Engine charge efficiency virtual perception method based on physical information neural network

Abstract

The application discloses an engine charge efficiency virtual perception method based on a physical information neural network, and comprises the following steps: step 1, constructing a physical model based on the first law of simplified physics; step 2, constructing a feedforward neural network model FNN for engine charge efficiency estimation; step 3, training the physical model first, and then training the feedforward neural network model FNN after fixing the parameters of the physical model; step 4, combining the physical model and the feedforward neural network model FNN to form a physical information neural network PDM; step 5, training the parameters of the feedforward neural network model FNN by using a comprehensive loss function of physical loss and data loss, realizing the fusion of the physical model and the feedforward neural network model FNN, and obtaining the trained physical information neural network PDM; and step 6, applying the trained physical information neural network PDM to predict the charge efficiency of the engine under different operating conditions. The application can effectively improve the prediction accuracy of the engine charge efficiency.

Description

Technical Field

[0001] This invention relates to the fields of mechanical engineering and data science, and in particular to a virtual sensing method for engine charging efficiency based on a physical information neural network. Background Technology

[0002] In the modern automotive industry, engine performance, thermal efficiency, and environmental adaptability are key indicators for measuring its technological advancement. Precise control of charging efficiency not only improves engine thermal efficiency and reduces fuel consumption but also effectively controls emissions, meeting increasingly stringent environmental regulations. Furthermore, effective management of charging efficiency is equally crucial for improving engine responsiveness and driving performance. In turbocharging systems, accurate estimation and control of charging efficiency directly affects the turbocharger's matching effect and the dynamic response characteristics of the entire system. Accurate control and optimization of charging efficiency are fundamental to achieving efficient and environmentally friendly engine design.

[0003] Over the past decade, researchers have made significant progress in modeling engine intake efficiency. These methods can be broadly categorized into three main types: physical model-based methods, data-based methods, and hybrid physical-data-driven methods. Physical models are based on rigorous physical theories and formulas, estimating volumetric efficiency by capturing the fundamental physical laws of the intake process. Kocher et al. (Kocher L, Koeberlein E, Van Alstine DG, et al. Physically based volumetric efficiency model for diesel engines utilizing variable intake valve actuation[J]. International Journal of Engine Research, 2012, 13(2): 169-184.) developed a physical model that estimates volumetric efficiency using the law of energy conservation and has been experimentally verified to be applicable to various engine configurations. Menzel et al. (Menzel G, Och SH, Mariani VC, et al. Multi-objective optimization of the volumetric and thermal efficiencies applied to a multi-cylinder internal combustion engine[J]. Energy Conversion and Management, 2020, 216: 112930.) developed a physical model for optimizing the volumetric efficiency of a multi-cylinder internal combustion engine. This model uses multi-objective optimization methods, including Non-Dominated Sorting Genetic Algorithm II (NSGA-II) and Multi-Objective Differential Evolution (MODE), to find the optimal valve timing settings. The study used a one-dimensional unsteady compressible gas flow model and a zero-dimensional combustion model, and solved and optimized them using numerical methods. While these methods have achieved some success in the past, they generally have limitations, particularly in terms of cost, time efficiency, and adaptability to dynamic operating conditions.

[0004] With the rapid development of big data and high-performance computing technologies, data-driven methods have opened up new avenues for optimizing engine performance. By analyzing massive amounts of engine operating data, machine learning algorithms can reveal complex data patterns and potential performance optimization opportunities. The literature (Uzun A. Air mass flow estimation of diesel engines using neural network[J]. Fuel,2014,117:833-838.) uses recurrent neural networks to predict air mass flow, considering crankshaft angle, load, and engine speed as inputs. The results show high accuracy, but are overly dependent on the quantity and quality of the data. The paper (Luján JM, Climent H, García-Cuevas LM, et al. Volumetric efficiency modelling of internal combustion engines based on a novel adaptive learning algorithm of artificial neural networks[J]. Applied Thermal Engineering, 2017, 123: 625-634.) proposes a novel adaptive learning algorithm that improves the limitations of the standard gradient descent method by increasing the update speed of hidden layer weights. By applying different learning rates between network layers, the learning speed of the network is improved. However, this method has only been validated within a limited range of operating conditions, and its generalization ability is not strong enough. The paper (Li J, Zhou Q, Williams H, et al. Fuzzy-tree-constructed data-efficient modelling methodology for volumetric efficiency of dedicated hybrid engines[J]. Applied Energy, 2022, 310: 118534.) proposes a method based on hierarchical fuzzy inference trees. The paper simplifies model construction by combining several low-dimensional fuzzy inference systems, and develops a Gaussian distribution-based resampling technique to maintain sample diversity by selecting a small number of diverse samples.The literature (Gao J, Liu X, Sun B, et al. In-cylinder air mass flow estimation of gasoline engines based on map self-learning[J]. Control Engineering Practice, 2023, 140: 105674.) proposes an estimation method based on MAP self-learning. This method eliminates periodic fluctuations in the signal by designing appropriate filtering methods and uses an adaptive observer to estimate the system state and volumetric efficiency. These data-driven optimization strategies not only provide real-time performance monitoring and prediction but also reveal performance improvement points that are difficult to discover using traditional methods. However, although data-driven methods have significant advantages in handling large-scale datasets and discovering nonlinear relationships, they often neglect the underlying physical principles. Sometimes, they may overlook the fundamental laws of engine operation due to over-reliance on data, thus limiting the generality and accuracy of the predictive model.

[0005] In summary, there is an urgent need for a new method that combines engine physical models with advanced data-driven technologies to achieve high-precision virtual perception of engine charging efficiency, thus supporting high-precision engine control. Therefore, this invention proposes an innovative solution that ingeniously combines the rigor of physical models with the flexibility of data-driven models (feedforward neural network models, FNN). This invention constructs a physical model based on a simplified first law of physics and combines it with an FNN model built using advanced machine learning techniques. It leverages the combined advantages of physical models and FNN models through a physical information neural network, thereby achieving high-precision prediction of engine charging efficiency. Summary of the Invention

[0006] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a virtual sensing method for engine charging efficiency based on a physical information neural network.

[0007] The technical solution adopted to achieve the purpose of this invention is:

[0008] A virtual sensing method for engine charging efficiency based on physical information neural networks includes the following steps:

[0009] Step 1: Construct a physical model based on the simplified first law of physics. The physical model estimates the engine charging efficiency based on the engine speed, intake manifold pressure, exhaust pressure, and intake manifold temperature.

[0010] Step 2: Construct a feedforward neural network model (FNN) for estimating engine charging efficiency to fully capture the nonlinear characteristics in engine operating data. The inputs of the feedforward neural network model (FNN) are engine speed, intake manifold pressure, exhaust pressure and intake manifold temperature, and the output is engine charging efficiency.

[0011] Step 3: First train the physical model from Step 1, fix its parameters, and then further train the feedforward neural network model FNN from Step 2.

[0012] Step 4: Combine the physical model from Step 1 and the feedforward neural network model (FNN) from Step 3 to construct the Physical Information Neural Network (PDM).

[0013] Step 5: Train the parameters of the feedforward neural network model FNN using the combined loss function of physical loss and data loss, thereby achieving the fusion of the physical model and the feedforward neural network model FNN, and obtaining the trained physical information neural network PDM.

[0014] Step 6: Apply the trained physical information neural network (PDM) to predict the engine's charging efficiency under different operating conditions.

[0015] In the above technical solution, the physical model in step 1 is:

[0016]

[0017] Where n is the engine speed, P im P is the intake manifold pressure. em T is the exhaust pressure. im For intake manifold temperature, θ1-θ 13 Here, η represents the physical model parameters, and η is the estimated value of the engine's charging efficiency.

[0018] In the above technical solution, the structure of the feedforward neural network model FNN in step 2 includes an input layer, multiple hidden layers and an output layer. The hidden layer consists of multiple linear layers and nonlinear activation functions. The linear part of each hidden layer is implemented by nn.Linear, and the LeakyReLU activation function is applied to the output of each linear part.

[0019] In the above technical solution, in step 3, the nonlinear least squares method is used to adjust the parameters θ1-θ of the physical model. 13 .

[0020] In the above technical solution, in step 3, when training the feedforward neural network model FNN, the mean squared error (MSE) is used as the loss function. The loss function calculates the difference between the predicted value and the actual measured value of the feedforward neural network model FNN, and the optimization algorithm uses the first Adam optimizer.

[0021] In the above technical solution, in step 3, after the physical model parameters are optimized, the values ​​of the physical parameters are locked using a freezing technique, and then the non-physical parameters of the feedforward neural network model FNN are trained. The performance of the feedforward neural network model FNN is further optimized through a data-driven method.

[0022] if θ∈Θ physics thenθ.require_grad = False

[0023] Where, Θ physics The parameters θ1-θ of all physical models 13 A set of.

[0024] In the above technical solution, the first Adam optimizer is:

[0025]

[0026] Where, θ next Let θ represent the parameters of the optimized feedforward neural network model FNN, and l represent the parameters of the feedforward neural network model FNN. r It is the learning rate, and Adam(·) represents the update function of the first Adam optimizer.

[0027] In the above technical solution, in step 5, the comprehensive loss function is:

[0028] L total =(1-λ) physics )·L data +λ physics ·L physics

[0029]

[0030] Where: N is the number of samples, y physics,i y is the prediction value of the physical model for the i-th sample. pred,i It is the prediction of the i-th sample by the feedforward neural network model FNN, y true,i L is the measured inflation efficiency of the i-th sample. data L represents the data loss, which is the mean square error between the predicted values ​​and the actual measured values ​​of the feedforward neural network model (FNN). physics The physical loss represents the mean square error between the predicted values ​​of the physical model and the actual measured values, λ. physics It is a weighting coefficient used to balance the two types of losses, and its value is between 0 and 1.

[0031] In the above technical solution, in step 5, the second Adam optimizer is used to train the fusion loss to minimize the overall error between the predicted value and the actual measured value.

[0032] In the above technical solution, the second Adam optimizer is Where, θ next Let θ represent the parameters of the optimized feedforward neural network model FNN, and l represent the parameters of the feedforward neural network model FNN. r It is the learning rate, and Adam(·) represents the update function of the second Adam optimizer.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] This invention innovatively combines the accuracy of physical models with the flexibility of data-driven models (feedforward neural network models, FNN). By integrating the advantages of different models, it aims to achieve high-precision prediction of engine charging efficiency, while reducing computational costs and improving the model's adaptability to complex environments. Through this method, engine charging efficiency can be virtually sensed in real time with lower cost and higher accuracy, thus providing strong technical support for the optimization and control of engine performance. Attached Figure Description

[0035] Figure 1 PDM architecture diagram;

[0036] Figure 2 Comparison of prediction results from physical models, FNN models, and PDM with measured data;

[0037] Figure 3 A schematic diagram illustrating the relative errors between the physical model, the FNN model, and the PDM and the measured data;

[0038] Figure 4 A schematic diagram illustrating the correlation between the physical model, the FNN model, the PDM, and the measured data;

[0039] Figure 5 Relative error distribution between physical model, FNN model and PDM and measured data. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0041] like Figure 1 As shown, a virtual sensing method for engine charging efficiency based on a physical information neural network includes the following steps:

[0042] Step 1: Construct a physical model based on the simplified first law of physics using domain knowledge. This physical model estimates the engine charging efficiency based on engine speed, intake manifold pressure, exhaust pressure, and intake manifold temperature.

[0043] The physical model is shown in equation (1):

[0044]

[0045] Where n is the engine speed, P im P is the intake manifold pressure. em T is the exhaust pressure. im For intake manifold temperature, θ1-θ 13 Here are the physical model parameters, and η is the estimated value of the engine charging efficiency;

[0046] Step 2: In order to fully capture the nonlinear characteristics in the engine operating data and establish an accurate mapping relationship between input parameters and charging efficiency, a feedforward neural network (FNN) model for charging efficiency estimation is constructed.

[0047] The feedforward neural network model (FNN) consists of an input layer, multiple hidden layers, and an output layer, specifically designed to handle multidimensional engine operating parameters. The model's input parameters include engine speed, intake manifold pressure, exhaust pressure, and intake manifold temperature (obtained through automatic feature engineering and contrastive learning). The hidden layers consist of multiple linear layers and nonlinear activation functions. The linear part of each hidden layer is implemented using `nn.Linear`, responsible for performing linear transformations in the feature space. Subsequently, the LeakyReLU activation function is applied to the output of each linear part to introduce nonlinearity, which helps the network learn more complex function mappings. The LeakyReLU design avoids the vanishing gradient problem during training, allowing the network to learn more effectively during optimization. The output parameter of the feedforward neural network model (FNN) is the engine charging efficiency.

[0048] Step 3, when training the physical model from Step 1, is based on engine operating parameters or historical data, specifically including inputting the intake manifold pressure P. im Intake manifold temperature T im Engine speed n and exhaust pressure P em Data information is used to adjust the parameters θ1-θ of the physical model using the nonlinear least squares method. 13 This is to enable it to accurately output an estimate of inflation efficiency;

[0049] In the physical model, the parameters θ1-θ 13Once determined, the next step is to train the feedforward neural network model (FNN) using Mean Squared Error (MSE) as the loss function. This function calculates the difference between the predicted values ​​of the FNN and the actual measured values ​​(data samples generated from bench tests). The optimization algorithm employs the first Adam optimizer, an adaptive learning rate optimization method. The Adam optimizer automatically adjusts the learning rate during training, thereby accelerating convergence and improving the prediction accuracy of the FNN.

[0050] Step 4: The physical model from Step 1 is combined with the feedforward neural network model (FNN) from Step 2 to form a Physical Information Data-Driven Model (PDM). The physical model from Step 1 is used to make a basic estimate of the inflation efficiency. The parameters θ1-θ of these feedforward neural network models are... 13 Adjustments and optimizations are made to learn latent patterns and nonlinear relationships in the data. This hybrid-driven modeling approach not only preserves the interpretability of the physical model but also enhances its ability to capture complex phenomena through data-driven adjustments.

[0051] Step 5: Train the parameters of the feedforward neural network model FNN using a combined loss function of physical loss and data loss to achieve the fusion of the physical model and the feedforward neural network model FNN, resulting in the trained physical information neural network model PDM. Use the second Adam optimizer to train the fusion loss, minimizing the overall error between the predicted value and the actual measurement value. Simultaneously consider data loss and physical loss, and adjust the relative weights between physical parameters and data-driven parameters to enhance the model's ability to capture complex data patterns while preserving the accuracy of the physical model.

[0052] The overall loss function is:

[0053] L total =(1-λ) physics )·L data +λ physics ·L physics (2)

[0054]

[0055] Where: N is the number of samples, y physics,i y is the prediction value of the physical model for the i-th sample. pred,i It is the prediction of the i-th sample by the feedforward neural network model FNN, y true,i L is the measured inflation efficiency of the i-th sample. dataThis represents data loss, specifically the mean squared error between the predicted and actual measured values ​​of the feedforward neural network (FNN) model. L physics λ represents the physical loss, which is the mean square error between the predicted values ​​of the physical model and the actual measured values. physics It is a weighting coefficient used to balance the two types of losses, and its value is between 0 and 1.

[0056] In step 5, the second Adam optimizer, as shown in formula (5), ensures that the feedforward neural network model FNN can be accurately adjusted in different stages of training, thereby improving the accuracy and robustness of the overall prediction model.

[0057]

[0058] Where, θ next Let θ represent the parameters of the optimized feedforward neural network model FNN, and l represent the parameters of the feedforward neural network model FNN. r It is the learning rate, and Adam(·) represents the update function of the second Adam optimizer.

[0059] Step 6: Apply the trained PDM to predict the charging efficiency of the engine under different operating conditions.

[0060] Example 2

[0061] This embodiment is a further optimization based on Embodiment 1.

[0062] In step 3, after optimizing the physical model parameters, the value of the physical parameters is locked by freezing technology, which is expressed as formula (6). Then, the non-physical parameters of the feedforward neural network model FNN are trained to ensure that physical knowledge is effectively utilized and remains consistent. At the same time, the model performance is further optimized by data-driven methods.

[0063] if θ∈Θ physics ,thenθ.require_grad=False (6)

[0064] Where, Θ physics The parameters θ1-θ of all physical models 13 A set;

[0065] The first Adam optimizer in step 3 is

[0066]

[0067] Where, θ next Let θ represent the parameters of the optimized feedforward neural network model FNN, and l represent the parameters of the feedforward neural network model FNN. r It is the learning rate, and Adam(·) represents the update function of the Adam optimizer.

[0068] Example 3

[0069] The physical model, the feedforward neural network model (FNN), and the physical information neural network model (PDM) were validated, and the results were obtained by... Figures 2-5 .

[0070] Figure 2 To illustrate the comparison between the estimates from the physical model, FNN model, and PDM model and the measured inflation efficiency values, the blue dots in the figure represent the measured inflation efficiency values, the green line represents the physical model estimate, the purple line represents the FNN model estimate, and the red line represents the PDM model estimate. It is clear from the figure that the PDM model's prediction curve most closely matches the actual measured value, indicating that it has the highest prediction accuracy.

[0071] Figure 3 The graph shows the relative errors between the estimates from the physical model, FNN model, and PDM model and the measured inflation efficiency. The green line represents the relative error between the physical model estimate and the measured inflation efficiency, the purple line represents the relative error between the FNN model estimate and the measured inflation efficiency, and the red line represents the relative error between the PDM model estimate and the measured inflation efficiency. The graph shows that the relative error of the PDM model is generally lower than that of the physical model and the FNN model, exhibiting a smaller fluctuation range. This means that the PDM model's predictions are more accurate at each measurement point.

[0072] Figure 4 The graph shows the correlation between the physical model, the FNN model, and the PDM model and the measured inflation efficiency. The black line represents the measured inflation efficiency, the green dots represent the physical model estimates, the purple dots represent the FNN model estimates, and the red dots represent the PDM model estimates. The graph shows that the PDM model's correlation plot (red dots) exhibits the tightest distribution, with all points closely clustered around the diagonal. This indicates the strongest linear relationship between the PDM model's predictions and the actual measurements.

[0073] Figure 5 The frequency distribution of relative errors for the physical model, neural network model, and PDM model is shown. The green bars represent the frequency distribution of relative errors between the physical model estimates and the measured inflation efficiency values, the purple bars represent the frequency distribution of relative errors between the FNN model estimates and the measured inflation efficiency values, and the red bars represent the frequency distribution of relative errors between the PDM model estimates and the measured inflation efficiency values.

[0074] The relative errors of the physical model estimates to the measured inflation efficiency values ​​accounted for 41.03% of the total estimates within ±1% and 71.79% within ±2%. The relative errors of the FNN model estimates to the measured inflation efficiency values ​​accounted for 78.21% of the total estimates within ±1% and 87.82% within ±2%. The relative errors of the PDM model estimates to the measured inflation efficiency values ​​accounted for 80.13% of the total estimates within ±1% and 94.23% within ±2%. In summary, the PDM model exhibits the most concentrated relative error distribution, significantly biased towards the zero error line, indicating a significant advantage in minimizing prediction errors. Compared to the physical model and the FNN model, the PDM model not only performs better in terms of error magnitude but also demonstrates superior prediction stability. This high generalization ability means that the PDM model can adapt to different operating conditions, providing accurate real-time predictions of engine inflation efficiency.

[0075] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A virtual sensing method for engine charging efficiency based on a physical information neural network, characterized in that, Includes the following steps: Step 1: Construct a physical model based on the simplified first law of physics. The physical model estimates the engine charging efficiency based on the engine speed, intake manifold pressure, exhaust pressure, and intake manifold temperature. Step 2: Construct a feedforward neural network model (FNN) for estimating engine charging efficiency to fully capture the nonlinear characteristics in engine operating data. The inputs of the feedforward neural network model (FNN) are engine speed, intake manifold pressure, exhaust pressure and intake manifold temperature, and the output is engine charging efficiency. Step 3: First train the physical model from Step 1, fix its parameters, and then further train the feedforward neural network model FNN from Step 2. Step 4: Combine the physical model from Step 1 and the feedforward neural network model (FNN) from Step 3 to construct the Physical Information Neural Network (PDM). Step 5: Train the parameters of the feedforward neural network model FNN using the combined loss function of physical loss and data loss, thereby achieving the fusion of the physical model and the feedforward neural network model FNN, and obtaining the trained physical information neural network PDM. Step 6: Apply the trained physical information neural network (PDM) to predict the engine's charging efficiency under different operating conditions.

2. The virtual sensing method for engine charging efficiency based on physical information neural network as described in claim 1, characterized in that, The physical model in step 1 is as follows: in, Engine speed, For intake manifold pressure, For exhaust pressure, Intake manifold temperature, - For physical model parameters, This is an estimate of the engine's charging efficiency.

3. The virtual sensing method for engine charging efficiency based on physical information neural network as described in claim 1, characterized in that, The structure of the feedforward neural network model FNN in step 2 includes an input layer, multiple hidden layers, and an output layer. The hidden layers consist of multiple linear layers and non-linear activation functions. The linear part of each hidden layer is implemented by nn.Linear, and the LeakyReLU activation function is applied to the output of each linear part.

4. The virtual sensing method for engine charging efficiency based on physical information neural network as described in claim 1, characterized in that, In step 3, the parameters of the physical model are adjusted using the nonlinear least squares method. - .

5. The virtual sensing method for engine charging efficiency based on a physical information neural network as described in claim 4, characterized in that, In step 3, when training the feedforward neural network model FNN, the mean squared error (MSE) is used as the loss function. The loss function calculates the difference between the predicted value and the actual measured value of the feedforward neural network model FNN. The optimization algorithm uses the first Adam optimizer.

6. The virtual sensing method for engine charging efficiency based on a physical information neural network as described in claim 5, characterized in that, In step 3, after optimizing the physical model parameters, a freezing technique is used to lock the values ​​of the physical parameters, and then the non-physical parameters of the feedforward neural network model FNN are trained. The performance of the feedforward neural network model FNN is further optimized through a data-driven approach. in, These are the parameters of all physical models. - A set of.

7. The virtual sensing method for engine charging efficiency based on physical information neural network as described in claim 5, characterized in that, The first Adam optimizer is: in, This represents the parameters of the optimized feedforward neural network model FNN. The parameters represent the feedforward neural network model FNN. It's the learning rate. This represents the update function of the first Adam optimizer. This indicates data loss.

8. The virtual sensing method for engine charging efficiency based on physical information neural network as described in claim 1, characterized in that, In step 5, the comprehensive loss function is: in: It is the sample size. It is a physical model for the first The predicted value for each sample, It is the feedforward neural network model FNN for the first Prediction for each sample It is the first Measured inflation efficiency values ​​for each sample. This represents the data loss, which is the mean squared error between the predicted values ​​and the actual measured values ​​of the feedforward neural network model (FNN). Physical loss represents the mean squared error between the predicted values ​​of the physical model and the actual measured values. It is a weighting coefficient used to balance the two types of losses, and its value is between 0 and 1.

9. The virtual sensing method for engine charging efficiency based on a physical information neural network as described in claim 1, characterized in that, In step 5, the second Adam optimizer is used to train the fusion loss to minimize the overall error between the predicted value and the actual measured value.

10. The virtual sensing method for engine charging efficiency based on a physical information neural network as described in claim 9, characterized in that, The second Adam optimizer is ,in, This represents the parameters of the optimized feedforward neural network model FNN. The parameters represent the feedforward neural network model FNN. It's the learning rate. This represents the update function of the second Adam optimizer.

Images