A method for synergistic prediction of combustion process and performance emissions in piston engines

Abstract

This invention relates to the field of piston engine combustion process management technology, specifically a method for collaborative prediction of piston engine combustion process and performance emissions. The method uses a set of engine crankshaft angle positions, corresponding cylinder pressure values, and operating condition information at corresponding times as input feature sequences. A sequence modeling method is used to construct a cylinder pressure prediction model to reconstruct the pressure curve within the engine cylinder. This invention analyzes the mean square error loss corresponding to different combustion stages and the physical constraint loss term based on the PV / T relationship, and assigns different weights to the compression, combustion, and expansion stages of the combustion process. This results in the reconstructed engine cylinder pressure curve from the cylinder pressure prediction model having high prediction accuracy in the compression, combustion, and expansion stages, while maintaining the thermodynamic consistency of the model during the compression process.

Description

Technical Field

[0001] This invention relates to the field of piston engine combustion process management technology, specifically a method for synergistic prediction of piston engine combustion process and performance emissions. Background Technology

[0002] During operation, reciprocating piston engines often exhibit complex relationships between their combustion process, power performance, and pollutant emissions. To achieve a balance between performance and emissions, engineers typically need to jointly optimize multiple parameters, including diesel injection timing, injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate.

[0003] Current technologies often rely on extensive bench testing for optimization analysis, but this approach is time-consuming and costly. In recent years, machine learning-based surrogate models have been used to replace high-fidelity models to accelerate parameter optimization. However, most existing surrogate models directly map control parameters to performance or emission indicators, often ignoring the intermediate physical information of the in-cylinder combustion process. This can easily lead to insufficient generalization ability of the model under non-training conditions. Furthermore, while existing technologies utilize neural networks to reconstruct cylinder pressure curves, they typically employ a single mean squared error loss, failing to consider the differences in deviations between different combustion stages (compression, combustion, and expansion) or introduce physical constraints to ensure thermodynamic consistency. Therefore, existing technologies have significant shortcomings. Summary of the Invention

[0004] The purpose of this invention is to provide a method for synergistic prediction of combustion process and performance emissions of piston engines, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for co-predicting the combustion process and performance emissions of a piston engine, comprising: S1. Obtain multiple control parameters of the engine and construct the engine's operating condition information at the corresponding time. S2. The engine crankshaft angle position, the cylinder pressure value corresponding to the crankshaft angle, and the corresponding operating condition information at the corresponding time are taken as the input feature sequence. The cylinder pressure prediction model is constructed using the sequence modeling method to reconstruct the pressure curve in the engine cylinder. S3. Extract combustion features from the reconstructed pressure curve inside the engine cylinder to obtain the combustion feature parameters of the engine; S4. Using the engine's operating condition information and the combustion characteristic parameters corresponding to the reconstructed engine cylinder pressure curve under the corresponding operating condition information as input, a nonlinear mapping model based on combustion characteristics is constructed to predict the engine's performance indicators and pollutant emission indicators.

[0006] Furthermore, the multiple control parameters of the engine in S1 include, but are not limited to, diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate. The engine's operating condition information at a given time is a summary set of each control parameter corresponding to the engine at that time.

[0007] This invention can also dynamically adjust the parameter types within the control parameters according to the engine model to be analyzed. The above five parameter types (diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate) are selected because these five parameter types are universally applicable to the analysis of any reciprocating piston engine. However, in order to further improve the analysis accuracy of this technical solution for a specific engine model, the parameter types within the selected control parameters can be adjusted.

[0008] Furthermore, the sequence modeling method includes, but is not limited to, attention-based neural network models (such as the Transformer algorithm), recurrent neural network models (such as the LSTM algorithm), or convolutional neural network models (such as the CNN algorithm).

[0009] Furthermore, the sequence modeling method is a Transformer-based neural network model.

[0010] Furthermore, if the sequence modeling method is a neural network model based on Transformer, the crankshaft angle position of the engine in the S2 input feature sequence is converted into the crankshaft angle position code of the engine. The crankshaft angle position code of the engine is a two-dimensional code that is a combination of the corresponding calculation results after taking the sine function and cosine function of the crankshaft angle value respectively. When training the Transformer-based neural network model, a piecewise weighted mean square error loss function is used to calculate the mean square error of different combustion stages of the engine and assign different weights to different combustion stages. The piecewise weighted mean square error loss function also includes a physical constraint term constructed based on the variance of the product of pressure, volume and temperature during the engine compression stroke.

[0011] Furthermore, specific training methods for Transformer-based neural network models include: Based on the actual values ​​of the verification operating conditions, the lower limit of the sample values ​​for the control parameters—diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate—was determined. The upper limit of the sample values ​​for the control parameters—diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate—was set with the absence of knocking as the upper limit. Sample value intervals corresponding to these parameters were constructed. A preset number of different values ​​were selected from the sample value intervals corresponding to each parameter type in the control parameters. A combination of one sample value corresponding to each parameter type in the control parameters was used as a type of operating condition information.

[0012] The calculation of the overall consistency variance of the engine's compression section using pressure, volume, and temperature during the compression stroke involves the following formulas: in, The overall consistency variance of the engine compression section when representing the corresponding operating condition information; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding cylinder pressure at that time; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding cylinder internal volume at that time; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding in-cylinder temperature; Var() represents the function for calculating variance; The loss weights corresponding to the mean square error calculated for the compression, combustion and expansion sections of the engine, and the overall consistency variance of the compression section, are all preset constants, and the loss weights are adjusted according to the actual dataset.

[0013] This invention divides the entire cylinder pressure curve into three physical stages according to the engine combustion process: the compression stage (from intake valve closing to diesel injection timing), the combustion stage (from diesel injection timing to 40°CA), and the expansion stage (from 40°CA to exhaust valve opening). The mean squared error (MSE) loss for each of these three physical stages is calculated to enhance the model's prediction accuracy at each stage. Since the cylinder pressure peak region in the combustion stage has the greatest impact on the overall curve reconstruction accuracy, it is given a higher weight in the total loss function. Loss weights are set according to the importance of each stage in predicting the cylinder pressure peak, with the weights for the compression, combustion, and expansion stages set to 1.5, 4.0, and 2.5, respectively (weights are adjustable based on the actual dataset). Furthermore, a physical constraint loss term based on the PV / T relationship is introduced in the compression stage to enhance the model's thermodynamic consistency during the compression process.

[0014] Furthermore, during the compression stroke, since neither combustion nor fuel injection occurs within the cylinder, the system can be approximated as a closed system, and the mass of the gas inside the cylinder remains constant. If the gas is approximated as an ideal gas, according to thermodynamic relationships, PV / T should remain approximately constant during this stage. Therefore, this embodiment applies a variance penalty (overall consistency variance of the compression segment) to the overall consistency of the compression segment within the same operating condition.

[0015] Furthermore, it also includes: based on the engine cylinder pressure curve obtained from S2 reconstruction, the apparent heat release rate curve is calculated according to the relationship between specific heat ratio, cylinder pressure, and cylinder volume as a function of crankshaft angle, for analysis of the engine combustion process; the calculation formulas involved are as follows: in, This represents the corresponding apparent heat release rate curve; , , and These represent the engine's specific heat ratio, in-cylinder pressure, in-cylinder volume, and crankshaft angle, respectively. The combustion process of the engine includes a compression phase (from intake valve closed to diesel injection timing), a combustion phase (from diesel injection timing to 40°CA), and an expansion phase (from 40°CA to exhaust valve open).

[0016] Furthermore, the combustion characteristic parameters of the engine in S3 include: maximum in-cylinder pressure, crankshaft angle position where the maximum in-cylinder pressure occurs, average in-cylinder pressure, rate of increase of maximum in-cylinder pressure, and rate of increase of minimum in-cylinder pressure. The performance indicators mentioned in S4 include average indicated pressure and brake fuel consumption rate, and the pollutant emission indicators include ammonia, nitrogen oxides and greenhouse gases.

[0017] Furthermore, the nonlinear mapping model based on combustion features in S4 includes, but is not limited to, random forest model, gradient boosting tree model, or neural network model (e.g., BP neural network model). The nonlinear mapping model based on combustion characteristics represents a model used to establish a nonlinear correspondence between engine operating parameters and combustion characteristic parameters and engine performance indicators and pollutant emission indicators.

[0018] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention analyzes the mean square error loss and physical constraint loss term based on PV / T relationship corresponding to different combustion stages, and assigns different weights to the compression, combustion and expansion stages of the combustion process, so that the cylinder pressure prediction model reconstructs the engine cylinder pressure curve with high prediction accuracy in the compression, combustion and expansion stages, and maintains the thermodynamic consistency of the model in the compression process. (2) The present invention uses the combustion features extracted from the reconstructed engine cylinder pressure curve and the corresponding engine operating condition information as inputs to the constructed nonlinear mapping model, which can accurately perceive the dynamic characteristics of the engine in the combustion process in terms of predicting performance indicators and pollutant emission indicators. Attached Figure Description

[0019] 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 a schematic diagram of the overall framework of the method for synergistic prediction of combustion process and performance emissions of piston engines according to the present invention; Figure 2 This is a schematic diagram of the overall structure of the Transformer model in the collaborative prediction method of piston engine combustion process and performance emissions of the present invention; Figure 3 This is a schematic diagram of the sample value range corresponding to each parameter type in the control parameters of the present invention, which is a method for synergistic prediction of combustion process and performance emissions of piston engines. Figure 4 This is a schematic diagram comparing cylinder pressure with experimental simulation results of Transformer model, LSTM model and CNN model under different operating conditions in an embodiment of the present invention, which is a method for coordinating prediction of combustion process and performance emissions of piston engine. Figure 5 This is a schematic diagram comparing cylinder pressure and AHRR of the model prediction results and experimental simulation results under the verification set operating conditions in an embodiment of the present invention, which is a method for synergistic prediction of combustion process and performance emissions of a piston engine. Figure 6 This is a schematic diagram comparing the cylinder pressure and AHRR of the model prediction results and experimental simulation results under the unknown operating condition in an embodiment of the method for synergistic prediction of combustion process and performance emissions of a piston engine according to the present invention. Figure 7 This is a schematic diagram comparing the prediction results of the random forest model with the actual values ​​in an embodiment of the method for synergistic prediction of combustion process, performance and emissions of a piston engine according to the present invention. Figure 8This is a schematic diagram showing the actual values ​​on the test set and the predicted values ​​of the random forest model in an embodiment of a method for coordinating prediction of combustion process and performance emissions of a piston engine according to the present invention. Figure 9 This is a schematic diagram comparing the prediction results of the BP neural network model with the actual values ​​in an embodiment of the method for synergistic prediction of combustion process, performance and emissions of a piston engine according to the present invention. Detailed Implementation

[0020] 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 only some 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.

[0021] Please see Figure 1 This embodiment provides a method for synergistic prediction of piston engine combustion process and performance emissions, including: S1. Obtain multiple control parameters of the engine and construct the engine's operating condition information at the corresponding time. The control parameters in S1 include diesel injection timing (diesel SOI), diesel injection pressure, urea injection timing (urea SOI), urea injection mass ratio (urea IMR), and exhaust gas recirculation rate (EGR rate). The engine's operating condition information at a given time is a summary set of each control parameter corresponding to the engine at that time.

[0022] In this embodiment, the parameter types within the control parameters can also be dynamically adjusted according to the engine model to be analyzed. The above five parameter types (diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate) are selected because these five parameter types are universally applicable to the analysis of any reciprocating piston engine. However, in order to further improve the analysis accuracy of this technical solution for a specific engine model, the parameter types within the selected control parameters can be adjusted.

[0023] S2. The engine crankshaft angle position, the cylinder pressure value corresponding to the crankshaft angle, and the corresponding operating condition information at the corresponding time are taken as the input feature sequence. The cylinder pressure prediction model is constructed using the sequence modeling method to reconstruct the pressure curve in the engine cylinder. The sequence modeling methods include, but are not limited to, attention-based neural network models (such as the Transformer algorithm), recurrent neural network models (such as the LSTM algorithm), or convolutional neural network models (such as the CNN algorithm).

[0024] In the current implementation, the cylinder pressure prediction module supports three sequence models: Transformer, LSTM, and CNN. These three models employ a unified data organization method, a unified input / output interface, a unified training process, and a unified loss design. Their main difference lies in the different ways they model the internal structure of the input sequence. Based on this unified framework, cylinder pressure prediction can be performed using different sequence backbone networks without altering the overall data flow.

[0025] Following cylinder pressure prediction, the existing code also includes cylinder pressure curve post-processing and downstream index prediction stages. The predicted cylinder pressure curve can be further used to calculate the apparent heat release rate, extract combustion characteristic parameters derived from cylinder pressure, and, in optional embodiments, input together with engine operating parameters into a nonlinear mapping model to predict indices such as NH3, NOx, GHG, and IMEP.

[0026] The sequence modeling method described is a Transformer-based neural network model. Transformer is a deep neural network model with self-attention as its core mechanism, such as... Figure 2 As shown, it completely abandons the recursive structure of traditional recurrent neural networks, and achieves efficient modeling of long-range dependencies in sequential data through a parallelized multi-head Self-Attention mechanism. This feature makes it widely used in tasks such as natural language processing, time series prediction, and complex nonlinear regression. The cylinder pressure curve is essentially a sequential data that varies with the crankshaft angle, exhibiting obvious sequential characteristics. Therefore, this invention uses the Transformer model to model and predict the engine's cylinder pressure curve.

[0027] If the sequence modeling method is a neural network model based on Transformer, the crankshaft angle position of the engine in the input feature sequence of S2 is converted into the crankshaft angle position code of the engine. The crankshaft angle position code of the engine is a two-dimensional code that is a combination of the corresponding calculation results after taking the sine function and cosine function of the crankshaft angle value respectively. The parameters used in the current training configuration include d_model=128, nhead=8, num_layers=4, dim_feedforward=256, and dropout=0.1. In some implementations, the number of layers, the latent space dimension, the number of attention heads, and the feedforward layer dimension can all be adjusted.

[0028] This implementation outputs a cylinder pressure sequence of the same length as the input crankshaft angle sequence, that is, it outputs a cylinder pressure prediction value at each crankshaft angle position, forming a complete cylinder pressure curve. The output tensor is in the form of [batch, sequence_length, 1].

[0029] In current implementations, the Transformer model shares training scripts with other models. The training process includes: Step 1: Construct a sequence of samples organized by operating conditions; based on the actual values ​​of the verification operating conditions, determine the lower limit of the sample values ​​for the control parameters diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate; with the absence of knocking as the upper limit of the sample values ​​for the control parameters diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate, construct the sample value intervals corresponding to the control parameters diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate respectively; in this embodiment, three different values ​​are selected in the sample value intervals corresponding to each parameter type in the control parameters, as shown in Figure 3; Step 2: Divide the training set and validation set according to the set ratio; Step 3: Perform multiple rounds of iterative training using the Adam optimizer; Step 4: Calculate the loss on the validation set after each round of training. Specifically, when training the Transformer-based neural network model, a piecewise weighted mean square error loss function is used to calculate the mean square error of different combustion stages of the engine and assign different weights to different combustion stages. The piecewise weighted mean square error loss function also includes a physical constraint term constructed based on the variance of the product of pressure, volume and temperature during the engine's compression stroke.

[0030] Step 5: Save the optimal model based on the verification loss.

[0031] It should be noted that since the Transformer lacks a loop structure, it must rely on position encoding to enable the model to perceive the sequence order. Therefore, this embodiment maps the crankshaft angle (CA) to a two-dimensional position code using sine and cosine functions. Through this mapping, each crankshaft angle position has a unique code, allowing the model to capture the pattern of cylinder pressure variation with crankshaft angle while preserving its periodicity and continuity. The crankshaft angle position code, together with the corresponding cylinder pressure value and constant input parameters under the same operating condition, constitutes the model's input feature sequence, used for reconstructing the pressure curve within the entire engine cylinder.

[0032] The calculation of the overall consistency variance of the engine's compression section using pressure, volume, and temperature during the compression stroke involves the following formulas: in, The overall consistency variance of the engine compression section when representing the corresponding operating condition information; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding cylinder pressure at that time; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding cylinder internal volume at that time; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding in-cylinder temperature; Var() represents the function for calculating variance; The loss weights corresponding to the mean square error calculated for the compression, combustion and expansion sections of the engine, and the overall consistency variance of the compression section, are all preset constants, and the loss weights are adjusted according to the actual dataset.

[0033] In this embodiment, the entire cylinder pressure curve is divided into three physical stages according to the engine combustion process: compression (from intake valve closing to diesel injection timing), combustion (from diesel injection timing to 40°CA), and expansion (from 40°CA to exhaust valve opening). The mean squared error (MSE) loss for each of these three physical stages is calculated to enhance the model's prediction accuracy at each stage. Since the cylinder pressure peak region in the combustion stage has the greatest impact on the overall curve reconstruction accuracy, it is given a higher weight in the total loss function. Loss weights are set according to the importance of each stage in predicting the cylinder pressure peak, with the weights for compression, combustion, and expansion stages set to 1.5, 4.0, and 2.5, respectively (weights are adjustable based on the actual dataset). Furthermore, a physical constraint loss term based on the PV / T relationship is introduced in the compression stage to enhance the model's thermodynamic consistency during compression.

[0034] Furthermore, during the compression stroke, since neither combustion nor fuel injection occurs within the cylinder, the system can be approximated as a closed system, and the mass of the gas inside the cylinder remains constant. If the gas is approximated as an ideal gas, according to thermodynamic relationships, PV / T should remain approximately constant during this stage. Therefore, this embodiment applies a variance penalty (overall consistency variance of the compression segment) to the overall consistency of the compression segment within the same operating condition.

[0035] In another embodiment, the sequence modeling method uses a recurrent neural network model (LSTM model). In this embodiment, the LSTM model uses the same data organization method as the Transformer model, that is, each sample corresponds to a crankshaft angle sequence under a complete operating condition; the input at each time step consists of control parameters and crankshaft angle features. In the current training implementation, the LSTM model constructs the input by adding normalized CA to the control parameters; in the existing implementation, the LSTM model consists of multiple long short-term memory networks, optionally with a bidirectional structure. The input sequence is sequentially input into the LSTM layers along the crankshaft angle direction, and the modeling of adjacent and longer-range relationships is achieved through the propagation of hidden states in the sequence. The hidden states at each time step output by the LSTM are then mapped to the cylinder pressure prediction values ​​at the corresponding crankshaft angle positions through a linear layer.

[0036] In the current training configuration, the LSTM model uses hidden_dim=128, num_layers=2, dropout=0.1, and bidirectional=True. In some implementations, the hidden layer dimension, number of layers, whether to use a bidirectional structure, and dropout can all be adjusted as needed.

[0037] The LSTM model outputs a cylinder pressure prediction sequence of the same length as the input crankshaft angle sequence, that is, it outputs the corresponding predicted cylinder pressure value at each crankshaft angle position. Its output tensor form is consistent with the Transformer model, which is [batch, sequence_length, 1].

[0038] In existing implementations, the LSTM model shares the same training script, data partitioning method, feature scaling method, and loss function design with the Transformer model. That is, the LSTM model only replaces the internal temporal modeling backbone network under the same data flow and supervision objectives.

[0039] This implementation method is also used to predict the complete cylinder pressure curve from operating parameters. The predicted cylinder pressure curve obtained by the model can be directly used for combustion process analysis, peak cylinder pressure extraction, and subsequent performance and / or emission-related index calculations. This implementation method is consistent with the implementation method of the Transformer model in terms of inventive concept, both revolving around the same input-output relationship and the same cylinder pressure reconstruction objective. The only difference is that LSTM achieves this mapping through a cyclic temporal structure, and therefore belongs to the alternative sequential modeling implementation method in this invention.

[0040] In another embodiment, the sequence modeling method is a convolutional neural network (CNN) model; the CNN model uses the same sample definition method and output target as the two models mentioned above. Each sample still corresponds to a crankshaft angle sequence of a complete working condition, and the input features consist of control parameters and crankshaft angle features. In the current training implementation, the CNN model constructs the input by adding normalized CA to the control parameters.

[0041] In existing implementations, the CNN model first transforms the input tensor from [batch, seq, features] to [batch, features, seq] so that it can perform one-dimensional convolution along the crankshaft angle axis. The model first performs input projection through 1x1 convolution, then performs sequence feature extraction through multiple residual convolution blocks, and finally obtains the cylinder pressure prediction value at each crankshaft angle position through output projection convolution.

[0042] Each residual convolutional block contains a convolutional layer, a batch normalization layer, an activation layer, dropout, and residual connections; and in the existing implementation, dilated convolution is used so that the receptive field gradually expands as the network layer deepens, in order to cover a longer range of crankshaft angle sequence relationships.

[0043] In the current training configuration, the CNN model uses channels=128, num_blocks=4, kernel_size=5, and dropout=0.1. In some implementations, the number of convolutional channels, the number of residual blocks, the kernel size, and dropout can be adjusted as needed.

[0044] The CNN model outputs the cylinder pressure prediction value at each crankshaft angle position, and in the output stage, the tensor is converted back to the [batch, sequence_length, 1] form to form a complete cylinder pressure curve.

[0045] In existing implementations, CNN models share the same training process as Transformer and LSTM models, including the same dataset construction method, the same input / output scaling method, the same training and validation set partitioning method, and the same piecewise loss function. The only difference is that the internal representation extraction of the cylinder compression sequence is performed by a one-dimensional convolutional network.

[0046] This implementation method is used to extract local and multi-scale sequence features along the crankshaft rotation direction and output a complete cylinder pressure curve. The obtained cylinder pressure results can also be used for subsequent apparent heat release rate calculation, combustion characteristic parameter analysis, and downstream performance and / or emission index prediction. This implementation method, along with the aforementioned Transformer and LSTM model implementation methods, represents different network structure schemes for achieving the same inventive concept. Their commonality lies in using operating parameters and crankshaft rotation features as inputs and a complete cylinder pressure curve as output; the difference lies only in the internal implementation of sequence feature extraction. Therefore, this CNN model implementation method is also an alternative sequence modeling implementation method in this invention. Figure 4 The figure shows a comparison of cylinder pressure between the model prediction results and experimental simulation results of the Transformer model, LSTM model and CNN model under different working conditions.

[0047] This embodiment also includes: based on the pressure curve inside the engine cylinder obtained by S2 reconstruction, calculating the apparent heat release rate curve according to the relationship between specific heat ratio, cylinder pressure, and cylinder volume with crankshaft angle, for analyzing the engine combustion process; the calculation formulas involved are as follows: in, This represents the corresponding apparent heat release rate curve; , , and These represent the engine's specific heat ratio, in-cylinder pressure, in-cylinder volume, and crankshaft angle, respectively. The combustion process of the engine includes a compression phase (from intake valve closed to diesel injection timing), a combustion phase (from diesel injection timing to 40°CA), and an expansion phase (from 40°CA to exhaust valve open).

[0048] Figure 5 A comparison of cylinder pressure and AHRR between the model predictions and experimental simulations under validation set conditions is presented. It can be seen that the cylinder pressure curve predicted by the model is in high agreement with the experimental results across the entire crankshaft rotation range. Since the AHRR curve is more sensitive to changes in combustion phase and combustion rate, it can more intuitively reflect the dynamic characteristics of the combustion process. Figure 5 It can be seen that the model can reconstruct the peak value and position of AHRR in the main combustion stage relatively well, and the prediction of the start and end positions of combustion is also relatively accurate.

[0049] To further verify the model's predictive ability under unknown parameter combinations, random interpolation was performed within the range of five control and emission reduction parameters to generate new operating conditions that did not appear in the original 243 combinations. Figure 6The comparison of cylinder pressure and AHRR between the model's predicted results and experimental simulation results under the unknown operating condition is shown. Compared with the validation set results, the curves in the validation set still maintain good smoothness and physical consistency, indicating that the introduced piecewise loss and the PV / T physical constraint of the compression segment improve the stability and physical rationality of the model in the continuous parameter space while ensuring prediction accuracy. Therefore, the reconstructed model constructed in this invention can not only achieve high-precision fitting on the validation set with known parameter combinations, but also maintain stable prediction performance under unknown operating conditions with random interpolation, demonstrating strong generalization ability.

[0050] S3. Extract combustion features from the reconstructed engine cylinder pressure curve to obtain the engine's combustion characteristic parameters. The combustion characteristic parameters of the engine in S3 include: maximum in-cylinder pressure ( ), the crankshaft angle at which the maximum cylinder pressure occurs ( ), average cylinder pressure ( ), Maximum in-cylinder pressure rise rate ( ) and minimum in-cylinder pressure rise rate ( ); S4. Using the engine's operating condition information and the combustion characteristic parameters corresponding to the reconstructed engine cylinder pressure curve under the corresponding operating condition information as input, a nonlinear mapping model based on combustion characteristics is constructed to predict the engine's performance indicators and pollutant emission indicators.

[0051] The performance indicators mentioned in S4 include average indicated pressure and brake fuel consumption rate, and the pollutant emission indicators include ammonia, nitrogen oxides and greenhouse gases (specifically NH3, NOx, GHG and IMEP).

[0052] The nonlinear mapping model based on combustion features in S4 includes, but is not limited to, random forest model, gradient boosting tree model or neural network model (e.g., BP neural network model). Random forest (RF) is an ensemble machine learning method that constructs multiple decision trees using random resampling cardinality and random node splitting cardinality, and obtains the final classification result through voting. Random forest models have the ability to analyze complex interactive classification features, exhibit good robustness to noisy data, and have a fast learning speed. Their variable importance measure can serve as a feature selection tool for high-dimensional data, and in recent years have been widely applied to various classification, prediction, feature selection, and outlier detection problems.

[0053] Given the outstanding advantages of random forests in analyzing complex interaction characteristics, handling noisy data, and measuring variable importance, they are well-suited for building high-precision engine replacement models.

[0054] In one specific embodiment, the nonlinear mapping model is a random forest model. Specifically, engine operating parameters are concatenated with combustion characteristic parameters extracted from the cylinder pressure curve to form a feature vector, which is then input into the random forest regression model to establish a nonlinear correspondence between the vector and output indicators such as NH3, NOx, GHG, and IMEP.

[0055] In existing implementations, the random forest model employs an ensemble learning approach based on decision trees to perform multi-output regression. Example default parameters in the current code include: 300 decision trees (this number ensures computational efficiency while providing sufficient diversity and generalization ability, avoiding overfitting, and enhancing model stability), no upper limit on tree depth (allowing each tree to grow freely according to the complexity of the data, thus better capturing nonlinear relationships and high-dimensional features), and a random seed of 42, with parallel training using available computing resources. This setting ensures the model has strong fitting ability, especially when dealing with complex combustion processes and emission data, helping to improve prediction accuracy. In some implementations, the number of trees, tree depth, sample splitting conditions, feature sampling methods, and parallel computing parameters can be adjusted according to the data scale and prediction objectives, and are not limited to the aforementioned numerical configurations.

[0056] The purpose of this embodiment is to improve the accuracy of engine performance and emission index prediction by utilizing the combustion state information represented by the cylinder pressure curve. It should be noted that the aforementioned random forest parameters are merely an example of a set of implementations in existing code, and this invention is not limited to this specific set of parameters.

[0057] The nonlinear mapping model based on combustion characteristics represents a model used to establish a nonlinear correspondence between engine operating parameters and combustion characteristic parameters and engine performance indicators and pollutant emission indicators.

[0058] In this embodiment, when evaluating the performance of the alternative model constructed using the random forest algorithm, goodness of fit (R²) is usually a key evaluation metric. R² represents the goodness of fit between the model's predicted values ​​and the actual values, and its value ranges from 0 to 1: the closer R² is to 1, the more accurate the prediction. The calculation formula is as follows: In the formula, n represents the number of datasets. and These represent the predicted and actual values ​​of the data, respectively. This represents the average of the predicted values.

[0059] Figure 7The graph shows a comparison between the predicted and actual values ​​of NH3, NOx, GHG, and IMEP by the random forest model. As can be seen from the graph, the model fits the output variables well, with R² values ​​for NH3, NOx, GHG, and IMEP being particularly high. 2 The values ​​are 0.983, 0.992, 0.963, and 0.985, respectively. This indicates that random forests can effectively capture the complex nonlinear relationships between input features and various emission and performance indicators. Figure 8 The actual values ​​of NH3, NOx, GHG, and IMEP on the test set were compared with the predicted values ​​of the random forest model. The figures show that the model can accurately predict the output variables under different input conditions, and the predicted values ​​are highly consistent with the actual values, indicating that the random forest alternative model is reliable in capturing the nonlinear relationship between input characteristics and emission and performance indicators. Overall, the random forest model can achieve high-precision predictions on the test set, providing an effective alternative tool for multi-objective optimization and control strategy formulation.

[0060] In another embodiment, the nonlinear mapping model based on combustion features is a feedforward neural network model (BP neural network model) trained based on the backpropagation algorithm. Specifically, the engine operating parameters and the combustion feature parameters extracted from the cylinder pressure curve are concatenated to form a feature vector, which is then input into the feedforward neural network regression model to establish a nonlinear correspondence between the engine operating parameters and output indicators such as NH3, NOx, GHG, and IMEP. Figure 9 The results show a comparison between the predictions of the BP neural network model for NH3, NOx, GHG, and IMEP and the actual values.

[0061] In the existing code implementation, the neural network model adopts a fully connected feedforward structure, contains two hidden layers, and updates parameters through the backpropagation algorithm.

[0062] The current code's default parameter examples include: hidden layer size (hidden_layer_sizes=(128, 64)), activation function (activation='relu'), optimizer (solver='adam'), and regularization coefficient (alpha=1e). -4 Initial learning rate learning_rate_init=1e -3 The maximum number of iterations is set to max_iter=2000, the batch size is set to 'auto', and the random seed is set to random_state=42.

[0063] In some implementations, the number of neurons in the two hidden layers can be further adjusted or determined through search based on the training results. The activation function, optimizer type, regularization parameter, learning rate, number of iterations, and batch size can also be changed according to actual needs, and are not limited to the aforementioned parameter configurations. Therefore, the above parameters are merely specific implementation examples in the current code and do not constitute a limitation on the scope of protection of this invention, nor do they change its technical essence as a "nonlinear mapping model based on combustion features".

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for synergistic prediction of combustion process and performance emissions in a piston engine, characterized in that, include: S1. Obtain multiple control parameters of the engine and construct the engine's operating condition information at the corresponding time. S2. The engine crankshaft angle position, the cylinder pressure value corresponding to the crankshaft angle, and the corresponding operating condition information at the corresponding time are taken as the input feature sequence. The cylinder pressure prediction model is constructed using the sequence modeling method to reconstruct the pressure curve in the engine cylinder. S3. Extract combustion features from the reconstructed pressure curve inside the engine cylinder to obtain the combustion feature parameters of the engine; S4. Using the engine's operating condition information and the combustion characteristic parameters corresponding to the reconstructed engine cylinder pressure curve under the corresponding operating condition information as input, a nonlinear mapping model based on combustion characteristics is constructed to predict the engine's performance indicators and pollutant emission indicators.

2. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 1, characterized in that, The engine control parameters in S1 include, but are not limited to, diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate. The engine's operating condition information at a given time is a summary set of each control parameter corresponding to the engine at that time.

3. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 1, characterized in that, The sequence modeling methods include, but are not limited to, attention-based neural network models, recurrent neural network models, or convolutional neural network models.

4. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 3, characterized in that, The sequence modeling method is a neural network model based on Transformer.

5. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 4, characterized in that, If the sequence modeling method is a neural network model based on Transformer, the crankshaft angle position of the engine in the input feature sequence of S2 is converted into the crankshaft angle position code of the engine. The crankshaft angle position code of the engine is a two-dimensional code that is a combination of the corresponding calculation results after taking the sine function and cosine function of the crankshaft angle value respectively. When training the Transformer-based neural network model, a piecewise weighted mean square error loss function is used to calculate the mean square error of different combustion stages of the engine and assign different weights to different combustion stages. The piecewise weighted mean square error loss function also includes a physical constraint term constructed based on the variance of the product of pressure, volume and temperature during the engine compression stroke.

6. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 5, characterized in that, Specific training methods for Transformer-based neural network models include: Based on the actual values ​​of the verification conditions, the lower limit of the sample values ​​for the control parameters—diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate—was determined. The upper limit of the sample values ​​for the control parameters—diesel injection timing, diesel injection pressure, urea injection timing, urea injection mass ratio, and exhaust gas recirculation rate—was set as the absence of knocking. Sample value intervals corresponding to these parameters were then constructed. A preset number of different values ​​were selected from the sample value intervals corresponding to each parameter type within the control parameters. The calculation of the overall consistency variance of the engine's compression section using pressure, volume, and temperature during the compression stroke involves the following formulas: in, The overall consistency variance of the engine compression section when representing the corresponding operating condition information; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding cylinder pressure at that time; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding cylinder internal volume at that time; This indicates the corresponding operating condition information and the crankshaft rotation angle is... The corresponding in-cylinder temperature; Var() represents the function for calculating variance; The loss weights corresponding to the mean square error calculated for the compression, combustion and expansion sections of the engine, and the overall consistency variance of the compression section, are all preset constants, and the loss weights are adjusted according to the actual dataset.

7. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 1, characterized in that, Also includes: Based on the pressure curve inside the engine cylinder obtained by S2 reconstruction, the apparent heat release rate curve is calculated according to the relationship between specific heat ratio, cylinder pressure, and cylinder volume with crankshaft angle, and is used to analyze the combustion process of the engine. The combustion process of the engine includes a compression section, a combustion section, and an expansion section.

8. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 1, characterized in that, The combustion characteristic parameters of the engine in S3 include: maximum in-cylinder pressure, crankshaft angle at which the maximum in-cylinder pressure occurs, average in-cylinder pressure, rate of increase of maximum in-cylinder pressure, and rate of increase of minimum in-cylinder pressure. The performance indicators mentioned in S4 include average indicated pressure and brake fuel consumption rate, and the pollutant emission indicators include ammonia, nitrogen oxides and greenhouse gases.

9. The method for synergistic prediction of combustion process and performance emissions of a piston engine according to claim 1, characterized in that, The nonlinear mapping model based on combustion features in S4 includes, but is not limited to, random forest model, gradient boosting tree model or neural network model; The nonlinear mapping model based on combustion characteristics represents a model used to establish a nonlinear correspondence between engine operating parameters and combustion characteristic parameters and engine performance indicators and pollutant emission indicators.

Images