Engine ventilation system performance prediction method based on physical information Gaussian process
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-12
AI Technical Summary
Existing neural network prediction models cannot accurately account for the spatial non-uniformity of transient flow and the influence of piston motion in engine intake systems. They also rely on a large amount of high-quality data, are prone to overfitting, and lack physical interpretation and universality.
A performance prediction method for the scavenging system of a two-stroke engine based on physical information Gaussian processes is adopted. Through a phased evaluation system, combined with the Gaussian process regression algorithm of the Bayesian framework and physical constraints, a multi-dimensional evaluation model is constructed, embedding the initial mean function and covariance function, and using the master coordinator to optimize the weights of the sub-learners to achieve global collaborative optimization.
Under conditions of small sample data, it achieves accurate prediction of the performance of ventilation systems, has strong physical interpretability and extrapolation ability, improves prediction accuracy and generalization ability, and is suitable for data-scarce and complex nonlinear systems.
Smart Images
Figure CN122021331A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of internal combustion engine simulation and calculation technology, specifically to a method for predicting the performance of an engine scavenging system based on a Gaussian process of physical information. Background Technology
[0002] Predictive models can be broadly categorized into three types based on their training processes: 1. Black-box models driven purely by data; 2. Purely white-box models driven purely by physics; and 3. Hybrid gray-box models combining data-driven and physics-driven approaches. Neural networks, currently the mainstream black-box predictive model, offer advantages such as simple mathematical models, low training costs, and good performance in handling complex nonlinearities, and have been widely applied in the internal combustion engine field. However, neural network predictive models generally face two key problems: 1. Lack of physical interpretability; 2. Strong dependence on datasets. Lack of physical interpretability means that models trained using neural networks cannot know whether their prediction process conforms to known physical laws, which significantly limits their universality. Strong dataset dependence indicates that a large amount of high-quality data is required as a training set during neural network training; if the dataset is insufficient, overfitting is very likely to occur.
[0003] For engine intake systems, a combination of steady-state intake port testing and simulation is typically used for design. However, this approach assumes steady-state flow, stationary pistons, and overly simplistic evaluation parameters. As a result, it cannot account for the spatial non-uniformity of transient flow, the actual piston movement, and the impact of the compression process on the flow. Furthermore, transient simulation calculations are very time-consuming, and there is currently no suitable evaluation system for intake systems. This often leads to overfitting when using neural network predictions.
[0004] Therefore, based on the transient evaluation system of the intake system of an opposed piston two-stroke engine, this invention proposes a fast prediction model for the performance of the scavenging system based on Gaussian process regression of physical information using sparse modeling. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the technical solution of the present invention is as follows:
[0007] A method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process using physical information includes the following contents and steps:
[0008] S1. A three-dimensional simulation model of the ventilation system based on steady-state airway test verification and adjustment;
[0009] S2. Parameterize the ventilation system, extract key parameter features, set the sampling range of each key parameter feature, sample to generate feature data sets, input them into the calibrated 3D simulation model for simulation calculation, and obtain the simulation performance dataset.
[0010] S3. Fit the explicit key parameter features and use relational parameters to represent them. Define the initial mean function by the linear summation of each relational parameter.
[0011] S4. Establish a multi-dimensional evaluation system for the performance of the ventilation system, divide the working process that affects the performance of the ventilation system into several sub-stages, and match the dominant physical equation of the dominant physical process according to each evaluation parameter of each sub-stage.
[0012] The simulation performance dataset is divided into a test set and a training set. For each evaluation parameter in each sub-stage, a dedicated sub-learner is built based on the training set. The sub-learner adopts a Gaussian process regression algorithm based on the Bayesian framework and embeds an initial mean function to build an independent Gaussian kernel. The regression yields a probabilistic performance prediction model with observation noise.
[0013] In the training of sub-learners, physical constraints are applied based on the matching dominant physical equations; the hyperparameters of Gaussian process regression are optimized using the maximum marginal likelihood method, and the negative log-likelihood function is derived; physical loss terms are constructed based on the physical equations of each sub-learner, and the physical loss terms are weighted and fused into the negative log-likelihood function to obtain the Gaussian process regression sub-loss function corresponding to each sub-learner.
[0014] S5. The main coordinator assigns weights to the outputs of each sub-learner through an adaptive weighted fusion strategy.
[0015] S6. Use the test set to evaluate the generalization performance and prediction accuracy of the performance prediction model, set the evaluation index threshold, load the adjustment strategy for the performance prediction model that does not reach the threshold, and repeat the construction and training process of the performance prediction model based on the adjusted configuration until the performance index is met.
[0016] According to one aspect of this disclosure, the sub-stages include a free exhaust stage, a scavenging stage, a post-exhaust stage, and a compression stage. In the multi-dimensional evaluation system, the performance evaluation of the free exhaust stage, the scavenging stage, and the post-exhaust stage includes gas exchange performance and flow field performance, while the performance evaluation of the compression stage includes flow field performance. The gas exchange performance includes scavenging efficiency, capture mass, and feed ratio, while the flow field performance includes eddy current ratio, tumble ratio around the x-axis, and tumble ratio around the y-axis.
[0017] According to one aspect of this disclosure, for the free exhaust stage, the dominant physical equations of the gas exchange performance include the continuity equation; for the scavenging stage and the post-exhaust stage, the dominant physical equations of the gas exchange performance include the continuity equation and the component transport equation; for the compression stage, the dominant physical equations of the gas exchange performance include the continuity equation and the gas state equation; for the entire stage, the flow field performance is constrained primarily by the continuity equation and secondarily by the gas state equation.
[0018] According to one aspect of this disclosure, the key parameter feature sampling in step S2 is initially sampled based on the Sobol uniform sampling method; after the initial sampling, a preliminary performance prediction model is trained based on the preliminary feature dataset and a performance evaluation is performed. If the evaluation result is not up to standard, a second sampling is initiated based on the Monte Carlo sampling method to supplement the preliminary feature dataset and obtain the final feature dataset.
[0019] According to one aspect of this disclosure, the covariance function can be selected from... The prediction accuracy and generalization performance of performance prediction models built based on the three kernel function structures (kernel function, RBF kernel function, or linear combination kernel function) are compared to determine the final covariance function structure used to define the Gaussian process; among them, the linear combination kernel function is... A linear combination of kernel functions and RBF kernel functions.
[0020] According to one aspect of this disclosure, the physical constraint points on which the physical constraints in step S4 are based are collected using a basic sampling method: taking the crankshaft angle as the time dimension, the fluid velocity amplitude of the ventilation system at each crankshaft angle is calculated, and each crankshaft angle is sorted according to the fluid velocity amplitude. Based on the sorting result, the number of spatial sampling points is allocated to the gradient of each crankshaft angle, and a spatial uniform sampling method is used to sample in the space corresponding to each crankshaft angle to obtain the basic physical constraint point set.
[0021] According to one aspect of this disclosure, physical constraint point sampling is based on a two-stage physical information driven strategy: the first stage implements basic sampling to obtain a basic set of physical constraint points; the second stage implements adaptive sampling based on the physical field, combining the training results of the sub-learner, and performs targeted sampling on the high-value regions and high-gradient regions of its main physical field to supplement the basic set of physical constraint points and obtain an enhanced set of physical constraint points.
[0022] According to one aspect of this disclosure, in step S2, the parameter characteristics of the ventilation system are optimized based on the Spearman correlation analysis method, and key parameter characteristics with strong correlation are selected.
[0023] According to one aspect of this disclosure, the initial mean function includes a flow field prior mean function and an air exchange 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 air exchange prior mean function is used for Gaussian process regression of the sub-learner related to air exchange performance.
[0024] Compared with the prior art, the performance prediction method of a two-stroke engine scavenging system based on a Gaussian process of physical information of the present invention has the following beneficial effects:
[0025] This prediction method divides the operation of the ventilation system into four stages and designs a two-layer learning architecture for multiple stages and performance parameters. The lower layer consists of multiple parallel dedicated sub-learners, each with a customized combination of mean function, Gaussian kernel, and physical constraints, which can accurately predict the comprehensive performance of the ventilation system under real transient operating cycles. The upper layer master coordinator drives global collaborative optimization to achieve system and global optima for each sub-learner. This method embeds structured priors and physical constraints into a Gaussian process regression framework, enabling the model to achieve good generalization ability and high prediction accuracy even with small sample data. At the same time, the model has strong physical interpretability and strong extrapolation ability, making it suitable for performance prediction of ventilation systems with scarce data, complex mechanisms, and strong nonlinearity. Attached Figure Description
[0026] Figure 1 This is a flowchart of the air inlet performance prediction method disclosed in this invention;
[0027] Figure 2 A comprehensive schematic diagram showing the air inlet structure and its structural characteristic parameters;
[0028] Figure 3 R-squared for deep neural networks, zero-mean GPR, and ensemble GPR under fixed-set-test-set partitioning. 2 value;
[0029] Figure 4 MAE values for deep neural networks, zero-mean GPR, and ensemble GPR under the fixed-set-test-set partitioning setting;
[0030] Figure 5 Ri for deep neural networks, zero-mean GPR, and ensemble GPR in a five-fold cross-validation setting 2 value;
[0031] Figure 6 MAE values for deep neural networks, zero-mean GPR, and ensemble GPR under five-fold cross-validation settings.
[0032] Attached image labels:
[0033] 1. Air inlet one; 2. Air inlet two; The upper tilt angle of air inlet 1; The lower tilt angle of air inlet 1; The upper tilt angle of air inlet 2; 1. Lower tilt angle of air inlet 2; W. Air inlet width; Total height of the air inlet; Height of the upper layer of the air inlet. Detailed Implementation
[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely the best embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] The term "embodiment" as used herein means that a particular method, step, or content described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0036] This embodiment provides a performance prediction method for a two-stroke engine scavenging system based on a Gaussian process of physical information. It utilizes a Gaussian process regression method to handle complex nonlinear physical processes under small sample training conditions, adapting to the design and operating characteristics of diesel engine scavenging systems. Figures 1-3 As shown, this embodiment combines Figure 2 The ventilation system shown illustrates the steps and content of the method for predicting the performance of the air inlet:
[0037] S1, based on such Figure 2 The air inlet structure shown in the ventilation system underwent a steady-state airway test. This air inlet structure employs a double-layer, double-structure design, comprising air inlet 11 and air inlet 2, which are evenly arranged in a ring around the airway circumference. Both air inlet 11 and air inlet 2 are double-layered structures, with the upper and lower inclination angles differing. Furthermore, the inclination angles of adjacent air inlets within the same layer are different; that is, the upper inclination angles of air inlet 11 and air inlet 22 are different, and the lower inclination angles of air inlet 11 and air inlet 22 are different.
[0038] Steady-state airway tests were conducted on 3-5 different ventilation systems with the above-mentioned air inlet structures to obtain test data including but not limited to steady-state eddy ratio and flow coefficient. Simulation calculations were performed on a three-dimensional model of the ventilation system with the same structural parameters. The three-dimensional simulation model of the ventilation system was verified and adjusted using the test data. If the simulation results differed significantly from the test data, the parameters in the turbulence model were modified for calibration. The turbulence model generally adopts the RANS model. If adjusting the parameters of the RANS model could not obtain the required simulation results, the LES model was used for simulation calculations.
[0039] S2. The ventilation system is structurally parameterized. Structural parameterization should ensure that the parameter features completely describe the ventilation system as much as possible. Considering that a large number of parameterized features can lead to the optimizer getting trapped in local optima during mean function embedding, an appropriate number of features are extracted from the above parameter features based on Spearman correlation analysis to screen out strongly correlated key parameter features. The number of key parameter features should not exceed seven. For example, in this embodiment, the key parameter feature of the air inlet structure is the upper tilt angle of air inlet-1. The lower tilt angle of air inlet 1 The upper tilt angle of air outlet 2 The lower tilt angle of air outlet 2 1. Air inlet width W; 2. Total air inlet height and the height of the upper air inlet ;
[0040] Based on the structural dimension design experience range, the sampling range of each key parameter feature is set. The data of the key parameter feature is initially sampled based on the Sobol uniform sampling method. Depending on the number of key parameter features, 32 or 64 sets of small sample data are collected. When there are many key parameter features, 64 sets of data are automatically collected. In this embodiment, there are 7 key parameter features, so the sampling amount of its key parameter features is 64 sets.
[0041] The sampled feature data set is input into the calibrated 3D simulation model for simulation calculation to obtain the simulation performance dataset, which includes, but is not limited to, eddy current ratio, scavenging efficiency, capture mass and feed ratio.
[0042] S3. Integrate explicit key parameter features based on experience or performance mechanism correlations, and use relational formulas to initially fit the relationships between different key parameter features. Describe the parameters using more representative relational parameters. The descriptive parameters of the ventilation system should be able to describe the ventilation system structure as comprehensively as possible. This embodiment uses a ventilation system with a double-layer, double-structure air inlet structure as an example. Figure 2 As shown, the relationship parameters of its ventilation system and their fitting equations are shown in Table 1:
[0043] Table 1. Comparison of Relationship Parameters and Their Fitting Formulas
[0044]
[0045] In the table, W represents the width of the air inlet; and These represent the total height of the air inlet and the upper height of the air inlet, respectively, and are used to define the height distribution ratio between the upper and lower layers.
[0046] The above relational parameters are linearly summed to construct a prior initial mean function that can be embedded in Gaussian process regression.
[0047] To accommodate the different correlations of different performance characteristics, such as ~ It mainly describes the airflow guidance characteristics. , and The main description focuses on flow capacity, and prior mean functions of the flow field are constructed to adapt to the flow field performance. and the prior mean function of ventilation to adapt to ventilation performance ;in,
[0048] (1)
[0049] (2)
[0050] in (x=1,2,...,8) are the weight hyperparameters in the mean function, which are optimized together with the hyperparameters in the Gaussian kernel function using the maximum marginal likelihood method.
[0051] S4. Establish a multi-dimensional evaluation system for the performance of the ventilation system, including evaluating the ventilation performance and flow field performance of the ventilation system. The main performance parameters of the ventilation performance include scavenging efficiency, air supply ratio and capture mass. The performance parameters that affect the flow field performance include eddy ratio, tumble ratio around the X-axis and tumble around the Y-axis.
[0052] The scavenging process of a diesel engine can be divided into the free exhaust stage, the scavenging stage, and the after-exhaust stage. The scavenging efficiency, capture quality, and air-fuel ratio used to evaluate the scavenging process only need to consider the above three stages. The influence of the scavenging system on the flow field mainly affects the spray and combustion. Therefore, for the three flow field performance parameters of eddy ratio, tumble ratio around the X-axis, and tumble around the Y-axis, the values must correspond to the in-cylinder values at about 1°CA before spraying. The flow field performance parameters also need to be evaluated in the compression stage.
[0053] Therefore, the working process that affects the performance of the ventilation system is divided into four stages: free exhaust stage, scavenging stage, post-exhaust stage and compression stage. Based on each evaluation parameter of each stage, the physical equation of the dominant physical process is matched.
[0054] The simulation performance dataset is divided into a test set and a training set. For each evaluation parameter in each stage, a dedicated sub-learner is constructed based on the training set. The sub-learner adopts a Gaussian process regression algorithm based on the Bayesian framework. According to the design of seven evaluation parameters and four stages in this embodiment, an independent Gaussian kernel function for each evaluation parameter is constructed based on time series segmentation. In total: 4 stages * 4 flow field performances + 3 stages * 3 ventilation performances = 25. The above-mentioned flow field prior mean function and ventilation prior mean function are used as non-stationary prior mean functions and embedded in the Gaussian process regression.
[0055] A Gaussian process is defined as an infinite-dimensional set of random variables over a continuous domain, where any finite number of random variables follow a joint Gaussian distribution, i.e.:
[0056] (3)
[0057] In equation (3), It is a mean function. It is the covariance function, i.e., the Gaussian kernel function;
[0058] Assuming each output parameter is defined by a non-stationary prior mean function, a zero-mean Gaussian process, and observation noise, a performance prediction model is established:
[0059] (4)
[0060] In equation (4), i represents the i-th characteristic parameter data point in the flow field. It is a zero-mean Gaussian random process. Let be the observation error of the i-th data point, which follows a normal distribution, i.e. ;
[0061] For the training set and test set The input, whose observed values and predicted target satisfy a joint Gaussian prior distribution, i.e.:
[0062] (5)
[0063] In equation (5), X is the input matrix of the training set. n is the number of samples in the training set; y is the observed data vector. ; The noise-free prediction target corresponding to the test point; Input matrix for the test set, m is the number of samples in the test set; The observation noise variance matrix;
[0064] Based on the joint distribution (5) and the Gaussian conditional distribution formula, the posterior predicted distribution of the function values at the test points can be derived:
[0065] (6)
[0066] In equation (6), the predicted mean As a point estimate, predicting covariance Uncertainty in quantitative forecasting:
[0067] (7)
[0068] (8)
[0069] The hyperparameters in Gaussian process regression are optimized using the maximum marginal likelihood method, corresponding to the negative log-likelihood function. :
[0070] (9)
[0071] In equation (9), This is the set of hyperparameters; the function consists of three parts: The term represents the deviation between the observed value and the model prediction value, and is used to penalize the error in fitting the performance prediction model; This term is used to penalize the complexity of the model and prevent overfitting of the performance prediction model; The term is used to ensure the normalization of the probability distribution;
[0072] To incorporate prior physical knowledge, a physical information Gaussian process regression framework is constructed, with the total loss function being a weighted sum of data loss and physical constraint loss:
[0073] (10)
[0074] In equation (10), For the total marginal likelihood loss, and To balance the parameters, + =1, For the marginal likelihood loss of the data, This represents the physical marginal likelihood loss.
[0075] For the gas exchange process and compression stage of the gas exchange system, consider the dominant physical equations as soft constraints. The dominant physical equations include, but are not limited to, the continuity equation L1, the component transport equation L2, and the gas state equation L3:
[0076] (11)
[0077] (12)
[0078] (13)
[0079] In equations (11), (12) and (13), u, v and w are the component velocities of the constraint point in the flow field in the x, y and z directions, respectively; The value represents the mass fraction of the fresh gas, and t represents the time interval of the stage. It is a velocity vector. denoted as the two-component diffusion coefficient of fresh gas and residual exhaust gas; P is the in-cylinder gas pressure, and T is the in-cylinder gas temperature. Let j be the mass fraction of component j. Let R be the molar mass of component j, and R be the universal gas constant. The density of the mixture;
[0080] The fresh gas is usually air, and the proportion of each component in the residual exhaust gas is calculated based on the actual air-fuel ratio, thereby obtaining the molar mass of the above components and the density of the mixture.
[0081] Different combinations and weights of physical constraints are set based on the physical characteristics of each stage:
[0082] For the free exhaust phase, the main physical equations for the gas exchange performance and flow field performance are continuity equations, therefore, their physical marginal likelihoods... for:
[0083] (14)
[0084] This stage uses the continuity equation as the main governing equation; therefore, the fusion weights... The influence of the gas state can also be ignored;
[0085] For the scavenging stage, the dominant physical equations for scavenging performance, tumble ratio around the x-axis, tumble ratio around the y-axis, and in-cylinder average TKE are the continuity equation and the component transport equation. Therefore, their physical marginal likelihoods... for:
[0086] (15)
[0087] Since the continuity equation is the main governing equation in this stage, and the component transport equation also plays a certain role, the fusion weights... Furthermore, their weights are similar;
[0088] For the after-exhaust stage, the dominant physical equations for scavenging performance, tumble ratio around the x-axis, tumble ratio around the y-axis, and in-cylinder average TKE are the continuity equation and the component transport equation. Therefore, its physical marginal likelihood... for:
[0089] (16)
[0090] In this stage, the continuity equation and the component transport equation play equally important roles, with weighted integration. ;
[0091] For the compression stage, the dominant physical equations for gas exchange performance and flow field performance are the continuity equation and the gas state equation. Therefore, its physical marginal likelihood... for:
[0092] (17)
[0093] At this stage, the continuity equation and the gas law equation have equally important influences, and the fusion weights are crucial. ;
[0094] For the eddy ratio, the physical equations throughout the entire process use the continuity equation as the primary constraint and the gas state equation as the secondary constraint. Therefore, its physical marginal likelihood... for:
[0095] (18)
[0096] This stage involves weight fusion. ;
[0097] For different evaluation objectives and stages, construct corresponding total negative logarithmic marginal likelihood functions. By minimizing this function, all hyperparameters in the model can be optimized. The joint learning process employs the L-BFGS-B algorithm, with hyperparameters... This includes, but is not limited to, the parameters of the Gaussian process kernel function, the mean function hyperparameter, the observation noise variance, the trade-off parameters, and the weighting coefficients.
[0098] S5. This example implementation adopts a two-layer ensemble learning architecture. The lower layer consists of multiple parallel sub-learners, each of which is trained and predicted independently. The upper layer is the main coordinator, responsible for global coordination and optimization. First, based on the real-time performance and uncertainty of each sub-learner, the fusion weights of each sub-learner are dynamically calculated and allocated, and the prediction results of each sub-learner are integrated into a unified global prediction. (19)
[0099] In equation (19), , , ... The loss weights are for each physical marginal likelihood function;
[0100] The master coordinator back-adjusts the loss weights of each sub-learner by minimizing the ensemble prediction error of all sub-learners on a shared validation set, thus guiding their parameter trade-offs. and fusion weight The redistribution guides each sub-learner to collaboratively optimize towards the overall optimal direction.
[0101] S6. Use the test set to evaluate the prediction accuracy and generalization ability of the trained prediction model. Evaluation metrics include, but are not limited to, mean error (MAE) and coefficient of determination (R²). 2 And predict the coverage area (PICP); preset performance thresholds for each indicator; if the model evaluation results do not reach the thresholds, analyze the potential reasons, load the corresponding adjustment strategies to optimize the model, and perform at least one of the following optimization adjustments:
[0102] 1. Initiate secondary sampling to supplement the preliminary feature dataset. Perform secondary sampling based on the Monte Carlo sampling method to increase the sampling density and supplement the preliminary feature dataset to form the final feature dataset used for simulation.
[0103] 2. Optimize the hyperparameters of the sub-learners, including but not limited to kernel function parameters, noise variance, and fusion weights;
[0104] 3. Adjust the various loss weights and trade-off parameters of the master coordinator;
[0105] 4. Change the Gaussian kernel function. In this embodiment, the kernel function can be selected as follows: Any one of the following kernel functions: kernel function, RBF kernel function, and linear combination kernel function, wherein the linear combination kernel function is: A linear combination of kernel functions and RBF kernel functions;
[0106] Based on the adjusted new configuration, the model is reconstructed and trained, and this process is iterated until the model meets the evaluation criteria.
[0107] Depend on Figures 3-6 As can be seen, for engineering prediction problems such as ventilation systems where data is scarce, mechanisms are complex, and nonlinearity is strong, the performance prediction model constructed based on the Gaussian structural regression framework of physical information in this embodiment exhibits higher prediction accuracy and generalization stability, regardless of whether the fixed set-test set partitioning setting is used or the 5-fold cross-validation setting is used. Furthermore, compared to models constructed using deep neural network models or conventional Gaussian process regression with zero-valued mean functions, the model's prediction performance is significantly improved through the embedding of structured mean functions and physical constraints. Specifically, the coefficient of determination R0... 2The improvement reached up to 20% (under the 5-fold cross-validation setting). Compared with the model constructed by the zero-value mean function, the mean absolute error (MAE) was only less than 37.71% of the comparison model, indicating that the performance prediction model based on physical information integrated Gaussian process regression captures data variability more effectively and has higher prediction accuracy.
[0108] As a further technical solution, the dominant physical equation in step S4 implements physical constraints based on the corresponding spatiotemporally sampled physical constraint points, and the sampling of physical constraint points adopts a two-stage physical information-driven strategy; wherein...
[0109] The first stage focuses on basic sampling, which is based on global spatiotemporal characteristics and aims to obtain a set of basic physical constraint points covering the entire computational domain. Using the crankshaft angle as the time coordinate, the entire stage is discretized with a preset time step. The fluid velocity amplitude of the flow field in the entire domain at each crankshaft angle is calculated. The crankshaft angles are sorted according to the magnitude of the fluid velocity amplitude, and more spatial sampling points are allocated to the crankshaft angles with higher velocity amplitudes based on the sorting results. This achieves the importance gradient allocation of sampling resources in the time dimension. Within the computational space threshold corresponding to each crankshaft angle, a spatial uniform sampling method is used to implement local uniform sampling according to the allocated number of points.
[0110] Supplemented by the second-stage adaptive sampling of the physical field, based on the prediction of the physical field by the sub-learner trained in the first stage, the dominant physical fields (such as velocity field and component concentration field) in different stages are calculated, thresholds are set, high numerical regions and high gradient regions under the physical field are identified, and adaptive encrypted sampling is implemented. The encryption strategy can be based on random sampling weighted by field strength or gradient magnitude, or local linear encryption is performed on the normal with the largest gradient. The newly added sampling points in the second stage are merged with the basic physical constraint point set in the first stage to form the final enhanced physical constraint point set.
[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that the various embodiments of this application can be implemented by means of software or software combined with necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware functions. Based on this understanding, the technical solution of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. The software product is stored in a storage medium and includes several instructions to cause a computer device, such as including but not limited to a personal computer, server, or network device, to execute all or part of the steps of the method described in any embodiment of this application.
[0112] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information, characterized in that, Includes the following steps and content: S1. A three-dimensional simulation model of the ventilation system based on steady-state airway test verification and adjustment; S2. Parameterize the ventilation system, extract key parameter features, set the sampling range of each key parameter feature, sample to generate feature data groups, input them into the calibrated three-dimensional simulation model for simulation calculation, and obtain simulation performance dataset; S3. Fit the explicit key parameter features and use relational parameters to represent them, and define the initial mean function by the linear summation of each relational parameter; S4. Establish a multi-dimensional evaluation system for the performance of the ventilation system, divide the working process that affects the performance of the ventilation system into several sub-stages, and match the dominant physical equation of the dominant physical process according to each evaluation parameter of each sub-stage. The simulation performance dataset is divided into a test set and a training set. For each evaluation parameter in each sub-stage, a dedicated sub-learner is constructed based on the training set. The sub-learner employs a Gaussian process regression algorithm based on a Bayesian framework and embeds the initial mean function to construct an independent Gaussian kernel, thereby obtaining a probabilistic performance prediction model with observation noise. In the training of the sub-learners, physical constraints are applied based on the matched dominant physical equations; the hyperparameters of Gaussian process regression are optimized using the maximum marginal likelihood method, and the negative log-likelihood function is derived; physical loss terms are constructed based on the dominant physical equations of each sub-learner, and the physical loss terms are weighted and fused into the negative log-likelihood function to obtain the Gaussian process regression sub-loss function corresponding to each sub-learner; S5. The main coordinator assigns weights to the outputs of each of the sub-learners through an adaptive weighted fusion strategy. S6. Use the test set to evaluate the generalization performance and prediction accuracy of the performance prediction model, set the evaluation index threshold, load the adjustment strategy for the performance prediction model that does not reach the threshold, and repeat the construction and training process of the performance prediction model based on the adjusted configuration until the performance index is met.
2. The method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information according to claim 1, characterized in that: The sub-stages include a free exhaust stage, a scavenging stage, a post-exhaust stage, and a compression stage. In the multi-dimensional evaluation system, the performance of the free exhaust stage, the scavenging stage, and the post-exhaust stage includes air exchange performance and flow field performance, and the performance of the compression stage includes flow field performance. The air exchange performance includes scavenging efficiency, capture mass, and feed ratio, and the flow field performance includes eddy ratio, tumble ratio around the x-axis, and tumble ratio around the y-axis.
3. The method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information according to claim 2, characterized in that: For the free exhaust phase, the dominant physical equation for the ventilation performance includes a continuity equation; for the scavenging phase and the post-exhaust phase, the dominant physical equation for the ventilation performance includes a continuity equation and a component transport equation. For the compression phase, the dominant physical equations for the gas exchange performance include the continuity equation and the gas state equation. For the entire stage, the flow field performance is based on the continuity equation as the primary constraint equation and the gas state equation as the secondary constraint equation.
4. The method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information according to claim 1, characterized in that: The key parameter feature sampling in step S2 is initially sampled based on the Sobol uniform sampling method; After the initial sampling, a preliminary performance prediction model is trained based on the preliminary feature dataset and its performance is evaluated. If the evaluation result is not up to standard, a second sampling based on the Monte Carlo sampling method is initiated to supplement the preliminary feature dataset and obtain the final feature dataset.
5. The method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information according to claim 1, characterized in that: Covariance function can be selected The prediction accuracy and generalization performance of the performance prediction models constructed based on the three kernel function structures—kernel function, RBF kernel function, or linear combination kernel function—are compared to determine the final covariance function structure used to define the Gaussian process; wherein, the linear combination kernel function is the... A linear combination of the kernel function and the RBF kernel function.
6. The method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information according to claim 1, characterized in that: In step S4, the physical constraint points applied by the physical constraints are collected based on the basic sampling method: taking the crankshaft angle as the time dimension, the fluid velocity amplitude of the ventilation system at each crankshaft angle is calculated, and each crankshaft angle is sorted according to the fluid velocity amplitude. Based on the sorting result, the number of spatial sampling points is allocated to the gradient of each crankshaft angle. The spatial uniform sampling method is used to sample in the space corresponding to each crankshaft angle to obtain the basic physical constraint point set.
7. The method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information according to claim 6, characterized in that: The physical constraint point sampling is based on a two-stage physical information driven strategy: the first stage implements the basic sampling to obtain the basic physical constraint point set; The second stage implements adaptive sampling based on the physical field. Combining the training results of the sub-learner, targeted sampling is performed on the high-value and high-gradient regions of its main physical field to supplement the basic physical constraint point set and obtain an enhanced physical constraint point set.
8. The method for predicting the performance of a two-stroke engine scavenging system based on a Gaussian process of physical information according to claim 1, characterized in that: In step S2, the parameter characteristics of the ventilation system are optimized based on the Spearman correlation analysis method, and the key parameter characteristics with strong correlation are selected.
9. The prediction method for the gas exchange system of a two-stroke engine based on a Gaussian process of physical information according to claim 2, characterized in that: The initial mean function includes 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 the flow field performance, and the ventilation prior mean function is used for Gaussian process regression of the sub-learner related to the ventilation performance.