Diesel engine thermal performance prediction based on digital simulation and extrapolation method for extreme working conditions

Abstract

The application discloses a diesel engine thermal performance prediction and limit working condition extrapolation method based on digital simulation, comprising the following steps: S1, obtaining multi-source associated data and preprocessing to form a basic sample set; S2, performing combined configuration, segment fitting and response parameter extraction to obtain thermal response parameters; S3, performing multi-working condition response deduction, calculating residual and iteratively calibrating model parameters to generate a thermal performance data set; S4, performing multi-variable kernel density estimation, extracting working condition boundary threshold and performing region marking to construct a conventional prediction sample domain and a limit extrapolation sample domain; S5, matching the threshold and calling corresponding models to perform thermal response calculation to generate thermal performance prediction and limit working condition extrapolation results; and S6, obtaining verification response data and performing incremental updating. The application can realize rapid prediction and performance deduction of diesel engine thermal performance and improve the accuracy of diesel engine performance evaluation results.

Description

Technical Field

[0001] The present invention relates to the technical field of digital simulation, and particularly to a method for predicting the thermal performance of a diesel engine based on digital simulation and extrapolating extreme operating conditions. Background Art

[0002] Predicting the thermal performance of a diesel engine is an important part in the design, development, and operation state evaluation of a diesel engine. Existing technologies usually calculate the power performance, fuel economy, and heat load response of a diesel engine under typical rotational speeds, loads, and environmental conditions based on a thermodynamic simulation model, and verify the design scheme in combination with the results of bench tests. This method can analyze the performance under established operating conditions, but in the process of multi-scheme design iteration, it often requires repeated modeling, calculation, and test verification, with a long analysis cycle and difficulty in meeting the rapid R & D requirements.

[0003] In recent years, establishing a rapid prediction model using digital simulation data has become an important direction for diesel engine performance analysis. However, existing methods usually use a limited number of conventional operating condition data as the modeling basis, with insufficient consideration of the response differences in different operating condition ranges, and it is difficult to simultaneously take into account the prediction accuracy under normal operating conditions and the extrapolation ability under extreme environments. Especially under boundary conditions such as low temperature, high altitude, or performance degradation, the actual number of available samples is small. Directly using a conventional prediction model for extended calculation is prone to amplification of prediction errors and difficult to reliably reflect the performance changes of a diesel engine under extreme operating conditions.

[0004] In addition, existing prediction models usually rely on initial simulation data or stage test data to complete calibration, with insufficient utilization of the performance change data obtained subsequently. When the operating conditions change or the performance state migrates, it is difficult for the model to timely correct the prediction deviation, affecting its applicability in full-cycle performance evaluation and boundary operation analysis.

[0005] Therefore, how to provide a method for predicting the thermal performance of a diesel engine based on digital simulation and extrapolating extreme operating conditions is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to propose a method for predicting the thermal performance of a diesel engine based on digital simulation and extrapolating extreme operating conditions. The present invention makes full use of thermal response digital simulation, kernel density boundary modeling, Gaussian process regression prediction, and generalized Pareto extrapolation techniques, and details the processing methods from simulation response parameter extraction, multi-operating condition performance data generation, conventional operating condition prediction and extreme operating condition extrapolation to verification deviation increment update, with the advantages of high prediction accuracy, strong extreme operating condition extrapolation ability, and good full-cycle adaptability.

[0007] The method for predicting the thermal performance of a diesel engine based on digital simulation and extrapolating extreme operating conditions according to an embodiment of the present invention includes the following steps: S1. Obtain multi-source correlation data on the thermodynamic performance of diesel engines and perform preprocessing to form a basic sample set; S2. Combine and configure the simulation input parameters in the basic sample set, perform digital simulation calculations, and perform piecewise fitting and response parameter extraction on the obtained simulation response sequence to obtain the thermal response parameters. S3. Embed the thermal response parameters into the thermal simulation model, configure the simulation boundary according to the operating condition parameters, perform multi-condition response deduction, calculate the residual between the deduced response and the verification response, iteratively calibrate the model parameters, and generate a thermal performance dataset. S4. Perform multivariate kernel density estimation on the operating parameters in the thermal performance dataset, extract the operating boundary thresholds based on the density estimation results, and label the corresponding samples to construct the conventional prediction sample domain and the extreme extrapolation sample domain. S5. Train a Gaussian process regression model based on the conventional prediction sample domain, train a generalized Pareto model based on the extreme extrapolation sample domain, match the parameters of the operating condition to be predicted with the operating condition boundary threshold, and call the corresponding model to perform thermal response calculation to generate thermal performance prediction and extreme operating condition extrapolation results. S6. Obtain the verification response data of the thermal performance prediction and extreme condition extrapolation results, and use the maximum likelihood estimation method to incrementally update the Gaussian process regression model and the generalized Pareto model.

[0008] Optionally, the preprocessing specifically includes: Simulation input parameters, operating condition parameters, and verification responses are extracted from multi-source correlated data. Null values ​​and outliers exceeding the valid value range of each parameter are detected in the corresponding data records, and the data records corresponding to the outliers are deleted. The simulation input parameters include configuration characterization parameters, control configuration parameters, and model configuration parameters; The operating parameters include operating status parameters and environmental boundary parameters; The verification response includes thermal response verification values ​​and performance response verification values ​​associated with the corresponding operating conditions; Based on the operating state change range and boundary value range corresponding to the operating condition parameters, the retained data records are labeled with operating condition types to obtain steady-state operating condition samples, transient operating condition samples and extreme operating condition samples; The simulation input parameters and operating condition parameters are set as input features, the verification response is set as output features, and the sample correspondence between input features and output features is established according to the operating condition type corresponding to each data record. The Z-score normalization method is used to scale the input and output features, and the scaled input and output features and corresponding working condition type labels are combined to form a basic sample set.

[0009] Optionally, S2 specifically includes: S21. Extract simulation input parameters from the basic sample set, and combine and sample the simulation input parameters according to the preset variation range and the Latin hypercube sampling method to form a simulation parameter combination. S22. Load each combination of simulation parameters into the digital simulation calculation process in sequence, perform response solving based on the corresponding combination of simulation parameters, and extract continuous response data from the solution results to form the simulation response sequence corresponding to each combination of simulation parameters. S23. The simulated response sequence is piecewise fitted using a three-Wiebe model. The fitting evaluation value between the fitted response and the simulated response sequence is calculated, and the piecewise fitting characteristics corresponding to each response segment are solved based on the fitting evaluation value. The piecewise fitting features include proportion, duration interval and shape. The three Wiebe models are response fitting models composed of three Wiebe sub-functions superimposed. Each Wiebe sub-function corresponds to a response segment, and the fitted response is calculated based on the piecewise fitting features of the corresponding response segment. S24. Extract the piecewise fitting features that have reached the preset fitting threshold, and record the piecewise fitting features according to the corresponding simulation parameter combinations to obtain the thermal response parameters.

[0010] Optionally, S23 specifically includes: S231. For each simulation response sequence, configure the initial segment fitting features corresponding to each response segment in the three Wiebe models according to the three response segments. S232. Based on the initial segment fitting features corresponding to each response segment, calculate the segment fitting value at each sequence position, and superimpose the segment fitting values ​​of the three response segments to form a fitted response sequence. S233. Based on the fitted response sequence and the corresponding simulated response sequence, match them according to the same sequence position, calculate the determination coefficient of the fitted response sequence relative to the simulated response sequence, and obtain the fitting evaluation quantity. S234. Based on the fitting evaluation value, iteratively adjust the segmented fitting features corresponding to each response segment, and update the fitted response sequence based on the adjusted segmented fitting features until the fitting evaluation value reaches the preset fitting threshold to obtain the segmented fitting result.

[0011] Optionally, S3 specifically includes: S31. Load the thermal response parameters into the thermal simulation model, and configure the simulation boundaries corresponding to each working condition according to the working condition parameters in the basic sample set. S32. Perform multi-condition response deduction based on the configured simulation boundary, match the deduced response obtained under each condition with the corresponding verification response, and calculate the residual between the corresponding response items. S33. Based on the comparison results between the residuals corresponding to each response item and the preset calibration threshold, screen out the response items that do not meet the calibration threshold, and determine the model parameters to be adjusted corresponding to the response items that do not meet the calibration threshold. S34. Iteratively adjust the parameters of the model to be adjusted, and re-execute the response deduction of the corresponding working condition based on the adjusted thermodynamic simulation model until the residuals corresponding to each response item meet the preset calibration threshold. Combine the calibrated simulation input parameters, working condition parameters and deduction response to generate a thermodynamic performance dataset.

[0012] Optionally, the thermal simulation model is a parameterized model of the entire machine constructed based on a one-dimensional thermodynamic calculation framework, wherein: The one-dimensional thermodynamic calculation framework is a simulation calculation framework that configures parameters and solves the correlation of each calculation object according to the calculation order of operating boundary input, thermodynamic response loading, state transfer calculation and performance response output. The whole machine parameterization model loads the thermal response parameters into the thermal response calculation process, loads the operating condition parameters into the operating boundary input process, and configures the model parameters corresponding to the state transmission calculation process based on the simulation input parameters in the basic sample set. When performing multi-condition response simulation, the calculation boundary corresponding to each condition is determined based on the operation boundary input process. The response change under the corresponding condition is obtained by using the thermodynamic response calculation process. The response change is then passed to the performance response output process step by step according to the state transfer calculation process to obtain the simulation response corresponding to each condition. The deduced response is the thermodynamic performance calculation result that corresponds to the verification response. It is used to calculate the residual between the deduced response and the verification response, and to iteratively calibrate the model parameters in the whole machine parameterized model based on the residual.

[0013] Optionally, S4 specifically includes: S41. Extract the standardized operating condition parameters corresponding to each sample from the thermal performance dataset, and combine the standardized operating condition parameters corresponding to the same sample according to the parameter dimension to form a set of operating condition parameter vectors. S42. Calculate the sample dispersion corresponding to each parameter dimension in the working condition parameter vector set. Use the Silverman bandwidth criterion to determine the kernel bandwidth corresponding to each parameter dimension, and construct the diagonal bandwidth matrix corresponding to the multivariate kernel density estimation based on each kernel bandwidth, where: The Silverman bandwidth criterion is as follows: based on the standard deviation of each parameter dimension, the number of samples in the operating condition parameter vector set, and the number of parameter dimensions, the bandwidth determination method for calculating the kernel bandwidth corresponding to each parameter dimension is used. S43. Based on the diagonal bandwidth matrix, the Gaussian kernel function is used to perform multivariate kernel density estimation on the set of working condition parameter vectors, calculate the distribution density value corresponding to each working condition parameter vector, and form a working condition distribution density sequence based on the distribution density value. S44. Construct a load condition classifier based on the load condition distribution density sequence, and use the load condition classifier to label the regions of samples in the thermal performance dataset, wherein: The working condition distribution density sequence is a density value sequence obtained by arranging the distribution density values ​​corresponding to each working condition parameter vector according to their numerical values. The working condition boundary threshold is: extracting a preset quantile on the low-density side of the working condition distribution density sequence, and setting the distribution density value corresponding to the preset quantile as the sample region division threshold; The working condition classifier is a region labeling rule that uses the comparison result between the distribution density value corresponding to each sample and the working condition boundary threshold as the classification condition. When the distribution density value corresponding to a sample is not lower than the working condition boundary threshold, the corresponding sample is marked as a regular prediction sample and included in the regular prediction sample domain. When the distribution density value corresponding to a sample is lower than the working condition boundary threshold, the corresponding sample is marked as a boundary working condition sample and included in the limit extrapolation sample domain.

[0014] Optionally, S42 specifically includes: S421. Split the working condition parameter vector set according to the parameter dimension, extract the standardized parameter values ​​of each parameter dimension in all samples, and form the parameter value sequence corresponding to each parameter dimension. S422. Based on the parameter value sequence corresponding to each parameter dimension, calculate the standard deviation of each parameter dimension, and obtain the number of samples and the number of parameter dimensions in the working condition parameter vector set. S423. Using the Silverman bandwidth criterion, calculate the kernel bandwidth corresponding to each parameter dimension based on the dimensional standard deviation, sample size, and number of parameter dimensions. S423. According to the order of each parameter dimension in the working condition parameter vector, set the diagonal terms of the matrix based on the kernel bandwidth corresponding to each parameter dimension, and set the off-diagonal terms in the matrix to zero to obtain the diagonal bandwidth matrix used for multivariate kernel density estimation.

[0015] Optionally, S5 specifically includes: S51. Extract operating parameters and corresponding thermodynamic response terms to be predicted from the conventional prediction sample domain. Use the operating parameters as training input and the thermodynamic response terms to be predicted as training output. Use the Matern kernel function to construct the corresponding Gaussian process regression model and use maximum likelihood estimation to determine the kernel function parameters of the Gaussian process regression model. S52. Extract the thermodynamic response terms to be extrapolated from the extreme extrapolation sample domain, determine the response threshold corresponding to each thermodynamic response term to be extrapolated, calculate the excess between the response value exceeding the response threshold and the corresponding response threshold, and based on the excess, use maximum likelihood estimation to determine the scale parameter and shape parameter of the corresponding generalized Pareto model. S53. Input the working condition parameters to be predicted into the working condition classifier constructed based on the working condition boundary threshold, and determine whether the working condition parameters to be predicted correspond to the conventional prediction sample domain or the extreme extrapolation sample domain according to the region label output by the working condition classifier. S54. Based on the sample domain corresponding to the operating condition parameters to be predicted, call the corresponding model to perform thermal response calculation: When the operating condition parameter to be predicted corresponds to the conventional prediction sample domain, the corresponding Gaussian process regression model is called to calculate the thermodynamic response term to be predicted and generate the thermodynamic performance prediction result. When the operating condition parameter to be predicted corresponds to the extreme extrapolation sample domain, the corresponding generalized Pareto model is called to calculate the thermodynamic response term to be extrapolated based on the preset extrapolation probability level, and the extreme operating condition extrapolation result is generated.

[0016] Optionally, S6 specifically includes: S61. Obtain the verification response data, and pair the verification response data with the thermal performance prediction results or the extreme condition extrapolation results according to the corresponding operating condition parameters and thermal response items to form a response verification record. S62. In each response verification record, calculate the response deviation between the verification response data and the corresponding result, mark the response verification record with the response deviation exceeding the preset deviation threshold as a deviation record, and extract the operating condition parameters and verification response data from the deviation record to form a deviation sample. S63. Based on the result type corresponding to the deviation sample, the deviation sample corresponding to the thermal performance prediction result is classified into the regular prediction deviation sample set, and the deviation sample corresponding to the extreme condition extrapolation result is classified into the extreme extrapolation deviation sample set. S64. Add the standard prediction bias sample set to the training samples of the Gaussian process regression model, and update the kernel function parameters of the Gaussian process regression model based on the added training samples. Add the limit extrapolation bias sample set to the training samples of the generalized Pareto model, and update the scale parameters and shape parameters of the generalized Pareto model based on the added training samples.

[0017] The beneficial effects of this invention are: First, this invention establishes a data foundation for multi-condition thermodynamic performance analysis through digital simulation and response parameter extraction, and uses verification responses to iteratively calibrate the simulation model, thereby improving the accuracy of performance extrapolation results.

[0018] Secondly, this invention enhances the performance extrapolation capability under extreme working conditions by modeling the working condition boundary and entrusting conventional prediction and extreme extrapolation to Gaussian process regression model and generalized Pareto model respectively.

[0019] Finally, this invention extracts deviation samples based on verification response data and updates the prediction model, enabling the model to adapt to new data and state changes, thus meeting the requirements for full-cycle thermodynamic performance evaluation of diesel engines. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of the diesel engine thermodynamic performance prediction and extreme condition extrapolation method based on digital simulation proposed in this invention. Figure 2 This is a flowchart of the multi-condition response deduction and model iteration calibration of the diesel engine thermodynamic performance prediction and extreme condition extrapolation method based on digital simulation proposed in this invention. Figure 3 This is a flowchart of the verification response analysis and model incremental update process for the diesel engine thermodynamic performance prediction and extreme condition extrapolation method based on digital simulation proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] like Figure 1 As shown, the method for predicting the thermodynamic performance of a diesel engine and extrapolating its extreme operating conditions based on digital simulation includes the following steps: S1. Obtain multi-source correlation data on the thermodynamic performance of diesel engines and perform preprocessing to form a basic sample set; S2. Combine and configure the simulation input parameters in the basic sample set, perform digital simulation calculations, and perform piecewise fitting and response parameter extraction on the obtained simulation response sequence to obtain the thermal response parameters. S3. Embed the thermal response parameters into the thermal simulation model, configure the simulation boundary according to the operating condition parameters, perform multi-condition response deduction, calculate the residual between the deduced response and the verification response, iteratively calibrate the model parameters, and generate a thermal performance dataset. S4. Perform multivariate kernel density estimation on the operating parameters in the thermal performance dataset, extract the operating boundary thresholds based on the density estimation results, and label the corresponding samples to construct the conventional prediction sample domain and the extreme extrapolation sample domain. S5. Train a Gaussian process regression model based on the conventional prediction sample domain, train a generalized Pareto model based on the extreme extrapolation sample domain, match the parameters of the operating condition to be predicted with the operating condition boundary threshold, and call the corresponding model to perform thermal response calculation to generate thermal performance prediction and extreme operating condition extrapolation results. S6. Obtain the verification response data of the thermal performance prediction and extreme condition extrapolation results, and use the maximum likelihood estimation method to incrementally update the Gaussian process regression model and the generalized Pareto model.

[0023] In this embodiment, the preprocessing specifically includes: Simulation input parameters, operating condition parameters, and verification responses are extracted from multi-source correlated data. Null values ​​and outliers exceeding the valid value range of each parameter are detected in the corresponding data records, and the data records corresponding to the outliers are deleted. Simulation input parameters include configuration characterization parameters, control configuration parameters, and model configuration parameters; Operating parameters include operating status parameters and environmental boundary parameters; The verification response includes thermodynamic response verification values ​​and performance response verification values ​​associated with the corresponding operating conditions; Based on the operating state change range and boundary value range corresponding to the operating condition parameters, the retained data records are labeled with operating condition types to obtain steady-state operating condition samples, transient operating condition samples and extreme operating condition samples; The simulation input parameters and operating condition parameters are set as input features, the verification response is set as output features, and the sample correspondence between input features and output features is established according to the operating condition type corresponding to each data record. The Z-score normalization method is used to scale the input and output features, and the scaled input and output features and corresponding working condition type labels are combined to form a basic sample set.

[0024] In this embodiment, S2 specifically includes: S21. Extract simulation input parameters from the basic sample set, and combine and sample the simulation input parameters according to the preset variation range and the Latin hypercube sampling method to form a simulation parameter combination. S22. Load each combination of simulation parameters into the digital simulation calculation process in sequence, perform response solving based on the corresponding combination of simulation parameters, and extract continuous response data from the solution results to form the simulation response sequence corresponding to each combination of simulation parameters. S23. The simulated response sequence is piecewise fitted using a three-Wiebe model. The fitting evaluation value between the fitted response and the simulated response sequence is calculated, and the piecewise fitting characteristics corresponding to each response segment are solved based on the fitting evaluation value. Piecewise fitting features include proportion, duration interval and shape. The three-Wiebe model is a response fitting model composed of three Wiebe sub-functions superimposed. Each Wiebe sub-function corresponds to a response segment, and the fitted response is calculated based on the piecewise fitting features of the corresponding response segment. S24. Extract the piecewise fitting features that have reached the preset fitting threshold, and record the piecewise fitting features according to the corresponding combination of simulation parameters to obtain the thermal response parameters. In this embodiment, the simulation input parameters corresponding to the basic configuration are used as the benchmark. The parameter sampling interval is set within a range of 10% above and below the benchmark value of each parameter, and the Latin hypercube sampling method is used to generate the simulation parameter combination within each parameter sampling interval. Each combination of simulation parameters is loaded into the digital simulation calculation process one by one. The maximum number of iterations is set to 50 and the convergence judgment value is 0.01 to obtain the continuous simulation response sequence corresponding to each combination of simulation parameters. The simulation response sequence is compared with the verification response under the same configuration. When the deviation of the peak response, peak response position or response duration interval is greater than 5%, the corresponding simulation input parameters are adjusted and the digital simulation calculation is re-executed. For the simulation response sequence that meets the deviation requirements, a three-Wiebe model is used for piecewise fitting. The proportion, duration, and shape of the corresponding response segments are solved by three Wiebe sub-functions, and the fitted responses of the three response segments are superimposed to form the overall fitted response. Calculate the coefficient of determination of the overall fitted response relative to the simulation response sequence. When the coefficient of determination reaches the preset fitting threshold, associate and record the piecewise fitting features obtained with the corresponding simulation parameter combinations to form the thermodynamic response parameters.

[0025] In this embodiment, S23 specifically includes: S231. For each simulation response sequence, configure the initial segment fitting features corresponding to each response segment in the three Wiebe models according to the three response segments. S232. Based on the initial segment fitting features corresponding to each response segment, calculate the segment fitting value at each sequence position, and superimpose the segment fitting values ​​of the three response segments to form a fitted response sequence. S233. Based on the fitted response sequence and the corresponding simulated response sequence, match them according to the same sequence position, calculate the determination coefficient of the fitted response sequence relative to the simulated response sequence, and obtain the fitting evaluation quantity. S234. Based on the fitting evaluation value, iteratively adjust the segmented fitting features corresponding to each response segment, and update the fitted response sequence based on the adjusted segmented fitting features until the fitting evaluation value reaches the preset fitting threshold to obtain the segmented fitting result. In this embodiment, three response segments are configured for each simulation response sequence. The initial proportions of the three response segments are set to be equal to 1, and the corresponding initial duration intervals and initial shapes are configured respectively. Based on the initial segment fitting features of each response segment, the segment fitting values ​​corresponding to each sequence position are calculated using three Wiebe sub-functions, and the three segment fitting values ​​are superimposed to form the fitted response sequence. According to the same sequence position, the fitted response sequence and the simulated response sequence are matched point by point, the determination coefficient between the two is calculated, and the determination coefficient is used as the fitting evaluation quantity. When the fit evaluation value is lower than 0.95, the proportion, duration interval and shape of the three response segments are iteratively adjusted, and the fitted response sequence and corresponding determination coefficients are recalculated until the fit evaluation value is not lower than 0.95. The proportion, duration, and shape corresponding to the current three response segments are used as the segmented fitting results to form the thermal response parameters.

[0026] like Figure 2 As shown, in this embodiment, S3 specifically includes: S31. Load the thermal response parameters into the thermal simulation model, and configure the simulation boundaries corresponding to each working condition according to the working condition parameters in the basic sample set. S32. Perform multi-condition response deduction based on the configured simulation boundary, match the deduced response obtained under each condition with the corresponding verification response, and calculate the residual between the corresponding response items. S33. Based on the comparison results between the residuals corresponding to each response item and the preset calibration threshold, screen out the response items that do not meet the calibration threshold, and determine the model parameters to be adjusted corresponding to the response items that do not meet the calibration threshold. S34. Iteratively adjust the parameters of the model to be adjusted, and re-execute the response deduction of the corresponding working condition based on the adjusted thermodynamic simulation model until the residuals corresponding to each response item meet the preset calibration threshold. Combine the calibrated simulation input parameters, working condition parameters and deduction response to generate a thermodynamic performance dataset.

[0027] In this embodiment, the thermal simulation model is a parameterized model of the entire machine constructed based on a one-dimensional thermodynamic calculation framework, wherein: The one-dimensional thermodynamic calculation framework is a simulation calculation framework that configures parameters and solves the correlation of each calculation object according to the calculation order of operating boundary input, thermodynamic response loading, state transfer calculation and performance response output. The whole-machine parameterized model loads the thermodynamic response parameters into the thermodynamic response calculation process, loads the operating condition parameters into the operating boundary input process, and configures the model parameters corresponding to the state transmission calculation process based on the simulation input parameters in the basic sample set. When performing multi-condition response simulation, the calculation boundary corresponding to each condition is determined based on the operation boundary input process. The response change under the corresponding condition is obtained by using the thermodynamic response calculation process. The response change is then passed to the performance response output process step by step according to the state transfer calculation process to obtain the simulation response corresponding to each condition. The deduced response is the thermodynamic performance calculation result that corresponds to the verification response. It is used to calculate the residual between the deduced response and the verification response, and to iteratively calibrate the model parameters in the parameterized model of the whole machine based on the residual.

[0028] In this embodiment, S4 specifically includes: S41. Extract the standardized operating condition parameters corresponding to each sample from the thermal performance dataset, and combine the standardized operating condition parameters corresponding to the same sample according to the parameter dimension to form a set of operating condition parameter vectors. S42. Calculate the sample dispersion corresponding to each parameter dimension in the working condition parameter vector set. Use the Silverman bandwidth criterion to determine the kernel bandwidth corresponding to each parameter dimension, and construct the diagonal bandwidth matrix corresponding to the multivariate kernel density estimation based on each kernel bandwidth, where: The Silverman bandwidth criterion is: based on the standard deviation of each parameter dimension, the number of samples in the operating parameter vector set, and the number of parameter dimensions, the bandwidth determination method for calculating the kernel bandwidth corresponding to each parameter dimension is used. S43. Based on the diagonal bandwidth matrix, the Gaussian kernel function is used to perform multivariate kernel density estimation on the set of working condition parameter vectors, calculate the distribution density value corresponding to each working condition parameter vector, and form a working condition distribution density sequence based on the distribution density value. S44. Construct a load condition classifier based on the load condition distribution density sequence, and use the load condition classifier to label the regions of samples in the thermal performance dataset, where: The working condition distribution density sequence is: a sequence of density values ​​obtained by arranging the distribution density values ​​corresponding to each working condition parameter vector according to their numerical values; The working condition boundary threshold is: extract a preset quantile on the low-density side of the working condition distribution density sequence, and set the distribution density value corresponding to the preset quantile as the sample region division threshold; The working condition classifier is: the comparison result between the distribution density value corresponding to each sample and the working condition boundary threshold is used as the region labeling rule for classification conditions; When the distribution density value corresponding to a sample is not lower than the working condition boundary threshold, the corresponding sample is marked as a regular prediction sample and included in the regular prediction sample domain. When the distribution density value corresponding to a sample is lower than the working condition boundary threshold, the corresponding sample is marked as a boundary working condition sample and included in the limit extrapolation sample domain.

[0029] In this embodiment, S42 specifically includes: S421. Split the working condition parameter vector set according to the parameter dimension, extract the standardized parameter values ​​of each parameter dimension in all samples, and form the parameter value sequence corresponding to each parameter dimension. S422. Based on the parameter value sequence corresponding to each parameter dimension, calculate the standard deviation of each parameter dimension, and obtain the number of samples and the number of parameter dimensions in the working condition parameter vector set. S423. Using the Silverman bandwidth criterion, calculate the kernel bandwidth corresponding to each parameter dimension based on the dimensional standard deviation, sample size, and number of parameter dimensions. S423. According to the order of each parameter dimension in the working condition parameter vector, set the diagonal terms of the matrix based on the kernel bandwidth corresponding to each parameter dimension, and set the off-diagonal terms in the matrix to zero to obtain the diagonal bandwidth matrix used for multivariate kernel density estimation.

[0030] In this embodiment, S5 specifically includes: S51. Extract operating condition parameters and corresponding thermodynamic response terms to be predicted from the conventional prediction sample domain. Use the operating condition parameters as training input and the thermodynamic response terms to be predicted as training output. Construct the corresponding Gaussian process regression model using the Matern kernel function, and determine the kernel function parameters of the Gaussian process regression model using maximum likelihood estimation. Specifically, this includes: According to the thermal response term to be predicted, corresponding samples are extracted from the conventional prediction sample domain. The operating parameters in each sample are combined into a training input vector, and the thermal response terms to be predicted corresponding to the same sample are combined into a training output sequence. The similarity between each training input vector is calculated using the Matern kernel function. The kernel matrix corresponding to the training sample is constructed based on the similarity between each training input vector. The kernel function parameters of the Matern kernel function include smoothness parameter and length scale parameter. Based on the kernel matrix and the training output sequence, the training distribution of the Gaussian process regression model is constructed, and noise parameters for characterizing training output perturbations are configured in the kernel matrix to obtain the Gaussian process regression model corresponding to the thermodynamic response term to be predicted. The log marginal likelihood of the training output sequence under the training distribution is used as the optimization objective. The smoothness parameter, length scale parameter and noise parameter are solved iteratively. The parameters corresponding to the optimization objective when the preset conditions are met are determined as the kernel function parameters of the Gaussian process regression model. S52. Extract the thermodynamic response terms to be extrapolated from the extreme extrapolation sample domain, determine the response thresholds corresponding to each thermodynamic response term to be extrapolated, calculate the exceedance between the response values ​​exceeding the response thresholds and the corresponding response thresholds, and based on the exceedance, use maximum likelihood estimation to determine the scale parameters and shape parameters of the corresponding generalized Pareto model, specifically including: According to the thermodynamic response term to be extrapolated, the corresponding response value is extracted from the extreme extrapolation sample domain, and extreme response sequences are formed according to the thermodynamic response term to which each response value belongs. The response values ​​in each limiting response sequence are statistically analyzed by quantiles, and the response values ​​corresponding to the preset high quantiles are extracted as the response thresholds of the corresponding thermodynamic response terms to be extrapolated. In each limiting response sequence, the response value that exceeds the corresponding response threshold is marked as the overthreshold response value. The difference between each overthreshold response value and the corresponding response threshold is calculated to form the overthreshold sequence corresponding to each thermodynamic response term to be extrapolated. Using the sequence of each transcendent quantity as the training data for the corresponding generalized Pareto model, we construct the likelihood functions corresponding to the scale parameter and shape parameter, and use maximum likelihood estimation to solve for the scale parameter and shape parameter to obtain the generalized Pareto model corresponding to each thermodynamic response term to be extrapolated. S53. Input the working condition parameters to be predicted into the working condition classifier constructed based on the working condition boundary threshold, and determine whether the working condition parameters to be predicted correspond to the conventional prediction sample domain or the extreme extrapolation sample domain according to the region label output by the working condition classifier. S54. Based on the sample domain corresponding to the parameters of the operating condition to be predicted, call the corresponding model to perform thermal response calculation: When the operating condition parameter to be predicted corresponds to the conventional prediction sample domain, the corresponding Gaussian process regression model is called to calculate the thermodynamic response term to be predicted and generate the thermodynamic performance prediction result. When the operating condition parameter to be predicted corresponds to the extreme extrapolation sample domain, the corresponding generalized Pareto model is called to calculate the thermodynamic response term to be extrapolated based on the preset extrapolation probability level, and the extreme operating condition extrapolation result is generated.

[0031] like Figure 3 As shown, in this embodiment, S6 specifically includes: S61. Obtain the verification response data, and pair the verification response data with the thermal performance prediction results or the extreme condition extrapolation results according to the corresponding operating condition parameters and thermal response items to form a response verification record. S62. In each response verification record, calculate the response deviation between the verification response data and the corresponding result, mark the response verification record with the response deviation exceeding the preset deviation threshold as a deviation record, and extract the operating condition parameters and verification response data from the deviation record to form a deviation sample. S63. Based on the result type corresponding to the deviation sample, the deviation sample corresponding to the thermal performance prediction result is classified into the regular prediction deviation sample set, and the deviation sample corresponding to the extreme condition extrapolation result is classified into the extreme extrapolation deviation sample set. S64. Add the regular prediction bias sample set to the training samples of the Gaussian process regression model, and update the kernel function parameters of the Gaussian process regression model based on the added training samples. Add the limit extrapolation bias sample set to the training samples of the generalized Pareto model, and update the scale parameters and shape parameters of the generalized Pareto model based on the added training samples. In this embodiment, verification response data is obtained according to a continuous 500-hour mixed working condition verification process, and verification response data is extracted every 50 hours and matched with the corresponding model results. During the extreme condition verification phase, equivalent altitude conditions of 3000m and 4500m and low temperature conditions of -20℃ and -30℃ were set respectively to obtain the verification response data corresponding to the extreme condition extrapolation results. For paired response verification records, deviation thresholds are set according to the type of thermal response item: for sequential thermal response items, when the correlation coefficient between the verification response data and the corresponding result is less than 0.95, it is marked as a deviation record; For proportional thermodynamic response terms, when the relative deviation exceeds 2%, it is marked as a deviation record; For temperature-type thermodynamic response items, when the absolute deviation exceeds 15℃, it is marked as a deviation record; The deviation samples corresponding to the thermal performance prediction results are added to the training samples of the Gaussian process regression model, and its kernel function parameters are updated. The deviation samples corresponding to the extrapolation results of the extreme conditions are added to the training samples of the generalized Pareto model, and its scale parameters and shape parameters are updated to form a prediction model that has completed incremental updates.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to the multi-condition thermodynamic performance prediction and extreme condition extrapolation scenario of a certain type of diesel engine. This diesel engine needs to have its performance changes analyzed under normal operation, equivalent low-pressure operation, low-temperature operation, and after continuous operation. Existing methods typically establish a prediction model based on normal operating condition data and then directly extend the prediction results to boundary conditions. This can easily lead to increased extrapolation deviations after a decrease in air pressure or temperature, and furthermore, after continuous operation causes changes in performance status, the original model struggles to adapt to the newly added response data in a timely manner.

[0033] In this scenario, simulation input parameters, operating parameters, and verification response data related to thermodynamic performance are acquired. Data records with null values ​​or values ​​outside the valid range are removed and scaled to form a basic sample set. A parameter sampling interval is set within a 10% range above and below the basic configuration. The Latin hypercube sampling method is used to generate simulation parameter combinations, and digital simulation calculations are performed. The obtained continuous simulation response sequence is piecewise fitted using a three-Wiebe model. When the coefficient of determination between the fitted response and the simulation response is not less than 0.95, the corresponding piecewise fitted features are recorded as thermodynamic response parameters.

[0034] Thermodynamic response parameters are loaded into the thermodynamic simulation model. Simulation boundaries are configured based on operating condition parameters, and multi-condition response extrapolation is performed. The model parameters are iteratively calibrated based on the residuals between the verification response and the extrapolated response, forming a thermodynamic performance dataset. Multivariate kernel density estimation is performed on the standard operating condition parameters. The bandwidth matrix is ​​configured using the Silverman bandwidth criterion. The 5% of samples from the low-density side are assigned to the extreme extrapolation sample domain, and the remaining 95% are assigned to the conventional prediction sample domain. A Gaussian process regression model is constructed for the conventional prediction sample domain, and a generalized Pareto model is constructed for the extreme extrapolation sample domain, with the extrapolation probability level set to 0.99.

[0035] During the verification process, normal rated conditions, equivalent 3000m and 4500m boundary conditions, and -20℃ and -30℃ low-temperature conditions were set, and a mixed-condition verification was performed continuously for 500 hours, with verification response data extracted every 50 hours. Before the model update, the extrapolated power reduction rate under the equivalent 4500m condition differed from the verification response by 3.70 percentage points, and the temperature-related response error under the -30℃ condition was 21.6℃. After adding the deviation samples to the corresponding model for incremental updates, the deviation of the power reduction rate under the equivalent 4500m condition decreased to 0.50 percentage points, and the temperature-related response error under the -30℃ condition decreased to 8.4℃. After the continuous 500-hour verification, the prediction deviations of the power reduction rate and fuel consumption change rate were 0.19 percentage points and 0.21 percentage points, respectively, indicating that the present invention can improve the adaptability and accuracy of normal operating condition prediction and extreme operating condition extrapolation.

[0036] Table 1. Results of Multi-condition Thermodynamic Performance Prediction and Extreme Condition Extrapolation Verification

[0037] As can be seen from the data in Table 1, the present invention can complete the prediction of thermodynamic performance and the extrapolation of extreme conditions based on conventional operating condition samples and boundary operating condition samples, respectively. By verifying the response data, the model is incrementally corrected, so that the updated prediction results and extrapolation results are closer to the actual response. This shows that the present invention has good multi-condition prediction capability, extreme state extrapolation capability, and model adaptability during continuous operation.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the thermodynamic performance of diesel engines and extrapolating extreme operating conditions based on digital simulation, characterized in that, Includes the following steps: S1. Obtain multi-source correlation data of diesel engine thermodynamic performance and perform preprocessing to form a basic sample set. The preprocessing includes: extracting simulation input parameters, operating condition parameters and verification responses from the multi-source correlation data. The simulation input parameters include configuration characterization parameters, control configuration parameters and model configuration parameters. The operating condition parameters include operating state parameters and environmental boundary parameters. The verification response includes thermodynamic response verification values ​​and performance response verification values ​​associated with the corresponding operating conditions. S2. Extract simulation input parameters from the basic sample set, combine and sample the simulation input parameters according to the preset variation range and Latin hypercube sampling method to form a simulation parameter combination, and perform digital simulation calculation. For each simulation response sequence, configure the initial segment fitting features corresponding to each response segment in the three Wiebe models according to the three response segments. The segment fitting features include proportion, duration interval and shape. Each Wiebe subfunction corresponds to a response segment. Based on the initial segment fitting features corresponding to each response segment, the segment fitting value at each sequence position is calculated, and the segment fitting values ​​of the three response segments are superimposed to form the fitted response sequence. Based on the fitted response sequence and the corresponding simulated response sequence, the matching is performed according to the positions of the same sequence, and the determination coefficient of the fitted response sequence relative to the simulated response sequence is calculated to obtain the fitting evaluation quantity. Based on the fitting evaluation value, the piecewise fitting features corresponding to each response segment are iteratively adjusted, and the fitted response sequence is updated based on the adjusted piecewise fitting features until the fitting evaluation value reaches the preset fitting threshold, thus obtaining the piecewise fitting result. Extract the piecewise fitting features that reach the preset fitting threshold, and associate and record the piecewise fitting features according to the corresponding combination of simulation parameters to obtain the thermodynamic response parameters. S3. Embed the thermodynamic response parameters into the thermodynamic simulation model, configure the simulation boundary according to the operating condition parameters, and perform multi-operating condition response deduction. Calculate the residual between the deduced response and the verification response, and iteratively calibrate the model parameters to generate a thermodynamic performance dataset. Specifically, S3 includes: S31. Load the thermal response parameters into the thermal simulation model, and configure the simulation boundaries corresponding to each working condition according to the working condition parameters in the basic sample set. S32. Perform multi-condition response deduction based on the configured simulation boundary, match the deduced response obtained under each condition with the corresponding verification response, and calculate the residual between the corresponding response items. S33. Based on the comparison results between the residuals corresponding to each response item and the preset calibration threshold, filter out the response items that do not meet the calibration threshold, and determine the model parameters to be adjusted corresponding to the response items that do not meet the calibration threshold. S34. Iteratively adjust the parameters of the model to be adjusted, and re-execute the response deduction of the corresponding working condition based on the adjusted thermodynamic simulation model until the residuals corresponding to each response item meet the preset calibration threshold. Combine the calibrated simulation input parameters, working condition parameters and deduction response to generate a thermodynamic performance dataset. S4. Perform multivariate kernel density estimation on the operating parameters in the thermal performance dataset, extract the operating condition boundary threshold based on the density estimation results. The operating condition boundary threshold is to extract a preset quantile on the low density side of the operating condition distribution density sequence, and set the distribution density value corresponding to the preset quantile as the sample region division threshold. Construct an operating condition classifier. The operating condition classifier is to use the comparison result between the distribution density value corresponding to each sample and the operating condition boundary threshold as the region labeling rule for classification conditions. When the distribution density value corresponding to a sample is not lower than the working condition boundary threshold, the corresponding sample is marked as a regular prediction sample and included in the regular prediction sample domain. When the distribution density value corresponding to a sample is lower than the working condition boundary threshold, the corresponding sample is marked as a boundary working condition sample and included in the limit extrapolation sample domain. S5. Train a Gaussian process regression model based on the conventional prediction sample domain, train a generalized Pareto model based on the extreme extrapolation sample domain, match the parameters of the operating condition to be predicted with the operating condition boundary thresholds, and call the corresponding model to perform thermal response calculations to generate thermal performance prediction and extreme operating condition extrapolation results. Specifically, S5 includes: S51. Extract operating parameters and corresponding thermodynamic response terms to be predicted from the conventional prediction sample domain. Use the operating parameters as training input and the thermodynamic response terms to be predicted as training output. Use the Matern kernel function to construct the corresponding Gaussian process regression model and use maximum likelihood estimation to determine the kernel function parameters of the Gaussian process regression model. S52. Extract the thermodynamic response terms to be extrapolated from the extreme extrapolation sample domain, determine the response threshold corresponding to each thermodynamic response term to be extrapolated, extract the response value corresponding to the preset high quantile as the response threshold of the corresponding thermodynamic response term to be extrapolated, calculate the excess between the response value exceeding the response threshold and the corresponding response threshold, and based on the excess, use maximum likelihood estimation to determine the scale parameter and shape parameter of the corresponding generalized Pareto model. S53. Input the operating condition parameter to be predicted into the operating condition classifier constructed based on the operating condition boundary threshold. The operating condition classifier performs multivariate kernel density estimation on the operating condition parameter to be predicted to calculate its distribution density value. Based on the region label output by the operating condition classifier, determine whether the operating condition parameter to be predicted corresponds to the conventional prediction sample domain or the extreme extrapolation sample domain. S54. Based on the sample domain corresponding to the operating condition parameters to be predicted, call the corresponding model to perform thermal response calculation: When the operating condition parameter to be predicted corresponds to the conventional prediction sample domain, the corresponding Gaussian process regression model is called to calculate the thermodynamic response term to be predicted and generate the thermodynamic performance prediction result. When the operating condition parameter to be predicted corresponds to the extreme extrapolation sample domain, the corresponding generalized Pareto model is called to calculate the thermodynamic response term to be extrapolated based on the preset extrapolation probability level, and the extreme operating condition extrapolation result is generated. S6. Obtain the verification response data of the thermal performance prediction and extreme condition extrapolation results, and use the maximum likelihood estimation method to incrementally update the Gaussian process regression model and the generalized Pareto model.

2. The method for predicting the thermodynamic performance of a diesel engine and extrapolating its extreme operating conditions based on digital simulation according to claim 1, characterized in that, The preprocessing specifically includes: Detect null values ​​and outliers that exceed the valid range of values ​​for each parameter in the corresponding data records, and delete the data records corresponding to the outliers. Based on the operating state change range and boundary value range corresponding to the operating condition parameters, the retained data records are labeled with operating condition types to obtain steady-state operating condition samples, transient operating condition samples and extreme operating condition samples; The simulation input parameters and operating condition parameters are set as input features, the verification response is set as output features, and the sample correspondence between input features and output features is established according to the operating condition type corresponding to each data record. The Z-score normalization method is used to scale the input and output features, and the scaled input and output features and corresponding working condition type labels are combined to form a basic sample set.

3. The method for predicting the thermodynamic performance of a diesel engine and extrapolating its extreme operating conditions based on digital simulation according to claim 1, characterized in that, The digital simulation calculations in S2 specifically include: Each combination of simulation parameters is loaded sequentially into the digital simulation calculation process. The response is solved based on the corresponding combination of simulation parameters, and continuous response data is extracted from the solution to form the simulation response sequence corresponding to each combination of simulation parameters.

4. The method for predicting the thermodynamic performance of a diesel engine and extrapolating its extreme operating conditions based on digital simulation according to claim 1, characterized in that, The thermal simulation model is a parameterized model of the entire machine built based on a one-dimensional thermodynamic calculation framework, wherein: The one-dimensional thermodynamic calculation framework is a simulation calculation framework that configures parameters and solves the correlation of each calculation object according to the calculation order of operating boundary input, thermodynamic response loading, state transfer calculation and performance response output. The whole machine parameterization model loads the thermal response parameters into the thermal response calculation process, loads the operating condition parameters into the operating boundary input process, and configures the model parameters corresponding to the state transmission calculation process based on the simulation input parameters in the basic sample set. When performing multi-condition response simulation, the calculation boundary corresponding to each condition is determined based on the operation boundary input process. The response change under the corresponding condition is obtained by using the thermodynamic response calculation process. The response change is then passed to the performance response output process step by step according to the state transfer calculation process to obtain the simulation response corresponding to each condition. The deduced response is the thermodynamic performance calculation result that corresponds to the verification response. It is used to calculate the residual between the deduced response and the verification response, and to iteratively calibrate the model parameters in the whole machine parameterized model based on the residual.

5. The method for predicting the thermodynamic performance of a diesel engine and extrapolating its extreme operating conditions based on digital simulation according to claim 1, characterized in that, S4 specifically includes: S41. Extract the standardized operating condition parameters corresponding to each sample from the thermal performance dataset, and combine the standardized operating condition parameters corresponding to the same sample according to the parameter dimension to form a set of operating condition parameter vectors. S42. Calculate the sample dispersion corresponding to each parameter dimension in the working condition parameter vector set. Use the Silverman bandwidth criterion to determine the kernel bandwidth corresponding to each parameter dimension, and construct the diagonal bandwidth matrix corresponding to the multivariate kernel density estimation based on each kernel bandwidth, where: The Silverman bandwidth criterion is as follows: based on the standard deviation of each parameter dimension, the number of samples in the operating condition parameter vector set, and the number of parameter dimensions, the bandwidth determination method for calculating the kernel bandwidth corresponding to each parameter dimension is used. S43. Based on the diagonal bandwidth matrix, the Gaussian kernel function is used to perform multivariate kernel density estimation on the set of working condition parameter vectors, calculate the distribution density value corresponding to each working condition parameter vector, and form a working condition distribution density sequence based on the distribution density value. S44. Construct a load condition classifier based on the load condition distribution density sequence, and use the load condition classifier to label the regions of samples in the thermal performance dataset, where: The operating condition distribution density sequence is a density value sequence obtained by arranging the distribution density values ​​corresponding to each operating condition parameter vector according to their numerical values.

6. The method for predicting the thermodynamic performance of a diesel engine and extrapolating its extreme operating conditions based on digital simulation according to claim 5, characterized in that, S42 specifically includes: S421. Split the working condition parameter vector set according to the parameter dimension, extract the standardized parameter values ​​of each parameter dimension in all samples, and form the parameter value sequence corresponding to each parameter dimension. S422. Based on the parameter value sequence corresponding to each parameter dimension, calculate the standard deviation of each parameter dimension, and obtain the number of samples and the number of parameter dimensions in the working condition parameter vector set. S423. Using the Silverman bandwidth criterion, calculate the kernel bandwidth corresponding to each parameter dimension based on the dimensional standard deviation, sample size, and number of parameter dimensions. S424. According to the order of each parameter dimension in the working condition parameter vector, set the diagonal terms of the matrix based on the kernel bandwidth corresponding to each parameter dimension, and set the off-diagonal terms in the matrix to zero to obtain the diagonal bandwidth matrix used for multivariate kernel density estimation.

7. The method for predicting the thermodynamic performance of a diesel engine and extrapolating its extreme operating conditions based on digital simulation according to claim 1, characterized in that, S6 specifically includes: S61. Obtain the verification response data, and pair the verification response data with the thermal performance prediction results or the extreme condition extrapolation results according to the corresponding operating condition parameters and thermal response items to form a response verification record. S62. In each response verification record, calculate the response deviation between the verification response data and the corresponding result, mark the response verification record with the response deviation exceeding the preset deviation threshold as a deviation record, and extract the operating condition parameters and verification response data from the deviation record to form a deviation sample. S63. Based on the result type corresponding to the deviation sample, the deviation sample corresponding to the thermal performance prediction result is classified into the regular prediction deviation sample set, and the deviation sample corresponding to the extreme condition extrapolation result is classified into the extreme extrapolation deviation sample set. S64. Add the standard prediction bias sample set to the training samples of the Gaussian process regression model, and update the kernel function parameters of the Gaussian process regression model based on the added training samples. Add the limit extrapolation bias sample set to the training samples of the generalized Pareto model, and update the scale parameters and shape parameters of the generalized Pareto model based on the added training samples.

Images