Electromechanical simulation model evaluation method and system based on AI analysis

Abstract

The invention belongs to the technical field of electromechanical system simulation, and particularly relates to an electromechanical simulation model evaluation method based on AI analysis, which comprises the following steps: data preprocessing, AI evaluation engine construction, dynamic index evaluation and closed-loop optimization. An electromechanical simulation model evaluation system based on AI analysis comprises the following layered architecture: a data preprocessing layer, an AI evaluation engine layer, a dynamic evaluation layer and a closed-loop optimization layer, one of kernel principal component analysis or a variational auto-encoder is introduced to carry out nonlinear feature extraction, and game Nash equilibrium is introduced to carry out model weight dynamic allocation. And sliding window anomaly detection is introduced to dynamically adjust the threshold. According to the invention, a'physical rule + AI reasoning 'dual-drive evaluation architecture of the electromechanical simulation model is initiated; the dynamic weight distribution algorithm improves the evaluation efficiency compared with a traditional method; and the three-level indexes converge step by step, so that comprehensive evaluation from local performance to global performance is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromechanical system simulation, and specifically relates to an electromechanical simulation model evaluation method and system based on AI analysis. Background Art

[0002] Electrical simulation model evaluation refers to verifying the degree of approximation and credibility of the simulation model to the dynamic characteristics, reliability and performance of the actual electromechanical system through quantitative or qualitative methods. Its core includes the following elements: Model verification: testing whether the model can accurately reflect the system behavior through physical rules (such as differential equations or data-driven methods (such as Bayesian networks); parameter impact analysis: evaluating the impact of uncertain parameters (such as material properties, load conditions) on simulation results, such as non-probabilistic measurement methods; reliability evaluation: quantifying the robustness of the model under extreme working conditions based on indicators such as failure time distribution; dynamic adaptability: dynamically adjusting model parameters in combination with real-time data (such as digital twin technology) to ensure the consistency of evaluation results with the actual system.

[0003] The process aims to ensure the engineering applicability of simulation models in scenarios such as design, optimization, and fault prediction through a combination of mathematical modeling, statistical analysis, and experimental verification.

[0004] Problems with existing technologies: Traditional electromechanical simulation model evaluation mostly relies on static indicators and manual experience, and has problems such as delayed response and poor adaptability to complex working conditions. In the prior art, CN112668223A proposes a lightweight model simulation method based on digital twins, but does not involve an AI-driven dynamic evaluation mechanism. At the same time, although the Bayesian network achieves an accuracy rate of 98% in fault diagnosis, it lacks the ability to evaluate the overall performance of the simulation model. Summary of the invention

[0005] The purpose of the present invention is to provide an electromechanical simulation model evaluation method and system based on AI analysis, which is the first electromechanical simulation model "physical rules + AI reasoning" dual-driven evaluation architecture; the dynamic weight allocation algorithm improves the evaluation efficiency compared with the traditional method; the three-level indicators converge step by step to achieve a comprehensive evaluation from local performance to global effectiveness.

[0006] The technical solution adopted by the present invention is as follows: An electromechanical simulation model evaluation method based on AI analysis includes the following steps: Step 1: Data preprocessing: Use parametric modeling technology to build a fully parametric geometric model of the electromechanical system, combine Simulink to generate a multi-condition simulation data set, use principal component analysis (PCA) to extract key feature parameters and establish a standardized input matrix; Step 2, AI Evaluation Engine Construction: Integrate the differential equation model of the electromechanical system based on physics, the Bayesian network dynamic inference module, and the lightweight digital twin model to form a hybrid-driven evaluation architecture; Step 3, Dynamic Index Evaluation: Calculate the comprehensive score through a three-level index system. The three-level indexes include the basic layer (model response speed, computing resource occupancy rate), the enhancement layer (multi-physical field coupling accuracy, robustness under extreme conditions), and the intelligent layer (AI prediction deviation rate from measured data, self-learning iteration efficiency); Step 4, Closed-loop Optimization: Dynamically adjust the model parameters according to the online safety assessment algorithm. When the deviation of key parameters exceeds the threshold, trigger the model reconstruction mechanism, and realize the dynamic weight allocation of evaluation indexes through the reinforcement learning framework.

[0007] The fault diagnosis accuracy rate of the Bayesian network dynamic inference module is ≥95%.

[0008] The real-time state update frequency of the lightweight digital twin model is ≤10ms.

[0009] An electromechanical simulation model evaluation system based on AI analysis includes the following hierarchical architecture: Data Preprocessing Layer: Configured to construct a full-parameterized model of the electromechanical system through parametric geometric modeling technology, generate a multi-condition simulation data set in combination with Simulink, and extract key feature parameters by using principal component analysis to establish a standardized input matrix; AI Evaluation Engine Layer: Configured to integrate the following modules to form a hybrid-driven architecture: Physical Rule Model: Describe the dynamic characteristics of multi-physical field coupling based on the differential equations of the electromechanical system; Bayesian Network Inference Module: Realize the prediction of the fault propagation path through a dynamic conditional probability table, and the diagnosis accuracy rate is ≥95%; Digital Twin Synchronization Unit: Support a real-time state update frequency of ≤10ms, and realize model lightweight through proper orthogonal decomposition; Dynamic Evaluation Layer: Configured to calculate the comprehensive score through a three-level index system, including: Basic Layer: Model response speed and computing resource occupancy rate; Enhancement Layer: Multi-physical field coupling accuracy and robustness under extreme conditions; Intelligent Layer: AI prediction deviation rate and self-learning iteration efficiency; Closed-loop Optimization Layer: Configured to dynamically allocate the weights of evaluation indexes through the reinforcement learning framework, and trigger the model reconstruction mechanism when the deviation of key parameters exceeds the threshold.

[0010] The closed-loop optimization module automatically pushes the optimization scheme to the digital twin model through the knowledge engineering library.

[0011] The dynamic evaluation module initializes the index weights using the fuzzy analytic hierarchy process and dynamically adjusts the weight allocation through a reinforcement learning framework.

[0012] In the data preprocessing layer, either kernel principal component analysis or variational autoencoder is introduced for non-linear feature extraction to solve the problem of limited feature extraction ability of principal component analysis for non-linear electromechanical systems.

[0013] In the AI evaluation engine layer, game theory Nash equilibrium is introduced for dynamic model weight allocation to solve the problem that the weight allocation of physical models, Bayesian networks, and digital twins all depends on static rules. Cross-modal contrast learning is introduced to align the feature space for integrating multi-modal data such as vibration, thermal imaging, and acoustic fingerprint for multi-modal data fusion evaluation. Generative adversarial networks are introduced to synthesize extreme working condition data to solve the problem of insufficient measured data.

[0014] In dynamic index evaluation, sliding window anomaly detection is introduced to dynamically adjust the threshold to solve the problem that the model reconstruction trigger threshold is a fixed value and cannot adapt to working condition changes.

[0015] In closed-loop optimization, constrained reinforcement learning is introduced to improve the reward function. A federated learning + model distillation architecture is introduced to alleviate the problem of high edge computing resource pressure. A security monitoring proxy layer is introduced to clarify the fault propagation isolation strategy.

[0016] The technical effects achieved by the present invention are as follows: (1) In the present invention, the linkage between parameterization and simulation directly drives the Simulink dynamic model through geometric parameters, avoiding manual repeated modeling; the cooperation between PCA and standardization, first dimension reduction and then normalization, not only retains the main features but also eliminates the influence of dimension; introducing kernel principal component analysis (KPCA) or variational autoencoder (VAE) for non-linear feature extraction, optimizing the limited feature extraction ability of non-linear electromechanical systems and enhancing non-linear feature extraction.

[0017] (2) The present invention is the first to create a dual-driven evaluation architecture of "physical rules + AI reasoning" for electromechanical simulation models, which is specifically manifested as follows: multi-physical field coupling modeling: accurate description of the dynamic characteristics of the electromechanical system is achieved by combining the elastic medium equation and the Maxwell equations; integration of multi-modal data such as vibration, thermal imaging, and soundprint to improve the comprehensiveness of the evaluation; dynamic reasoning and lightweight collaboration: Bayesian networks handle uncertainty, and digital twin models achieve real-time performance through POD reduction; game theory Nash equilibrium is used to dynamically allocate model weights to compensate for the lack of dynamic collaboration and optimize the multi-model collaboration mechanism; adversarial networks are generated to synthesize extreme working condition data, thereby enhancing data diversity; and the automation, credibility, and adaptability of AI in the evaluation process are significantly improved.

[0018] (3) The dynamic weight allocation algorithm of the present invention improves the evaluation efficiency by 40% compared with the traditional method. The specific performance is as follows: dynamic and static weight fusion: combining AHP static weight and fuzzy dynamic correction to balance expert experience and real-time data feedback; multi-level scoring mechanism: through the step-by-step convergence of three-level indicators, a comprehensive evaluation from local performance to global efficiency is achieved; sliding window statistics are used to update the threshold in real time to adapt to changes in working conditions.

[0019] (4) The present invention combines PID control with reinforcement learning to achieve dual-time scale control of "real-time adjustment-long-term optimization"; dynamic weight adaptation: through Q-learning, the weight distribution is directly linked to the evaluation index reward to improve the robustness of the system; at the same time, safety constraints are added and the fault propagation isolation strategy can be clarified. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0022] like Figure 1 As shown in the figure, a method for evaluating electromechanical simulation models based on AI analysis is shown in the figure. In data preprocessing, the specific process is as follows: Step 1: Parametric Geometry Modeling The mathematical relationship between the geometric characteristics and physical parameters of the electromechanical system is defined through parametric modeling technology, such as: geometric parameters: size; physical parameters: mass, damping coefficient; , where M is the parameterized geometric model and f(.) is the parameter mapping function.

[0023] Step 2: Generation of Simulink Multi - operating Condition Simulation Datasets Configure the dynamic model of the electromechanical system in Simulink, set multiple groups of operating condition parameters (such as load, rotational speed, temperature), and run the simulation to generate a time - series dataset; , where is the simulation data vector at the t - th moment under the s - th operating condition, and S is the total number of operating conditions.

[0024] Step 3: Feature Extraction by Principal Component Analysis (PCA) Standardization: Eliminate the difference in dimensions , where and are the mean and standard deviation of the j - th feature respectively; Calculation of the covariance matrix: , where Z is the standardized data matrix and N is the number of samples; Eigenvalue decomposition: , where the eigenvectors corresponding to the first K largest eigenvalues are selected to form the projection matrix ; Dimensionality reduction mapping: ; Based on the above - mentioned feature extraction methods, the current PCA is a linear dimensionality reduction method, and its ability to extract features of non - linear electromechanical systems (such as friction and contact non - linearity) is limited. Therefore, kernel principal component analysis (KPCA) or variational auto - encoder (VAE) is introduced here for non - linear feature extraction: ; , where (VAE loss function).

[0025] Step 4: Construction of the Standardized Input Matrix Normalize the data after PCA dimensionality reduction to ensure that each feature is at the same magnitude; , The final input matrix .

[0026] According to the above content, the linkage between parameterization and simulation directly drives the Simulink dynamic model through geometric parameters, avoiding manual repeated modeling; the cooperation between PCA and standardization, first reducing dimensions and then normalizing, not only retains the main features but also eliminates the influence of dimensions; introducing kernel principal component analysis (KPCA) or variational auto - encoder (VAE) for non - linear feature extraction optimizes the situation where the ability to extract features of non - linear electromechanical systems is limited and enhances non - linear feature extraction.

[0027] In the construction of the AI ​​evaluation engine, the specific process is as follows: Step 1: Physics-based differential equation modeling of electromechanical systems Establish the multi-physics field coupling equations of the electromechanical system, including mechanical dynamics equations, electromagnetic field equations and control equations: Among them, the mechanical dynamics equations (elastic medium equations): , where ρ is density, u is displacement field, σ is stress tensor, and f is external force; Electromagnetic field equations (Maxwell's equations): , where E is the electric field intensity, B is the magnetic induction intensity, and J is the current density; Simultaneous solution: The differential equations are solved by time-domain simulation and numerical integration (such as Runge-Kutta method) is used to solve the state variables: .

[0028] Step 2: Multimodal data fusion evaluation Align feature space using cross-modal contrastive learning (CLIP): ,in , are visual and text features respectively; By integrating multimodal data such as vibration, thermal imaging, and voiceprint, the comprehensiveness of the evaluation is improved and multimodal data fusion evaluation is achieved.

[0029] Step 3: Bayesian network dynamic reasoning module construction Node definition: The state variables of the electromechanical system (such as temperature and vibration) are used as nodes, and the fault modes are used as hidden variables; Conditional Probability Table (CPT): learns the conditional probability relationship between nodes based on historical data; Among them, the Bayesian network structure is: , where Parents(Xi) represents the parent node set of node Xi; Dynamic Inference: Update the posterior probability through variational approximate inference: .

[0030] Step 4: Lightweight digital twin model construction Geometry lightweighting: compressing 3D model data through mesh simplification and parameterization; Dynamic synchronization: establish data mapping relationship between physical entity and virtual model; Among them, the model is reduced in order: the dominant mode is extracted using proper orthogonal decomposition (POD): ,in is the basis function, is the time-varying coefficient; Real-time synchronization equations: ,in is the mapping function, is the synchronization error.

[0031] Step 5: Hybrid drive architecture integration Data coupling: The differential equation output is used as the input of the Bayesian network, and the digital twin model provides real-time status feedback; Joint optimization: Coordinate the outputs of physical and data-driven models through dynamic weight assignment; Among them, the hybrid drive equation is: ,in is the dynamic weight coefficient, satisfying ; Weight update rule (reinforcement learning framework): , where η is the learning rate and R is the evaluation indicator reward function.

[0032] Based on the above process, since the weight allocation of the physical model, Bayesian network and digital twin all rely on static rules, the game theory Nash equilibrium is also used to dynamically allocate model weights: ,in Provide real-time contribution to each model to make up for the lack of dynamic coordination.

[0033] Step 6: Introducing Generative Adversarial Networks Constructing conditional GAN ​​(CGAN) to generate multi-physics coupling data: , where c is the operating condition vector; By generating adversarial networks to synthesize extreme working condition data, the problem of insufficient measured data can be solved, thereby enhancing data diversity.

[0034] According to the above content, multi-physics field coupling modeling: through the combination of elastic medium equations and Maxwell equations, the dynamic characteristics of the electromechanical system can be accurately described; the integration of multi-modal data such as vibration, thermal imaging, and soundprint can improve the comprehensiveness of the evaluation; dynamic reasoning and lightweight collaboration: Bayesian networks handle uncertainty, and the digital twin model achieves real-time performance through POD reduction; the game theory Nash equilibrium is used to dynamically allocate model weights to make up for the lack of dynamic collaboration and optimize the multi-model collaboration mechanism; the adversarial network is generated to synthesize extreme working condition data, thereby enhancing data diversity; Significantly improve the automation, credibility and adaptability of AI in the evaluation process.

[0035] In the evaluation of dynamic indicators, the specific process is as follows: Step 1: Index standardization Perform dimensionless processing on the third-level indicators to eliminate the differences in dimensions and magnitudes: score standardization (applicable to normally distributed data): , where and are the mean and standard deviation of the j-th indicator respectively; Min-Max standardization (applicable to non-normally distributed data): .

[0036] Step 2: Weight assignment Use the Analytic Hierarchy Process (AHP) to determine the static weights: · Construct a judgment matrix , where represents the importance ratio of indicator i relative to indicator j; · Calculate the weight vector , where is the maximum eigenvalue; Use fuzzy comprehensive evaluation to dynamically adjust the weights: · Define the membership function , and map the indicator values to the fuzzy set; · Calculate the dynamic weight correction coefficient through the fuzzy relation matrix R: .

[0037] Step 3: Comprehensive score calculation Sum the weights level by level according to the third-level indicator hierarchy to generate the final score; Single-level indicator score: , where is the weight, is the standardized value; Global comprehensive score: , where is the hierarchical weight (basic layer ( ), enhancement layer ( ), intelligent layer ( )), ( ) is the score of each layer.

[0038] Step 4: Dynamic feedback optimization Trigger the model parameter adjustment or reconstruction mechanism according to the scoring results: Parameter adaptive update: , where θ is the model parameter and η is the learning rate; Model reconstruction trigger condition: , where is a preset threshold (e.g., the key parameter deviation rate ≥ 5%).

[0039] Based on the above, the model reconstruction trigger threshold is a fixed value. To adapt to the changes in working conditions (such as sudden load changes), the threshold is dynamically adjusted here based on the sliding window anomaly detection: , where k is the adjustment coefficient, and the threshold is updated in real time through this sliding window statistic.

[0040] According to the above, the mathematical expression example of the third-level index Basic layer: Model response speed: First Token latency; Computational resource occupancy rate: Memory peak value; Enhanced layer: Multi-physical field coupling accuracy: ; Robustness under extreme working conditions: , where is the stability indicator function; Intelligent layer: AI prediction deviation rate: ; Self-learning iteration efficiency: , where ΔZ is the score improvement amount and Δt is the iteration time; Combining the above technologies, dynamic and static weight fusion: Combining the AHP static weight and fuzzy dynamic correction to balance expert experience and real-time data feedback; Multi-level scoring mechanism: Realizing the comprehensive evaluation from local performance to global efficiency through the gradual convergence of the third-level indicators; Using the sliding window statistic to update the threshold in real time to adapt to the changes in working conditions.

[0041] In the closed-loop optimization, the specific process is as follows: Step 1: Online safety assessment and deviation monitoring Collect the state data of the electromechanical system (such as temperature, vibration, current) in real time through sensors, compare it with the predicted value of the digital twin model, and calculate the key parameter deviation; Deviation calculation: , where is the measured value, is the predicted value of the digital twin model; Safety threshold determination: , where is the preset safety threshold (e.g., the insulation aging parameter deviation threshold is set to 5%).

[0042] Step 2: Dynamic adjustment of model parameters Based on the deviation feedback, the model parameters are adjusted through control algorithms (such as PID or reinforcement learning) to restore the system to a stable state; PID control adjustment: , where are the proportional, integral, and differential coefficients; Gradient descent optimization: , where η is the learning rate, is the loss function (such as mean squared error).

[0043] Step 3: Reinforcement learning dynamic weight allocation Through the reinforcement learning framework (such as Q-learning or PPO algorithm), the three-level weight allocation is dynamically adjusted according to real-time evaluation indicators; Reward function design: , where α, β, γ are the reward coefficients, is the three-level score; Weight update rule (taking Q-learning as an example): , where s is the state (current weight distribution), a is the action (weight adjustment strategy), and γ is the discount factor; Based on the above, the current reward function does not explicitly include safety constraints (such as overload protection), so constraint reinforcement learning (CRL) is also introduced here: , where is the safety cost function, ϵ is the constraint upper limit, which improves the reinforcement learning reward function.

[0044] Step 4: Trigger the model reconstruction mechanism When the deviation of the key parameters continuously exceeds the threshold, call the predefined rules in the knowledge engineering library to reconstruct the model structure or parameters; Reconstruction decision function: , where is the maximum allowable number of overlimit times; Model reconstruction strategy: parameter correction, ; structure optimization, adjust the model complexity through robust optimization algorithms.

[0045] Based on the above, the real-time update frequency range of the digital twin model is ≤10ms, but the requirements of this range pose a great pressure on the edge computing resources. Therefore, a federated learning + model distillation architecture is adopted: The edge nodes run lightweight student models (such as MobileNet) The teacher model is trained in the cloud and the edge model is optimized through knowledge distillation: , using this architecture to relieve the great pressure on the edge computing resources.

[0046] In addition, the fault propagation isolation strategy is not clearly defined in this closed-loop optimization. Therefore, a safety monitoring agent layer is also introduced: Predict the fault propagation path based on LSTM; Cut off the abnormal subsystem through the isolation valve: .

[0047] According to the above content: Closed-loop feedback fusion: Combine PID control and reinforcement learning to achieve dual-time-scale control of "real-time adjustment - long-term optimization"; Dynamic weight adaptation: Directly associate weight allocation with evaluation index rewards through Q-learning to improve the system's robustness; At the same time, add safety constraints; Alleviate the situation of heavy pressure on edge computing resources; And be able to clarify the fault propagation isolation strategy.

[0048] An electromechanical simulation model evaluation system based on AI analysis includes the following hierarchical architecture: Data preprocessing layer: Configured to build a fully parametric model of the electromechanical system through parametric geometric modeling technology, generate a multi-condition simulation data set in combination with Simulink, and extract key feature parameters using principal component analysis to establish a standardized input matrix; AI evaluation engine layer: Configured to form a hybrid drive architecture by integrating the following modules: Physical rule model: Describe the multi-physical field coupling dynamic characteristics based on the differential equations of the electromechanical system; Bayesian network inference module: Realize the prediction of the fault propagation path through a dynamic conditional probability table, and the diagnostic accuracy rate ≥ 95%; Digital twin synchronization unit: Support a real-time status update frequency ≤ 10ms, and achieve model lightweight through proper orthogonal decomposition; Dynamic evaluation layer: Configured to calculate the comprehensive score through a three-level index system, including: Basic layer: Model response speed and computing resource occupancy rate; Enhanced layer: Multi-physical field coupling accuracy and robustness under extreme conditions; Intelligent layer: AI prediction deviation rate and self-learning iteration efficiency; Closed-loop optimization layer: Configured to dynamically allocate evaluation index weights through a reinforcement learning framework, and trigger the model reconstruction mechanism when the deviation of key parameters exceeds the threshold.

[0049] The closed-loop optimization module automatically pushes the optimization plan to the digital twin model through the knowledge engineering library.

[0050] The dynamic evaluation module initializes the index weights using the fuzzy analytic hierarchy process and dynamically adjusts the weight allocation through a reinforcement learning framework.

[0051] In the data preprocessing layer, either kernel principal component analysis or variational autoencoder is introduced for non-linear feature extraction to solve the problem that the principal component analysis has limited ability to extract features of non-linear electromechanical systems.

[0052] In the AI evaluation engine layer, game theory Nash equilibrium is introduced for dynamic allocation of model weights to solve the problem that the weight allocation of physical models, Bayesian networks and digital twins all depends on static rules. Cross-modal contrast learning is introduced to align the feature space for integrating multi-modal data such as vibration, thermal imaging and voiceprint for multi-modal data fusion evaluation. Generative adversarial network is introduced to synthesize extreme working condition data to solve the problem of insufficient measured data.

[0053] In the dynamic index evaluation, sliding window anomaly detection is introduced to dynamically adjust the threshold to solve the problem that the model reconstruction trigger threshold is a fixed value and cannot adapt to the change of working conditions.

[0054] In the closed-loop optimization, constrained reinforcement learning is introduced to improve the reward function. Federated learning + model distillation architecture is introduced to alleviate the problem of high pressure on edge computing resources. A security monitoring proxy layer is introduced to clarify the fault propagation isolation strategy.

[0055] The working principle of this system: Specifically refer to the specific steps of the above-mentioned method for evaluating an electromechanical simulation model based on AI analysis.

[0056] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. The structures, devices and operation methods not specifically described and explained in the present invention are implemented according to the conventional means in the art without special description and limitation.

Claims

1. A method for evaluating an electromechanical simulation model based on AI analysis, characterized in that: The following steps are involved: Step 1: Data preprocessing: Use parametric modeling technology to build a fully parametric geometric model of the electromechanical system, combine Simulink to generate a multi-condition simulation data set, use principal component analysis (PCA) to extract key feature parameters and establish a standardized input matrix; Step 2: AI evaluation engine construction: Integrate the physics-based electromechanical system differential equation model, Bayesian network dynamic reasoning module and lightweight digital twin model to form a hybrid drive evaluation architecture; Step 3: Dynamic indicator evaluation: Calculate the comprehensive score through a three-level indicator system, which includes the basic layer (model response speed, computing resource occupancy), the enhanced layer (multi-physics field coupling accuracy, extreme working condition robustness) and the intelligent layer (AI prediction and measured data deviation rate, self-learning iteration efficiency). Step 4: Closed-loop optimization: Dynamically adjust model parameters according to the online security assessment algorithm. When the deviation of key parameters exceeds the threshold, the model reconstruction mechanism is triggered, and the dynamic weight allocation of evaluation indicators is realized through the reinforcement learning framework.

2. The method according to claim 1, characterized in that The fault diagnosis accuracy of the Bayesian network dynamic reasoning module is ≥95%.

3. The method according to claim 1, characterized in that The real-time status update frequency of the lightweight digital twin model is ≤10ms.

4. An electromechanical simulation model evaluation system based on AI analysis, characterized in that: It includes the following layered architecture: Data preprocessing layer: It is configured to build a fully parametric model of the electromechanical system through parametric geometric modeling technology, generate multi-condition simulation data sets in combination with Simulink, and use principal component analysis to extract key characteristic parameters to establish a standardized input matrix; AI evaluation engine layer: It is configured to integrate the following modules to form a hybrid drive architecture: Physical rule model: describes the dynamic characteristics of multi-physical field coupling based on the differential equations of the electromechanical system; Bayesian network reasoning module: Fault propagation path prediction is achieved through dynamic conditional probability tables, with a diagnostic accuracy of ≥ 95%; Digital twin synchronization unit: supports real-time status update frequency ≤ 10ms, and achieves model lightweighting through intrinsic orthogonal decomposition; Dynamic evaluation layer: configured to calculate the comprehensive score through a three-level indicator system, including: Basic layer: model response speed and computing resource occupancy rate; Enhanced layer: multi-physics field coupling accuracy and extreme working condition robustness; Intelligence layer: AI prediction deviation rate and self-learning iteration efficiency; Closed-loop optimization layer: It is configured to dynamically assign evaluation indicator weights through a reinforcement learning framework and trigger a model reconstruction mechanism when the deviation of key parameters exceeds a threshold.

5. The electromechanical simulation model evaluation system based on AI analysis according to claim 4, characterized in that: The closed-loop optimization module automatically pushes the optimization solution to the digital twin model through the knowledge engineering library.

6. The electromechanical simulation model evaluation system based on AI analysis according to claim 4, characterized in that: The dynamic evaluation module uses fuzzy analytic hierarchy process to initialize indicator weights and dynamically adjusts weight distribution through a reinforcement learning framework.

7. The electromechanical simulation model evaluation system based on AI analysis according to claim 4, characterized in that: In the data preprocessing layer, either kernel principal component analysis or variational autoencoder is introduced for nonlinear feature extraction to solve the problem that principal component analysis has limited feature extraction capabilities for nonlinear electromechanical systems.

8. The electromechanical simulation model evaluation system based on AI analysis according to claim 4, characterized in that: In the AI ​​evaluation engine layer, the Nash equilibrium of game theory is introduced to dynamically allocate model weights to solve the problem that the weight allocation of physical models, Bayesian networks, and digital twins all rely on static rules; Introducing cross-modal contrast learning to align feature space, integrating vibration, thermal imaging, and voiceprint multimodal data for multimodal data fusion evaluation; A generative adversarial network is introduced to synthesize extreme working condition data to solve the problem of insufficient measured data.

9. The electromechanical simulation model evaluation system based on AI analysis according to claim 4, characterized in that: In the dynamic indicator evaluation, sliding window anomaly detection is introduced to dynamically adjust the threshold to solve the problem that the model reconstruction trigger threshold is a fixed value and cannot adapt to changes in working conditions.

10. The electromechanical simulation model evaluation system based on AI analysis according to claim 4, characterized in that: In closed-loop optimization, constrained reinforcement learning is introduced to improve the reward function; Introducing the federated learning + model distillation architecture to alleviate the problem of heavy pressure on edge computing resources; A security monitoring agent layer is introduced to clarify the fault propagation isolation strategy.

Images