Search

CN-121980686-A - Training method and device for modeling of aeroengine driven by mechanism and data fusion

CN121980686ACN 121980686 ACN121980686 ACN 121980686ACN-121980686-A

Abstract

The invention provides a training method for modeling an aeroengine driven by mechanism and data fusion, which relates to the technical field of aeroengine modeling simulation, and comprises the steps of decomposing a simulation model of the aeroengine into a plurality of components containing unknown parameters to be identified, and associating the components by adopting a component integration algorithm to form a complete simulation model which can be differentiated from end to end; the method comprises the steps of firstly estimating unknown parameters of components in the highest hierarchy according to actual measurement output data, generating a training target of the next lower hierarchy based on calculation results of the estimated hierarchy to sequentially complete estimation of the unknown parameters of the lower hierarchy, and optimizing and updating the unknown parameters by carrying out gradient back propagation in the complete simulation model with the initial estimated values as starting points.

Inventors

  • LIU KAI
  • ZHU JIHONG

Assignees

  • 清华大学

Dates

Publication Date
20260505
Application Date
20260327

Claims (10)

  1. 1. A training method for modeling an aeroengine driven by mechanism and data fusion is characterized by comprising the following steps: (a) Decomposing a simulation model of an aeroengine into a plurality of components containing unknown parameters to be identified, and associating the components by adopting a component integration algorithm to form a complete simulation model which can be differentiated from end to end, wherein the components comprise a mechanism model component and/or a proxy model component, the mechanism model component is built based on a physical mechanism, and the proxy model component is built based on a data driving method; (b) Based on the hierarchical division of the simulation model, carrying out parameter pre-estimation by adopting a strategy of progressive from top to bottom layer by layer to obtain an initial estimated value of the unknown parameter, wherein the strategy comprises the steps of firstly estimating the unknown parameter of a component in the highest hierarchy according to actual measurement output data, and generating a training target of the next lower hierarchy based on the calculation result of the estimated hierarchy so as to sequentially finish the estimation of the unknown parameter of the lower hierarchy; (c) And optimizing and updating the unknown parameters by taking the initial estimated value as a starting point and carrying out gradient back propagation in the complete simulation model.
  2. 2. The method of claim 1, wherein in the step (b), generating a training target of a next lower hierarchy based on the calculation result of the estimated hierarchy includes: And taking the calculation result of the estimated level as the known output of the estimated level, reversely calculating the model input required for generating the known output according to the simulation model, and taking the model input as the training target of the next lower level.
  3. 3. The method of claim 1, wherein in step (b), estimating the unknown parameters of the components in the highest hierarchy based on the measured output data comprises: Unknown parameters of components within the highest hierarchy are adjusted by system identification or machine learning such that, given model input at that hierarchy, the predicted output of the simulation model matches the measured output data.
  4. 4. The method of claim 1, wherein prior to performing step (b), the method further comprises: One of a plurality of physical dependency orders is selected and a hierarchical division of the simulation model is determined according to an inverse of the selected order, the physical dependency order including a causal relationship chain, an energy transfer path, or a mass transfer path.
  5. 5. The method of claim 1, wherein the data-driven method employed by the proxy model component comprises at least one of: a polynomial regression model, a support vector regression model, a gaussian process regression model, or a neural network-based model; a machine learning model based on data feature mapping, preset basis function learning, or kernel function learning.
  6. 6. The method of claim 1, wherein the specific form of the component integration algorithm comprises at least one of: defining an explicit function mapping of data transfer or algebraic operations between components; A numerical solver for solving an implicit equation composed of coupling relationships between components; A machine learning model for fitting transfer relationships between components.
  7. 7. The method of claim 6, wherein when the component integration algorithm comprises a numerical solver for solving an implicit equation, the implicit equation comprises an algebraic equation, a normal differential equation, a differential algebraic equation, or a partial differential equation.
  8. 8. The method of claim 1, wherein the optimizing the updating of the unknown parameters in step (c) comprises: An optimization control method for limiting the update amplitude of the parameter is introduced when calculating the loss function, and comprises the step of introducing a regularization term for restraining the unknown parameter.
  9. 9. A mechanism and data fusion driven aeroengine modeling training device, comprising: The model construction module is used for decomposing a simulation model of the aeroengine into a plurality of components containing unknown parameters to be identified, and associating the components by adopting a component integration algorithm to form a complete simulation model which can be differentiated from end to end; The parameter pre-estimation module is used for carrying out parameter pre-estimation by adopting a strategy of progressive from top to bottom layer by layer based on the hierarchical division of the simulation model to obtain an initial estimation value of the unknown parameter, wherein the strategy comprises the steps of firstly estimating the unknown parameter in the highest hierarchy according to actual measurement output data, and generating a training target of the next lower hierarchy based on the calculation result of the estimated hierarchy so as to sequentially complete the estimation of the unknown parameter of the lower hierarchy; And the parameter fine tuning module is used for optimizing and updating the unknown parameters by taking the initial estimated value as a starting point and carrying out gradient back propagation in the complete simulation model.
  10. 10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements a method of training the mechanism-and-data fusion driven aeroengine modeling of any of claims 1-8 when the computer program is executed.

Description

Training method and device for modeling of aeroengine driven by mechanism and data fusion Technical Field The invention relates to the technical field of aeroengine modeling simulation, in particular to a training method and device for aeroengine modeling driven by mechanism and data fusion. Background In modern scientific research and engineering design, it is important to build a simulation model capable of accurately reflecting a real physical process, and the simulation model has irreplaceable values for performance analysis, state prediction and optimization control of a system, especially in the high-precision fields of aeroengines, chemical processes and the like. Currently, there are two main technical routes for modeling aeroengines. The first is the mechanism modeling (white box modeling) based on the physical rule, which has the advantages that the model has clear physical meaning and good interpretability, but when the physical effect which is unknown or difficult to describe accurately exists in the system, the model precision is often limited. The second is data-driven modeling (black box modeling), which utilizes methods such as machine learning to directly learn the input-output relationship of the system from the measured data, and although the complex nonlinear system can be approximated with higher precision, the problems of poor physical interpretability, weak generalization capability and the like are generally existed. To combine the advantages of the two, the mechanism clear part is described by a mechanism model, and the mechanism fuzzy part is replaced by a data-driven proxy model, so that a mechanism-data fusion modeling (gray box modeling) strategy is generated, and the method becomes a hot spot of current research. However, the existing mechanism-data fusion model still has a core and urgent problem to be solved in training and optimizing, namely, joint optimization is extremely difficult. Such fusion models typically deeply couple multiple mechanism components with data-driven components to form a monolithic model with deep hierarchy and large parameter dimensions. If the random initialized parameters are directly adopted to perform the end-to-end joint optimization, the loss function of the whole model is highly non-convex, and then the uncertainty and the measurement error of the model are overlapped, so that the optimization process is extremely easy to fall into a local optimal solution, and the training failure or the model precision is far from reaching the expectations. Therefore, there is a strong need in the art for a systematic training method that can provide high quality initial values of parameters for such complex fusion models and ensure stable convergence of the global optimization process. Disclosure of Invention The invention provides a training method and a training device for modeling an aeroengine driven by mechanism and data fusion, which are used for solving the technical problems that in the prior art, the model structure is complex, the measurement error is poor, the parameter space is highly non-convex, the joint optimization is difficult, and the training process is easy to fall into a local optimal solution, so that the modeling of the aeroengine with high precision and high stability is realized. The invention provides a training method for modeling an aeroengine driven by mechanism and data fusion, which comprises the following steps: (a) Decomposing a simulation model of an aeroengine into a plurality of components containing unknown parameters to be identified, and associating the components by adopting a component integration algorithm to form a complete simulation model which can be differentiated from end to end, wherein the components comprise a mechanism model component and/or a proxy model component, the mechanism model component is built based on a physical mechanism, and the proxy model component is built based on a data driving method; (b) Based on the hierarchical division of the simulation model, carrying out parameter pre-estimation by adopting a strategy of progressive from top to bottom layer by layer to obtain an initial estimated value of the unknown parameter, wherein the strategy comprises the steps of firstly estimating the unknown parameter of a component in the highest hierarchy according to actual measurement output data, and generating a training target of the next lower hierarchy based on the calculation result of the estimated hierarchy so as to sequentially finish the estimation of the unknown parameter of the lower hierarchy; (c) And optimizing and updating the unknown parameters by taking the initial estimated value as a starting point and carrying out gradient back propagation in the complete simulation model. Optionally, in step (b), generating a next lower level training target based on the estimated level calculation includes taking the estimated level calculation as a known output of the estimated level, and back-c