Diesel engine performance and emission prediction method and system based on transient working condition

Abstract

The invention discloses a diesel engine performance and emission prediction method and system based on transient working conditions, and relates to the technical field of engine performance and emission parameter prediction. Multi-channel high-resolution time sequence data and performance and emission parameters under a typical transient condition are collected as a training data source, key input variables are screened through a random forest algorithm, multi-signal fusion and nonlinear dimension reduction are performed by using a one-dimensional convolutional neural network, a multi-output feedforward neural network model is constructed, and the multi-output feedforward neural network model is constructed. And introducing a Bayesian optimization algorithm to automatically optimize hyper-parameters of the neural network, and finally evaluating the performance of the model by adopting a root-mean-square error, a mean absolute error and a goodness-of-fit index. The system comprises a diesel engine simulation module, a feature screening and signal fusion module, a neural network prediction module, a hyper-parameter optimization module and a performance evaluation module which are connected through data stream interfaces to realize data transmission and processing.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of engine performance and emission parameter prediction, and particularly relates to a diesel engine performance and emission prediction method based on transient operating conditions. BACKGROUND

[0002] At present, diesel engines, as the core power devices in the fields of transportation and engineering machinery, provide high-efficiency output while also facing serious pollution emission problems. In order to achieve the energy-saving and emission-reducing goals of diesel engines, accurately predicting their performance and emission behavior under typical dynamic operating conditions has become a key problem in modern engine research and development. Especially under the transient operating condition cycle represented by FTP75, the engine operating state is complex and variable, and the traditional steady-state modeling method has been difficult to meet the prediction needs under the conditions of nonlinearity and dynamic characteristics.

[0003] In the prior art, diesel engine performance and emission modeling mainly focuses on the following two directions: (1) Static modeling method based on experimental bench data Most studies collect data such as engine speed, fuel injection quantity, intake pressure, etc. under different steady-state operating conditions on the engine test bench, and construct traditional polynomial regression models, support vector machines (SVM) or artificial neural networks (ANN) prediction models. These models are based on a small number of input variables to predict engine output power, thermal efficiency, and carbon monoxide, hydrocarbon, etc. emission parameters, and have certain modeling accuracy.

[0004] (2) Deep learning prediction method based on a single signal source With the development of deep learning, some studies have begun to explore the use of convolutional neural networks (CNN), long short-term memory networks (LSTM), etc. to realize the modeling of engine time series data. For example, engine speed and fuel injection pressure, etc. are used as inputs to predict NOx, CO, etc. emission factors. Some studies also introduce neural network optimization methods (such as grid search) to improve model efficiency.

[0005] Although the above methods have promoted the development of diesel engine modeling technology to some extent, there are still the following deficiencies and technical bottlenecks: (1) Strong adaptability to steady-state operating conditions, poor modeling capability under transient operating conditions Most current models are based on steady-state experiments and lack the ability to adapt to typical transient cycle operating conditions, and cannot effectively reflect the actual response characteristics of the engine during acceleration, deceleration, cold start, etc. high-frequency nonlinear dynamic processes. Such models often exhibit poor generalization ability and prediction lag, limiting their application in real operating conditions.

[0006] (2) Insufficient input signal fusion, with redundancy and information loss problems Most existing methods directly use multiple original sensor signals as neural network inputs without effective feature compression and nonlinear fusion, resulting in high input dimension, large model complexity, poor training stability, and easy overfitting. In addition, traditional linear dimension reduction methods are difficult to extract deep nonlinear interaction between input variables, affecting the prediction accuracy.

[0007] (3) Low efficiency of neural network hyperparameter optimization, lack of intelligent optimization mechanism During the training process of neural network, the model performance is highly sensitive to network structure (number of layers, number of nodes) and training strategy (activation function, weight initialization, number of training rounds, etc.). The current mainstream optimization method still uses grid search or empirical method to tune parameters, which is difficult to deal with high-dimensional parameter space, has high training cost, slow convergence speed, and may miss the global optimal solution.

[0008] (4) Low utilization of simulation data, lack of physical consistency Although GT-POWER and other one-dimensional simulation software can provide high-precision engine operation data, most existing researches still rely on limited experimental data to train the model, and do not fully exploit the large-scale, dynamic, and high-time-resolution data resources provided by simulation models, resulting in lack of physical consistency and engineering adaptability of the model.

[0009] Through comprehensive analysis, the existing diesel engine performance and emission prediction technology has not realized effective modeling of nonlinear response mechanism under transient operating conditions, especially in the aspects of multi-variable signal fusion and intelligent optimization of neural network. Therefore, it is urgent to develop a modeling method combining high-precision simulation data, multi-signal fusion feature extraction, and intelligent hyperparameter optimization mechanism to improve the prediction accuracy, stability, and dynamic adaptability of the model. SUMMARY

[0010] The purpose of the present application is to provide a diesel engine performance and emission prediction method based on transient operating conditions, which addresses the outstanding problems in the prior art, such as poor dynamic operating condition adaptability, weak input signal fusion capability, low model structure optimization efficiency, and lack of physical consistency. A systematic method is proposed, which integrates simulation data driving, multi-signal fusion feature extraction, deep neural network modeling, and Bayesian optimization, has the significant advantages of high modeling accuracy, strong generalization ability, and high computational efficiency, and is widely applicable to diesel engine online monitoring, control strategy optimization, and emission prediction.

[0011] To achieve the above purpose, the present application provides a diesel engine performance and emission prediction method based on transient operating conditions, the steps are as follows: S1, simulation model construction and data acquisition: based on GT-POWER software to build a one-dimensional simulation model of the target diesel engine, and based on the standard dynamic working condition to set the simulation boundary condition and control strategy, to obtain multi-channel high-resolution time series data and thermal efficiency, effective power, torque, main emission concentration as training data source; S2, input feature selection and multi-signal fusion: using random forest algorithm to evaluate the importance of the initial selected multi-input variables, and extracting the key variables highly related to the target output; using one-dimensional convolutional neural network to compress and nonlinearly fuse the original high-dimensional time series input signal, to generate a low-dimensional fusion feature vector; S3, deep neural network structure design and construction: constructing a multi-output feedforward neural network model, taking the low-dimensional fusion feature vector compressed by one-dimensional convolutional neural network as input, and outputting six key prediction targets including thermal efficiency, torque, power, carbon monoxide, carbon dioxide and hydrocarbon; S4, Bayesian optimization: introducing Bayesian optimization algorithm to automatically optimize the number of hidden layers, the number of neurons, the type of activation function, the weight initialization method and the maximum number of training rounds of the neural network; S5, model evaluation and deployment: using root mean square error, mean absolute error and goodness of fit index to evaluate the fitting accuracy and generalization ability of the model under dynamic cycle.

[0012] Preferably, in S1, the multi-channel high-resolution time series data includes engine speed, injection pressure, intake and exhaust temperature / pressure, intake air volume, fuel injection volume, in-cylinder pressure and EGR rate.

[0013] Preferably, in S2, the structure of the one-dimensional convolutional neural network is 3 layers of 1D convolution, each layer has 64→32→4 channels, the kernel size is 1, and the activation function is ReLU.

[0014] Preferably, the Bayesian optimization algorithm guides the network structure iterative update through Gaussian surrogate model and expected improvement function EI.

[0015] A diesel engine performance and emission prediction system based on transient working condition, comprising: A diesel engine simulation module: for building a four-cylinder turbocharged diesel engine model on the GT-POWER platform, setting typical transient working condition boundary conditions, and outputting multi-channel engine operation data, the data output by the diesel engine simulation module is transmitted to the feature selection and signal fusion module; A feature selection and signal fusion module: containing a random forest evaluation unit and a one-dimensional convolutional neural network, for selecting key variables and nonlinear compression and fusion of input signals, the fused features are used as the input of the neural network prediction module; Neural network prediction module: a feedforward neural network composed of an input layer, a hidden layer and an output layer, taking fused features as input and outputting performance and emission parameter prediction values; Hyperparameter optimization module: a Bayesian optimization algorithm is used to optimize the neural network structure and training parameters, and the neural network prediction module and the hyperparameter optimization module interact bidirectionally, and structure search and model evaluation are performed in the optimization process; Performance evaluation module: compare the prediction results with the simulation data, and output the root mean square error, mean absolute error and goodness of fit evaluation index.

[0016] Preferably, the feature screening and signal fusion module first screens 8 optimal input variables through the random forest evaluation unit, and then performs feature compression processing on the 8-dimensional input signal through the one-dimensional convolutional neural network to output a 4-dimensional fused feature vector.

[0017] Preferably, the optimal structure of the neural network prediction module is that the input layer has 4 nodes, the number of nodes in the hidden layer is 26-89-87-36, the output layer has 6 nodes, the activation function combination is Tanh-ReLU-ReLU-Tanh, and the weight initialization adopts the Glorot method.

[0018] The diesel engine performance and emission prediction method and system based on transient operating conditions proposed in the application form a complete diesel engine performance and emission prediction system for typical dynamic operating conditions by fusing high-precision physical simulation data, multi-signal fusion dimension reduction technology, deep neural network modeling and Bayesian optimization strategy. Compared with the prior art, the application has the following outstanding advantages and technical effects: (1) Improve prediction accuracy and operating condition adaptability The application constructs a high-precision diesel engine simulation model on the GT-POWER platform, and collects multi-channel, high-resolution dynamic data under typical transient operating conditions, so that the constructed neural network model has good dynamic response learning ability. The prediction results show that the optimized neural network has achieved more than 0.95 high fitting accuracy in multiple output dimensions such as power, torque, CO and CO2, which is significantly better than the traditional model based on empirical method to configure hyperparameters, and solves the problem of insufficient prediction ability of existing methods under transient operating conditions. 2

[0019] (2) Reduce model dimension and computational complexity ​By combining the random forest algorithm to select input variables and using a one-dimensional convolutional neural network (1D-CNN) to perform channel compression and nonlinear fusion on high-dimensional input features, the original high-dimensional input data is effectively reduced to a low-dimensional feature representation. While maintaining good prediction performance, the scale of input data and the total number of neural network parameters are reduced by about 50%, significantly reducing training time and system resource overhead, and providing better deployment flexibility and resource adaptability.

[0020] (3) Improve modeling efficiency and automation A Bayesian optimization algorithm is employed to automatically search for network structure and training parameters, replacing traditional parameter tuning methods that rely on manual experience or grid search. This achieves efficient and intelligent optimization and rapid convergence of neural network models. The optimization process only requires dozens of iterations to find the globally optimal structure, significantly reducing model development cycle and debugging costs, and supporting rapid engineering deployment.

[0021] (4) Enhance the generalization ability and stability of the model This invention achieves collaborative prediction and error balance control of multiple output targets of engine performance and emissions through multi-objective joint modeling and distribution consistency training strategy. The optimized model shows a stable fitting trend in different operating ranges and different output dimensions, has good generalization ability, and can adapt to a wider range of operating conditions and equipment types.

[0022] (5) Facilitates integration with application and control decision support The constructed predictive model is compact and flexible in deployment, and can be integrated into engine controllers (ECUs), intelligent diagnostic systems, or online optimization platforms. It can acquire engine operating status in real time and predict emission trends, providing data support for intelligent control strategy optimization, energy conservation and emission reduction scheduling, and vehicle operation management. This helps to reduce pollution emissions and improve energy efficiency from the source, responding to the national "dual carbon" target.

[0023] (6) Possesses good engineering replicability and versatility. This invention is constructed and implemented using a general simulation platform (GT-POWER) and an open algorithm framework (MATLAB). The method has a clear structure, strong versatility, and is easy to extend to other types of internal combustion engine systems or new energy power platforms. It has good engineering promotion value and industrial application prospects.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0026] Figure 1 The function module structure block diagram of the diesel engine performance and emission prediction system of the embodiment of the present application; Figure 2 The structure schematic diagram of the diesel engine simulation model of the embodiment of the present application; Figure 3 The structure schematic diagram of the one-dimensional convolutional neural network of the embodiment of the present application; Figure 4 The structure schematic diagram of the neural network prediction model of the embodiment of the present application; Figure 5 The flow chart of the Bayesian optimization of the embodiment of the present application; Figure 6 The comparison curve diagram of the predicted value and the true value of the thermal efficiency of the embodiment of the present application; Figure 7 The comparison curve diagram of the predicted value and the true value of the torque of the embodiment of the present application; Figure 8 The comparison curve diagram of the predicted value and the true value of the power of the embodiment of the present application; Figure 9 The comparison curve diagram of the predicted value and the true value of carbon monoxide (CO) of the embodiment of the present application; Figure 10 The comparison curve diagram of the predicted value and the true value of carbon dioxide (CO2) of the embodiment of the present application; Figure 11 The comparison curve diagram of the predicted value and the true value of hydrocarbon (HC) of the embodiment of the present application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0028] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0029] EMBODIMENT As Figure 1The diesel engine performance and emission prediction system based on FTP75 transient working condition shown performs a diesel engine performance and emission prediction method based on FTP75 transient working condition.

[0030] The diesel engine performance and emission prediction system based on FTP75 transient working condition specifically comprises: The diesel engine simulation module is used to build a four-cylinder turbocharged diesel engine model on a GT-POWER platform, set typical FTP75 working condition boundary conditions, and output multi-channel engine operation data. The feature screening and signal fusion module comprises a random forest evaluation unit and a one-dimensional convolutional neural network, and is used for screening key variables and input signals, nonlinear compression and fusion of the input signals, and fusion of the input signals.

[0031] The neural network prediction module is a feedforward neural network composed of an input layer, a hidden layer and an output layer, takes the fused features as input, and outputs performance and emission parameter prediction values. The hyperparameter optimization module optimizes the neural network structure and training parameters using a Bayesian optimization algorithm. The performance evaluation module compares the prediction results with the simulation data and outputs root mean square error, mean absolute error and goodness-of-fit evaluation indexes.

[0032] A diesel engine performance and emission prediction method based on FTP75 transient working condition comprises the following steps: S1, simulation model construction and data acquisition: based on GT-POWER software, a one-dimensional simulation model of the target diesel engine is constructed, which includes the intake and exhaust systems, combustion chamber, turbocharging and control system, etc. Figure 2 The diesel engine simulation model structure is shown.

[0033] Based on the FTP75 standard dynamic working condition, simulation boundary conditions and control strategies are set, and multi-channel high-resolution time series data and thermal efficiency, effective power, torque and main emission concentration are obtained as training data sources.

[0034] S2, input feature screening and multi-signal fusion: Utilize random forest algorithm to evaluate the importance of the initial 10 input variables, extract the 8 optimal variables highly correlated with the target output, which are engine speed, intake pressure, exhaust pressure, exhaust temperature, intake volume, fuel injection volume, in-cylinder pressure and EGR rate; adopt one-dimensional convolutional neural network to compress the channel and nonlinearly fuse the original high-dimensional time series input signal, to generate a low-dimensional fusion feature vector; the structure of one-dimensional convolutional neural network is as shown in Figure 3

[0035] S3, design and construction of deep neural network structure: construct a multi-output feedforward neural network model, the structure is as shown in Figure 4 S4, Bayesian optimization: the process is as shown in Figure 5 The following optimization variables are configured by using the bayesopt function of MATLAB: number of hidden layers (1-5), number of nodes (10-100), type of activation function (ReLU, Tanh, Sigmoid), initialization method (He, Glorot, Narrow-Normal) and maximum training rounds (50-300), and the objective function is set as the weighted combination of MAE and MSE on the validation set. The optimization process is as shown in Figure 5

[0036] The optimal structure of the neural network prediction module is 4 nodes in the input layer, 26-89-87-36 nodes in the hidden layer, 6 nodes in the output layer, the activation function combination is Tanh-ReLU-ReLU-Tanh, the weight initialization adopts Glorot method. Training rounds: 129, Mini-Batch Size=32; optimization tools: MATLAB 2024a Deep Learning Toolbox + Statistics and Machine Learning Toolbox.

[0037] ​​​​S5, model evaluation and deployment: the root mean square error, mean absolute error and goodness of fit index are used to evaluate the fitting accuracy and generalization ability of the model under FTP75 dynamic cycle. Specifically, it includes: predicting on the test set data, outputting the true value and predicted value comparison curve, and calculating the root mean square error RMSE, mean absolute error MAE, goodness of fit R 2 Figures 6-11

[0038] Figure 6 The heat efficiency prediction value and true value comparison curve is shown, and the obtained evaluation index values are MAE = 1.50887, RMSE = 3.4351, R 2 = 0.9488; Figure 7 The torque prediction value and true value comparison curve is shown; the obtained evaluation index values are MAE = 1.8386, RMSE = 3.4657, R 2 = 0.9896; Figure 8 The power prediction value and true value comparison curve is shown; the obtained evaluation index values are MAE = 0.2574, RMSE = 0.5185, R 2 = 0.9917; Figure 9 The carbon monoxide (CO) prediction value and true value comparison curve is shown; the obtained evaluation index values are MAE = 0.0344, RMSE = 0.0459, R 2 = 0.9953; Figure 10 The carbon dioxide (CO2) prediction value and true value comparison curve is shown; the obtained evaluation index values are MAE = 0.0887, RMSE = 0.1179, R 2 = 0.9960; Figure 11 The hydrocarbon (HC) prediction value and true value comparison curve is shown. The obtained evaluation index values are MAE = 0.0004, RMSE = 0.0006, R 2 = 0.9949.

[0039] From the above results, the final results of the embodiment show that the R 2 of each output variable is greater than 0.94, the model training time is reduced by about 30%, the model size is compressed by more than 50%, and the model has good deployment efficiency and accuracy balance.

[0040] ​​Under this optimal embodiment, the proposed method exhibits high fitting accuracy and excellent generalization ability on the test set data of real FTP75 working conditions, and the training time is shortened by about 35%, which can realize the rapid discrimination of diesel engine performance state and the early warning of emission trend in engineering application, and is especially suitable for intelligent control, on-board state diagnosis and energy-saving strategy making.

[0041] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art according to actual conditions to meet different specific actual needs. However, it is obvious to those skilled in the art that the specific details do not have to be used to implement the present application. In other examples, in order to avoid confusion of the present application, well-known components, structures or parts are not specifically described, and are within the technical scope defined in the claims of the present application.

[0042] Changes and variations made by those skilled in the art without departing from the spirit and scope of the present application shall be within the scope of protection of the appended claims of the present application. In the above description, a large number of specific details are set forth in order to provide a thorough understanding of the present application. However, it is obvious to those skilled in the art that the specific details do not have to be used to implement the present application. In other examples, in order to avoid confusion of the present application, well-known technologies are not specifically described, such as specific construction details, operating conditions and other technical conditions.

[0043] In this paper, specific examples are used to illustrate the principles and implementation of the present application. The above examples are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in specific implementation and application range. In summary, the content of this specification should not be understood as a limitation of the present application.

Claims

1. A method for predicting diesel engine performance and emissions based on transient operating conditions, characterized in that, The steps are as follows: S1. Simulation Model Construction and Data Acquisition: A one-dimensional simulation model of the target diesel engine is constructed based on GT-POWER software. Simulation boundary conditions and control strategies are set based on standard dynamic operating conditions. Multi-channel high-resolution time-series data, thermal efficiency, effective power, torque, and concentration of major emissions are acquired as training data sources. S2. Input Feature Filtering and Multi-Signal Fusion: The Random Forest algorithm is used to evaluate the importance of the initially selected multiple input variables and extract key variables that are highly correlated with the target output; a one-dimensional convolutional neural network is used to perform channel compression and nonlinear fusion on the original high-dimensional time-series input signal to generate a low-dimensional fused feature vector. S3. Deep Neural Network Structure Design and Construction: Construct a multi-output feedforward neural network model, using the low-dimensional fused feature vector compressed from a one-dimensional convolutional neural network as input, and outputting six key prediction targets, including thermal efficiency, torque, power, carbon monoxide, carbon dioxide, and hydrocarbons. S4. Bayesian Optimization: Introducing the Bayesian optimization algorithm to automatically optimize the number of hidden layers, number of neurons, activation function type, weight initialization method, and maximum number of training rounds in a neural network; S5. Model Evaluation and Deployment: The root mean square error, mean absolute error, and goodness-of-fit index are used to evaluate the model's fitting accuracy and generalization ability under dynamic cycling.

2. The method for predicting diesel engine performance and emissions based on transient operating conditions according to claim 1, characterized in that: In S1, multi-channel high-resolution timing data includes engine speed, injection pressure, intake and exhaust temperature / pressure, intake volume, injection volume, cylinder pressure, and EGR rate.

3. The method for predicting diesel engine performance and emissions based on transient operating conditions according to claim 1, characterized in that: In S2, the structure of the one-dimensional convolutional neural network is 3 layers of 1D convolution, with the number of channels in each layer being 64→32→4, the kernel size being 1, and the activation function being ReLU.

4. The method for predicting diesel engine performance and emissions based on transient operating conditions according to claim 1, characterized in that: The Bayesian optimization algorithm guides the iterative update of the network structure through a Gaussian surrogate model and the expectation boosting function EI.

5. A diesel engine performance and emission prediction system based on transient operating conditions, characterized in that, include: Diesel engine simulation module: used to build a four-cylinder turbocharged diesel engine model on the GT-POWER platform, set typical transient operating condition boundary conditions, and output multi-channel engine operating data. The data output by the diesel engine simulation module is transmitted to the feature filtering and signal fusion module. Feature selection and signal fusion module: includes a random forest evaluation unit and a one-dimensional convolutional neural network, used to select key variables and perform nonlinear compression and fusion of input signals. The fused features are used as input to the neural network prediction module. Neural network prediction module: A feedforward neural network consisting of an input layer, a hidden layer and an output layer, which takes fused features as input and outputs predicted values ​​of performance and emission parameters; Hyperparameter optimization module: The Bayesian optimization algorithm is used to optimize the neural network structure and training parameters. The neural network prediction module and the hyperparameter optimization module interact bidirectionally, and structure search and model evaluation are performed during the optimization process. Performance evaluation module: Compares the prediction results with the simulation data, and outputs the root mean square error, mean absolute error and goodness of fit evaluation indicators.

6. The diesel engine performance and emission prediction system based on transient operating conditions according to claim 5, characterized in that, The feature selection and signal fusion module first selects eight optimal input variables through a random forest evaluation unit, and then performs feature compression processing on the eight-dimensional input signal through a one-dimensional convolutional neural network to output a four-dimensional fused feature vector.

7. The diesel engine performance and emission prediction system based on transient operating conditions according to claim 5, characterized in that, The optimal structure of the neural network prediction module is as follows: 4 nodes in the input layer, 26-89-87-36 nodes in the hidden layer, 6 nodes in the output layer, Tanh-ReLU-ReLU-Tanh activation function combination, and Glorot method for weight initialization.

Images