A model and method for predicting gasoline fuel spray characteristics of a gdi engine

Abstract

The application discloses a GDI engine gasoline fuel spray characteristic prediction model and method, comprising: a feature extraction layer, a time sequence modeling layer, a feature weighting layer and an output layer; the feature extraction layer is a two-layer full connection structure, comprising a first full connection layer and a second full connection layer; the time sequence modeling layer receives a feature sequence output by the feature extraction layer and comprises a forget gate, an input gate and an output gate; the feature weighting layer is located behind the time sequence modeling layer, the feature weighting layer introduces an attention mechanism and is used for secondary processing of a sequence hidden state output by the time sequence modeling layer; the output layer is connected with the feature weighting layer, the prediction of a spray characteristic parameter is realized by mapping the weighted feature sequence; the model can learn the dependent relationship behind the development of a spray from the start of injection to the complete spray formation process, and the prediction performance reflects the physical authenticity of the spray characteristics.

Description

Technical Field

[0001] This invention relates to the field of engine technology, and in particular to a predictive model and method for gasoline fuel spray characteristics of GDI engines. Background Technology

[0002] Direct injection technology, by precisely controlling the fuel spray process to improve combustion efficiency, has become the mainstream technology in the gasoline engine field. Fuel spray characteristics are a key factor affecting the combustion organization, energy conversion efficiency, and pollutant emissions of internal combustion engines; their quality directly determines the formation effect of the air-fuel mixture, flame propagation speed, and combustion temperature distribution. Accurate prediction of spray characteristics not only provides a scientific basis for the optimized design of fuel injectors but is also an important foundation for developing efficient fuel injection strategies and achieving intelligent engine control.

[0003] In traditional studies of gasoline spray characteristics, researchers primarily rely on empirical formulas and computational fluid dynamics (CFD) simulations. While empirical formulas can describe the macroscopic behavior of spray within a specific range, their applicability and generalizability are significantly limited when facing complex environments with multiple operating conditions and coupled parameters. Although CFD models can provide a more in-depth description of the physical process, their modeling process is complex, computationally expensive, and highly dependent on numerous physical assumptions and empirical parameters, making it difficult to achieve rapid and efficient predictions in engineering practice. Furthermore, because the spray process involves coupling across multiple length and time scales from the micrometer scale to the combustion chamber scale, it exhibits strong nonlinearity and uncertainty, often making it difficult for traditional methods to achieve an ideal balance between accuracy and efficiency.

[0004] In recent years, data-driven methods, represented by deep learning, have demonstrated significant advantages in modeling complex nonlinear systems, providing new research ideas for spray characteristic prediction. Although some studies have attempted to capture the relationships between spray parameters using machine learning algorithms, existing prediction methods often struggle to simultaneously consider the nonlinear mapping of static operating conditions and the long- and short-term dependencies in the dynamic evolution process when processing spray evolution data with significant temporal characteristics. Especially under extreme operating conditions such as high pressure and variable temperature, effectively identifying and weighting the key characteristic information affecting spray development remains a challenge in the current technological field. Summary of the Invention

[0005] This application provides a prediction scheme for spray characteristics based on a deep learning heterogeneous model, through a gasoline fuel spray characteristic prediction model and method for GDI engines. In existing internal combustion engine development, traditional empirical formulas suffer from limited applicability and insufficient prediction accuracy when dealing with strongly nonlinear spray behavior under multiple operating conditions and high-dimensional parameter coupling. While numerical simulation methods based on computational fluid dynamics offer a certain depth of physical description, their computational process involves complex mesh generation and multiphase flow coupling, resulting in lengthy computation cycles. Furthermore, the models are highly sensitive to boundary conditions and empirical parameter settings, making it difficult to meet the design requirements of real-time control or rapid iteration. In addition, existing prediction methods based on single machine learning algorithms often neglect the deep correlation between static operating characteristics and dynamic evolution laws when processing spray evolution data with significant time-series characteristics, leading to prediction biases at key nodes in spray development.

[0006] This application provides a gasoline fuel spray characteristic prediction model for a GDI engine, which includes, in order of feature flow direction: feature extraction layer, temporal modeling layer, feature weighting layer, and output layer. The feature extraction layer is a two-layer fully connected structure, including a first fully connected layer and a second fully connected layer; the input vector of the feature extraction layer is a 4-dimensional vector, the first fully connected layer maps the 4-dimensional input to 128 dimensions, the second fully connected layer expands the 128 dimensions to 256 dimensions, and then compresses it to a 128-dimensional feature sequence through linear transformation; The temporal modeling layer receives the feature sequence output by the feature extraction layer, and includes a forget gate, an input gate, and an output gate. The feature weighting layer is located after the temporal modeling layer. The feature weighting layer introduces an attention mechanism to perform secondary processing on the sequence hidden state output by the temporal modeling layer. The output layer is connected to the feature weighting layer, and the spray characteristic parameters are predicted by mapping the weighted feature sequence.

[0007] Preferably, the input vector is: injection pressure, ambient pressure, ambient temperature, and injection pulse width.

[0008] Preferably, the first fully connected layer is followed by a first normalization layer, a first activation function, and a Dropout layer in sequence; the second fully connected layer is followed by a second normalization layer and a second activation function in sequence.

[0009] Preferably, the first layer activation function and the second layer activation function are LeakyReLU, and their negative slope parameter is set to 0.2.

[0010] Preferably, the dropout layer has a drop rate of 0.2.

[0011] Preferably, the attention mechanism calculates the score of the hidden state at each time step, generates a weight distribution using the Softmax function, and sums the weights with the corresponding hidden states to generate a context vector containing global temporal key information.

[0012] Preferably, the spray characteristic parameters are spray penetration distance and spray cone angle.

[0013] This application also proposes a method for predicting the gasoline fuel spray characteristics of a GDI engine using the aforementioned GDI engine gasoline fuel spray characteristic prediction model, including: Step S1: First, construct a spray acquisition system and collect raw spray data. The spray acquisition system consists of a fixed container body, a fuel supply system, an environmental pressure regulation system, an environmental temperature regulation system, a fuel injection control system, and a high-speed camera system. Step S2: Image preprocessing and feature parameter extraction are performed. Background subtraction is used to eliminate the interference of brackets and window shadows inside the fixed volume chamber. A random sampling consistency algorithm is used to randomly extract a subset from the edge pixel set for model fitting. The distance from the remaining points to the fitted model is calculated. Points with a distance exceeding the preset deviation value are identified as outliers and removed. A dataset containing injection pressure, ambient pressure, temperature, injection pulse width and corresponding time series is constructed. Step S3: Construct a prediction model for gasoline fuel spray characteristics of GDI engine; including a feature extraction layer, a temporal modeling layer, a feature weighting layer, and an output layer; Step S4: Divide the preprocessed dataset from step S2 into a training set and a test set in an 8:2 ratio; use the coefficient of determination R0. 2 Mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators.

[0014] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. This application adopts a feature extraction layer consisting of two fully connected layers, batch normalization, LeakyReLU activation function and Dropout layer, which solves the technical problem of low input parameter dimension and insufficient information density, making it difficult to deeply map static working condition parameters, and realizes the efficient expansion of injection pressure, ambient pressure, ambient temperature and injection pulse width into high-dimensional feature sequences.

[0015] 2. This application adopts an LSTM temporal modeling layer containing a forget gate, an input gate, and an output gate, as well as a feature weighting layer based on an attention mechanism. This solves the problem of gradient vanishing in the spray evolution process and the technical problem that traditional models cannot distinguish the importance of different stages, and achieves accurate capture of the dynamic dependency between spray penetration distance and cone angle.

[0016] 3. This application uses the RANSAC random sampling consensus algorithm to iteratively fit and remove outliers from the edge pixels of the image, which solves the technical problems of false edges and abnormal data interference caused by droplet splashing and light refraction in the experimental image. It improves the physical realism of the spray penetration distance and cone angle parameters, and enhances the generalization ability and robustness of the model.

[0017] 4. The GDI engine gasoline fuel spray characteristic prediction model proposed in this application solves the technical problem that traditional simulation methods rely on computer empirical formulas, which lead to limited applicability under various nonlinear operating conditions. Through the design of feature extraction layer, time series modeling layer and feature weighting layer, it improves the inference time of a single prediction task. It can be integrated into the control algorithm of the engine control unit or used as a rapid screening tool in the injector design stage, shortening the iteration cycle of fuel injection strategy optimization and hardware structure improvement. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the spray collection system of this application; Figure 2 This is a schematic diagram of the image preprocessing workflow of this application; Figure 3 This is a schematic diagram of the RANSAC algorithm flow in this application; Figure 4 This is a schematic diagram of the structure of the GDI engine gasoline fuel spray characteristic prediction model of this application; Figure 5 This is a flowchart of the data processing for the GDI engine gasoline fuel spray characteristic prediction model in this application; Figure 6 This is a training loss curve for the model in this application; Figure 7 The figure shows a comparison between the four models and the actual experimental data under the conditions of chamber pressure of 0.1 MPa, injection pressure of 3 MPa, temperature of 20℃, and injection pulse width of 0.1 ms. Figure 8 The figure shows a comparison between the four models and real experimental data under the conditions of chamber pressure of 1 MPa, injection pressure of 3 MPa, temperature of 20℃, and injection pulse width of 0.1 ms. Detailed Implementation

[0019] This invention provides a prediction model and method for gasoline fuel spray characteristics of GDI engines. By constructing a high-precision experimental acquisition system and an advanced deep learning heterogeneous network, it solves the problems of low computational efficiency and insufficient accuracy of traditional physical models under complex operating conditions. To better understand the above technical solution, the following will describe it in detail with reference to the accompanying drawings and specific implementation methods.

[0020] like Figure 1 As shown, in step S1, a spray acquisition system is first constructed and raw spray data is collected. The spray acquisition system consists of a fixed container body, a fuel supply system, an environmental pressure regulation system, an environmental temperature regulation system, a fuel injection control system, and a high-speed camera system.

[0021] The fixed container body serves as the reaction chamber, and its interior is designed as a sealed space resistant to high pressure and high temperature. The fuel supply system uses a high-pressure fuel pump to pressurize the gasoline medium required for the experiment to preset injection pressures of 3 MPa, 5 MPa, 7 MPa, 8 MPa, and 9 MPa, respectively, and then delivers the gasoline medium to the injector installed on the top of the fixed container through pipelines. The environmental pressure regulation system simulates the combustion chamber pressure environment near the top dead center of an engine piston by filling the fixed container body with high-pressure nitrogen or other inert gases to 0.1 MPa, 0.2 MPa, 0.3 MPa, 0.5 MPa, and 1 MPa, respectively. The environmental temperature regulation system integrates an electric heating rod and a high-precision thermocouple feedback module, and maintains the gas temperature inside the fixed container at a constant state of 20℃, 30℃, 40℃, 50℃, 60℃, 70℃, and 80℃ through a closed-loop control algorithm. The fuel injection control system, based on dSPACE, precisely controls the opening time and duration of the fuel injector 8, i.e., the fuel injection pulse width. In this embodiment, the fuel injection pulse width is set at 0.1ms, 0.2ms, 0.3ms, 0.5ms, and 1ms. A high-speed camera system captures images of the entire process of the fuel injector's ejection moment and subsequent evolution.

[0022] Step S2 involves image preprocessing and feature parameter extraction after acquiring the original image. For example... Figure 2 and Figure 3 As shown, the image processing program runs on a computer terminal. First, it eliminates interference from the brackets and window shadows inside the fixed-volume chamber using background subtraction. To separate the outline of the gasoline spray 7 from the complex background noise, the program binarizes the image. In the edge detection stage, this invention introduces the Random Sample Consensus (RANSAC) algorithm. This algorithm randomly selects a subset from the edge pixel set for model fitting and calculates the distance from the remaining points to the fitted model. Points with a distance exceeding a preset deviation value are identified as outliers and removed. This processing logic effectively solves the problem of false edges caused by fuel droplet splashing or light refraction, ensuring that the extracted macroscopic characteristic parameters have extremely high physical authenticity. The parameters extracted in this embodiment include the spray penetration distance (STP) and the spray cone angle (SCA). The spray penetration distance (STP) is defined as the farthest geometric distance between the gasoline spray tip and the injector nozzle outlet, and the spray cone angle (SCA) is defined as the angle formed by the spray edge line and the nozzle central axis at the half-penetration distance position. By processing experimental examples, a complete dataset containing injection pressure, ambient pressure, temperature, injection pulse width, and corresponding time series was constructed.

[0023] Step S3: Construct a prediction model for the gasoline fuel spray characteristics of the GDI engine. For example... Figure 4 , Figure 5 As shown in the figure, the GDI engine gasoline fuel spray characteristic prediction model proposed in this embodiment, namely the MLP-LSTM-Attention model, includes a feature extraction layer, a temporal modeling layer, a feature weighting layer, and an output layer.

[0024] The feature extraction layer employs a multilayer perceptron (MLP) architecture, with its input vector consisting of four dimensions: injection pressure, ambient pressure, ambient temperature, and injection pulse width. Specifically, the feature extraction layer is configured as a two-layer fully connected structure. The first fully connected layer maps the 4-dimensional input to a 128-dimensional high-dimensional space, followed by batch normalization to adjust the data distribution and prevent gradient vanishing during training. The activation function used is LeakyReLU, with a negative slope parameter set to 0.2 to preserve feature information in negative value regions. A Dropout layer with a dropout rate of 0.2 is then connected to improve the model's generalization performance. The second fully connected layer further expands the feature space to 256 dimensions, then compresses it to 128 dimensions through linear transformation, providing rich semantic representations for subsequent temporal processing.

[0025] The temporal modeling layer follows the feature extraction layer and uses a Long Short-Term Memory (LSTM) network to model the evolution of the spray over time. This layer receives the feature sequence output from the feature extraction layer and contains forget gates, input gates, and output gates to capture the long- and short-term dependencies in the spray's evolution over time. The forget gate discards redundant temporal information based on the current input and the hidden state from the previous time step; the input gate updates the spray state information at the current time step; and the output gate extracts features that significantly influence subsequent spray behavior at the current time step. This model characterizes the dynamic growth of the spray penetration distance over time and the transient pulsation characteristics of the spray cone angle.

[0026] The feature-weighted layer introduces an attention mechanism, which follows the temporal modeling layer and is used to further process the sequence hidden states output by the temporal modeling layer. The attention mechanism calculates the score of the hidden state at each time step, generates a weight distribution using the Softmax function, and then sums the weights with the corresponding hidden states to generate a context vector containing key global temporal information. This mechanism enables the model to automatically focus on the critical historical moments that have the greatest impact on the prediction results when predicting the spray characteristics at a specific time, thereby further suppressing the interference of irrelevant noise on the prediction results.

[0027] Step S4 involves parameter optimization and performance evaluation. In this embodiment, the dataset is divided into two parts: 80% for training and 20% for testing.

[0028] During model training, the coefficient of determination R is used. 2 Mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation metrics. Generally, R... 2 The closer the values ​​are to 1, the smaller the MAE and RMSE, the better the model fit, the higher the accuracy, and the smaller the difference between the predicted and measured values. The calculation formulas are shown below: In the formula, For predicted values, This is the actual value. The average of the actual values. This represents the number of data points.

[0029] To verify the effectiveness of the MLP-LSTM-Attention model, a gasoline fuel spray dataset was input into this model and other comparative models. These compared models included Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGB), and Random Forest (RF), three commonly used machine learning algorithms. To ensure the fairness and comparability of the experimental results, all models used the same data preprocessing methods, training and test set partitioning, and consistent input features for training and testing. Based on this, R... 2 MAE and RMSE were used as evaluation indicators, and the final results are shown in the table below.

[0030] Table 3 Performance comparison of MLP-LSTM-Attention, MLP, XGB, and RF: The MLP-LSTM-Attention fusion prediction model proposed in this invention has a determination coefficient R0 2 The accuracy reached 0.9632, with a mean absolute error (MAE) of only 0.169 and a root mean square error (RMSE) of 0.221. Compared to the single MLP model (R²=0.8856), the extreme gradient boosting (XGB) model (R²=0.8643), and the random forest (RF) model (R²=0.8252), the model of this invention exhibits a significant advantage in prediction accuracy, achieving an accurate understanding of the macroscopic characteristics of spraying.

[0031] like Figure 6 As shown, the training loss curve indicates that the dataset is large enough to train the system and provide accurate predictions of spray penetration distance.

[0032] To quantitatively compare predicted spray characteristics with experimental values, spray tip penetration predicted by different models was plotted and compared with experimental data at two different chamber pressures. Figure 7 As shown, this is a comparison chart of four models and actual experimental data under the conditions of chamber pressure of 0.1 MPa, injection pressure of 3 MPa, temperature of 20℃, and injection pulse width of 0.1 ms; Figure 8 The figure shows a comparison between four models and real experimental data under the conditions of chamber pressure of 1 MPa, injection pressure of 3 MPa, temperature of 20℃, and injection pulse width of 0.1 ms. As can be seen from the figure, the MLP-LSTM-Attention model established in this study can learn the dependencies behind spray development from the start of injection to the complete spray formation process, exhibiting the best predictive performance and showing broad application prospects in future spray characteristic prediction.

[0033] The embodiments described herein are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made to the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention. Although preferred embodiments of the present invention have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the present invention. Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A predictive model for gasoline fuel spray characteristics of a GDI engine, characterized in that, The layers, arranged in order of feature flow, are: feature extraction layer, temporal modeling layer, feature weighting layer, and output layer. The feature extraction layer is a two-layer fully connected structure, including a first fully connected layer and a second fully connected layer; the input vector of the feature extraction layer is a 4-dimensional vector, the first fully connected layer maps the 4-dimensional input to 128 dimensions, the second fully connected layer expands the 128 dimensions to 256 dimensions, and then compresses it to a 128-dimensional feature sequence through linear transformation; The temporal modeling layer receives the feature sequence output by the feature extraction layer, including a forget gate, an input gate, and an output gate; The feature weighting layer is located after the temporal modeling layer. The feature weighting layer introduces an attention mechanism to perform secondary processing on the sequence hidden state output by the temporal modeling layer. The output layer is connected to the feature weighting layer, and the spray characteristic parameters are predicted by mapping the weighted feature sequence.

2. The GDI engine gasoline fuel spray characteristic prediction model as described in claim 1, characterized in that, The input vector is: injection pressure, ambient pressure, ambient temperature, and injection pulse width.

3. The GDI engine gasoline fuel spray characteristic model as described in claim 1, characterized in that, The first fully connected layer is followed by a first normalization layer, a first activation function, and a Dropout layer in sequence; the second fully connected layer is followed by a second normalization layer and a second activation function in sequence.

4. The GDI engine gasoline fuel spray characteristic prediction model as described in claim 3, characterized in that, The first and second activation functions use LeakyReLU, with a negative slope parameter set to 0.

2.

5. The GDI engine gasoline fuel spray characteristic prediction model as described in claim 3, characterized in that, The dropout rate of the Dropout layer is 0.

2.

6. The GDI engine gasoline fuel spray characteristic prediction model as described in claim 1, characterized in that, The attention mechanism calculates the score of the hidden state at each time step, generates a weight distribution using the Softmax function, and sums the weights with the corresponding hidden states to generate a context vector containing global temporal key information.

7. The GDI engine gasoline fuel spray characteristic prediction model as described in claim 1, characterized in that, The spray characteristic parameters are spray penetration distance and spray cone angle.

8. A method for predicting gasoline fuel spray characteristics of a GDI engine using a gasoline fuel spray characteristic prediction model for any one of claims 1-7, characterized in that, include: Step S1: First, construct a spray acquisition system and collect raw spray data. The spray acquisition system consists of a fixed container body, a fuel supply system, an environmental pressure regulation system, an environmental temperature regulation system, a fuel injection control system, and a high-speed camera system. Step S2: Image preprocessing and feature parameter extraction are performed. Background subtraction is used to eliminate the interference of brackets and window shadows inside the fixed volume chamber. A random sampling consistency algorithm is used to randomly extract a subset from the edge pixel set for model fitting. The distance from the remaining points to the fitted model is calculated. Points with a distance exceeding the preset deviation value are identified as outliers and removed. A dataset containing injection pressure, ambient pressure, temperature, injection pulse width and corresponding time series is constructed. Step S3: Construct a prediction model for the gasoline fuel spray characteristics of the GDI engine; It includes a feature extraction layer, a temporal modeling layer, a feature weighting layer, and an output layer; Step S4: Divide the preprocessed dataset from step S2 into a training set and a test set in an 8:2 ratio; use the coefficient of determination R0. 2 Mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators.

Images