Engineering experiment data prediction and optimization method and system based on ai algorithm

Abstract

The application provides an engineering experiment data prediction and optimization method and system based on an AI algorithm, and belongs to the technical field of the cross of artificial intelligence, calculation engineering and experimental design.The method comprises the following steps: performing multi-fidelity fusion processing on high-fidelity experiment data and low-fidelity simulation data, generating a fusion experiment simulation data set, constructing an initial physical constraint model, embedding a preset engineering system physical mechanism constraint condition, and generating a physical constraint embedded model; and the engineering experiment data prediction and optimization method based on the AI algorithm improves the precision of performance prediction of a complex engineering system, makes the prediction result more in line with the actual situation, and provides a reliable basis for experiment optimization.

Description

Technical Field

[0001] This invention proposes a method and system for predicting and optimizing engineering experimental data based on AI algorithms, belonging to the interdisciplinary field of artificial intelligence, computational engineering, and experimental design. Background Technology

[0002] In many engineering fields such as aerospace, energy, and civil engineering, engineering experiments are crucial for the accurate understanding and optimization of system performance. However, current methods for predicting and optimizing engineering experiment data face numerous challenges that urgently need to be addressed.

[0003] While high-fidelity experiments provide accurate data, their high cost results in sparse sample sizes, making it difficult to comprehensively cover various operating conditions of complex engineering systems. Low-fidelity simulations, on the other hand, are fast but have significant biases. The failure to effectively integrate these two approaches leads to inefficient data utilization. Most mainstream AI models are black-box models, lacking physical consistency, and their predictions are prone to instability when exceeding the training domain, making it difficult to guarantee the reliability of prediction results.

[0004] Traditional experimental designs often employ static strategies, failing to dynamically select the most valuable new experimental points based on model uncertainties, resulting in a waste of experimental resources. Furthermore, existing methods focus only on correlations between data, ignoring causality and failing to answer counterfactual questions, thus limiting the reliability of optimization. In addition, publicly available solutions often focus on isolated techniques, failing to organically combine multi-fidelity fusion, physical constraint embedding, causal reinforcement, and active learning loops to build an engineering intelligent experimental operating system with understanding, prediction, optimization, and verification capabilities. Essentially, they remain more accurate fitters, unable to achieve interpretable, interventionist, and self-evolving experimental agents, and thus unable to meet the high demands of performance prediction and optimization for complex engineering systems. Summary of the Invention

[0005] This invention provides a method and system for predicting and optimizing engineering experimental data based on AI algorithms, in order to solve the problems mentioned in the background section above: The present invention proposes an AI-based method for predicting and optimizing engineering experimental data, the method comprising: S1. Perform multi-fidelity fusion processing on high-fidelity experimental data and low-fidelity simulation data to generate a fused experimental simulation dataset, construct an initial physical constraint model, embed the preset physical mechanism constraint conditions of the engineering system, and generate a physical constraint embedding model. S2. Based on the physical constraint embedding model, perform uncertainty quantification analysis to determine the distribution of uncertainty in model prediction; adopt an active learning strategy to dynamically select new experimental points, conduct targeted experiments to obtain supplementary experimental data, integrate the supplementary experimental data into the fusion experimental simulation dataset for model iterative updates, and generate an optimized physical constraint embedding model. S3: Introduce a causal reasoning mechanism into the optimized physical constraint embedding model to construct a causal enhanced prediction model; learn the causal relationship between engineering system parameters and performance to generate a causal relationship graph; use the causal relationship graph to perform counterfactual reasoning to predict the change in engineering system performance when a certain parameter is changed, and generate counterfactual prediction data; S4. Optimize the multi-objective parameters of the engineering system based on counterfactual prediction data, determine the optimal parameter combination, conduct experimental verification, and generate experimental verification data; compare and analyze the experimental verification data with the counterfactual prediction data to evaluate the accuracy of the causal enhancement prediction model and generate model evaluation data; revise the causal enhancement prediction model based on the model evaluation data to generate a revised causal enhancement prediction model. S5. Utilize the modified causal enhanced prediction model to perform high-precision prediction of the performance of complex engineering systems and generate performance prediction data; formulate engineering experiment optimization schemes based on the performance prediction data and generate engineering experiment optimization data; dynamically adjust engineering experiments based on the engineering experiment optimization data and generate engineering system performance optimization data.

[0006] The present invention proposes a system for implementing the AI ​​algorithm-based engineering experimental data prediction and optimization method as described above, the system comprising: Fusion processing module: Performs multi-fidelity fusion processing on high-fidelity experimental data and low-fidelity simulation data to generate a fused experimental simulation dataset; constructs an initial physical constraint model, embeds preset physical mechanism constraints of the engineering system, and generates a physical constraint embedding model; Iterative update module: Based on the physical constraint embedding model, perform uncertainty quantification analysis to determine the uncertainty distribution of model prediction; adopt an active learning strategy to dynamically select new experimental points, conduct targeted experiments to obtain supplementary experimental data, integrate the supplementary experimental data into the fusion experimental simulation dataset for model iterative update, and generate an optimized physical constraint embedding model. The factual reasoning module introduces a causal reasoning mechanism into the optimized physical constraint embedding model to build a causal enhanced prediction model; it learns the causal relationship between engineering system parameters and performance to generate a causal relationship graph; it uses the causal relationship graph to perform counterfactual reasoning to predict the change in engineering system performance when a certain parameter is changed, and generates counterfactual prediction data. Experimental Verification Module: Based on counterfactual prediction data, optimize the multi-objective parameters of the engineering system to determine the optimal parameter combination; conduct experimental verification and generate experimental verification data; compare and analyze the experimental verification data with the counterfactual prediction data to evaluate the accuracy of the causal reinforcement prediction model and generate model evaluation data; revise the causal reinforcement prediction model based on the model evaluation data and generate a revised causal reinforcement prediction model. Dynamic Adjustment Module: Utilizes a modified causal enhancement prediction model to perform high-precision predictions of the performance of complex engineering systems, generating performance prediction data; formulates engineering experiment optimization schemes based on the performance prediction data, generating engineering experiment optimization data; and dynamically adjusts engineering experiments based on the engineering experiment optimization data, generating engineering system performance optimization data.

[0007] The beneficial effects of this invention are as follows: The AI-based method for predicting and optimizing engineering experimental data improves the accuracy of performance prediction for complex engineering systems, making the prediction results more consistent with actual conditions and providing a reliable basis for experimental optimization; it reduces the cost of high-fidelity experiments by using multi-fidelity fusion and active learning strategies to obtain more effective information with limited experiments, reducing unnecessary resource consumption; it enhances the interpretability and operability of the model by introducing a causal reasoning mechanism, enabling a clear understanding of the causal relationship between parameters and performance, facilitating precise adjustments to the experiment; it reduces the risk of prediction instability outside the training domain due to black-box modeling by embedding physical mechanism constraints, ensuring the stability of model predictions; it avoids the limitation that learning correlations alone cannot answer counterfactual questions, accurately predicting changes in results after parameter changes; it achieves both high-precision prediction and multi-objective parameter optimization, providing core technical support for next-generation intelligent experimental design in aerospace, energy, civil engineering, and other fields, promoting the intelligent and efficient development of engineering experiments. Attached Figure Description

[0008] Figure 1 This is a diagram illustrating the steps of the method described in this invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0009] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0010] One embodiment of the present invention, such as Figure 1 As shown, an AI algorithm-based method for predicting and optimizing engineering experimental data includes the following steps: S1. Perform multi-fidelity fusion processing on high-fidelity experimental data and low-fidelity simulation data to generate a fused experimental simulation dataset; construct an initial physical constraint model based on the fused experimental simulation dataset, embed the preset physical mechanism constraint conditions of the engineering system, and generate a physical constraint embedding model. S2. Based on the physical constraint embedding model, perform uncertainty quantification analysis to determine the uncertainty distribution of the model prediction; based on the uncertainty distribution, adopt an active learning strategy to dynamically select new experimental points, conduct targeted experiments to obtain supplementary experimental data, integrate the supplementary experimental data into the fusion experimental simulation dataset for model iterative updates, and generate an optimized physical constraint embedding model. S3: Introduce a causal reasoning mechanism into the optimized physical constraint embedding model to construct a causal enhanced prediction model; learn the causal relationship between engineering system parameters and performance through the causal enhanced prediction model to generate a causal relationship graph; use the causal relationship graph to perform counterfactual reasoning to predict the change in engineering system performance when a certain parameter is changed, and generate counterfactual prediction data. S4. Optimize the multi-objective parameters of the engineering system based on counterfactual prediction data to determine the optimal parameter combination; conduct experimental verification based on the optimal parameter combination to generate experimental verification data; compare and analyze the experimental verification data with the counterfactual prediction data to evaluate the accuracy of the causal enhancement prediction model and generate model evaluation data; revise the causal enhancement prediction model based on the model evaluation data to generate a revised causal enhancement prediction model. S5. Utilize a modified causal enhanced prediction model to perform high-precision predictions of the performance of complex engineering systems and generate performance prediction data; formulate engineering experiment optimization schemes based on the performance prediction data and generate engineering experiment optimization data; dynamically adjust engineering experiments based on the engineering experiment optimization data to achieve continuous optimization of the performance of complex engineering systems and generate engineering system performance optimization data.

[0011] The working principle of the above technical solution is as follows: By fusing high- and low-fidelity data at multiple fidelities, the value of the data can be fully explored and the data utilization rate improved. Combined with physical constraint embedding, the model becomes more closely aligned with engineering practice, avoiding prediction distortion caused by purely data-driven models deviating from physical mechanisms. Active learning dynamically selects experimental points, reducing invalid experiments, lowering experimental costs, and avoiding resource waste caused by blind experiments. A causal reasoning mechanism is introduced to clarify the intrinsic relationship between parameters and performance, improving prediction reliability, reducing optimization bias caused by spurious correlations, and preventing optimization directions from deviating from actual needs. Through multi-objective optimization and experimental verification and correction, the model's prediction accuracy and parameter optimization rationality are improved. Combined with closed-loop continuous optimization, the performance stability of the engineering system is enhanced, avoiding performance fluctuations caused by rigid optimization schemes, while simultaneously improving the efficiency of engineering experiments and the sustainability of optimization effects.

[0012] In one embodiment of the present invention, S1 includes: S11. Collect high-fidelity experimental data measured at the engineering site and low-fidelity simulation data simulated by simulation software. Perform outlier removal, missing value completion and data normalization preprocessing on the two types of data respectively to generate standardized high-fidelity dataset and standardized low-fidelity dataset. S12. A weighted fusion algorithm is used to perform multi-fidelity fusion processing on the standardized high-fidelity dataset and the standardized low-fidelity dataset. Dynamic weights are assigned according to the credibility of the two types of data to generate a fused experimental simulation dataset. The fused experimental simulation dataset takes into account both accuracy and efficiency. S13. Based on the key variables and operating rules of the engineering system in the fusion experimental simulation dataset, construct a mechanism-based parameterized initial physical constraint model and clarify the input and output dimensions and core parameter range of the model. S14. Extract the physical mechanism constraints of the engineering system, including mass conservation and energy conservation. Transform the constraints into mathematical constraint equations and integrate them into the loss function of the initial physical constraint model to generate a physical constraint embedding model that satisfies the actual operation rules of the engineering system.

[0013] The working principle of the above technical solution is as follows: By preprocessing high-fidelity experimental data and low-fidelity simulation data separately, outliers are removed, missing values ​​are filled in, and normalization is performed, which effectively improves data quality and avoids dirty data interfering with subsequent processing results. A weighted fusion algorithm is adopted, and dynamic weights are allocated according to data reliability. This leverages the accuracy advantage of high-fidelity data while improving processing efficiency with low-fidelity data, reducing the limitations of a single data source. An initial physical constraint model is constructed based on the fused data and engineering operation laws. Then, physical mechanism constraints such as mass conservation and energy conservation are incorporated, making the model more closely aligned with engineering reality. This prevents subsequent model training from deviating from physical laws, laying a reliable data and model foundation for the entire prediction and optimization process.

[0014] In one embodiment of the present invention, step S12 includes: Analyze the feature differences and data credibility between standardized high-fidelity datasets and standardized low-fidelity datasets, and generate data feature and credibility analysis results; Based on the data characteristics and credibility analysis results, a dynamic weight allocation model is built to generate dynamic weight allocation rules for the dataset. A weighted fusion algorithm is used to perform multi-fidelity fusion calculations on standardized high-fidelity datasets and standardized low-fidelity datasets according to the dynamic weight allocation rules of the datasets, and to generate intermediate datasets for fusion experimental simulation. Data consistency verification and feature complementarity optimization are performed on the intermediate dataset of the fusion experimental simulation to generate a fusion experimental simulation dataset that balances accuracy and efficiency.

[0015] The working principle of the above technical solution is as follows: By first analyzing the feature differences and reliability of the two types of standardized data, the subsequent weight allocation can be more targeted, avoiding poor fusion results caused by blind weight allocation. Based on the analysis results, a dynamic weight model is built to generate allocation rules, which can fully leverage the accuracy advantage of high-fidelity data while improving processing efficiency with the help of low-fidelity data, reducing the limitations of a single data source. After fusion calculation, consistency verification and feature complementarity optimization are carried out to eliminate data conflicts, make up for feature shortcomings, enhance the integrity and reliability of fused data, and avoid biases in fused data affecting subsequent model construction. The final fused experimental simulation dataset can balance accuracy and efficiency, providing high-quality data support for subsequent processes.

[0016] In one embodiment of the present invention, S2 includes: S21. Based on the physical constraint embedded model, the Monte Carlo simulation method is used to carry out uncertainty quantification analysis, calculate the mean, variance and confidence interval of the model prediction results, and obtain the model prediction uncertainty distribution. S22. Combining the characteristics of uncertainty distribution with actual engineering needs, an active learning strategy is adopted to construct an experimental point selection evaluation function, prioritizing the selection of regions with high uncertainty and sensitivity to model performance improvement, and dynamically determining the location and number of new experimental points. S23. Based on the new experimental site planning scheme, build a special experimental platform, keep the experimental environment variables consistent, carry out targeted experimental tests, collect real-time data during the experiment, and obtain supplementary experimental data after data verification. S24. Standardize the supplementary experimental data according to the preprocessing standard in S11, integrate it into the original fusion experimental simulation dataset to form an updated dataset, retrain the physical constraint embedding model based on the dataset, adjust the model parameters until convergence, and generate an optimized physical constraint embedding model.

[0017] The working principle of the above technical solution is as follows: Uncertainty quantification analysis is conducted through Monte Carlo simulation, clearly presenting the mean, variance, and confidence interval of the model predictions. This makes the distribution of model uncertainty intuitively verifiable, avoiding the blind selection of experimental points due to a lack of direction. An evaluation function is constructed by combining uncertainty characteristics with engineering requirements, prioritizing the selection of new experimental points in key sensitive areas to reduce invalid experiments and lower experimental costs and resource consumption. The consistency of experimental environment variables is controlled, and the collected data is verified to ensure the reliability of supplementary experimental data and avoid data bias interfering with model optimization. After standardization, the supplementary data is integrated into the original dataset to retrain the model. Parameters are adjusted until convergence to generate an optimized model, significantly improving the model's prediction accuracy and stability. This avoids excessive prediction bias caused by insufficient initial model performance, laying a solid foundation for subsequent causal inference and parameter optimization.

[0018] In one embodiment of the present invention, step S22 includes: Extract the core features of the uncertainty distribution in the model prediction, including peak regions and dispersion; identify the key elements in actual engineering needs, including parameter constraint range and experimental cost threshold; and generate a fusion analysis result of features and needs. Based on the results of feature and demand fusion analysis, an active learning strategy is adopted to build an experimental point selection evaluation function, incorporating two core indicators: uncertainty reduction efficiency and model performance improvement gain, to generate an initial evaluation function. By setting the screening threshold and priority rules of the evaluation function, a complete experimental point selection and judgment system is formed. Through this system, the parameter space is scanned across the entire domain, and a candidate region set with high uncertainty and high performance gain is output. By combining the parameter range information of the candidate region set, matching the feasible conditions of the engineering experiment, dynamically planning the specific coordinate location of the new experimental point, balancing the experimental coverage density and cost investment to determine the number of experimental points, and generating the planning results of the new experimental point.

[0019] The working principle of the above technical solution is as follows: By extracting the core features of uncertainty distribution and sorting out the key elements of engineering requirements, a deep integration of features and requirements is achieved, avoiding the problem of experimental point selection being out of touch with engineering reality. An evaluation function is built based on the fusion results, incorporating efficiency indicators for uncertainty reduction and gain indicators for model performance improvement, making the selection criteria for experimental points more precise and reducing ineffective screening of low-value regions. A complete judgment system scans the entire parameter space, comprehensively capturing candidate regions with high uncertainty and high performance gain, avoiding the omission of key experimental regions. By combining candidate region information with experimental feasibility conditions, the location of experimental points is dynamically planned and coverage density and cost are balanced, reducing resource waste caused by blind experiments and ensuring that new experimental points accurately point to the core direction of model performance improvement, providing scientific guidance for the efficient conduct of subsequent supplementary experiments.

[0020] In one embodiment of the present invention, S3 includes: S31. Introduce a structural causal model (SCM) into the optimized physical constraint embedding model to construct a causal inference mechanism, clarify the potential causal links between engineering system parameters and performance indicators, and build a causal enhancement prediction model. S32. Input the updated fusion experimental simulation dataset, mine the causal dependencies between variables through the causal discovery algorithm, eliminate false correlation links, and generate a causal relationship graph, which includes direct causality and indirect causality. S33. Based on the causal relationship graph, set up a counterfactual reasoning scenario, keep other parameters unchanged, and perform virtual intervention on the target parameter. The virtual intervention includes parameter value increase / decrease and type replacement to simulate the system operation state after parameter change. S34. Use the causal enhancement prediction model to extrapolate the state of the system after intervention, calculate the magnitude, trend and optimal range of performance index changes, and generate counterfactual prediction data in multiple scenarios.

[0021] The working principle of the above technical solution is as follows: By introducing a structural causal model to construct an inference mechanism, the potential causal links between engineering system parameters and performance indicators are clarified, allowing the prediction model to move beyond data correlation, enhancing the model's interpretability, and avoiding misleading false correlations caused by purely data-driven approaches. Causal discovery algorithms are used to uncover true dependencies and eliminate false links, generating a graph containing both direct and indirect causal relationships, improving the accuracy of causal relationship identification, and reducing the interference of erroneous associations on subsequent inference. Other parameters are fixed while virtually intervening with the target parameter to simulate its operating state, without requiring real modifications to the engineering system, reducing experimental risks and costs, and avoiding potential system failures or performance losses that could result from real intervention. Multi-scenario counterfactual prediction data is generated through model deduction, accurately presenting the magnitude and trend of parameter changes on performance, providing a comprehensive and reliable basis for subsequent optimization, preventing optimization directions from deviating from core requirements, and improving the scientific rigor and effectiveness of subsequent parameter optimization.

[0022] In one embodiment of the present invention, S33 includes: Extract core parameter nodes, causal link strength, and parameter interaction rules from the causal relationship graph to generate parameter association analysis results; Based on the results of parameter correlation analysis, the scenario boundaries of counterfactual reasoning are defined, scenario categories are divided according to parameter intervention types, and a multi-dimensional counterfactual reasoning scenario framework is generated. In each counterfactual reasoning scenario, the current value state of non-target parameters is locked to maintain the stability of the parameter interaction environment and generate a fixed parameter set for the scenario; For target parameters in a fixed parameter set of a scenario, perform virtual intervention operations, including gradient parameter value increment / decretion, multi-type parameter replacement, and generate a parameter intervention instruction set; Input the parameter intervention instruction set into the system operation simulation module to deduce the dynamic response process of the system after parameter changes and generate simulated system operation status data.

[0023] The working principle of the above technical solution is as follows: By extracting core parameter nodes and association rules from the causal relationship graph, accurate parameter association analysis results are generated, providing a clear basis for subsequent counterfactual inference and avoiding the confusion caused by scenario settings deviating from actual parameter associations. Based on the analysis results, scenario boundaries are delineated and categorized to construct a multi-dimensional inference scenario framework, which can comprehensively cover key intervention scenarios and reduce the possibility of missing important parameter change scenarios. Non-target parameters are locked to maintain the stability of the interaction environment, eliminate interference from irrelevant parameters, improve the accuracy of inference results, and avoid inference bias caused by parameter cross-influence. Virtual interventions such as gradient addition, subtraction, and multi-type replacement are performed on target parameters, making parameter change simulation more detailed and comprehensive, and reducing the one-sidedness caused by a single intervention method. Intervention commands are input into the simulation module to deduce dynamic responses without actually modifying the engineering system, reducing experimental risks and costs, avoiding system failures that may be caused by real operations, and the generated simulation data can provide reliable support for subsequent performance prediction.

[0024] In one embodiment of the present invention, step S4 includes: S41. Based on counterfactual prediction data, determine multiple optimization objectives of the engineering system, including maximizing efficiency, minimizing cost, and minimizing error. Establish a multi-objective optimization mathematical model and set objective weights and constraint boundaries. S42. The non-dominated sorting genetic algorithm (NSGA-Ⅲ) is used to solve the multi-objective optimization model, iteratively searching for Pareto optimal solutions in the parameter space, and selecting the parameter combination with the best overall performance. S43. Build an experimental verification platform according to the optimal parameter combination, control the experimental conditions to be consistent with the simulation scenario, carry out multiple sets of parallel experiments, record various performance data during the experiment, and generate experimental verification data. S44. Using evaluation indicators, including mean absolute error and coefficient of determination, to conduct a dimension-by-dimensional comparative analysis of experimental verification data and counterfactual prediction data, quantify model prediction bias, and evaluate the accuracy of the causal enhancement prediction model. S45. Based on the model evaluation data, locate the parameter range with large prediction deviation, adjust the link weights in the causal relationship graph, optimize the model's loss function and inference logic, and retrain the model to generate a corrected causal enhanced prediction model.

[0025] The working principle of the above technical solution is as follows: Multiple optimization objectives are determined based on counterfactual prediction data, and a mathematical model is established to balance the needs of maximizing efficiency, minimizing cost, and minimizing error, avoiding performance imbalances in the engineering system caused by single-objective optimization. A non-dominated sorting genetic algorithm is used to iteratively search for Pareto optimal solutions, accurately selecting the parameter combination with the best overall performance, improving the scientific nature of parameter optimization, and reducing resource waste caused by blind optimization. Multiple parallel experiments are conducted under controlled experimental conditions consistent with the simulation scenario to ensure the reliability of experimental verification data and avoid distortion of verification results caused by environmental interference. Multi-dimensional comparative analysis of multiple evaluation indicators quantifies model prediction bias, making model accuracy assessment more precise and reducing correction direction deviations caused by evaluation ambiguity. Based on the evaluation data, the deviation range is located, and the model link weights and inference logic are adjusted. A corrected model is retrained to significantly improve model prediction accuracy, preventing excessive model bias from affecting subsequent engineering optimization effects, and providing reliable model support for improving the performance of the engineering system.

[0026] In one embodiment of the present invention, S42 includes: S421. Analyze the parameter space dimension, objective function coupling relationship and constraint boundary conditions of the multi-objective optimization mathematical model to generate basic information for model solution; S422. Based on the basic information of model solution, initialize the core parameters of the non-dominated sorting genetic algorithm, including population size, number of iterations, crossover probability and mutation probability, and generate the initial configuration parameter set of the algorithm. S423. According to the initial configuration parameter set, randomly generate an initial population in the parameter space, encode the individuals in the population, and generate an initial parameter population encoding set. S424. Perform selection, crossover, and mutation genetic operations on the initial parameter population encoding set to generate the offspring population encoding set and complete a single iteration of evolution. S425. Merge the parent and offspring population encoding sets, perform non-dominated sorting to divide the frontier level, calculate the crowding distance of individuals in each level, and select high-quality individuals to form a new generation population. S426. Determine whether the new generation population meets the convergence condition. If not, return to S424 to continue iterating. If it does, output multiple Pareto optimal solutions to form a Pareto optimal solution set. S427. Combine multi-objective weights to perform comprehensive performance ranking of the Pareto optimal solution set, eliminate performance overlap and boundary invalid solutions, and select the parameter combination with the best comprehensive performance.

[0027] The working principle of the above technical solution is as follows: By analyzing the parameter space and constraints of the multi-objective optimization model, complete basic information for solution calculation is generated, making the algorithm configuration more closely match the model characteristics and avoiding solution deviations caused by blind parameter initialization. Based on this basic information, the core parameters of the genetic algorithm are accurately initialized to ensure the stability of the algorithm's operation and reduce iterative oscillations caused by improper parameter settings. Genetic operations generate a progeny population and fuse and select high-quality individuals, comprehensively exploring the potential optimal solutions in the parameter space and avoiding overlooking key high-quality solutions. A convergence judgment mechanism is set up; it continues iterating if the conditions are not met, and stops outputting once they are met, reducing the time loss caused by invalid iterations. Combining multi-objective weights to sort Pareto optimal solutions, overlapping and invalid solutions are eliminated, accurately selecting the parameter combination with the best overall performance, improving the targeting of parameter optimization, and avoiding poor engineering optimization effects caused by ambiguity in optimal solution selection.

[0028] In one embodiment of the present invention, S45 includes: Analyze the model evaluation data, extract the prediction deviation value, deviation distribution trend and deviation peak position of each parameter interval, and generate the deviation parameter interval location result; Based on the deviation parameter interval localization results, the associated links in the corresponding causal relationship map are identified, and the link weight values ​​are adjusted according to the deviation magnitude gradient to generate the adjusted causal relationship map. By combining the adjusted causal relationship graph, the model loss function is reconstructed, the variable association calculation rules and feature extraction paths in the inference logic are optimized, and the optimized model structure parameters are generated. Call the updated fusion experimental simulation dataset, divide it into training and validation sets, and configure the threshold for the number of iterations for model retraining and the convergence criteria. Load the optimized model structure parameters, input the training set to start the retraining of the causal augmentation prediction model, and use the validation set to monitor the changes in the model's prediction accuracy in real time. Training stops when the model's prediction accuracy reaches the convergence criterion, the trained model is output, and a modified causal augmentation prediction model is generated.

[0029] The working principle of the above technical solution is as follows: By analyzing the model evaluation data, deviation features of each parameter range are extracted to generate accurate deviation localization results, giving the model correction a clear direction and avoiding poor correction effects caused by blind adjustments. Based on the localization results, the causal link weights are adjusted gradients to generate an optimized causal relationship graph, making the causal associations more consistent with actual data patterns and reducing prediction bias caused by imbalanced link weights. The loss function is reconstructed and the inference logic is optimized using the new graph to generate suitable model structure parameters, improving the rationality of model inference and avoiding the problem of the loss function being disconnected from the inference logic. Training and validation sets are divided and training parameters are configured to standardize the model retraining process and reduce overfitting or underfitting caused by improper parameter settings. Parameters are loaded for training, and the accuracy is monitored in real time using the validation set. Training stops upon convergence, reducing the time wasted on invalid iterations while ensuring sufficient model training. Finally, a corrected causal enhanced prediction model is generated, significantly improving the model's prediction accuracy, enhancing its adaptability to engineering systems, and preventing excessive model bias from affecting the scientific nature of subsequent optimization decisions.

[0030] In one embodiment of the present invention, step S5 includes: S51. Collect real-time operation data and historical operating condition data of complex engineering systems, standardize the data according to the preprocessing standard, input the modified causal enhancement prediction model to perform high-precision prediction calculations, and generate performance prediction data covering short-term, medium-term and long-term. S52. Based on the performance prediction data and actual engineering needs, formulate a multi-dimensional engineering experiment optimization plan, including parameter adjustment plan, experimental process optimization plan and equipment configuration improvement plan, and generate engineering experiment optimization data. S53. Based on the engineering experiment optimization data, dynamically adjust the relevant parameters, including experimental equipment parameters, data acquisition frequency and environmental control standards, and establish a real-time monitoring mechanism for the experimental process to track the execution effect of the optimization scheme. S54. Continuously collect optimized experimental data and feed it back to the modified causal enhancement prediction model for secondary iterative optimization. Dynamically adjust the experimental optimization scheme to achieve closed-loop continuous optimization of the performance of complex engineering systems and generate engineering system performance optimization data.

[0031] The working principle of the above technical solution is as follows: By collecting real-time operating data and historical operating condition data of complex engineering systems, standardizing them according to standards, and inputting them into the correction model for calculation, performance prediction data covering short, medium, and long term is generated. This improves the comprehensiveness and accuracy of performance prediction, avoiding biased judgments caused by single data sources or short-term predictions. Multi-dimensional optimization schemes are formulated by combining the prediction data with actual engineering needs, making the schemes more aligned with actual operating requirements and preventing optimization directions from deviating from engineering goals. Based on the optimization data, key elements such as equipment parameters and acquisition frequency are dynamically adjusted, and a real-time monitoring mechanism is established to track the effect, enabling timely detection of adjustment deviations and avoiding system performance fluctuations caused by improper parameter settings. Continuous collection of optimized data is used to feedback the model for secondary iterations, dynamically adjusting the optimization scheme to achieve closed-loop continuous optimization. This enhances the stability and sustainability of engineering system performance, reduces resource waste caused by repeated optimization, and avoids performance stagnation caused by rigid optimization schemes.

[0032] In one embodiment of the present invention, S53 includes: Analyze engineering experimental optimization data, extract the types of parameters that need to be adjusted, the range of adjustment gradients, and the priority of parameter adjustment, and generate a list of dynamic parameter adjustments; Based on the dynamic parameter adjustment list, the experimental equipment parameters, data acquisition frequency, and environmental control standards are adjusted in a gradient manner according to priority, generating an adjusted parameter configuration set. Identify key performance indicators during the experiment, determine monitoring dimensions and data acquisition intervals, and generate a real-time monitoring plan for the experiment. Deploy the monitoring modules corresponding to the monitoring scheme, connect to the real-time data stream of the experimental system, build a data transmission channel, and generate a real-time monitoring data link; By collecting operational data of the experimental system after adjustments through real-time monitoring of the data link, comparing the changes in performance indicators before and after the adjustments, and generating tracking data on the execution effect of the optimization plan.

[0033] The working principle of the above technical solution is as follows: By analyzing engineering experimental optimization data, the type, gradient, and priority of adjustment parameters are extracted to generate a clear dynamic adjustment list, providing a clear basis for parameter adjustments and avoiding parameter configuration chaos caused by blind operation. Gradient adjustments are executed sequentially according to priority, ensuring the optimization effect of key parameters first, improving adjustment efficiency, and reducing performance fluctuations in the engineering system caused by disordered adjustments. Key performance indicators are identified to determine monitoring dimensions and intervals, generating a precise real-time monitoring plan to avoid monitoring directions deviating from core needs or redundant monitoring data. A monitoring module is deployed to build a real-time data link, ensuring smooth data transmission and preventing data interruptions that would prevent timely monitoring of adjustment effects. Data is collected and compared to observe performance changes before and after adjustments, generating intuitive effect tracking data. This allows for timely detection of deviations in the execution of the optimization plan, avoiding undetected problems of inadequate adjustments and providing reliable data support for subsequent dynamic optimization.

[0034] One embodiment of the present invention, such as Figure 2 As shown, a system for implementing the AI ​​algorithm-based engineering experiment data prediction and optimization method described above is provided, the system comprising: Fusion processing module: Performs multi-fidelity fusion processing on high-fidelity experimental data and low-fidelity simulation data to generate a fused experimental simulation dataset; Constructs an initial physical constraint model based on the fused experimental simulation dataset, embeds preset physical mechanism constraints of the engineering system, and generates a physical constraint embedding model; Iterative update module: Based on the physical constraint embedding model, perform uncertainty quantification analysis to determine the uncertainty distribution of the model prediction; based on the uncertainty distribution, adopt an active learning strategy to dynamically select new experimental points, conduct targeted experiments to obtain supplementary experimental data, integrate the supplementary experimental data into the fusion experimental simulation dataset for model iterative update, and generate an optimized physical constraint embedding model. The factual reasoning module introduces a causal reasoning mechanism into the optimized physical constraint embedding model to construct a causal enhanced prediction model; it learns the causal relationship between engineering system parameters and performance through the causal enhanced prediction model to generate a causal relationship graph; and it uses the causal relationship graph to perform counterfactual reasoning to predict the change in engineering system performance when a certain parameter is changed, generating counterfactual prediction data. Experimental Verification Module: Based on counterfactual prediction data, the module optimizes the multi-objective parameters of the engineering system to determine the optimal parameter combination; conducts experimental verification based on the optimal parameter combination to generate experimental verification data; compares and analyzes the experimental verification data with the counterfactual prediction data to evaluate the accuracy of the causal reinforcement prediction model and generates model evaluation data; and corrects the causal reinforcement prediction model based on the model evaluation data to generate a corrected causal reinforcement prediction model. Dynamic Adjustment Module: Utilizes a modified causal enhancement prediction model to perform high-precision predictions of the performance of complex engineering systems, generating performance prediction data; formulates engineering experiment optimization plans based on the performance prediction data, generating engineering experiment optimization data; and dynamically adjusts engineering experiments based on the engineering experiment optimization data to achieve continuous optimization of the performance of complex engineering systems, generating engineering system performance optimization data.

[0035] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting and optimizing engineering experimental data based on AI algorithms, characterized in that, The method includes: S1. Perform multi-fidelity fusion processing on high-fidelity experimental data and low-fidelity simulation data to generate a fused experimental simulation dataset, construct an initial physical constraint model, embed the preset physical mechanism constraint conditions of the engineering system, and generate a physical constraint embedding model. S2. Based on the physical constraint embedding model, perform uncertainty quantification analysis to determine the distribution of uncertainty in model prediction; adopt an active learning strategy to dynamically select new experimental points, conduct targeted experiments to obtain supplementary experimental data, integrate the supplementary experimental data into the fusion experimental simulation dataset for model iterative updates, and generate an optimized physical constraint embedding model. S3: Introduce a causal reasoning mechanism into the optimized physical constraint embedding model to construct a causal enhanced prediction model; learn the causal relationship between engineering system parameters and performance to generate a causal relationship graph; use the causal relationship graph to perform counterfactual reasoning to predict the change in engineering system performance when a certain parameter is changed, and generate counterfactual prediction data; S4. Optimize the multi-objective parameters of the engineering system based on counterfactual prediction data, determine the optimal parameter combination, conduct experimental verification, and generate experimental verification data; compare and analyze the experimental verification data with the counterfactual prediction data to evaluate the accuracy of the causal enhancement prediction model and generate model evaluation data; revise the causal enhancement prediction model based on the model evaluation data to generate a revised causal enhancement prediction model. S5. Utilize the modified causal enhanced prediction model to perform high-precision prediction of the performance of complex engineering systems and generate performance prediction data; formulate engineering experiment optimization schemes based on the performance prediction data and generate engineering experiment optimization data; dynamically adjust engineering experiments based on the engineering experiment optimization data and generate engineering system performance optimization data.

2. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 1, characterized in that, S1 includes: S11. Collect high-fidelity experimental data measured at the engineering site and low-fidelity simulation data simulated by simulation software. Perform outlier removal, missing value completion and data normalization preprocessing on the two types of data respectively to generate standardized high-fidelity dataset and standardized low-fidelity dataset. S12. A weighted fusion algorithm is used to perform multi-fidelity fusion processing on the standardized high-fidelity dataset and the standardized low-fidelity dataset. Dynamic weights are assigned according to the credibility of the two types of data to generate a fusion experimental simulation dataset. S13. Based on the key variables and operating rules of the engineering system in the fusion experimental simulation dataset, construct a mechanism-based parameterized initial physical constraint model and clarify the input and output dimensions and core parameter range of the model. S14. Extract the physical mechanism constraints of the engineering system, transform the constraints into mathematical constraint equations, integrate them into the loss function of the initial physical constraint model, and generate a physical constraint embedding model that satisfies the actual operation rules of the engineering system.

3. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 1, characterized in that, S2 includes: S21. Based on the physical constraint embedded model, the Monte Carlo simulation method is used to carry out uncertainty quantification analysis, calculate the mean, variance and confidence interval of the model prediction results, and obtain the model prediction uncertainty distribution. S22. Combining the characteristics of uncertainty distribution with actual engineering needs, an active learning strategy is adopted to construct an experimental point selection evaluation function, prioritizing the selection of regions with high uncertainty and sensitivity to model performance improvement, and dynamically determining the location and number of new experimental points. S23. Based on the new experimental site planning scheme, build a special experimental platform, keep the experimental environment variables consistent, carry out targeted experimental tests, collect real-time data during the experiment, and obtain supplementary experimental data after data verification. S24. Standardize the supplementary experimental data according to the preprocessing standard in S11, integrate it into the original fusion experimental simulation dataset to form an updated dataset, retrain the physical constraint embedding model based on the dataset, adjust the model parameters until convergence, and generate an optimized physical constraint embedding model.

4. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 3, characterized in that, S22 includes: Extract the core features of the uncertainty distribution predicted by the model, sort out the key elements in the actual needs of engineering, and generate the fusion analysis results of features and needs; Based on the results of feature and demand fusion analysis, an active learning strategy is adopted to build an evaluation function for experimental point selection, incorporating two core indicators: efficiency in reducing uncertainty and gain in improving model performance, to generate an initial evaluation function. By setting the screening threshold and priority rules of the evaluation function, a complete experimental point selection and judgment system is formed. Through this system, the parameter space is scanned across the entire domain, and a candidate region set with high uncertainty and high performance gain is output. By combining the parameter range information of the candidate region set, matching the feasible conditions of the engineering experiment, dynamically planning the specific coordinate location of the new experimental point, balancing the experimental coverage density and cost investment to determine the number of experimental points, and generating the planning results of the new experimental point.

5. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 1, characterized in that, The S3 includes: S31. In the optimized physical constraint embedding model, a structural causal model is introduced to construct a causal inference mechanism, clarify the potential causal links between engineering system parameters and performance indicators, and build a causal enhancement prediction model. S32. Input the updated fusion experimental simulation dataset, use the causal discovery algorithm to mine the causal dependencies between variables, eliminate spurious correlations, and generate a causal relationship graph. S33. Based on the causal relationship graph, set up a counterfactual reasoning scenario, keep other parameters unchanged, and perform virtual intervention on the target parameters to simulate the system operation state after the parameter changes; S34. Use the causal enhancement prediction model to extrapolate the state of the system after intervention, calculate the magnitude, trend and optimal range of performance index changes, and generate counterfactual prediction data in multiple scenarios.

6. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 5, characterized in that, S33 includes: Extract core parameter nodes, causal link strength, and parameter interaction rules from the causal relationship graph to generate parameter association analysis results; Based on the results of parameter correlation analysis, the scenario boundaries of counterfactual reasoning are defined, scenario categories are divided according to parameter intervention types, and a multi-dimensional counterfactual reasoning scenario framework is generated. In each counterfactual reasoning scenario, the current value state of non-target parameters is locked to maintain the stability of the parameter interaction environment and generate a fixed parameter set for the scenario; For target parameters in a fixed parameter set for a scenario, perform virtual intervention operations and generate a parameter intervention instruction set; Input the parameter intervention instruction set into the system operation simulation module to deduce the dynamic response process of the system after parameter changes and generate simulated system operation status data.

7. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 1, characterized in that, The S4 includes: S41. Based on counterfactual prediction data, determine multiple optimization objectives of the engineering system, establish a multi-objective optimization mathematical model, and set objective weights and constraint boundaries; S42. The non-dominated sorting genetic algorithm is used to solve the multi-objective optimization model, iteratively searching for Pareto optimal solutions in the parameter space, and selecting the parameter combination with the best overall performance. S43. Build an experimental verification platform according to the optimal parameter combination, control the experimental conditions to be consistent with the simulation scenario, carry out multiple sets of parallel experiments, record various performance data during the experiment, and generate experimental verification data. S44. Using evaluation indicators, conduct a dimension-by-dimensional comparative analysis of experimental verification data and counterfactual prediction data to quantify model prediction bias and evaluate the accuracy of the causal enhancement prediction model. S45. Based on the model evaluation data, locate the parameter range with large prediction deviation, adjust the link weights in the causal relationship graph, optimize the model's loss function and inference logic, and retrain the model to generate a corrected causal enhanced prediction model.

8. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 7, characterized in that, S42 includes: S421. Analyze the parameter space dimension, objective function coupling relationship and constraint boundary conditions of the multi-objective optimization mathematical model to generate basic information for model solution; S422. Based on the basic information of model solution, initialize the core parameters of the non-dominated sorting genetic algorithm and generate the initial configuration parameter set of the algorithm. S423. According to the initial configuration parameter set, randomly generate an initial population in the parameter space, encode the individuals in the population, and generate an initial parameter population encoding set. S424. Perform selection, crossover, and mutation genetic operations on the initial parameter population encoding set to generate the offspring population encoding set and complete a single iteration of evolution. S425. Merge the parent and offspring population encoding sets, perform non-dominated sorting to divide the frontier level, calculate the crowding distance of individuals in each level, and select high-quality individuals to form a new generation population. S426. Determine whether the new generation population meets the convergence condition. If not, return to S424 to continue iterating. If it does, output multiple Pareto optimal solutions to form a Pareto optimal solution set. S427. Combine multi-objective weights to perform comprehensive performance ranking of the Pareto optimal solution set, eliminate performance overlap and boundary invalid solutions, and select the parameter combination with the best comprehensive performance.

9. The method for predicting and optimizing engineering experimental data based on AI algorithms according to claim 1, characterized in that, The S5 includes: S51. Collect real-time operation data and historical operating condition data of complex engineering systems, standardize the data according to the preprocessing standard, input the modified causal enhancement prediction model to perform high-precision prediction calculations, and generate performance prediction data covering short-term, medium-term and long-term. S52. Combine performance prediction data with actual engineering needs to formulate a multi-dimensional engineering experiment optimization plan and generate engineering experiment optimization data; S53. Based on the engineering experiment optimization data, dynamically adjust the relevant parameters, establish a real-time monitoring mechanism for the experimental process, and track the implementation effect of the optimization scheme; S54. Continuously collect optimized experimental data and feed it back to the modified causal enhancement prediction model for secondary iterative optimization, dynamically adjust the experimental optimization scheme, and generate engineering system performance optimization data.

10. A system for implementing the AI ​​algorithm-based engineering experimental data prediction and optimization method as described in claim 1, characterized in that, The system includes: Fusion processing module: Performs multi-fidelity fusion processing on high-fidelity experimental data and low-fidelity simulation data to generate a fused experimental simulation dataset; constructs an initial physical constraint model, embeds preset physical mechanism constraints of the engineering system, and generates a physical constraint embedding model; Iterative update module: Based on the physical constraint embedding model, perform uncertainty quantification analysis to determine the uncertainty distribution of model prediction; adopt an active learning strategy to dynamically select new experimental points, conduct targeted experiments to obtain supplementary experimental data, integrate the supplementary experimental data into the fusion experimental simulation dataset for model iterative update, and generate an optimized physical constraint embedding model. The factual reasoning module introduces a causal reasoning mechanism into the optimized physical constraint embedding model to build a causal enhanced prediction model; it learns the causal relationship between engineering system parameters and performance to generate a causal relationship graph; it uses the causal relationship graph to perform counterfactual reasoning to predict the change in engineering system performance when a certain parameter is changed, and generates counterfactual prediction data. Experimental Verification Module: Based on counterfactual prediction data, optimize the multi-objective parameters of the engineering system to determine the optimal parameter combination; conduct experimental verification and generate experimental verification data; compare and analyze the experimental verification data with the counterfactual prediction data to evaluate the accuracy of the causal reinforcement prediction model and generate model evaluation data; revise the causal reinforcement prediction model based on the model evaluation data and generate a revised causal reinforcement prediction model. Dynamic Adjustment Module: Utilizes a modified causal enhancement prediction model to perform high-precision predictions of the performance of complex engineering systems, generating performance prediction data; formulates engineering experiment optimization schemes based on the performance prediction data, generating engineering experiment optimization data; and dynamically adjusts engineering experiments based on the engineering experiment optimization data, generating engineering system performance optimization data.

Images