the L1 norm, This represents a physical consistency regularization term. This represents the inversion operator from radiation to temperature. Indicates in the region Temperature order mapping function within, This represents the gradient operator along the tangent of the isothermal surface. This indicates the rank constraint evaluation region.

[0032] It should be noted that style bridging refers to performing visual domain transformation on the pixel-level appearance features (such as contrast, noise texture, edge sharpness, and grayscale distribution) of the synthesized infrared image while keeping the physical content of the image unchanged, so as to make it closer to the imaging style of the target real camera. It is a mapping function from the synthetic domain to the real domain, usually using an encoder-decoder structure, responsible for converting the synthetic code value map into an image with real domain visual features; It is an inverse mapping generator from the real domain to the synthetic domain, used to ensure the bidirectional reversibility of the transformation process. The adversarial loss consists of a discriminator that attempts to distinguish... The generated image and the real image are compared, while G learns to generate realistic images that are difficult for the discriminator to recognize. This creates an adversarial game between the two to improve generation quality. Cyclic consistency loss forces that the image generated by G, after being transformed by F, return to the original input, avoiding information loss and pattern collapse. Identity loss keeps the output stable when the input is close to the target domain, preventing over-correction. Physical consistency regularization term. The process includes two constraints: maintaining temperature order and smoothing isothermal surfaces. The temperature order mapping function sorts all pixels within a specified region by temperature value and encodes them as a rank vector, ensuring that the relative temperature order of hotspots remains strictly consistent before and after the transformation. The isothermal surface tangential gradient operator calculates the gradient of the temperature field along the isotherm direction, penalizing irregular temperature distortions generated during the transformation process and maintaining thermodynamic smoothness. The radiance-to-temperature inversion operator is used to infer the temperature field from the code value, serving as a bridge between visual transformation and physical constraints. This step is optional and should only be enabled when there is a significant difference between the synthetic and real domain perceptions; otherwise, subsequent adaptive training can proceed directly.

[0033] Understandably, although steps S1 to S3 have constructed physically consistent synthetic data, subtle differences may still exist between the synthetic image and the real image at the perceptual level. These differences might include specific camera color response curves, compression artifacts, or unique noise patterns. While these differences do not affect physical reversibility, they increase the difficulty of subsequent domain adaptation. Style bridging is used to visually modify the synthetic image to narrow the domain gap; however, purely data-driven style transfer may distort temperature order in pursuit of visual realism, such as reducing pixel values ​​in high-temperature hotspot areas to match the grayscale distribution of the real domain, leading to… The estimated systematicity is low. Physical regularity. Forced by temperature order constraints The output retains the original temperature order after inversion, ensuring that severe defects are not downgraded during transformation; isothermal surface smoothing constraints prevent the generation of temperature abrupt changes that violate the laws of heat conduction. Weights Balancing visual realism and physical fidelity, this step can be skipped when the differences are small. This achieves a balance between appearance adaptation and physical conservation, ensuring that the synthesized data maintains high physical credibility while possessing a visual style that seamlessly integrates with real data. This reduces the burden on subsequent domain adaptation while eliminating the risk of physical distortion.

[0034] Preferably, step S4 includes: Pixel-level and target-level physical tags and metadata are directly derived from the digital twin, wherein the tag set Satisfying the relation: ; Metadata collection Satisfying the relation: ; in, Indicates the temperature field distribution. Indicates the temperature difference distribution. This represents a defect segmentation mask. This represents the equivalent heat source distribution map. Indicates the imaging distance. Indicates the imaging perspective; Latin hypercube sampling or stratified sampling methods were used to cover the parameter space of machine type, distance, viewing angle, operating condition, and season, and different The levels are resampled and balanced to construct a source domain training set and a target domain unlabeled set.

[0035] It should be noted that pixel-level physical labeling refers to assigning a corresponding physical attribute truth value to each pixel in an image, such as temperature field distribution. It is a two-dimensional matrix with the same size as the image resolution, where each element is the absolute temperature of the device surface corresponding to that pixel, in Kelvin; temperature difference distribution. It is a temperature difference matrix relative to the environmental reference temperature or the neighborhood mean, directly quantifying the severity of defects; defect segmentation mask. It is a binary image, with 1-valued pixels marking thermal defect areas and 0-valued pixels representing normal background; the equivalent heat source distribution map H is a two-dimensional matrix characterizing the distribution of heat power density inside the equipment, used to trace the cause of defects. Target-level labels refer to the overall labeling of connected thermal defect areas with attributes such as maximum temperature, average temperature difference, and defect type. Metadata set It is a set of parameters describing the imaging conditions, including the imaging distance. The spatial resolution and atmospheric attenuation are determined, measured in meters; imaging angle. The data includes azimuth and elevation angles, affecting surface emissivity and occlusion relationships; material information records the surface material type of the equipment, such as silicone rubber or aluminum alloy; model information identifies the infrared camera model and noise characteristics; load information records current and voltage conditions; and meteorological information includes ambient temperature, humidity, wind speed, and solar irradiance. Latin hypercube sampling is a stratified random sampling technique that divides each parameter dimension into several equally probable intervals, independently and randomly sampling and combining them to ensure uniform coverage of the parameter space and a sample size far smaller than that of grid sampling. Stratified sampling involves independent sampling according to preset intervals, such as dividing distances into intervals of 10-30 meters, 31-50 meters, etc., with a fixed number of samples in each interval to ensure representative samples in each subspace. Resampling equalization is an adjustment made to address the uneven distribution of samples at different temperature levels. By oversampling a minority of severe defects or undersampling a majority of minor defects, the proportion of samples at each level in the training set is balanced, preventing the model from biased towards frequent categories. The source domain training set consists of synthetic images with complete physical labels, while the target domain unlabeled set consists of real infrared images collected on-site. The two are paired to support subsequent unsupervised domain adaptation.

[0036] Understandably, traditional infrared inspection relies on manual delineation of defect areas and temperature estimation on images, which is time-consuming, labor-intensive, and highly subjective, severely limiting the construction of large-scale datasets. This invention utilizes physical quantities such as temperature field and heat source distribution, accurately calculated during simulation using a digital twin, to directly derive pixel-level and target-level labels, achieving automated labeling with 100% accuracy and reducing manual labeling costs and errors. Generating data only under a single operating condition makes it difficult for the model to generalize to different shooting distances, camera models, or seasonal weather conditions; therefore, a systematic sampling of the parameter space is necessary. Latin hypercube or stratified sampling ensures that the training set covers high-dimensional combinations of camera model, distance, viewing angle, load, and season, exposing the model to diverse scenarios during training and improving cross-domain robustness. In real-world inspections, defect severity exhibits a long-tail distribution, with minor defects far outnumbering severe ones. Direct training leads to low recall rates for severe defects. Resampling equalization, by manually adjusting the sample distribution, forces the model to learn the discriminative features of each temperature difference level equally, significantly improving grading accuracy. The final source domain training set has complete physical labels and is evenly distributed, while the unlabeled target domain set consists of real images. Together, they support subsequent unsupervised domain adaptation, enabling low-cost and high-fidelity transfer from virtual simulation to real-world scenarios.

[0037] Preferably, step S5 includes: Using the synthetic data with physical tags as the source domain Using real infrared data as the target domain Through feature extraction network Simultaneously outputs defect segmentation and temperature difference. The regression and classification results show that the overall loss function satisfies the following relationship: ; in, This includes classification cross-entropy loss, Dice loss, and... The source domain supervised loss of the regression Huber loss, Indicates domain adversarial loss, This represents the statistical distribution matching term of the maximum mean difference. This represents a physical consistency regularization term. , and Indicates the weighting coefficient. and These represent samples from the source and target domains, respectively. This represents a feature extraction network; The domain adversarial loss Satisfying the relation: ; in, Discriminator for the domain, Represents the expectation operator. and These represent samples from the source and target domains, respectively. The statistical distribution matching item Satisfying the relation: ; in, Represents the kernel mapping function. Denotes the Hilbert space norm of the regenerating kernel. and These represent the number of samples in the source domain and the target domain, respectively. and They represent the first The source domain sample and the first Features of each target domain sample; The physical consistency regularization term Satisfying the relation: ; in, , and This represents the regularization weight coefficient. Indicates the temperature order maintenance item. Indicates the isothermal surface smoothing term. This represents the radiation field or equivalent quantity obtained through inversion. Represents the distance-dependent point spread function / instantaneous field-of-view convolution kernel; Use metadata Conditional modulation is performed using characteristic linear modulation or conditional normalization methods to characterize equipment, attitude, and weather differences.

[0038] It should be noted that the source domain refers to the synthetic data distribution with complete physical labels, while the target domain refers to the real data distribution with missing or sparse labels. Domain adaptation aims to generalize the model trained in the source domain to the target domain through feature alignment and distribution matching. Feature extraction network It is a convolutional neural network with an encoder-decoder structure, in which These represent trainable weights responsible for extracting high-level semantic features from the input image. Defect segmentation is a pixel-level binary classification task, outputting a mask of the defect region; temperature difference... Regression predicts the temperature difference value for each pixel or each target, belonging to continuous value estimation tasks; the grading result classifies defects into discrete levels such as minor, moderate, and severe. Source domain supervised loss. The model's primary task loss on synthetic data is the classification cross-entropy loss, which is used for defect level classification. The loss is used to optimize the cross-union ratio of the segmentation mask. Loss used for Regression is insensitive to outliers, improving robustness. Domain adversarial loss. Implemented using a gradient inversion layer, the domain discriminator D attempts to distinguish whether features originate from the source or target domain, while the feature extractor learns to confuse the discriminator, forcing the feature distributions of the source and target domains to align. Maximum mean difference (MMD) is a nonparametric statistic measuring the difference between two distributions. It calculates the mean distance after transforming features to a high-dimensional space through kernel mapping, constraining the second-order statistic matching between the source and target domains. Physical consistency regularization is also employed. It is a set of constraints, temperature order maintenance terms. Ensure the relative temperature order of hotspots matches the physical truth to prevent temperature sorting errors during domain alignment; isothermal surface smoothing term. The system penalizes discontinuous abrupt changes in isotherms in the predicted temperature field to maintain thermodynamic smoothness. The reprojection error term constrains the prediction results to conform to the camera's imaging physics by comparing the difference between the predicted radiation field and the radiation field convolved with the camera's PSF. Metadata M is injected into the network through Feature Linear Modulation (FiLM) or Conditional Normalization (CondNorm), enabling the model to dynamically adjust feature statistics according to different operating conditions, thus achieving conditional prediction.

[0039] Understandably, the technical challenge of step S5 lies in the significant domain shift between the source domain synthetic data and the target domain real data, such as differences in camera noise characteristics and atmospheric attenuation. Directly applying the source domain model can lead to a sharp drop in target domain performance. Traditional domain adaptation only aligns feature distributions, which can easily sacrifice the accuracy of physical quantities, such as distorting the temperature scale to align the distribution. This invention addresses this problem through a triple mechanism: domain adversarial loss and MMD joint constraints achieve statistical distribution alignment, ensuring the model's adaptability to the target domain appearance; physical consistency regularization embeds thermodynamic laws into the training objective, ensuring that the alignment process does not disrupt the intrinsic structure of the temperature field; metadata modulation explicitly decouples the influence of different variables, enabling the model to learn to distinguish between physically explainable domain differences and camera style differences that need to be adapted. The resulting adaptive model can both understand the visual features of the target domain and maintain the ability to quantitatively estimate ΔT, achieving a unity of appearance adaptation and physical fidelity.

[0040] Preferably, step S6 includes: Generate pseudo-tags in the target domain. Through the quantile threshold of uncertainty or residual To control the quality of pseudo-tags, the accept function must satisfy the following relation: ; The pseudo-labeled dataset satisfies the following relation: ; in, The pseudo-labels representing the target domain samples. This represents the uncertainty measurement function. This represents a threshold based on quantiles. Indicates an indicator function, This represents the accepted pseudo-labeled dataset; Gradually relax threshold parameters using a course-based learning approach. To expand the pseudo-label dataset And through weights The pseudo-label supervision loss is accumulated into the overall training objective, and the update relation satisfies: ; in, This represents the loss weight for pseudo-label supervision. This represents the supervised loss for false-labeled samples. This indicates the overall training objective.

[0041] It's important to note that pseudo-labels refer to the model's predictions for unlabeled data in the target domain. These pseudo-labels are used as temporary labels for supervised learning, enabling self-supervised fine-tuning in the target domain. Uncertainty measures are functions that evaluate prediction reliability. They can be calculated using prediction entropy (measuring the uncertainty of the classification probability distribution), MC Dropout variance (statistically predicting fluctuations through multiple random forward propagations), or prediction residuals (comparing model output consistency with its neighborhood). A higher value indicates less confidence in the model's prediction of that sample. Quantile threshold. It is based on the uncertainty distribution The quality control threshold set by the quantile only accepts the quantile with the lowest uncertainty. Proportional samples, for example Only the most reliable 10% of samples are retained. The acceptance function is a binary selection function; it outputs 1 when the sample uncertainty is less than or equal to the threshold, indicating that the sample is accepted into the pseudo-label set; otherwise, it outputs 0 to reject it. The pseudo-label dataset consists of all accepted samples and their pseudo-labels, and its quality directly determines the self-training effect. Curriculum learning is a training strategy that mimics human cognitive processes. Initially, only the most reliable simple samples are used, gradually increasing the number of complex and difficult samples, expanding the training set from easy to difficult to avoid early noise contamination of the model. The threshold parameter is gradually relaxed. This means dynamically adjusting quality control standards, from conservative to lenient, gradually expanding the size of the pseudo-label set to cover more target domain patterns. Weighting To control the contribution of pseudo-label supervision loss to the total loss and prevent low-quality pseudo-labels from overwhelming the source domain supervision signal, a small value (e.g., 0.1-0.5) is typically set. Accumulated supervision loss adds the pseudo-label loss to the overall training objective, allowing the model to learn target domain-specific features while retaining source domain knowledge.

[0042] Understandably, although step S5 has completed multi-target domain adaptive training, there are still distributions in the target domain not covered by synthetic data (such as special weather conditions or rare defect morphologies). Directly applying a static model would lead to performance bottlenecks. This invention initiates a self-supervised fine-tuning loop for the target domain by generating pseudo-labels on the model itself and using the high-confidence portion for self-training. This mechanism enables the model to continuously learn the unique patterns of the target domain (such as the unique background thermal radiation distribution of a substation), rather than being limited to the transfer of knowledge from the source domain. By using quantile threshold filtering and curriculum-based relaxation, the cumulative effect of pseudo-label noise is effectively suppressed, ensuring stable convergence of the self-training process. If fine-tuning is performed directly on the full target domain data, the model may collapse due to fitting low-quality predictions; however, through confidence filtering and gradual expansion, a progressive knowledge fusion from "transfer" to "adaptation" and then to "internalization" is achieved, gradually improving the learning ability of fine-grained features of the target domain.

[0043] Preferably, step S7 includes: During the deployment phase, when performing lightweight testing on unseen target domain data, an adaptive approach is adopted. This is achieved by freezing backbone network parameters and updating only normalized layer parameters and small output header parameters to minimize prediction entropy, optimizing the objective to satisfy the following relation: ; in, This indicates the normalization layer and small output header parameters to be updated. Indicates category The predicted probability, This represents the summation over all categories. This indicates taking the expectation of the samples in the target domain; Uncertainty is estimated using heteroscedastic regression combined with Monte Carlo Dropout or deep ensemble methods. The heteroscedastic regression loss function satisfies the following relationship: ; in, Represents the actual value. This represents the model's predicted value. Indicates the prediction variance. Represents square Norm; predicted mean and cognitive uncertainty Obtained through statistical analysis of multiple forward propagations; Output set and model version information and timestamp, among which This represents a temperature difference distribution map. This represents a defect segmentation mask. This represents a temperature distribution map. This represents the uncertainty distribution diagram. Represents metadata.

[0044] It's important to note that adaptive testing refers to a lightweight migration mechanism that updates some model parameters online for each test sample or small batch of data after model deployment, allowing it to quickly adapt to the current data distribution. Lightweight updates involve adjusting only modules with a very small number of parameters, such as the statistics of the batch normalization layer and the classification little head, avoiding the computational overhead and catastrophic forgetting caused by full model fine-tuning. Prediction entropy is an information-theoretic metric that calculates the uncertainty of the model's output probability distribution. Higher entropy indicates greater model uncertainty; minimizing entropy forces the model's statistics to match the current domain data distribution. Heteroscedastic regression is a regression method that models the variation of observed noise variance with input. The loss function includes the prediction error divided by the variance and the logarithm of the variance, allowing the model to predict noise levels while fitting the mean. Monte Carlo Dropout keeps the Dropout layer active during inference, constructing a prediction distribution through multiple random forward propagations, using statistical variance as an estimate of cognitive uncertainty. Deep ensemble trains multiple models with different initializations to estimate uncertainty through ensemble variance. Model version information records algorithm iteration identifiers, and timestamps record the inference time; these two constitute key metadata for result traceability.

[0045] It is understandable that step S6 has already adapted the model to the target domain through self-training, but the real inspection environment still has dynamic changes (such as camera thermal drift, sudden rain and fog, and sudden load increases), making it difficult for static model parameters to respond in real time. This invention updates only the BN layer statistics during testing, aiming to minimize the prediction entropy, enabling the model to match the current data distribution online and quickly compensate for gradual environmental changes. Freezing the backbone network avoids computational delays and source domain knowledge forgetting caused by large-scale parameter updates, meeting the real-time requirements of edge computing scenarios such as UAV inspections. Heteroscedastic regression and MC Dropout / deep ensemble provide two uncertainty estimation approaches: heteroscedastic regression models observation noise (random uncertainty), and MC Dropout estimates cognitive uncertainty of model parameters (cognitive uncertainty). The combination of the two outputs a spatially differentiated uncertainty map, identifying low-confidence areas. This online adaptation and risk quantification mechanism upgrades the model output from deterministic results to probabilistic diagnosis, enabling maintenance personnel to prioritize the review of high-uncertainty areas, achieving an upgrade from black-box prediction to risk-aware decision-making.

[0046] Preferably, step S8 includes: Add an audit label to the output results, the audit label Satisfying the relation: ; in, Indicates the model identifier and version number. Indicates the site identifier. Indicates the imaging distance. Indicates the imaging perspective. Indicates ambient temperature. Indicates relative humidity. Indicates wind speed. Indicates calibration marking, Represents a timestamp; When comparing across different models or seasons, physical consistency indicators are included, such as temperature order-related indicators, reprojection error, and the proportion of equivalent temperature difference noise. When the physical consistency indicators do not reach a preset threshold, a prompt for review or minor calibration is generated.

[0047] It should be noted that the audit tag is a structured metadata string attached to each diagnostic conclusion, used for post-event traceability and accountability. The model identifier and version number record algorithm iteration information, such as... This facilitates tracking the model's change history. The site identifier shows the geographical location of the positioning device, such as... Imaging distance and viewing angle record the shooting geometric parameters for reprojection verification. Ambient temperature, relative humidity, and wind speed constitute the meteorological context, explaining the heat dissipation boundary conditions. The calibration identifier is associated with the validity period of the camera radiometric calibration curve to ensure the accuracy of radiometric inversion. The timestamp records the diagnostic time, supporting time-series analysis. Physical consistency indicators are statistical quantities that quantify the conformity between the model output and physical laws: the temperature order correlation index calculates the rank correlation coefficient between the predicted temperature ranking and the inverted temperature ranking, reflecting the correctness of the relative positions of hotspots; reprojection error measures the difference between the predicted radiation field reprojected by the camera model and the observed code value, assessing radiation reversibility; the equivalent temperature difference noise ratio compares the predicted noise level with the camera's nominal noise, verifying the accuracy of noise modeling. The "requires verification or light calibration" prompt is an automatic early warning mechanism, triggered when any indicator falls below a preset threshold, prompting maintenance personnel to manually intervene or initiate a light calibration process.

[0048] Understandably, the power industry has extremely high reliability requirements for intelligent diagnostic systems. Any conclusion must be traceable, auditable, and explainable. This invention binds each diagnosis to a complete physical-geometric-environmental-model context by embedding audit tags, forming an immutable chain of evidence. When a missed detection or false alarm occurs, it can quickly pinpoint whether the issue stems from a model version problem, camera calibration failure, abnormal environmental parameters, or insufficient training data, avoiding blame-shifting. Real-time monitoring of physical consistency indicators enables online assessment of model health. When the temperature order correlation is too low or the reprojection error is too large, it indicates that the model may be physically distorted due to drastic environmental changes. This automatically generates a verification prompt, forcing secondary confirmation and effectively preventing silent failure of the AI ​​system. This design elevates the AI ​​model from a black-box tool to a trusted industrial component, meeting the compliance requirements of power safety regulations.

[0049] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0050] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for adaptive infrared synthetic data and domain based on digital twins, characterized in that, Includes the following steps: S1: Construct a digital twin of the power equipment and align the parameters of the digital twin with the statistical characteristics of the real infrared sample through reverse calibration; S2: Perform multiphysics thermal simulation based on the digital twin to generate a temperature field; S3: Based on the temperature field, perform radiative transfer and imaging chain synthesis to generate a synthetic infrared image; S4: Automatically label the synthesized infrared image to obtain synthesized data with physical tags; S5: Using the synthetic data with physical labels as the source domain and the real infrared data as the target domain, perform multi-target domain adaptive training to obtain an adaptive model; S6: Use the adaptive model to generate confidence pseudo-labels for the target domain data, and perform stepwise self-training based on the confidence pseudo-labels to update the adaptive model; S7: During testing, adaptive testing is performed on data from unseen target domains, and the uncertainty is output. S8: Attach audit tags to the output results to form a traceable chain of evidence.

2. The infrared synthetic data and domain adaptation method based on digital twins according to claim 1, characterized in that, Step S1 includes: Construct a digital twin including geometry, materials, operating conditions, and camera response parameters; parameter set Including material thermal parameters such as thermal conductivity ,density Specific heat capacity directional emissivity Atmospheric extinction coefficient And camera response and noise parameters; Based on the statistical characteristics of several real infrared samples Alignment Synthesis Statistical Features Optimal parameters Satisfying the relation: ； in, Represents the distribution distance metric. Indicates in the parameter The statistical characteristics of the synthetic data are as follows: Represents the statistical characteristics of a real sample; First, fix the thermal parameters and adjust the radiation and imaging chain parameters, then fine-tune the heat source and boundary coefficient until the distribution distance is reached. Less than the threshold .

3. The infrared synthetic data and domain adaptation method based on digital twins according to claim 1, characterized in that, Step S2 includes: Under given load and environmental boundary conditions, the heat diffusion equation and boundary heat transfer equation are solved to generate a temperature field, wherein the heat diffusion equation satisfies the following relationship: ； The boundary heat transfer equation satisfies the following relationship: ； in, Indicates the density of the material. Indicates specific heat capacity. Represents the temperature field. Indicates time, Indicates thermal conductivity. Indicates the body's heat source. Represents the spatial derivative along the normal direction. Indicates the convective heat transfer coefficient. Indicates ambient temperature. Represents the Stefan-Boltzmann constant. Indicates surface emissivity; The heat diffusion equation is solved discretically using an implicit time-progression method, and the discretized form satisfies the following relationship: ； in, Indicates the time step. Represents the heat capacity matrix. Represents the thermal conductivity and boundary term stiffness matrix. Represents the equivalent heat source vector. and These represent the temperature vectors at the current time step and the next time step, respectively. By locally increasing the contact resistance Or reduce thermal conductivity Injecting defects in a manner that generates multi-level temperature differences The sample.

4. The infrared synthetic data and domain adaptation method based on digital twin according to claim 1, characterized in that, Step S3 includes: Radiative transfer and imaging chain synthesis are performed based on the temperature field, wherein atmospheric transmittance Satisfying the relation: ； Path radiation Satisfying the relation: ； in, Indicates the distance of propagation Atmospheric transmittance below Indicates the extinction coefficient. Indicates path radiation. This represents the bandwidth-averaged Planck radiation. Indicates the equivalent atmospheric temperature; Radiation reaching the detector Satisfying the relation: ； Camera code value Satisfying the relation: ； Among them, total noise Satisfying the relation: ； in, This indicates the amount of radiation reaching the detector. Indicates the angle of incidence Directional emissivity below Indicates environmental reflected radiation. Indicates camera code value, Represents the quantization operator. and These represent camera gain and bias, respectively. Indicates as distance Scaled IFOV / PSF convolution kernels, This represents the convolution operation. Indicates total noise. This indicates non-uniform response noise. This indicates noise due to unevenness in the dark field. Represents equivalent temperature difference noise. This represents point-by-point multiplication; Output a synthesized code value map and a complete physical tag that match the target camera specifications.

5. The infrared synthetic data and domain adaptation method based on digital twin according to claim 1, characterized in that, After step S3, the following is included: While maintaining temperature order and relative temperature difference ΔT, the appearance style of the synthesized infrared image is mapped to the real domain, where the mapping G from the synthesized domain to the real domain and the mapping F from the real domain to the synthesized domain are defined. Style bridging is achieved by minimizing the following objective function: ； Among them, physical consistency regularization term Satisfying the relation: ； in, This represents the mapping function from the synthetic domain to the real domain. This represents the mapping function from the real domain to the synthetic domain. Indicating resistance to loss, , and represent the weight coefficients of the cycle consistency loss, identity loss, and physical consistency regularization term, respectively. Describing the L1 norm, This represents a physical consistency regularization term. This represents the inversion operator from radiation to temperature. Indicates in the region Temperature order mapping function within, This represents the gradient operator along the tangent of the isothermal surface. This indicates the rank constraint evaluation region.

6. The infrared synthetic data and domain adaptation method based on digital twins according to claim 1, characterized in that, Step S4 includes: Pixel-level and target-level physical tags and metadata are directly derived from the digital twin, wherein the tag set Satisfying the relation: ； Metadata collection Satisfying the relation: ； in, Indicates the temperature field distribution. Indicates the temperature difference distribution. This represents a defect segmentation mask. This represents the equivalent heat source distribution map. Indicates the imaging distance. Indicates the imaging perspective; Latin hypercube sampling or stratified sampling methods were used to cover the parameter space of machine type, distance, viewing angle, operating condition, and season, and different The levels are resampled and balanced to construct a source domain training set and a target domain unlabeled set.

7. The infrared synthetic data and domain adaptation method based on digital twins according to claim 1, characterized in that, Step S5 includes: Using the synthetic data with physical tags as the source domain Using real infrared data as the target domain Through feature extraction network Simultaneously outputs defect segmentation and temperature difference. The regression and classification results show that the overall loss function satisfies the following relationship: ； in, This includes classification cross-entropy loss, Dice loss, and... The source domain supervised loss of the regression Huber loss, Indicates domain adversarial loss, This represents the statistical distribution matching term of the maximum mean difference. This represents a physical consistency regularization term. , and Indicates the weighting coefficient. and These represent samples from the source and target domains, respectively. This represents a feature extraction network; The domain adversarial loss Satisfying the relation: ； in, Discriminator for the domain, Represents the expectation operator. and These represent samples from the source and target domains, respectively. The statistical distribution matching item Satisfying the relation: ； in, Represents the kernel mapping function. Denotes the Hilbert space norm of the regenerating kernel. and These represent the number of samples in the source domain and the target domain, respectively. and They represent the first The source domain sample and the first Features of each target domain sample; The physical consistency regularization term Satisfying the relation: ； in, , and This represents the regularization weight coefficient. Indicates the temperature order maintenance item. Indicates the isothermal surface smoothing term. This represents the radiation field or equivalent quantity obtained through inversion. Represents the distance-dependent point spread function / instantaneous field-of-view convolution kernel; Use metadata Conditional modulation is performed using characteristic linear modulation or conditional normalization methods to characterize equipment, attitude, and weather differences.

8. The infrared synthetic data and domain adaptation method based on digital twins according to claim 1, characterized in that, Step S6 includes: Generate pseudo-tags in the target domain. Through the quantile threshold of uncertainty or residual To control the quality of pseudo-tags, the accept function must satisfy the following relation: ； The pseudo-labeled dataset satisfies the following relation: ； in, The pseudo-labels representing the target domain samples. This represents the uncertainty measurement function. This represents a threshold based on quantiles. Indicates an indicator function, This represents the accepted pseudo-labeled dataset; Gradually relax threshold parameters using a course-based learning approach. To expand the pseudo-label dataset And through weights The pseudo-label supervision loss is accumulated into the overall training objective, and the update relation satisfies: ； in, This represents the loss weight for pseudo-label supervision. This represents the supervised loss for false-labeled samples. This indicates the overall training objective.

9. The infrared synthetic data and domain adaptation method based on digital twin according to claim 1, characterized in that, Step S7 includes: During the deployment phase, when performing lightweight testing on unseen target domain data, an adaptive approach is adopted. This is achieved by freezing backbone network parameters and updating only normalized layer parameters and small output header parameters to minimize prediction entropy, optimizing the objective to satisfy the following relation: ； in, This indicates the normalization layer and small output header parameters to be updated. Indicates category The predicted probability, This represents the summation over all categories. This indicates taking the expectation of the samples in the target domain; Uncertainty is estimated using heteroscedastic regression combined with Monte Carlo Dropout or deep ensemble methods. The heteroscedastic regression loss function satisfies the following relationship: ； in, Represents the actual value. This represents the model's predicted value. Indicates the prediction variance. Represents square Norm; predicted mean and cognitive uncertainty Obtained through statistical analysis of multiple forward propagations; Output set and model version information and timestamp, among which This represents a temperature difference distribution map. This represents a defect segmentation mask. This represents a temperature distribution map. This represents the uncertainty distribution diagram. Represents metadata.

10. The infrared synthetic data and domain adaptation method based on digital twin according to claim 1, characterized in that, Step S8 includes: Add an audit label to the output results, the audit label Satisfying the relation: ； in, Indicates the model identifier and version number. Indicates the site identifier. Indicates the imaging distance. Indicates the imaging perspective. Indicates ambient temperature. Indicates relative humidity. Indicates wind speed. Indicates calibration marking, Represents a timestamp; When comparing across different models or seasons, physical consistency indicators are included, such as temperature order-related indicators, reprojection error, and the proportion of equivalent temperature difference noise. When the physical consistency indicators do not reach a preset threshold, a prompt for review or minor calibration is generated.

Images