A method for accelerating multi-scenario simulation of data-real fusion testing.

Abstract

This invention discloses a method for accelerating multi-scenario simulation of all elements in data-real fusion testing, belonging to the field of digital testing and verification of high-end equipment. The method includes: analyzing the test object, proposing test requirements / objectives and performance indicators, and formulating a preliminary test implementation plan; constructing a geometric and mechanistic model of the test object based on the analysis results; conducting multi-dimensional model co-simulation to achieve multi-dimensional simulation data acquisition and collection; collecting physical data and historical data, and combining them with simulation data to achieve the fusion and enhancement of data-real test data; using data dimensionality reduction methods to reduce the dimensionality of high-dimensional data to extract key features; constructing a deep learning simulation acceleration network based on the acquired dimensionality-reduced data and the test object model, and training and optimizing it; inputting the full-element simulation model of the test object and the full-scenario test model to achieve accelerated simulation and result prediction across all elements and scenarios. This invention achieves accelerated simulation across all elements and scenarios, improving testing efficiency.

Description

Technical Field

[0001] This invention belongs to the field of digital testing and verification of high-end equipment, specifically involving a multi-scenario accelerated simulation method for full-element testing that integrates digital and real data. Background Technology

[0002] With technological advancements, complex equipment is becoming increasingly sophisticated in function and performance. To achieve these functions and performance targets, the demand for testing complex equipment is constantly growing. While traditional physical testing methods offer high accuracy, they are also costly and time-consuming. Pure digital simulation methods, although capable of shortening testing time to some extent, are typically limited by the computational power of the simulator and the efficiency of processing high-dimensional data, resulting in excessively long simulation times in complex scenarios.

[0003] Currently, research on data-physical fusion testing has gradually developed, achieving higher-precision test object model construction by combining digital simulation and physical test data. However, how to efficiently utilize simulation data and physical data for fusion, and further accelerate simulation through deep learning networks, remains a challenge. Existing methods have limitations in dimensionality reduction and accelerated simulation prediction for high-dimensional simulation data, and struggle to achieve good scalability in multi-domain model simulations. Summary of the Invention

[0004] This invention discloses a data-real fusion test system and method for accelerating multi-scenario simulation across all elements, aiming to solve the problems of excessively long simulation times and difficulties in data fusion in existing technologies. While ensuring test accuracy, it achieves accelerated simulation and solution capabilities for the corresponding test objects and integrates existing multi-scenario test models, thereby significantly improving the efficiency and reliability of equipment testing.

[0005] The technical problem solved by this invention is achieved through the following technical solution: a method for accelerating multi-scenario simulation of data-real fusion testing, the method comprising:

[0006] Step 1: Design a test object feature model building module. This module will build a full-feature model of the test object. The specific implementation is as follows:

[0007] Analyze the test object, propose test requirements or objectives and performance indicators, and formulate a preliminary test implementation plan, including a detailed decomposition of the test objectives; based on the analysis results, construct a geometric model that reflects the geometric shape characteristics of the test object and a multi-domain mechanism model of the test object's operation process;

[0008] Step 2: Design a data fusion and enhancement module for the test object element model. This module utilizes the constructed test object element model and physical data to achieve data fusion and enhancement. The specific implementation is as follows:

[0009] Joint simulation of geometric and mechanistic models is performed to acquire and collect multidimensional simulation data; physical data and historical data are collected and combined with simulation data to achieve the fusion and enhancement of data-real fusion test data; dimensionality reduction methods are used to reduce the dimensionality of high-dimensional data in order to extract key features and obtain the dataset.

[0010] Step 3: Design a joint simulation acceleration module for the test object's element scene model. This module combines the full-element simulation acceleration model with the full-scene model required for testing to achieve joint simulation acceleration. The specific implementation is as follows:

[0011] For the geometric model of the test object, a symbolic distance field is used to represent the geometric model in a transformed form. The geometric model, symbolic distance field, and various operating parameters are used as training input datasets and input into a deep learning simulation acceleration network for training and optimization. After training, the deep learning simulation acceleration network can accept dimensionality-reduced data input and predict simulation results. Combined with an online learning mechanism, it automatically adapts to the characteristics of different test objects and dynamically optimizes the simulation process while maintaining simulation accuracy. Finally, the full-element deep simulation acceleration model and the full-scene test model are input to achieve joint simulation acceleration and result prediction of the full-element full-scene test model with data-real fusion.

[0012] The present invention has the following beneficial effects: the method disclosed in the present invention can significantly shorten the modeling and simulation time of the test object and improve the testing efficiency. At the same time, through the application of deep learning networks, the simulation speed can be significantly accelerated while ensuring the simulation accuracy, thus providing an efficient and fast solution for the testing of complex equipment. Attached Figure Description

[0013] Figure 1 This is a flowchart of a multi-scenario accelerated simulation method for full-element data-real fusion testing according to the present invention. Detailed Implementation

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings.

[0015] This invention discloses a method for accelerating multi-scenario simulation of all elements in data-real fusion testing, such as... Figure 1 As shown, it includes:

[0016] Step 1: Design the test object element model building module. This module is designed to build a model of all elements of the specific application object in the data-real fusion test. The specific implementation is as follows:

[0017] Analyze the test object, propose test requirements / objectives and performance indicators, and formulate a preliminary test implementation plan, including a detailed decomposition of the test objectives; based on the analysis results, construct a geometric model that reflects the geometric shape characteristics of the test object and a multi-domain mechanism model of the test object's operation process;

[0018] Step 2: Design a module for fusing and enhancing real-world data of the test object's feature model. This module utilizes the constructed full-feature model and physical data to achieve data fusion and enhancement. The specific implementation is as follows:

[0019] Multi-dimensional model co-simulation operation including geometric and mechanistic models is carried out to achieve multi-dimensional simulation data acquisition and collection; physical data acquisition and historical data collection are carried out, and combined with simulation data to achieve the fusion and enhancement of data-real fusion test data; data dimensionality reduction methods are used to reduce the dimensionality of high-dimensional data in order to extract key features and obtain datasets;

[0020] Step 3: Design a joint simulation acceleration module for the test object element scene model. This module combines the object element simulation acceleration model with the full scene model required for testing to achieve joint simulation acceleration for testing. The specific implementation is as follows:

[0021] Based on the acquired dataset, the symbolic distance field is used for data transformation and multidimensional representation. The data is then input into a constructed deep learning simulation acceleration network for training and optimization. After training, the deep learning simulation acceleration network can accept dimensionality-reduced data input and quickly predict simulation results, replacing traditional numerical simulation solvers. Combined with an online learning mechanism, it automatically adapts to the characteristics of different test objects and dynamically optimizes the simulation process while maintaining simulation accuracy. Finally, the geometric model, the multi-domain mechanism model of the test object's operation process, and the full-scene test model are input to achieve joint simulation acceleration and result prediction of the full-element, full-scene model of the data-real fusion test.

[0022] The multi-domain mechanism model of the test object is constructed by combining fluid dynamics modeling, structural mechanics modeling, parametric modeling, and modeling methods based on two-dimensional modeling language: the fluid dynamics model is based on incompressible fluid model, turbulence model, heat conduction model, etc.; the structural mechanics model is constructed by finite element method (FEM), including nonlinear behavior simulation and dynamic response prediction; parametric modeling includes adaptive adjustment of geometry and automatic calibration of simulation parameters; the modeling tool based on two-dimensional modeling language is used for coupled simulation of different physical fields to achieve accurate simulation of multi-physics fields.

[0023] Multidimensional model co-simulation, including geometric and mechanistic models, is achieved through the following methods: calling different types of models using a common simulation interface (FMI standard); interconnecting modules through communication protocols (such as TCP / IP, OPC UA); and using model exchange methods to achieve data exchange between multiple platforms based on simulation models.

[0024] The fusion of data from both simulation and physical sources specifically includes: data fusion by jointly processing simulation data and physical acquisition data to eliminate noise and errors; data fusion by extracting features from multi-source data and establishing a unified feature representation; and data fusion by jointly analyzing simulation results and actual physical data to achieve the fusion of test data results.

[0025] The data dimensionality reduction method specifically includes the following steps: extracting the main patterns from the simulation dataset, decomposing the high-dimensional data into a finite number of pattern vectors; retaining the main changing patterns, and reducing computational complexity.

[0026] Deep learning simulation acceleration networks utilize dimensionality-reduced simulation data and model data as datasets to train the constructed deep neural network structure. By adjusting network parameters and combining physical mechanisms to construct the network's loss function, and through multiple training and iterations, the stability of the deep learning network is ensured.

[0027] More specifically, the following illustrative examples are given.

[0028] First, a detailed analysis of the test object is conducted, including the definition of test requirements, performance indicators, and test objectives, and a preliminary test plan is developed. Test object analysis is the foundation of the entire process; its purpose is to refine test objectives and provide targeted solutions to ensure the rationality of the test plan and the accuracy of its execution.

[0029] Next, based on the characteristics of the test object, its geometric model and mechanistic model are constructed. The geometric model provides the spatial structural foundation for physical simulation, while the mechanistic model combines knowledge from various fields of physics, including fluid mechanics and structural mechanics, to provide a more realistic description of physical behavior for subsequent simulation operations.

[0030] Furthermore, this invention employs parametric modeling technology, which enables model optimization and automatic calibration in various scenarios through adaptive adjustment of geometric shapes and physical parameters, thereby improving the versatility and accuracy of the model.

[0031] To improve the efficiency of model simulation, this invention employs co-simulation of multi-dimensional models. Through data exchange and communication protocols between various simulation tools, it ensures the interoperability of simulation models in different environments. This co-simulation method utilizes a common simulation interface (FMI standard) and communication protocols (such as TCP / IP and OPC UA) to achieve efficient interconnection and invocation of simulation modules, greatly enhancing the flexibility and scalability of simulation. Furthermore, cross-platform simulation technology improves the compatibility and scalability between different simulation tools.

[0032] During the acquisition of simulation data, the system also simultaneously collects physical test data and historical data. Through multi-source data fusion, it achieves joint processing of simulation data, physical data, and historical data. Data-real fusion not only improves data reliability by removing noise and errors, but also further enhances test accuracy through feature-level and decision-level data fusion methods.

[0033] To address the computational burden of high-dimensional data, this invention employs data dimensionality reduction techniques, particularly based on intrinsic orthogonal decomposition, to decompose high-dimensional data into a finite number of pattern vectors. This dimensionality reduction operation preserves the main variation patterns in the data while reducing computational complexity, thereby effectively improving the training efficiency of deep learning models.

[0034] Regarding simulation acceleration, this invention proposes a solution for accelerating networks using deep learning simulations. By using a dimensionality-reduced dataset to train the deep learning network and constructing a loss function based on physical mechanisms, the network can maintain stability after multiple training and iterations.

[0035] By replacing the traditional numerical simulation solver with this deep learning model, we can not only accelerate the prediction of simulation results, but also achieve adaptive tuning through online learning mechanisms, dynamically optimizing the simulation process to adapt to different test objects and complex scenarios.

[0036] Based on the acquired dataset, the symbolic distance field is used to transform and represent the data in multiple dimensions.

[0037] The data is input into the constructed deep learning simulation acceleration network and trained and optimized.

[0038] Once trained, the deep learning simulation acceleration network can accept dimensionality-reduced data input and quickly predict simulation results, replacing traditional numerical simulation solvers.

[0039] Finally, the full-element deep simulation acceleration model and the full-scenario test model are input to achieve joint simulation acceleration and result prediction of the full-element full-scenario test model with data-real integration.

[0040] To better understand the implementation of this invention, the following uses aircraft aerodynamic simulation as an example to describe the specific implementation method of this invention in detail with reference to specific technical processes.

[0041] First, we analyze the test objects and define the simulation requirements:

[0042] In aircraft aerodynamic simulation, the first step is to conduct a comprehensive analysis of the test object (aircraft) to clarify the requirements and objectives for its aerodynamic characteristic testing. This includes, but is not limited to, the following:

[0043] Fluid dynamics characteristics, such as lift, drag, and pressure distribution;

[0044] Aerodynamic performance of an aircraft under different flight conditions (such as takeoff, cruise, and landing).

[0045] Structural mechanical properties, such as airfoil deformation and fatigue strength.

[0046] By analyzing the above requirements, the corresponding performance indicators (such as lift-to-drag ratio, pressure coefficient, etc.) were determined and a test plan was formulated, which was refined to the data acquisition and simulation scenarios under different flight conditions.

[0047] Then, the geometric and mechanistic models of the test objects are constructed:

[0048] In this step, a geometric model is constructed based on the aircraft's geometry, such as the wings, fuselage, and tail, and a mechanistic model is built by combining fundamental theories of fluid mechanics and structural mechanics. The specific process includes:

[0049] Fluid dynamics modeling: describing aerodynamic phenomena based on incompressible fluid equations (such as the Navier-Stokes equations) and turbulence models (such as the k-ε model);

[0050] Structural mechanics modeling: The aircraft structural model is constructed using the finite element method (FEM) to simulate the stress and strain distribution under external loads (such as aerodynamic forces, gravity, etc.);

[0051] Parametric modeling: By adaptively adjusting geometric parameters, aerodynamic simulation can be achieved under different wing structures, angles of attack, and other conditions.

[0052] These models can be co-simulated using simulation platforms such as CFD (Computational Fluid Dynamics) tools and FEM tools, and can achieve model interoperability through common interfaces such as the FMI standard.

[0053] Based on the completed model construction, multi-dimensional model co-simulation and data acquisition are performed:

[0054] Multidimensional simulation models are connected via TCP / IP or OPC UA protocols to achieve data exchange and synchronization between different simulation modules. The aerodynamic performance of the aircraft under different flight conditions is simulated and data is acquired in the following ways:

[0055] Computational fluid dynamics (CFD) simulations provide data such as airflow distribution and pressure distribution;

[0056] Structural simulation provides stress and deformation data;

[0057] Parametric modeling adjusts the geometry of the aircraft and automatically optimizes the simulation parameters.

[0058] These data will serve as the foundation for subsequent data fusion and enhancement.

[0059] By fusing physical data, historical data, and simulation data, the integrity and consistency of the data are ensured. The specific data fusion process includes:

[0060] Data layer fusion: performs preliminary processing on simulation data and physically acquired data to eliminate noise and errors;

[0061] Feature layer fusion: Extracting key features from different data sources to form a unified feature representation;

[0062] Decision-level fusion: Based on the joint analysis of simulation results and physical data, the model is optimized and the simulation results are adjusted.

[0063] By fusing aerodynamic performance data at different angles of attack, more accurate lift and drag models can be obtained. The fused data can improve the prediction accuracy of aircraft aerodynamic characteristics.

[0064] The data is fused with credible physical data, historical data and simulation data.

[0065] After multidimensional simulation data acquisition, high-dimensional datasets are often generated, and the processing efficiency of these datasets is relatively low.

[0066] The intrinsic orthogonal decomposition (POD) technique is used to reduce the dimensionality of high-dimensional data. The specific process is as follows:

[0067] Data pattern extraction: By performing SVD decomposition on the flow field data, the high-dimensional pressure and velocity field data are decomposed into finite pattern vectors;

[0068] Pattern filtering: Retain the most representative primary patterns and remove secondary patterns with low contribution to significantly reduce data dimensionality;

[0069] Dimensionality reduction and reconstruction: The flow field is reconstructed using the main mode vectors, maintaining simulation accuracy with less computational cost.

[0070] The specific formula is as follows:

[0071] ,

[0072] in, Represents the flow field. For spatial mode, is the time coefficient, and r is the number of patterns retained.

[0073] Using the dimensionality-reduced data, a deep learning simulation acceleration network is constructed. This invention employs a Transformer network structure and effectively captures the long-range dependencies of aerodynamic characteristics through an attention mechanism. The specific steps are as follows:

[0074] Data preprocessing: Standardize the dimensionality-reduced aerodynamic and physical property data;

[0075] Network Design: Based on the Transformer encoder, a deep neural network is constructed for predicting aerodynamic characteristics. The input is the dimensionality-reduced pattern vector, and the output is the predicted flow field.

[0076] Training and optimization: Through multiple rounds of training and iteration, combined with actual aerodynamic data, the network parameters are optimized; loss functions are used in conjunction with physical mechanisms to ensure the physical consistency of the model.

[0077] ,

[0078] in, For the predicted results, Based on actual simulation results, This is a regularization term.

[0079] The trained Transformer network can quickly predict aerodynamic characteristics such as flow field distribution and pressure distribution when new flight conditions (such as flight speed and angle of attack) are input. Compared with traditional simulation tools, this model can significantly accelerate the calculation while maintaining the same accuracy as traditional simulation methods.

[0080] The full-scene test model includes high-fidelity scene models of various elements required for aircraft operation, such as terrain, weather, particles, trees, oceans, people, and buildings.

[0081] This deep learning acceleration network reduces the computation time for traditional simulations of aircraft aerodynamic characteristics from hours to minutes or even less for real-time predictions. This significantly improves simulation efficiency, making it particularly suitable for large-scale aerodynamic simulation tasks with multiple operating conditions.

[0082] The test full-scenario model and the aircraft full-element model are integrated in a three-dimensional operating environment. The complete test process is carried out by setting specific simulation targets and tasks, and the test results are used to analyze whether the various performance indicators and functions of the test object meet the test requirements.

[0083] In summary, this invention, through the combination of deep learning, data fusion, and dimensionality reduction techniques, can effectively accelerate the simulation process of complex test objects such as aircraft, and is applicable to the testing needs of multiple fields and objects.

[0084] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0085] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for testing full-factor multi-scenario accelerated simulation of a number-real hybrid, the method comprising: The method comprises: ​ Step 1, design a test object element model construction module, which realizes full-element model construction for the test object, and the specific implementation is as follows: Analyze the test object, propose test requirements or targets, performance indicators, and formulate a preliminary test implementation scheme, including refinement and decomposition of the test target; according to the analysis results, construct a geometric model reflecting the geometric shape characteristics of the test object and a multi-field mechanism model reflecting the operation process of the test object; Step 2, design a test object element model numerical-real data fusion and enhancement module, which uses the constructed test object element model and physical data to realize data fusion and enhancement, and the specific implementation is as follows: Perform joint simulation of the geometric model and the mechanism model to realize multi-dimensional simulation data acquisition and collection; collect physical data and historical data, and combine the simulation data to realize fusion and enhancement of the numerical-real fusion test data; use data dimension reduction method to reduce high-dimensional data to extract key features to obtain a data set; Step 3, design a test object element scene model joint simulation acceleration module, which combines a full-element simulation acceleration model and a test full-scene model to realize joint simulation acceleration of the test, and the specific implementation is as follows: For the geometric model of the test object, use the signed distance field to perform geometric model transformation representation; input the geometric model, the signed distance field, and various operation parameters as training input data set into the deep learning simulation acceleration network and perform training and optimization; the trained deep learning simulation acceleration network realizes input of the reduced data and prediction of the simulation result; combine the online learning mechanism to automatically adapt to the characteristics of different test objects, dynamically optimize the simulation process under the premise of maintaining the simulation accuracy; finally, input the full-element deep simulation acceleration model and the test full-scene model to realize joint simulation acceleration and result prediction of the numerical-real fusion test full-element full-scene model; The test object is an airplane, and the requirements and targets of the aerodynamic characteristic test are determined, including but not limited to the following contents: fluid mechanics characteristics, including lift, drag, and pressure distribution.

2. The method of claim 1, wherein, In step 1, the multi-field mechanism model of the test object is modeled by fluid mechanics modeling, structural mechanics modeling, parameterized modeling, and two-dimensional modeling language-based modeling means, the fluid mechanics modeling is based on incompressible fluid model, turbulence model, and heat conduction model; The structural mechanics modeling is constructed by the finite element method, including nonlinear behavior simulation and dynamic response prediction; the parameterized modeling includes adaptive adjustment of geometric shape and automatic calibration of simulation parameters; The two-dimensional modeling language-based modeling is used for coupled simulation of different physical fields to realize accurate simulation of multiple physical fields.

3. The method of claim 2, wherein, In step 1, the joint simulation of the geometric model and the mechanism model is realized by the following ways: using a general simulation interface to call different types of models; realizing interconnection of modules through a communication protocol; using model exchange to realize model-based data exchange between multiple platforms.

4. The method of claim 3, wherein, In step 2, the number of real fusion specifically includes: through the joint processing simulation data and physical acquisition data, eliminating noise and error data fusion; through the extraction of multi-source data characteristics, the establishment of a unified feature representation data fusion; through the joint analysis of simulation results and actual physical data consistency characteristics, the realization of test results data fusion.

5. The method of claim 4, wherein the method is a hybrid test full-factor multi-scenario acceleration simulation method. In step 2, the data dimension reduction method specifically includes the following steps: extracting the main mode with weight greater than 95% from the simulation data set, removing the secondary mode with low weight, and decomposing the high-dimensional data into a limited number of mode vectors.

6. The method of claim 1, wherein the method is a hybrid test full-factor multi-scenario acceleration simulation method. The deep learning simulation acceleration network uses the reduced simulation data and model data as a data set, adjusts the network parameters, constructs the loss function of the network, trains the constructed deep learning simulation acceleration network structure, and through multiple training and iteration, ensures the generalization ability of the deep learning network.

Images