Physical isogeny single-mode driving AI virtual multi-source signal generation method, system and platform

Abstract

The application discloses a kind of physical homologous single mode modal driving AI virtual multi-source signal generation method, system and platform, method obtains target object single source physical signal, establishes physical constraint model based on inherent physical characteristics, constructs differentiable physical compliance item L_physics With data fitting loss L_data form dynamic weighting joint optimization target, training generative AI learning complies with the cross-modal mapping of physical coupling mechanism, generates physical homologous, space-time synchronous virtual multi-source signal from single source signal, and completes physical consistency verification and adaptive calibration.The application and traditional PINN equation solving exist essential difference, realize high credibility, low cost, strong synchronous cross-modal signal generation, suitable for industrial nondestructive testing, rail transit, aerospace and other high safety level detection scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and physical signal processing technology, specifically to a method, system, and platform for generating AI virtual multi-source signals driven by a physically homogeneous single-modal signal. It is particularly suitable for testing scenarios with stringent requirements for signal reliability, hardware cost, and deployment efficiency, such as industrial non-destructive testing, rail transportation, aerospace, and power equipment. Background Technology

[0002] Existing multimodal signal acquisition suffers from two unavoidable technical drawbacks: (1) Multiple hardware sensors are costly, complex to deploy, asynchronous in timing, and spatially misaligned, and are easily subject to environmental and structural limitations; (2) Pure data-driven AI generation does not embed physical constraints during the training phase. The output signal violates physical laws, is unreliable and inexplicable, and cannot be used in high-security scenarios. Existing physical information neural networks (PINNs) are mainly used to solve forward / inverse problems of physical equations, and are not designed for the task of "generating another synchronous multimodal signal from a single mode". Therefore, they cannot directly solve the needs of low-cost, synchronous, and reliable virtual multi-source generation in industrial detection. The essential difference between it and PINN This invention differs fundamentally from physical information neural networks. Existing technologies such as PINN are mainly used to solve for the distribution of a single physical field under known physical equations and boundary conditions, and their physical residual constraints are used to approximate the exact solution of the equation. This invention aims to solve the specific task of cross-modal signal generation. The physical constraints are not used to solve a single physical field, but to force the transformation relationship between different physical modes to conform to the field coupling mechanism. The constraint objectives, mathematical expressions, optimization objectives and application scenarios are all fundamentally different. Therefore, the core of this invention lies in systematically embedding complex cross-modal physical coupling relationships into a general generative AI framework for the first time in the form of a differentiable, optimizable, and trainable loss function. This solves the specific technical problem of "directly constraining and guiding cross-modal signal generation with physical laws," and is not a simple application or extension of methods such as PINN. Summary of the Invention

[0003] 1. Technical problems to be solved (1) The problems of high cost, difficult deployment, and inability to strictly collect data synchronously from the same source by multiple hardware sensors; (2) Multi-physics coupling brings problems such as electromagnetic interference, signal distortion, and low signal-to-noise ratio; (3) The problem of difficulty in deploying multiple sensors in extreme scenarios such as confined spaces, high temperatures and high pressures; (4) Pure data-driven AI generates problems such as lack of physical constraints, unreliable signals, and unusable for high-reliability detection; (5) The problem that physical laws are difficult to compute, quantify, and forcibly embed into the AI ​​training process; (6) Virtual multi-source signals cannot achieve physical homogeneity, spatiotemporal synchronization, and physical reasonable generation; (7) The existing PINN is only used for solving equations and does not have the ability to generate virtual data across modes. 2. Quantitative Modeling Methods for Physical Laws This invention transforms inherent, unified physical laws into computable constraints, including one or more of the following: (1) Partial Differential Equation (PDE) Residual Method: Discretize the physical control equations and use the residuals as loss terms; (2) Conservation constraint method: The conservation laws of mass, energy, momentum, etc. are expressed as equations or inequalities; for example, the conservation of energy can be quantified as: L_physics_E = |Σ(E_input) − Σ(E_virtual)|²; (3) Constitutive relation embedding method: The field coupling, stress-strain and other relationships are used as network output constraints. 3. Overview of Technical Solution Implementation The core process of this invention is as follows: Single-modal signal acquisition → Physical constraint modeling → Construction of joint loss function → Generative AI training → Virtual multi-source signal generation → Physical consistency verification and adaptive calibration. Its core lies in transforming abstract physical laws into trainable mathematical constraints, forcing the model to learn the physical relationships between modes, rather than simply fitting the statistical distribution of data. 4. Beneficial effects (1) Hardware cost reduction of 80%–95%: Based on the comparison of the total cost of ownership (TCO) of the present invention (single sensor + algorithm) and high-end multi-sensor fusion system under the same detection performance; (2) Synchronization accuracy improved by 1–2 orders of magnitude: virtual signals are naturally homogeneous and have no timing offset; (3) Signal-to-noise ratio improvement of 3–10 dB: derived from the statistical average value of actual measurements in the examples; (4) Virtual signal credibility ≥ 95%: average matching degree of virtual and real signal features; (5) Generalization ability improved by more than 50%: On the new defect / new working condition test set, the performance degradation of the present invention is more than 50% lower than that of the pure data-driven baseline model; (6) Physical constraints force embedded training, and the signals are interpretable and verifiable; (7) It can be used in extreme scenarios such as confined spaces, high temperature and high pressure. 5. Physical consistency verification and adaptive calibration The consistency verification module calculates the theoretical physical parameters of the generated virtual signal and jointly backpropagates the error with the model input. The verification result serves as a trigger signal to guide the next round of data acquisition or model fine-tuning, achieving closed-loop adaptive calibration. 6 Systems and Platforms A physical homogeneous single-modal driven AI virtual multi-source signal generation system includes: a data acquisition module, a physical law constraint modeling module, an AI training and inference module, a virtual multi-source generation module, a physical consistency verification module, and a feedback calibration module. Each module is connected sequentially according to the data flow direction. The feedback calibration module feeds back the verification error to the AI ​​training and inference module to achieve closed-loop iterative optimization. An AI virtual multi-source signal generation platform, comprising the above-mentioned system. Attached Figure Description Figure 1 is a schematic diagram of the method flow of the present invention. 1—Single modal data acquisition, 2—Physical constraint construction, 3—Joint optimization training, 4—Virtual multi-source generation, 5—Physical consistency verification, 6—Feedback optimization and calibration output. The steps are connected sequentially in time to form a complete signal generation and verification process. Figure 2 is a schematic diagram of the system structure of the present invention. 1—Data acquisition module, 2—Physical law constraint modeling module, 3—AI training and inference module, 4—Virtual multi-source generation module, 5—Physical consistency verification module, 6—Data calibration module, 7—Feedback optimization module, 8—Output application module. Each module is connected in sequence according to the data flow direction. The output of the feedback optimization module is fed back to the AI ​​training and inference module, forming a closed-loop iterative optimization structure. Figure 3 is a schematic diagram of the physical constraint loss calculation and backpropagation path of the present invention. 1—Physical constraint modeling, 2—Loss function construction, 3—Gradient backpropagation, 4—Model parameter update. Each link is connected in sequence according to the data flow, which clarifies the mandatory guiding role of physical constraints on model training. Detailed Implementation Example 1: Rail Inspection (Magnetic Homologous Mechanism) Input: Single magnetic field signal Physical modeling: Maxwell's equations + magnetoacoustic coupling Physical constraints: Frequency domain transfer function H(f) Physical loss: L_physics = ||S_virtual−H(f)・S_magnetic||² AI Model: Adaptively Modified Conditional Generative Adversarial Network (CGAN) Joint loss: L_total = L_adv + α・L_rec + β・L_physics Training strategy: Gradually increase β from 0.01 to 0.3, first fit the data and then strengthen the physical constraints. Dataset: 1000 sets of standard defect data, split 8:2; the test set includes new defects not found in the training set, used to verify generalization ability. Baseline Comparison: Identical to the present invention in structure and hyperparameters, except that pure data-driven CGAN based on L_physics is removed. Generalization capability definition: the percentage of performance degradation on a new defect test set. Results: The attenuation of the present invention is ≤5%, the attenuation of the baseline model is ≥55%, and the generalization ability is improved by more than 50%; the defect detection rate is 98.2%, and the physical consistency is 100%. Physical consistency verification: The theoretical signal is obtained by semi-analytical solution of Maxwell's equations. The correlation coefficient and energy error are calculated. If the threshold is not met, incremental training is performed. Example 2: Industrial Pipeline Inspection (Acoustic-Electrical Coupling) Input: Single acoustic signal Physical modeling: Wave equation + acoustic-electric coupling relationship Results: Positioning error ≤ 0.5mm, physical consistency 100%, signal-to-noise ratio improved by 6dB. Example 3: Rotating Component Inspection (Vibration, Acoustics, Electromagnetic Correlation) Input: Single vibration signal Physical modeling: dynamic equations + multi-field coupling relationships Results: Fault identification accuracy was 97.5%, and the cross-device generalization performance decreased by ≤5%.

Claims

1. A method for generating cross-modal virtual signals with enhanced physical law constraints, characterized in that, include: Acquire a single-source physical signal of the target object; Establish a physical constraint model between the single-source physical signal and at least one target modal signal, based on the inherent physical characteristics of the target object; The physical constraint model is quantized into differentiable physical law compliance terms L_physics; Construct a joint optimization objective L_total, which is a weighted sum of the data fitting loss term L_data and L_physics using dynamically adjusted weight coefficients λ(t). L_total = L_data + λ(t)・L_physics; The joint optimization objective L_total is used to drive the training of the generative artificial intelligence model, enabling it to learn the physically constrained mapping relationship from the source signal to the target modal signal; Using the trained model, a virtual target modal signal that is physically homologous to and spatiotemporally synchronized with the source signal is generated from a single source physical signal; The virtual signal is subjected to physical consistency verification and adaptive calibration.

2. The method according to claim 1, characterized in that, The physical law compliance term L_physics is constructed in at least one of the following ways: The residual is derived from the coupling relationship model F(x,y)=0 between the first mode signal x and the second mode signal y described by the control equation; The energy difference or quality difference between the source signal and the generated signal calculated based on the conservation law; Alternatively, the constraints can be used as boundary conditions for the neural network output and embedded in the network's forward computation process.

3. The method according to claim 1, characterized in that, The weighting coefficient λ(t) follows the following strategy: In the early stages of training, λ(t) approaches 0, and optimization is mainly based on L_data; During the training process, λ(t) increases monotonically according to the physical consistency index of the validation set to strengthen the adherence to physical laws.

4. The method according to claim 1, characterized in that, The physical constraint model is the frequency domain transfer function H(f) from the magnetic field signal to the eddy current signal or ultrasonic signal at the target defect, derived from Maxwell's equations.

5. The method according to claim 4, characterized in that, The physical laws obeyed are as follows: L_physics=||S_virtual−H(f)・S_input||².

6. The method according to claim 1, characterized in that, The generative artificial intelligence model is an adaptively modified conditional generative adversarial network (CGAN), where the physical constraint loss directly affects the generator and participates in backpropagation.

7. A physically homogeneous single-mode driven AI virtual multi-source signal generation system, characterized in that, include: The module includes: data acquisition module, physical law constraint modeling module, AI training and inference module, virtual multi-source generation module, physical consistency verification module, and feedback calibration module. The modules are connected in sequence, and the output of the feedback calibration module is fed back to the AI ​​training and inference module to form a closed-loop iterative optimization.

8. An AI virtual multi-source signal generation platform, characterized in that, It includes the system described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of claim 1.

Images