Simulation system modeling lightweight substitution method for rapid verification of whole process of high-end equipment manufacturing

Abstract

The invention discloses a simulation system modeling lightweight substitution method for high-end equipment manufacturing whole process rapid verification. The method comprises the steps that 1, multi-field simulation data and test data of typical parts of high-end equipment are collected; 2, performing normalization and standardization preprocessing on simulation data, and constructing a data set; step 3, fusing model network architecture design and initialization training of a physical mechanism, introducing a physical information neural network and embedding a physical equation constraint operator; 4, reducing the model calculation amount based on a parameter compression and self-adaptive advanced exit mechanism of similarity evaluation; and step 5, constructing a lightweight simulation system substitution model, exporting and integrating the model into a system model through a standard format, and supporting rapid verification of the whole process of high-end equipment manufacturing. According to the method, deep learning and a physical mechanism are fused, the calculation time is remarkably shortened while the calculation precision and the generalization ability are improved, the system-level verification period is shortened, and agile development of aerospace and other high-end equipment is effectively supported.

Description

Technical Field

[0001] This invention relates to the field of high-end equipment simulation verification, and in particular to a lightweight alternative method for modeling simulation systems for rapid verification of the entire high-end equipment manufacturing process. Background Technology

[0002] This section provides only background information relevant to this disclosure and is not necessarily prior art.

[0003] With the increasing complexity, interdisciplinary nature, and system coupling of aerospace equipment, traditional systems engineering-based development models face a series of severe challenges in the model verification stage. The performance verification of aerospace products, such as engines, airframe structures, propulsion systems, and their subsystems, typically relies on high-fidelity finite element and finite volume simulation models. These high-precision simulation models can accurately describe physical phenomena such as structural mechanics and fluid dynamics, but their computational complexity is extremely high; a single full-field simulation often takes hours or even days to complete. In the early design stages of equipment development, the need for frequent iterative design verification is difficult to meet, making high-frequency optimization and agile iteration a bottleneck.

[0004] Furthermore, in current engineering practice, system models and physical simulation models are often disconnected. System models focus on predicting overall system behavior, such as dynamic response, thermal balance, and control logic, while physical simulation models focus on detailed analysis of local or single-domain aspects. The lack of uniformity in data formats, interface standards, and verification processes prevents simulation verification during the design phase from forming an efficient closed loop, thus limiting system-level performance optimization and verification. This model disconnect not only increases verification costs but also prolongs the development cycle and reduces engineering efficiency.

[0005] Therefore, there is an urgent need for a lightweight simulation alternative that integrates physical constraints with data-driven characteristics. This alternative should be able to approximate complex numerical simulation models while maintaining high accuracy and physical consistency, and also achieve rapid, low-cost system-level verification. The method should support generalized predictions under multiple physics and operating conditions, and possess the ability to integrate with system models, thereby providing effective support for agile development, rapid iteration, and joint optimization of aerospace equipment. This will not only significantly reduce verification costs and development cycles but also improve design decision efficiency and product performance reliability, meeting the urgent needs of modern aerospace equipment development for efficient, accurate, and intelligent verification. Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by this invention is to address the shortcomings of existing technologies by providing a lightweight alternative method for simulation system modeling for rapid verification of the entire process of high-end equipment manufacturing, comprising the following steps:

[0007] Step 1: Multi-domain simulation and test data collection for typical components. For typical components of the equipment, multi-domain simulation data is collected, including but not limited to simulation data in fields such as structural mechanics, fluid mechanics, and thermodynamics. The typical components are those in high-end equipment that have a representative and critical impact on the overall performance, safety, or reliability of the system. Taking aero-engines as an example, typical components include one or more of the following: compressor rotor blades, turbine disks, combustion chamber casings, etc.

[0008] Step 2: Simulation data normalization and standardization preprocessing, and dataset construction. The original data with different dimensions and units obtained in Step 1 are normalized to eliminate the impact of unit differences on model training and improve the model convergence speed and stability.

[0009] Step 3: Design and initialization training of the surrogate model network architecture that integrates physical mechanisms. The surrogate model is designed based on the physical information neural network, and constraint operators based on the physical model are embedded. The model initialization and training are completed through composite loss function and optimization strategy.

[0010] Step 4: Based on the parameter compression and adaptive early exit mechanism of similarity evaluation, the initialized proxy model obtained in Step 3 is lightweighted to significantly reduce the amount of computation while ensuring the accuracy and performance of the model.

[0011] Step 5: Construct a lightweight simulation system alternative model. Export the optimized model obtained in Step 4 into a standard format (such as ONNX) for easy deployment into the MBSE tool. The final lightweight simulation system alternative model is then implemented and integrated into the system model via an API interface, supporting rapid verification and joint optimization throughout the entire high-end equipment manufacturing process.

[0012] Furthermore, the multi-domain simulation data acquisition based on typical components described in step 1 specifically includes:

[0013] Step 1-1: Define and execute simulation parameters for typical components. Define the geometric model and boundary conditions of typical components, such as the surface parameters and inlet air velocity of aero-engine blades. This is done through mesh generation (mesh density not less than 10). 6 The node performs finite element or finite volume simulation to generate raw simulation data. Simultaneously, the raw simulation data file incorporates experimental data calibration, including test data from real-world scenarios, to ensure the physical authenticity of the simulation data. Taking the aerospace field as an example, the experimental data calibration includes key performance indicators acquired from wind tunnel experiments, such as deformation, stress, and temperature distribution.

[0014] Steps 1-2: Integrate data from structural, fluid, and thermal fields into a unified dataset, labeling it with parameter and response tags, and recording metadata, including but not limited to simulation timestamps and software versions. Specifically, a unified data storage format is adopted:

[0015]

[0016] in, Indicates input features, This represents the simulation output response. Represents a timestamp. Indicates domain label, Metadata information is represented and stored in a unified file format. The preferred format is .h5 or .csv. N represents the total number of data samples.

[0017] Furthermore, the simulation data normalization and standardization preprocessing, as well as the dataset construction described in step 2, specifically include:

[0018] Step 2-1-1: Divide the multi-domain data according to specific domains, and sort the data subsets of each domain in ascending order;

[0019] Step 2-1-2: Calculate the lower quartile Q1, upper quartile Q3, and interquartile range IQR = Q3 - Q1 of the data subset; where Q1 is the 25th percentile in ascending order of the dataset, and Q3 is the 75th percentile.

[0020] Step 2-1-3: For features in each data subset Values ​​that meet the cleaning criteria are considered outliers and removed. The specific cleaning criteria are as follows:

[0021]

[0022] Step 2-2: Normalize the data samples:

[0023]

[0024] With standardization:

[0025]

[0026] Among them, X min and X max The minimum and maximum values ​​of the sample are given, and μ and σ are the mean and standard deviation, respectively. These values ​​are calculated separately for each domain to avoid cross-domain bias.

[0027] Steps 2-3: Divide the dataset into training, validation and test sets according to the ratio, and introduce cross-validation mechanism to avoid model overfitting; the ratio of the data to training, validation and test sets is 7:2:1.

[0028] Furthermore, the design and initialization training of the model network architecture that integrates physical mechanisms described in step 3 specifically includes:

[0029] Step 3-1: Introduce a physical information neural network. The network consists of an input layer, hidden layers, and an output layer: 1) Input layer. It receives multidimensional parameters from structural, thermal, and fluid fields, represented as:

[0030]

[0031] Where, x i The system comprises three layers: 1) **Hidden Layer:** Characteristic variables describing the system state, including temperature, stress, displacement, velocity, strain rate, and load, covering simulation parameter requirements for the entire high-end equipment manufacturing process; 2) **Hidden Layer:** Composed of multiple fully connected networks, with physical embedding units introduced to improve the activation units of traditional neural networks. This is used to capture complex cross-domain coupling relationships and physical mechanism constraints. The hidden layer includes 10 fully connected networks, each with 128 nodes; 3) **Output Layer:** Outputs the corresponding response variables, such as stress distribution, temperature field, or fluid pressure field. Constraint operators based on physical partial differential equations are embedded within the network, ensuring that the network prediction results physically satisfy equilibrium equations, continuity equations, and energy conservation constraints.

[0032] Step 3-2: By designing a joint training loss function to balance data fitting and physical consistency, the data fitting error, physical residual loss and weight regularization term are integrated to constrain the model to satisfy the physical equations;

[0033] Step 3-3: Initialize weights using random seeds and design a hybrid optimization strategy. In the initial training phase, use the first-order gradient descent optimizer ADMA; in the fine-tuning phase, when the loss decreases slowly, replace it with the second-order Newton-based optimizer L-BFGS to prevent overfitting and gradient explosion. Simultaneously, introduce an early stopping strategy to automatically terminate training when the validation error has not decreased for several consecutive rounds.

[0034] Furthermore, the composite training loss function in step 3-2 is expressed as:

[0035]

[0036] in, For data fitting error, This represents the residual error of the physical equation. For weight regularization, The loss balance parameter is an adjustable parameter, and its initial value is determined using a grid search.

[0037] The hybrid optimization strategy in step 3-3 includes: using the Adam optimizer for fast convergence in the early stage of training, and switching to the L-BFGS optimizer for fine optimization when the loss function decreases gradually in the later stage; at the same time, an early stopping strategy is adopted, and training is automatically terminated when the validation set loss does not decrease for several consecutive rounds.

[0038] Furthermore, the parameter compression and adaptive early exit mechanism based on similarity evaluation described in step 4 specifically includes:

[0039] Step 4-1: The row / column similarity of the weight matrix is ​​calculated using cosine similarity and Euclidean distance. K-means clustering is then used to group the weight matrices, and low-contribution parameters are pruned. Specifically, the cosine similarity and Euclidean distance are weighted and summed to form the similarity index of the fused weight matrix. Based on the K-means clustering algorithm, the weight matrix is ​​grouped into three groups: Group-1 represents important weight matrices, Group-2 represents relatively unimportant weight matrices, and Group-3 represents low-contribution weight matrices. For the low-contribution weight matrices in Group-3, matrix units with an average activation value < 0.1 are set to a weight of 0, while units with an activation value >= 0.1 are quantized from 16-bit floating-point representation to 8-bit representation, with the numerical range quantized from [-65, 504, 65504] to [-128, 127] to further reduce computational complexity.

[0040] Step 4-2: To address the computational needs of the surrogate model on samples of varying complexity, an adaptive early exit mechanism is introduced. To further improve the model's computational efficiency, multi-level early exit nodes are introduced into the surrogate model's network structure. When the network's prediction results at a shallower layer already meet the confidence threshold, the computation process can be terminated early. This mechanism dynamically adjusts the computational depth based on sample difficulty, thereby achieving adaptive allocation of computational load.

[0041] Step 4-3: Determine the optimal configuration of the confidence threshold and early exit layer through offline calibration experiments, so that the model can significantly reduce the average computation time while maintaining accuracy. Perform grid parameter search optimization using the validation set until convergence to the optimal threshold and layer configuration parameters that balance computational accuracy and computational efficiency.

[0042] Furthermore, in step 4-2, the adaptive early exit mechanism adds an early exit node classifier to the hidden layer; the intermediate result is g(i), and the confidence level is represented as... If confidence level Exceeding the predetermined threshold If the threshold value is not specified, the calculation will exit early from the current layer and the final result will be output. The method was determined through Pareto multi-objective optimization search and offline validation set calibration, with the dual objectives of computational accuracy and efficiency. The value range is [0.85, 0.99];

[0043] In step 4-3, the offline calibration experiment uses 1000 samples of the validation set to perform grid search optimization until it converges to Pareto optimality, which balances computational accuracy and computational efficiency.

[0044] Furthermore, the lightweight simulation system alternative model construction described in step 5 specifically includes:

[0045] Step 5-1: Encapsulate the converged lightweight alternative model into a unified format such as ONNX to ensure cross-platform compatibility. Define standard input / output interfaces to ensure seamless integration with system-level simulation platforms such as Matlab, Modelica, and Simulink;

[0046] Step 5-2: Embed the alternative model into the product system model, and achieve data exchange and parallel operation through the interface layer. The system automatically calls the alternative model for calculation during simulation to achieve rapid evaluation.

[0047] Beneficial effects:

[0048] The method of this invention integrates deep neural networks with physical mechanisms to model, replacing the traditional finite element, i.e., finite volume simulation model with high computational complexity. This can reduce model prediction errors while significantly shortening computation time.

[0049] The method of this invention enables joint training across multiple domains, and the alternative model can adapt to different working conditions, geometric and material parameters, and has good scenario transferability.

[0050] By integrating lightweight simulation alternative models and physical mechanism models through the method of this invention, the system-level verification cycle can be significantly shortened, the design iteration efficiency of high-end complex equipment can be improved, and rapid verification of the entire manufacturing process of high-end equipment such as aerospace can be effectively supported. Attached Figure Description

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0052] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0053] This invention discloses a lightweight alternative method for simulation system modeling for rapid verification of the entire high-end equipment manufacturing process, such as... Figure 1 As shown, it includes the following steps:

[0054] Step 1: Collection of multi-domain simulation and test data for typical components. For typical components of aero-engines, multi-domain simulation data is collected, including simulation data from fields such as structural mechanics, fluid mechanics, and thermodynamics.

[0055] This embodiment takes the compressor rotor blade of an aero-engine as an example, and first establishes a three-dimensional geometric model. The blade geometry adopts a parametric definition method, and the main parameters include: 1) chord length C (unit: mm); 2) twist angle θ (unit: °); 3) blade height H (unit: mm); 4) blade thickness ratio t / C; 5) installation angle. 6) Intake airflow velocity (Unit: m / s); 7) Inlet temperature (Unit: K); 8) Material parameters E, v, p, representing elastic modulus, Poisson's ratio, and density, respectively. Based on the three-dimensional geometric model definition, multi-domain physical field coupling simulation is performed. First, in the structural field, the finite element method is used to calculate the stress and strain distribution of the blades under centrifugal and thermal loads; in addition, the finite volume method is used in the fluid field to calculate the pressure and velocity distribution of the flow field; finally, in the thermal field, the heat transfer equation is used... Solve for the temperature field distribution.

[0056] The simulation model mesh uses a hybrid hexahedral / tetrahedral grid with a mesh density of no less than 10. 6 There are 10 nodes. A steady-state solution is used, and the convergence criterion is set to a residual of less than 10. −5 Meanwhile, key performance indicators, including deformation, stress, and temperature distribution, can be collected through blade experiments such as wind tunnel tests to calibrate the simulation model.

[0057] Based on the obtained structural field, fluid field, and thermal field output data, data from multiple fields are fused to construct a unified dataset. Specifically, a unified data storage format is adopted:

[0058]

[0059] in, This represents the input features (geometry, boundary conditions, loads, materials, etc.). This represents the simulation output response (stress, temperature, efficiency, etc.). Represents a timestamp (e.g., "2025-01-01 10:00:00"), Indicates the domain label (structural field=0, fluid field=1, thermal field=2). This represents metadata information (simulation software version). A multidimensional label file is generated for each sample using a unified data storage format, resulting in a unified .h5 / .csv file format for data storage.

[0060] Step 2: Perform normalization and standardization preprocessing on the unified dataset to form a standard dataset that can be used for training. Normalize the original data with different dimensions and units to eliminate the impact of unit differences on model training and improve the model's convergence speed and stability.

[0061] Data cleaning was performed on the collected unified dataset. Box plot statistical methods were used to detect and remove outliers, and domain-specific cleaning was performed on multi-domain data. First, quartiles Q1 and Q3 were calculated, with IQR = Q3 - Q1, and thresholds ranging from Q1 - 1.5 × IQR to Q3 + 1.5 × IQR. If feature x... j satisfy:

[0062]

[0063] Values ​​with outliers are considered outliers and removed. Q1 is the 25th percentile in ascending order of the entire dataset, meaning 25% of the data are less than or equal to this value; Q3 is the 75th percentile in the entire dataset, meaning 75% of the data are less than or equal to this value. Simultaneously, differentiated cleaning rules should be applied to data from different domains. First, for structural data, spurious stress peaks caused by mesh distortion should be removed; for fluid data, non-physical values ​​in localized backflow regions should be removed; for thermal data, values ​​with temperatures lower than the ambient temperature should be removed.

[0064] The cleaned data samples are then normalized and standardized, and this process needs to be calculated separately for each domain to avoid cross-domain bias. The normalization formula is as follows:

[0065]

[0066] in, For the current sample value, and These are the minimum and maximum values ​​among all samples, respectively. For example, in a sample column of chord lengths, the minimum value is 50 mm, the maximum value is 100 mm, and the current sample is 60 mm. Then, the normalized calculation result X' = (60-50) / (100-50) = 0.2;

[0067] Furthermore, the standardized calculation process is as follows:

[0068]

[0069] in, and These are the sample mean and sample standard deviation, respectively. Standardization needs to be calculated separately for each field. For example, in the structural field, a batch of samples has a mean of 500 MPa, a standard deviation of 50 MPa, and a sample size of 480 MPa. The standardized result is X'' = (480-500) / 50 = -0.4. In the fluid dynamics field, a batch of samples has a mean of 250 m / s, a standard deviation of 20 m / s, and a sample size of 240 m / s. The standardized result is X'' = (240-250) / 20 = -0.5. After normalization and standardization, the sample values ​​range from [0,1], with a mean of 0 and a variance of 1.

[0070] Furthermore, the normalized and standardized data are divided into training, validation, and test sets in a 7:2:1 ratio, and a cross-validation mechanism is introduced to avoid model overfitting. Specifically, a 5-fold cross-validation mechanism is introduced, dividing the training set into five folds and dynamically switching the training set as the validation set. The mean squared error (MSE) is set as the average performance metric, calculated using the following formula:

[0071]

[0072] Where n is the number of samples, The true value Predicted values ​​are used to quantify the fitting error.

[0073] Step 3: Design the model network architecture that integrates physical mechanisms and initialize training.

[0074] Given the complexity and sparsity of multi-domain simulation data, to avoid insufficient generalization and physical inconsistencies caused by relying solely on data-driven models, a neural network architecture integrating physical mechanism constraints is proposed as a simulation proxy model. The overall architecture of the proxy model includes an input layer, several hidden layers, and an output layer. The input layer receives a preprocessed physical quantity input vector, represented as:

[0075]

[0076] Where, x i These represent characteristic variables describing the system state, including temperature, stress, displacement, velocity, strain rate, and load. In the hidden layer design, a 10-layer fully connected network is defined, with 128 nodes per layer, and physical embedding units are introduced to improve the activation units of traditional neural networks. Each embedding unit not only receives the output of the previous layer but also incorporates residuals from domain-specific physical equations as constraints.

[0077]

[0078] in, This is the output of the l-th layer. and For trainable weights and biases, For activation function, These are the physical constraint weighting coefficients. The physical residual term measures the deviation between the predicted results and the fundamental physical equations. The physical residual term is defined as follows:

[0079]

[0080] Among them, F( ) represents the physical governing equations (such as the continuity and momentum conservation equations in fluid mechanics).

[0081] Based on this, to balance data fitting and physical consistency, a joint training loss function is designed, which integrates the data fitting error and the physical residual loss to constrain the model to satisfy the physical equations, expressed as:

[0082]

[0083] in, For data fitting error, This represents the residual error of the physical equation. This is the weight regularization term. These are loss balancing parameters used to control the weights of each error term, for example... .

[0084] To minimize the joint training loss function L, weights are first initialized using random seeds, and a hybrid optimization strategy is designed. In the initial training phase, a first-order gradient descent optimizer (Adma) is used; during the fine-tuning phase, when the loss decreases slowly, it is replaced with a second-order Newton-based optimizer (L-BFGS) to prevent overfitting and gradient explosion. Simultaneously, an early stopping strategy is introduced, automatically terminating training when the validation error fails to decrease for several consecutive rounds.

[0085] Assuming a training iteration cycle of 2000 epochs, the first 500 epochs use the first-order gradient descent Adam optimizer for fast convergence. When the loss function L decreases to a set threshold (e.g., <1e-4), the system switches to the second-order Newton-L-BFGS optimizer for fine-tuning and ensuring training performance. If the threshold is not reached after the initial 500 epochs, Adam training continues until a switching condition or early stopping is triggered. Furthermore, an early stopping strategy is introduced to automatically stop training before the model overfits the training data, thereby obtaining a model with better generalization ability. Specifically, during each training iteration, the loss function L in the validation set is monitored; if the loss does not decrease for 10 consecutive epochs, the training process is terminated early.

[0086] Step 4: Based on similarity evaluation, parameter compression and adaptive early exit mechanism are used to lightweight the surrogate model after initialization, which significantly reduces the amount of computation while ensuring the accuracy and performance of the model.

[0087] To reduce the number of parameters and computational overhead in the surrogate model, this invention performs similarity measurement and redundancy compression on the network parameters after training. The row / column similarity of the weight matrix is ​​calculated using a fusion of cosine similarity and Euclidean distance metrics, assuming the network parameter set is as follows:

[0088]

[0089] For any two-layer parameter matrix Define cosine similarity:

[0090]

[0091] in, Represents the weight vector. This is expressed as the L2 norm of the weight vector. Furthermore, the Euclidean distance between the two parameter matrices is defined as:

[0092]

[0093] The cosine similarity and Euclidean distance are weighted and summed to form a fused weight matrix similarity index. Based on the weight matrix similarity feature values, the K-means clustering algorithm is used to cluster the weight matrix into three groups, defined as follows: 1) Group-1 represents important weight matrices that should not be compressed; 2) Group-2 represents weight matrices between high and low contribution and should be retained; 3) Group-3 represents low-contribution weight matrices, which are potential targets for pruning and compression. After obtaining the clustering results, the low-contribution weights in Group-3 are pruned and quantized. Specifically, for matrix units with an average activation value < 0.1, their weights are set to 0. In addition, the remaining units with activation values ​​>= 0.1 are quantized from the original 16-bit floating-point representation (numerical range [-65, 504, 65504]) to an 8-bit representation (numerical range [-128, 127]) to further reduce computational overhead.

[0094] To further improve the efficiency of the model during the inference phase, multi-level early exit nodes are introduced into the network structure of the surrogate model. When the network's prediction results at a shallower layer meet the confidence threshold, the computation process can be terminated early. This mechanism dynamically adjusts the computation depth based on sample difficulty, thereby achieving adaptive allocation of computational load. For example, early exit node classifiers are added to layers 3-7 of the hidden layers (a total of 10 hidden layers). When the surrogate model performs forward propagation computation, when it reaches layers 3-7, its corresponding intermediate result is g(i). The confidence is calculated based on the early exit node classifier of the current layer, as shown below:

[0095]

[0096] in, The normalized exponential function is defined as follows:

[0097]

[0098] If the confidence level c exceeds a predetermined threshold θ (e.g., 0.92), the computation will exit prematurely from the current layer, and the final result will be output, significantly reducing the computation steps. For example, when the computation reaches the 5th hidden layer, if the confidence level c is calculated to be 0.97 (>0.92) based on the intermediate results and the confidence formula, the computation process will exit directly, and the computation of layers 6-10 will not be executed; the computation result will be output directly.

[0099] To determine the optimal configuration of the confidence threshold and early exit layer, and to ensure a balance between accuracy and computational efficiency while maintaining model accuracy, offline grid search optimization was performed using 1000 samples on the validation set until convergence to Pareto optimality that balances computational accuracy and computational efficiency.

[0100] Step 5: Construction of a Lightweight Simulation System Alternative Model. The converged lightweight simulation system alternative model is encapsulated in a unified format such as ONNX to ensure cross-platform compatibility. Standard input / output interfaces are defined to ensure seamless integration with system-level simulation platforms such as Matlab, Modelica, and Simulink. Finally, the alternative model is embedded into the product system model, enabling data exchange and parallel execution through API interfaces. During simulation, the system automatically calls the lightweight alternative model for calculations, avoiding highly complex finite element or finite volume calculations and achieving rapid evaluation.

[0101] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a lightweight alternative method for rapid verification of the entire high-end equipment manufacturing process, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0102] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MCU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0103] This invention provides a lightweight alternative method for simulation system modeling in the context of rapid verification throughout the entire high-end equipment manufacturing process. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A lightweight alternative method for modeling simulation systems for rapid verification of the entire high-end equipment manufacturing process, characterized in that: Includes the following steps: Step 1: Collect multi-domain simulation data and test data of typical components. The typical components are those in high-end equipment that have a representative and critical impact on the overall system performance, safety or reliability. Collect multi-domain simulation data of typical components in the equipment as the data basis for the entire manufacturing process. The scope of the domains includes, but is not limited to, structural mechanics, fluid mechanics and thermodynamics. Step 2: Normalization and standardization preprocessing of simulation data and construction of dataset. The original data of different dimensions and units obtained in Step 1 are normalized and training set, validation set and test set are constructed. Step 3: Design and initialization training of the surrogate model network architecture that integrates physical mechanisms. The surrogate model is designed based on the physical information neural network, and constraint operators based on the physical model are embedded. The model initialization and training are completed through composite loss function and optimization strategy. Step 4: Based on similarity evaluation, parameter compression and adaptive early exit mechanism are used to lightweight the initialized proxy model obtained in Step 3; Step 5: Lightweight simulation system alternative model construction. Export the optimized model obtained in Step 4 into a standard format that is easy to deploy into model-based systems engineering tools, thereby realizing a lightweight alternative to simulation system modeling.

2. The method according to claim 1, characterized in that, Step 1, which involves the acquisition of multi-domain simulation data based on typical components, specifically includes: Step 1-1: Define the geometric model and boundary conditions of typical components, perform finite element or finite volume simulation through mesh generation, and generate the original simulation data file; incorporate experimental data for calibration, including test data from actual scenarios; Steps 1-2: Merge data from multiple domains into a unified dataset, label parameters and responses, and record metadata; the metadata includes, but is not limited to, simulation timestamps and software versions, and the original simulation data adopts a unified data storage format D: in, Indicates input features, This represents the simulation output response. Represents a timestamp. Indicates domain label, Metadata information is represented and stored in a unified file format, where N is the total number of data samples.

3. The method according to claim 1, characterized in that, Step 2, which involves simulation data normalization and standardization preprocessing, as well as dataset construction, specifically includes: Step 2-1: Use script tools to clean the data, and use statistical methods to divide the multi-domain data according to specific domains and remove outliers; Step 2-2: Normalize and standardize the cleaned data samples according to their respective domains: Among them, X min and X max Let μ and σ be the minimum and maximum values ​​of the sample, respectively, and let μ and σ be the mean and standard deviation. These values ​​are calculated separately for each domain to avoid cross-domain bias. For the original sample, The sample obtained after normalization. The sample obtained after standardization; Steps 2-3 involve dividing the processed dataset into training, validation, and test sets proportionally, and introducing a cross-validation mechanism to avoid model overfitting.

4. The method according to claim 3, characterized in that, The statistical method described in step 2-1 is the box plot method, which includes the following steps: Step 2-1-1: Divide the multi-domain data according to specific domains, and sort the data subsets of each domain in ascending order; Step 2-1-2: Calculate the lower quartile Q1, upper quartile Q3, and interquartile range IQR = Q3 - Q1 of the data subset; where Q1 is the 25th percentile in ascending order of the dataset, and Q3 is the 75th percentile. Step 2-1-3: For features in each data subset If the cleaning conditions are met, the value is considered an outlier and is removed. The specific cleaning conditions are as follows: The cross-validation mechanism in steps 2-3 involves quantifying the fitting error through 5-fold cross-validation. The training set is divided into 5 folds, and the training set is dynamically switched for validation. The mean squared error (MSE) is set as the average performance metric. Where n is the number of samples. For the true value, These are predicted values.

5. The method according to claim 1, characterized in that, Step 3, which involves the design and initialization training of the model network architecture incorporating physical mechanisms, specifically includes: Step 3-1: Construct a physical information neural network, including an input layer, a hidden layer, and an output layer; the input layer receives multi-dimensional parameters from multiple domains, the hidden layer consists of multiple fully connected networks and introduces residual connections to capture cross-domain coupling relationships, and the output layer outputs the corresponding response variables; Step 3-2: Balance data fitting and physical consistency by designing a composite training loss function. The composite training loss function integrates data fitting error, physical residual loss and weight regularization term to constrain the model to satisfy the physical equation. Step 3-3: Initialize weights using random seeds, design a hybrid optimization strategy to prevent overfitting and gradient explosion; and terminate training when the validation error has not decreased for several consecutive rounds by introducing an early stopping strategy.

6. The method according to claim 5, characterized in that, The composite training loss function in step 3-2 is expressed as follows: in, For data fitting error, This represents the residual error of the physical equation. For weight regularization, The loss balance parameter is an adjustable parameter, and its initial value is determined using a grid search. The hybrid optimization strategy in step 3-3 includes: using the Adam optimizer for fast convergence in the early stage of training, and switching to the L-BFGS optimizer for fine optimization when the loss function decreases gradually in the later stage; at the same time, an early stopping strategy is adopted, and training is automatically terminated when the validation set loss does not decrease for several consecutive rounds.

7. The method according to claim 1, characterized in that, The parameter compression and adaptive early exit mechanism based on similarity evaluation in step 4 specifically includes: Step 4-1: Calculate the row / column similarity of the weight matrix using a fusion index of cosine similarity and Euclidean distance, group the weights using the K-means clustering algorithm, and prune low-contribution parameters in the weight matrix; quantize the weight parameters to 8 bits to reduce computational cost. Step 4-2: Insert multiple classifier exit nodes into the model's hidden layer to implement an adaptive early exit mechanism; when the prediction confidence of a sample in a shallower layer exceeds the threshold, the inference process is terminated early. Step 4-3: Determine the confidence threshold and the optimal configuration for early exit layer through offline calibration experiments.

8. The method according to claim 7, characterized in that, The network parameter set in step 4-1 Represented as The cosine similarity between any two layer parameter matrices is expressed as: in , Let ||·|| be the weight vector, and ||·|| be the L2 norm of the weight vector; the Euclidean distance is expressed as... The cosine similarity and Euclidean distance are weighted and summed to form a fused weight matrix similarity index. The weight matrix is ​​then grouped according to its contribution based on the K-means clustering algorithm, with matrix units with an average activation value <0.1 having a weight of 0.

9. The method according to claim 7, characterized in that, In step 4-2, the adaptive early exit mechanism adds an early exit node classifier to the hidden layer; the intermediate result is g(i), and the confidence level is expressed as... If confidence level Exceeding the predetermined threshold If the threshold value is not specified, the calculation will exit early from the current layer and the final result will be output. The method was determined through Pareto multi-objective optimization search and offline validation set calibration, with the dual objectives of computational accuracy and efficiency. The value range is [0.85, 0.99]. In step 4-3, the offline calibration experiment uses 1000 samples of the validation set to perform grid search optimization until it converges to Pareto optimality that balances computational accuracy and computational efficiency.

10. The method according to claim 1, characterized in that, The lightweight simulation system alternative model construction described in step 5 includes: encapsulating the optimized model in ONNX format; defining standard input / output interfaces to ensure compatibility with the system-level simulation platform; and automatically calling the alternative model for calculation during the simulation process to achieve rapid verification of the entire high-end equipment manufacturing process.

Images