CN-122021331-A - Engine ventilation system performance prediction method based on physical information Gaussian process

CN122021331ACN 122021331 ACN122021331 ACN 122021331ACN-122021331-A

Abstract

The invention discloses an engine ventilation system performance prediction method based on a physical information Gaussian process, which comprises the steps of calibrating a three-dimensional simulation model, extracting dominant key parameter characteristics, sampling to generate a characteristic data set, inputting the simulation model to simulate to obtain a performance data set, linearly summing relation parameters describing related key dominant parameter characteristics, constructing an initial mean function, establishing a multidimensional evaluation system and a division stage, constructing a dedicated sub-learner for each evaluation parameter in each stage, embedding the initial mean function, applying physical constraint by training the sub-learner based on a matched physical equation, adaptively distributing weights by a main coordinator, evaluating the performance of the model, and embedding the structured priori and the physical constraint into a Gaussian process regression frame by the method, so that the model still can obtain better generalization capability and higher prediction accuracy under small sample data, has strong physical interpretation and extrapolation, and is suitable for ventilation system performance prediction with sparse data, complex mechanism and strong nonlinearity.

Inventors

QIN JING
LIANG YONGSEN
LI SHANGSHU
PEI YIQIANG
Su liwang
Bo Yaqing

Assignees

天津大学

Dates

Publication Date: 20260512
Application Date: 20260206

Claims (9)

1. A two-stroke engine ventilation system performance prediction method based on a physical information Gaussian process is characterized by comprising the following steps and contents: s1, verifying and adjusting a three-dimensional simulation model of a ventilation system based on a steady-state airway test; S2, parameterizing the ventilation system, extracting key parameter characteristics, setting sampling ranges of the key parameter characteristics, sampling to generate a characteristic data set, and inputting the characteristic data set to perform simulation operation based on the calibrated three-dimensional simulation model to obtain a simulation performance data set; S3, fitting the dominant key parameter features, representing by using relation parameters, and defining an initial mean function by linear summation of the relation parameters; s4, establishing a multi-dimensional evaluation system of the performance of the ventilation system, dividing the working process affecting the performance of the ventilation system into a plurality of sub-stages, and matching a dominant physical equation of a dominant physical process according to each evaluation parameter of each sub-stage; Dividing the simulation performance data set into a test set and a training set, and constructing a dedicated sub-learner based on the training set aiming at each evaluation parameter of each sub-stage; the sub learner adopts a Bayesian process regression algorithm based on a Bayesian framework, embeds the initial mean function to construct an independent Gaussian kernel, and carries out regression to obtain a probabilistic performance prediction model with observation noise; Optimizing the hyper-parameters of Gaussian process regression by adopting a maximum marginal likelihood method, deducing to obtain a negative log-likelihood function, respectively constructing physical loss terms based on the dominant physical equation of each sub-learner, and carrying out weighted fusion on the physical loss terms into the negative log-likelihood function to obtain a Gaussian process regression sub-loss function corresponding to each sub-learner; S5, the main coordinator distributes weights for the output of each sub-learner through a self-adaptive weighted fusion strategy; S6, evaluating the generalization performance and the prediction precision of the performance prediction model by using the test set, setting an evaluation index threshold, loading an adjustment strategy for the performance prediction model which does not reach the threshold, and repeating the construction and training processes of the performance prediction model based on the adjusted configuration until the performance index is met.
2. The method according to claim 1, wherein the sub-phases include a free exhaust phase, a scavenging phase, a post exhaust phase, and a compression phase, and the performances of the free exhaust phase, the scavenging phase, and the post exhaust phase include a ventilation performance and a flow field performance, and the performance of the compression phase includes a flow field performance, and the ventilation performance includes a scavenging efficiency, a capturing quality, and a gas supply ratio, and the flow field performance includes a swirl ratio, a tumble ratio around an x-axis, and a tumble ratio around a y-axis.
3. The method for predicting performance of a two-stroke engine ventilation system based on a physical information Gaussian process according to claim 2, wherein the dominant physical equation of the ventilation performance comprises a continuity equation for the free exhaust stage, the dominant physical equation of the ventilation performance comprises a continuity equation and a component transport equation for the scavenging stage and the post exhaust stage, the dominant physical equation of the ventilation performance comprises a continuity equation and a gas state equation for the compression stage, and the continuity equation is used as a primary constraint equation and the gas state equation is used as a secondary constraint equation for the flow field performance for the full stage.
4. The method for predicting the performance of the two-stroke engine ventilation system based on the Gaussian process of physical information according to claim 1 is characterized in that the key parameter feature sampling in the step S2 is performed primarily based on a Sobol uniform sampling mode, a preliminary performance prediction model is obtained through training based on a preliminary feature dataset after the primary sampling, performance evaluation is performed, and if the evaluation result does not reach the standard, the method based on Monte-Carlo sampling is started to perform secondary sampling, and the preliminary feature dataset is supplemented to obtain a final feature dataset.
5. A two-stroke engine ventilation system performance prediction method based on a physical information Gaussian process according to claim 1, wherein the covariance function is selected from the group consisting of A kernel function, an RBF kernel function or a linear combination kernel function, wherein the linear combination kernel function is the one used for defining a Gaussian process by comparing the prediction precision and the generalization performance of the performance prediction model constructed based on the three kernel function structures A linear combination of a kernel function and the RBF kernel function.
6. The method for predicting the performance of the two-stroke engine ventilation system based on the Gaussian process of physical information of claim 1, wherein the physical constraint points applied by physical constraints in the step S4 are collected based on a basic sampling method, a crank angle is taken as a time dimension, fluid speed amplitude values of the ventilation system under each crank angle are calculated, the crank angles are ordered according to the fluid speed amplitude values, space sampling points are distributed for each crank angle gradient according to an ordering result, and a space uniform sampling method is adopted to sample in a space corresponding to each crank angle, so that a basic physical constraint point set is obtained.
7. The two-stroke engine ventilation system performance prediction method based on the physical information Gaussian process of claim 6 is characterized in that the physical constraint point sampling is based on a two-stage physical information driving strategy, wherein the basic sampling is implemented in a first stage to obtain a basic physical constraint point set, the self-adaptive sampling based on a physical field is implemented in a second stage, the sub-learner training result is combined, the target sampling is carried out on a high-value region and a high-gradient region of a main physical field of the sub-learner training result, and the basic physical constraint point set is supplemented to obtain an enhanced physical constraint point set.
8. The method for predicting the performance of a two-stroke engine ventilation system based on a physical information Gaussian process according to claim 1, wherein in step S2, the parameter characteristics of the ventilation system are optimized based on spearman correlation analysis method, and the key parameter characteristics with strong correlation are screened.
9. The method for predicting a two-stroke engine gas ventilation system based on a physical information Gaussian process of claim 2, wherein the initial mean function comprises a flow field prior mean function and a ventilation prior mean function, the flow field prior mean function is used for Gaussian process regression of the sub-learner related to flow field performance, and the ventilation prior mean function is used for Gaussian process regression of the sub-learner related to ventilation performance.

Description

Engine ventilation system performance prediction method based on physical information Gaussian process Technical Field The invention relates to the technical field of simulation calculation of internal combustion engines, in particular to an engine ventilation system performance prediction method based on a physical information Gaussian process. Background The prediction model can be mainly divided into three types according to different training processes, namely, 1 a black box model based on pure data driving, 2a pure white box model based on pure physical driving, and 3a mixed ash box model based on the combination of the data driving and the physical driving. The neural network is used as a current mainstream black box prediction model, has the advantages of simple mathematical model, low training cost, good complex nonlinear processing effect and the like, and is widely applied to the field of internal combustion engines. But the neural network prediction model has two major key problems, namely 1 no physical interpretation and 2 strong dependence on a data set. The lack of physical interpretation indicates that the model trained using the neural network cannot know whether the prediction process accords with the known physical rule, which limits the universality to a great extent. The strong data set dependence indicates that a large amount of high quality data is required as a training set in the process of training the neural network, and once the data set is insufficient, the over-fitting phenomenon is very easy to occur. For an engine air inlet system, a method combining a steady-state air passage test and simulation is generally adopted for design, but the method assumes that the flow is steady, the piston is static and the evaluation parameters are too simple, so that the method cannot consider the space non-uniformity of transient flow, the actual piston movement and the influence of a compression process on the flow, when transient simulation calculation is adopted, the simulation calculation time cost is very high, and an applicable air inlet system evaluation system still does not exist at present, and the phenomenon of fitting is easy to occur when a neural network is predicted; Therefore, the invention provides a ventilation system performance rapid prediction model based on sparse modeling physical information Gaussian process regression based on a transient evaluation system of an air inlet system of an opposed-piston two-stroke engine. Disclosure of Invention The invention aims to overcome the defects of the prior art and provide a two-stroke engine ventilation system performance prediction method based on a physical information Gaussian process so as to solve the problems in the background art. In order to achieve the above object, the technical scheme of the present invention is as follows: A two-stroke engine ventilation system performance prediction method based on a physical information Gaussian process comprises the following contents and steps: s1, verifying and adjusting a three-dimensional simulation model of a ventilation system based on a steady-state airway test; S2, parameterizing the air exchange system, extracting key parameter characteristics, setting sampling ranges of the key parameter characteristics, sampling to generate a characteristic data set, and inputting the characteristic data set to perform simulation operation based on the calibrated three-dimensional simulation model to obtain a simulation performance data set; S3, fitting dominant key parameter features, representing by using relation parameters, and defining an initial mean function by linear summation of the relation parameters; S4, establishing a multi-dimensional evaluation system of the performance of the ventilation system, dividing the working process affecting the performance of the ventilation system into a plurality of sub-stages, and matching a dominant physical equation of a dominant physical process according to each evaluation parameter of each sub-stage; Dividing the simulation performance data set into a test set and a training set, and constructing a dedicated sub-learner based on the training set aiming at each evaluation parameter of each sub-stage; Optimizing the hyper-parameters of the Gaussian process regression by adopting a maximum marginal likelihood method, deducing to obtain a negative log-likelihood function, respectively constructing physical loss terms based on the physical equations of all the sub-learners, weighting and fusing the physical loss terms into the negative log-likelihood function, and obtaining the Gaussian process regression sub-loss function corresponding to each sub-learner; S5, the main coordinator distributes weights for the output of each sub-learner through a self-adaptive weighted fusion strategy; s6, evaluating the performance prediction model by using the test set to perform generalization performance and prediction accuracy, setting an evaluation inde