Physical covariant shared residual latent variable based airborne radar multi-fidelity twin method

Abstract

The application discloses a kind of based on physical covariant sharing residual error hidden variable airborne radar multi-fidelity twin method;Through white platinum π theorem dimensionless realization cross-platform physical covariant and absolute parameter hiding;Through shared-residual hidden variable decoupling and binding physical semantic subspace, realize multi-fidelity fusion and explainable degradation attribution;Through physical guide variation information bottleneck channel and conditional discrete diffusion probability model channel cooperation, realize high-precision performance prediction and scarce sample generation;Through hierarchical physical constraint and cross-fidelity cycle consistency;Through three-dimensional uncertainty-driven active sampling, and using elastic weight solidification, experience replay and low-rank adaptation combined continuous learning strategy, support online evolution;Through shapley value attribution output five-level maintenance decision, form closed loop;Solve airborne radar digital twin in multi-fidelity fusion bias, physical illusion, cross-platform migration difficulty and online catastrophic forgetting and other problems, engineering can be landed.

Description

Technical Field

[0001] This invention relates to the field of radar health management and digital twin technology, which can be extended to other radar systems. Specifically, it relates to an airborne radar multi-fidelity twin method and system based on physical covariance shared residual latent variables. Background Technology

[0002] Airborne radar is a core sensor for modern combat aircraft, and its detection performance is affected by a combination of factors, including transmit / receive channel gain drift, T / R component degradation, array deformation, changes in the thermal environment, and complex electromagnetic interference. Establishing a high-precision radar digital twin model is of great significance for residual performance assessment, residual lifetime prediction, and system maintenance decisions.

[0003] The existing technology has the following shortcomings: The contradiction between high-fidelity simulation and real-time performance: electromagnetic full-wave simulation has high accuracy but is time-consuming and difficult to support online decision-making; low-fidelity models have good real-time performance but insufficient accuracy.

[0004] Multi-fidelity data is difficult to integrate effectively: high-fidelity samples are scarce, while low-fidelity samples are abundant but have systematic biases. Simple weighted fusion can easily introduce biases, causing the model to fail under critical operating conditions.

[0005] Data-driven models suffer from "physical illusions": pure neural network models may output results that violate fundamental physical laws such as radar equations and energy conservation, reducing the model's credibility under extreme conditions.

[0006] Cross-platform / cross-model migration is difficult: There are large differences in parameters between different radar systems and different carrier platforms, the model migration cost is high, and the original parameters are often classified, which limits the general application of the model.

[0007] Difficult to support continuous online evolution: Performance drift occurs continuously during field service, offline trained models have difficulty tracking new data distribution in a timely manner, and there is a catastrophic forgetting problem during incremental updates.

[0008] Lack of explainable degradation attribution and maintenance decision-making loop: Existing methods mostly stop at "performance prediction", making it difficult to locate the specific physical components that cause performance degradation and unable to provide actionable and evidence-based maintenance recommendations.

[0009] To address the aforementioned issues, a digital twin modeling method for airborne radar is needed that balances physical interpretability, cross-fidelity generalization capability, online evolution capability, and engineering feasibility. Summary of the Invention

[0010] The technical problem to be solved by this invention is to provide a method for airborne radar multifidelity twin based on physical covariant shared residual latent variables, which addresses issues such as multifidelity fusion bias, physical illusion, difficulties in cross-platform migration, online catastrophic forgetting, and lack of interpretable basis for maintenance decisions in airborne radar digital twins.

[0011] To solve the above-mentioned technical problems, the present invention first accesses radar data of at least two fidelities, and performs dimensionless transformation on the original physical quantities based on Buckingham's π theorem, so that data from different platforms, different frequency bands and different fidelity sources can be modeled in a unified physical covariance space, while hiding classified absolute physical parameters.

[0012] In the feature fusion stage, this invention processes modal data such as radar echo, range-Doppler map, track points and state parameters respectively through multimodal encoders, and adjusts the output of each encoder by using continuous fidelity embedding vectors, thereby forming fidelity-aware fused features.

[0013] In the latent variable modeling stage, this invention decouples the fused features into shared latent variables and residual latent variables. The shared latent variables represent the physical structures shared across fidelity levels, while the residual latent variables represent fidelity differences, component degradation, and environmental disturbances. Furthermore, they are bound to physical semantic subspaces such as antenna / array, T / R components, receiving channels, transmitting channels, environment and interference, and overall system health.

[0014] In the prediction and generation stage, this invention constructs a physically guided variational information bottleneck channel and a conditional diffusion generation channel. The strength of the information bottleneck is adjusted by the physical consistency residual, and radar equation residual guidance is introduced in the diffusion denoising process, so that performance prediction, scarce high-fidelity sample generation and uncertainty estimation are mutually verified.

[0015] During the training constraint phase, this invention embeds hard constraints, strong soft constraints, and weak soft constraints into the training target in layers, and combines them with cross-fidelity cycle consistency constraints to suppress situations where the model output violates basic physical laws such as radar equations, energy conservation, and probability normalization.

[0016] In the online application phase, this invention uses three-dimensional uncertainty estimation based on accidental uncertainty, cognitive uncertainty, and physical residual uncertainty to drive active sampling and continuous online learning. Based on the Shapley value contribution analysis of each physical semantic subspace of the residual latent variables, it outputs hierarchical maintenance decisions.

[0017] The present invention also provides a system for implementing the above method. The system includes a data access and dimensionless preprocessing module, a multimodal coding and fidelity fusion module, a shared-residual latent variable decoupling module, a physical-guided variational information bottleneck and conditional diffusion dual-channel generation module, a physical constraint and cyclic consistency training module, and an uncertainty estimation and maintenance decision module. The modules work together to complete multi-fidelity radar data processing, performance prediction, sample generation, degradation attribution, and maintenance decision output.

[0018] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: 1. Physical covariance and cross-platform migration: Through dimensionless preprocessing of Buckingham's π theorem, physical covariance migration is achieved across platforms and frequency bands, and absolute parameters are effectively hidden in classified scenarios.

[0019] 2. Physically Interpretable Degradation Attribution: The shared-residual latent variable decoupling structure enables the model to avoid bias propagation when fusing multi-fidelity data and provides a clear degradation attribution path through the physical semantic subspace.

[0020] 3. Dual-channel synergistic enhancement: The synergistic effect of the Physics-Guided Variational Information Bottleneck (PG-VIB) and the Conditional Discrete Diffusion Probabilistic Model (CDDPM) ensures both the accuracy of performance prediction and the quantification of uncertainty, while also compensating for insufficient coverage of extreme operating conditions by generating scarce high-fidelity samples.

[0021] 4. Suppress physical illusions: The hierarchical physical constraint mechanism effectively suppresses the "physical illusions" of neural networks, ensuring that the model output always meets the basic physical laws of radar.

[0022] 5. Supports online continuous evolution: The continuous learning strategy combining Elastic Weight Consolidation (EWC), experience replay, and Low-Rank Adaptation (LoRA) supports incremental learning during long-term field service, avoiding catastrophic forgetting.

[0023] 6. Assessment-Attribution-Decision Closed Loop: The combination of three-dimensional uncertainty estimation and five-level maintenance decision-making forms a complete engineering closed loop with strong feasibility. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall architecture of the method of the present invention.

[0025] Figure 2 This is a schematic diagram of the physical covariance dimensionless preprocessing of the present invention.

[0026] Figure 3 This is a diagram of the shared-residual latent variable decoupling structure of the present invention.

[0027] Figure 4 This is a schematic diagram illustrating the dual-channel collaborative generation of the Physically Guided Variational Information Bottleneck (PG-VIB) and the Conditional Discrete Diffusion Probability Model (CDDPM) of this invention.

[0028] Figure 5 This is a schematic diagram illustrating the hierarchical physical constraints and cross-fidelity cyclic consistency of the present invention.

[0029] Figure 6 This is a closed-loop diagram of the three-dimensional uncertainty-driven online continuous learning and five-level maintenance decision-making of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0031] like Figures 1 to 6 As shown, a multi-fidelity twin method for airborne radar based on physically covariant shared residual latent variables includes the following steps: S1: Access radar data of at least two fidelities and perform dimensionless transformation on the original physical quantities based on Buckingham's π theorem to obtain dimensionless samples; S2: Encode each modality data separately using a multimodal encoder, and inject the continuous fidelity embedding vector into each encoder after adaptive layer normalization to obtain fused features; S3: Decouple the fused features into shared latent variables z s With residual latent variable z r The residual latent variable z r It is further divided into multiple subspaces bound to physical semantics, and triple decoupling constraints of mutual information minimization, fidelity adversarial, and scene invariance are imposed. S4: Construct a Physics-Guided Variational Information Bottleneck (PG-VIB) channel and a Conditional Discrete Diffusion Probabilistic Model (CDDPM) generation channel. The two channels are mutually verified through consistency loss. The KL divergence weights of the PG-VIB channel are... Adaptive adjustment based on fidelity level and physical consistency residual; S5: Embed hierarchical physical constraints and cross-fidelity cycle consistency constraints in the training objective; S6: Based on the three-dimensional uncertainty estimation of accidental uncertainty, cognitive uncertainty and physical residual uncertainty, active sampling and online continuous learning are driven, and based on the Shapley value contribution analysis of each subspace of the residual latent variables, hierarchical maintenance decisions are output.

[0032] Step S1: Multi-source data access and physical covariance dimensionless preprocessing. Access at least two data sources of varying fidelity, including but not limited to: electromagnetic full-wave simulation, ray tracing / physical optics simulation, system-level behavioral simulation, semi-physical injection test data, and field measurement data. Perform dimensionless transformation on each fidelity data based on Buckingham's π theorem: Let the original set of physical quantities be... By selecting a basis of characteristic dimensions (such as wavelength λ, reference distance R0, and reference power P0), dimensionless quantities are constructed: , , , The radar equations are rewritten in a dimensionless form, allowing the model to be trained in a dimensionless space. This preprocessing serves two purposes: first, it enables physical covariance and transfer generalization across platforms and frequency bands; second, it hides absolute parameters and retains only physical structure information in classified application scenarios.

[0033] Step S2: Multimodal Adaptive Encoding and Fidelity-Aware Fusion. To address the multimodal characteristics of radar data, multiple dedicated encoders are set up: an IQ temporal encoder (1D convolution + temporal attention), a range-Doppler map encoder (2D convolution), a point / track encoder (graph neural network or Transformer), and a state parameter encoder (MLP). Continuous fidelity embedding is introduced. ,in This represents the fidelity level, which is injected into each encoder through adaptive layer normalization, enabling the model to support inputs of any fidelity level.

[0034] Step S3: Decoupling of shared-residual latent variables. Map the encoder output to two sets of latent variables: shared latent variable z.s Characterizes the shared physical structure information across fidelity levels (such as radar equation structure, waveform parameters, and geometric relationships); residual latent variable z r Characterizing fidelity differences, individual degradation, and environmental disturbances, it is further subdivided into several physical semantic subspaces, including the antenna / array subspace z. ant T / R component subspace z tr Receive channel subspace z rx Launch channel subspace z tx Environment and Interference Subspace z env and the whole machine health subspace z health Apply triple decoupling constraints: minimize mutual information. ; Fidelity resistance constraint: The discriminator cannot distinguish from z s Identify fidelity levels; Scene invariance constraint: z-values ​​of different fidelities within the same scene. s They should be consistent. env Corresponding to the external electromagnetic environment spectrum characteristics, z health Corresponding to a comprehensive index of system-level performance degradation.

[0035] Step S4: Physically Guided Variational Information Bottleneck (PG-VIB) and Conditional Discrete Diffusion Probability Model (CDDPM) dual-channel generation. Channel A (PG-VIB backbone) is used for performance prediction and uncertainty estimation. The KL divergence weight β in its loss function is adaptively adjusted, and the calculation formula is as follows: , where R phys For physical consistency residuals, β0, γ f η and η are hyperparameters. Channel B (Conditional Diffusion Model CDDPM) is used to generate scarce high-fidelity samples. Guided by low-fidelity data, fidelity embedding, and scene conditions, a physical consistency discriminator is embedded during the denoising process. The radar equation residual is injected into the sampling process as a guiding term. The calculation formula is: The consistency loss L between the two channels cross Mutual verification: The PG-VIB output of the sample generated by diffusion should be consistent with the original low-fidelity prediction in terms of physical dimensions.

[0036] Step S5: Layered physical constraints and cross-fidelity iterative consistency training. Layered physical constraints include: hard constraints (forced to meet conditions through model structure: energy conservation, non-negativity of probability, integral normalization); strong soft constraints (high-weight penalties: radar equation structure, CFAR threshold relationship, coherent accumulation gain); weak soft constraints (low-weight regularization: antenna pattern smoothness, noise spectrum shape prior). The weights of each layer are adaptively adjusted according to fidelity. The cross-fidelity cycle consistency constraint is: .

[0037] Step S6: Uncertainty-driven online continuous learning and maintenance decision-making. Three-dimensional uncertainty estimation: random uncertainty. (Data noise), cognitive uncertainty (Model posterior) Physical residual uncertainty (Degree of violation of physical constraints). Total uncertainty. It is used to actively select flight test / experiment conditions, trigger online updates, and output prediction confidence intervals. The continuous learning strategy adopts a three-pronged approach: elastic weight solidification, experience replay, and low-rank adaptation fine-tuning. The decision-making closed loop is based on the Shapley value for the residual z of each physical subspace. r Perform performance attribution and output five levels of maintenance recommendations.

[0038] The total loss function is: .

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0040] Example 1: Detection distance prediction based on multi-fidelity fusion This embodiment demonstrates the specific implementation process of predicting airborne radar detection range using the method of the present invention.

[0041] Data preparation: approximately 50,000 electromagnetic full-wave simulation samples (high fidelity, f=1.0), approximately 500,000 ray tracing simulation samples (medium fidelity, f=0.6), approximately 5 million system-level simulation samples (low fidelity, f=0.3), and approximately 2,000 field measurement samples (measured, f=1.0).

[0042] Implementation steps: First, all samples are subjected to Buckingham π dimensionless transformation using module M1, with wavelength λ, reference distance R0, and reference power P0 as the basis, transforming all physical quantities into a dimensionless space. Then, the IQ data, RD map, track, and state parameters are processed by the multimodal encoder of module M2, and fidelity sensing is achieved through fidelity embedding f. Module M3 is used to decouple the fused features into shared latent variables z. s (Common physical structures such as radar acquisition equations) and residual hidden variable z r (Capturing fidelity differences and individual degradation). For scarce field-measured samples, data augmentation is performed through the conditional diffusion generation channel in module M4 to generate high-fidelity synthetic samples consistent with the distribution of the measured samples. Joint training is then conducted under hierarchical physical constraints in module M5, where the radar equation structure is applied as a strong soft constraint with high weights, and energy conservation is embedded as a hard constraint in the network structure. The final output is the predicted detection range and its three-dimensional uncertainties (chance, cognition, and physical residuals).

[0043] Results: Compared to the pure low-fidelity model, the root mean square error of the detection range prediction is reduced by about 45%; compared to the pure data-driven model, the residual of the radar equation is reduced to a negligible level, and the prediction results still satisfy physical laws under extreme conditions of high signal-to-noise ratio.

[0044] Example 2: Early Degradation Detection and Maintenance Decisions for T / R Components This embodiment demonstrates the process of using the present invention for degradation attribution and maintenance decision-making.

[0045] During field service, the system continuously collects radar echoes and status data. The residual latent variable z is obtained through decoupling via module M3. r Further decomposed into various physical semantic subspaces, where z tr Corresponding to the T / R component status.

[0046] The system monitors z in real time tr The change in the Shapley value contribution of the subspace. When z is found tr When the Shapley value continuously increases and exceeds a preset threshold, it is determined that the T / R component has early performance drift. According to a predefined five-level decision table: if only z... tr An anomaly with a performance degradation between 5% and 15% triggers an L2-level "retest confirmation"; if z tr Abnormal amplitude intensified and accompanied by z tx If an anomaly is detected, it is classified as a Level 3 "channel calibration"; if simultaneously z ant With z tr z tx If all abnormalities are found, the process is escalated to Level L4, requiring "component replacement".

[0047] Results: Compared with traditional alarm methods based on fixed thresholds, this method can issue an early warning when the performance of the T / R component drops by about 8%, and provide clear attribution and maintenance suggestions, avoiding the maintenance decision difficulties caused by "black box" prediction.

[0048] In this embodiment, the five-level maintenance decision table is shown in Table 1: Table 1 L1 Performance degradation <5% Delayed maintenance, encrypted monitoring L2 5%~15% and with high uncertainty Retest confirmed L3 T / R component residual abnormality Channel calibration L4 Multi-channel consistency degradation Replace components L5 Multi-subspace joint anomaly Return to factory for in-depth repair Example 3: Cross-platform model migration This embodiment demonstrates the portability of the method of the present invention across different radar platforms.

[0049] A digital twin model, fully trained on type A radar (such as an X-band airborne active phased array radar), will be transferred to type B radar (such as a C-band shipborne radar). Since all training takes place in dimensionless space, the physical nature of both radars is described by the same dimensionless radar equations; therefore, z s The encoder can be directly reused.

[0050] Specific steps: Freeze the M1 dimensionless module, M2 encoder body, M3 decoupling network, and M4 generator network in the pre-trained model; on a small number of measured samples (approximately 500) from the B-type radar, fine-tune only the low-rank adaptation (LoRA) layer in the M6 ​​continuous learning module. The fine-tuning process only requires updating approximately 0.5% of the parameters.

[0051] Results: The accuracy of the transferred model in predicting the detection range on the Type B radar is comparable to that of training from scratch on the same radar, while the training time is reduced by more than 90% and the data requirement is reduced by 95%.

[0052] Example 4: Evaluation of Anti-interference Performance under Complex Electromagnetic Interference This embodiment demonstrates the performance evaluation capability of the present invention in adversarial scenarios.

[0053] The interference type (e.g., suppression, deception), interference power (relative to signal-to-interference ratio), and interference direction (relative to the radar main lobe angle) are used as input conditions for the conditional diffusion generation channel of module M4. The diffusion model starts with low-fidelity, interference-free echoes and generates high-fidelity echo samples under diverse interference scenarios under conditional guidance, covering interference parameter combinations that are difficult to traverse in conventional simulations.

[0054] z obtained through decoupling from module M3 env In subspaces, the impact of disturbances on each physical subsystem is quantitatively analyzed. For example, suppression disturbances mainly contribute to z-axis. env , and for z tr The contribution is relatively small; deceptive interference may simultaneously affect z. tx and z rx .

[0055] Results: The system will eventually output an anti-interference performance evaluation report, which includes the detection range reduction curve under various interference conditions, the interference identification accuracy, and targeted countermeasures (such as "suggest switching to frequency agility mode").

[0056] Explanation of alternative implementation methods: Those skilled in the art should understand that the solutions disclosed in this invention can be implemented in various equivalent ways, and the following alternative solutions all fall within the protection scope of this invention: Encoder structure: CNN, Transformer, GNN, MLP are all acceptable, and there is no specific structure limit.

[0057] Generative models: DDPM, Flow Matching, VAE, and GAN can all be used as equivalent implementations of channel B.

[0058] Continuous learning: EWC can be replaced by similar regularization methods such as SI and MAS, and LoRA can be replaced by Adapter and Prefix-Tuning.

[0059] Number of fidelity levels: ≥2 is sufficient, not limited to the 4 levels in the example.

[0060] Radar systems: Pulse Doppler, phased array, AESA, SAR, and early warning / fire control radars are all applicable.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, modifications, equivalent substitutions, or improvements can be made to the technical solutions described in the foregoing embodiments without departing from the spirit and principles of the present invention, and the technical features in each embodiment can be combined according to actual needs; all such modifications, substitutions, improvements, and combinations should be included within the protection scope of the present invention.

Claims

1. A multi-fidelity twin method for airborne radar based on physically covariant shared residual latent variables, characterized in that, Includes the following steps: S1: Access radar data of at least two fidelities and perform dimensionless transformation on the original physical quantities based on Buckingham's π theorem to obtain dimensionless samples; S2: Encode each modality data separately using a multimodal encoder, and inject the continuous fidelity embedding vector into each encoder after adaptive layer normalization to obtain fused features; S3: Decouple the fused features into shared latent variables z s With residual latent variable z r The residual latent variable z r It is further divided into multiple subspaces bound to physical semantics, and triple decoupling constraints of mutual information minimization, fidelity adversarial, and scene invariance are imposed. S4: Construct the Physically Guided Variational Information Bottleneck (PG-VIB) channel and the Conditional Discrete Diffusion Probability Model (CDDPM) generation channel. The two channels are mutually verified through consistency loss. The KL divergence weights of the PG-VIB channel are specified in the provided text. Adaptive adjustment based on fidelity level and physical consistency residual; S5: Embed hierarchical physical constraints and cross-fidelity cycle consistency constraints into the training objective; S6: Based on the three-dimensional uncertainty estimation of accidental uncertainty, cognitive uncertainty and physical residual uncertainty, active sampling and online continuous learning are driven, and the hierarchical maintenance decision is output based on the Shapley value contribution analysis of each subspace of the residual latent variables.

2. The airborne radar multi-fidelity twin method based on physically covariant shared residual latent variables according to claim 1, characterized in that, The dimensionless transformation described in step S1 is based on wavelength. Based on the reference range R0 and reference power P0, the radar equations are rewritten in a dimensionless form to conceal absolute physical parameters in classified scenarios.

3. The airborne radar multi-fidelity twin method based on physically covariant shared residual latent variables according to claim 1, characterized in that, The residual latent variable z mentioned in step S3 r The subspace includes at least the antenna / array subspace z. ant T / R component subspace z tr Receive channel subspace z rx Launch channel subspace z tx Environment and Interference Subspace z env and the whole machine health subspace z health Several of them.

4. The airborne radar multi-fidelity twin method based on physically covariant shared residual latent variables according to claim 1, characterized in that, The KL divergence weights of the variational information bottleneck mentioned in step S4 Satisfy the following adaptive formula: in, For fidelity levels, R phys For physical consistency residuals, β0, γ f , This is a hyperparameter.

5. The airborne radar multi-fidelity twin method based on physically covariant shared residual latent variables according to claim 1, characterized in that, The conditional diffusion generation channel described in step S4 introduces a physical consistency discriminator in the denoising step, injecting the radar equation residual as a guiding term into the sampling process.

6. The airborne radar multi-fidelity twin method based on physically covariant shared residual latent variables according to claim 1, characterized in that, The hierarchical physical constraints mentioned in step S5 include: hard constraints enforced by the model structure, strong soft constraints applied in the form of high-weight penalties, and weak soft constraints applied in the form of low-weight regularization; and the weights of each layer constraint are adaptively adjusted according to the fidelity level; the hard constraints include, but are not limited to: enforcing non-negativity of probabilities and integral normalization through the Softmax layer, and enforcing energy conservation through network output constraints.

7. The airborne radar multi-fidelity twin method based on physically covariant shared residual latent variables according to claim 1, characterized in that, The online continuous learning described in step S6 adopts a strategy that combines Elastic Weights Fixed (EWC), Experience Replay, and Low-Rank Adaptive (LoRA) fine-tuning.

8. The airborne radar multi-fidelity twin method based on physically covariant shared residual latent variables according to claim 1, characterized in that, The maintenance decision in step S6 is based on the Shapley value contribution analysis of each subspace of the residual latent variables, and outputs a graded maintenance recommendation with five levels.

9. A system for implementing the airborne radar multifidelity twin method based on physically covariant shared residual latent variables as described in any one of claims 1 to 8, characterized in that, include: The data access and dimensionless preprocessing module is used to access radar data of at least two fidelities and perform dimensionless transformation on the original physical quantities based on Buckingham's π theorem to obtain dimensionless samples. The multimodal coding and fidelity fusion module is used to encode each modal data separately through a multimodal encoder, and inject the continuous fidelity embedding vector into each encoder after adaptive layer normalization to obtain fused features; The shared-residual latent variable decoupling module is used to decouple the fused features into shared latent variables z. s With residual latent variable z r The residual latent variable z r It is further divided into multiple subspaces bound to physical semantics, and triple decoupling constraints of mutual information minimization, fidelity adversarial, and scene invariance are imposed. The physical-guided variational information bottleneck and conditional diffusion dual-channel generation module is used to construct the physical-guided variational information bottleneck channel and the conditional diffusion generation channel. The two channels are mutually verified through consistency loss. The KL divergence weight β of the variational information bottleneck is adaptively adjusted with the fidelity level and the physical consistency residual. The Physical Constraints and Cyclic Consistency Training Module is used to embed hierarchical physical constraints and cross-fidelity cyclic consistency constraints into the training objective. The uncertainty estimation and maintenance decision module is used to estimate three-dimensional uncertainty based on accidental uncertainty, cognitive uncertainty and physical residual uncertainty, drive active sampling and online continuous learning, and determine the contribution of each subspace to performance degradation based on the Shapley value contribution analysis of each physical semantic subspace of the residual latent variables, and output graded maintenance suggestions containing five levels.

Images