Gas injection quality prediction method based on data fusion

Abstract

The invention discloses a gas injection quality prediction method based on data fusion, and belongs to the field of multipurpose fuel injection systems.The prediction method comprises the following steps that diesel branch pressure fluctuation models are constructed, and the diesel branch pressure fluctuation models comprise an electrical simulation model, a fitting model, an online sensing iteration model and a fitting iteration model of a hydraulic system; acquiring and processing a data set; constructing a neural network fitting model; constructing a real-time injection quantity data fusion algorithm to obtain a final predicted value; and calculating a fusion value by synchronously operating a neural network fitting model and a diesel branch pressure fluctuation model and combining weights. Compared with an existing prediction value obtained through single diesel branch pressure fluctuation, the prediction precision and the adaptability of the system are improved, and the real-time response capacity and robustness of the system are remarkably enhanced.

Description

Technical Field

[0001] This invention relates to the field of multi-purpose fuel injection systems for vehicles, aviation, and other applications, specifically a gas injection quality prediction method based on data fusion. Background Technology

[0002] Multi-purpose fuel injection systems serve as a key bridge for the transition of traditional internal combustion engines to low-carbon technologies. Technically, they have achieved compatibility with multiple fuels such as hydrogen, ammonia, and methanol, breaking the dependence on a single energy source. Furthermore, compared to "single-fuel dedicated engines" or "directly switching to new energy power (such as pure electric or fuel cells)," these systems offer advantages in initial investment, operation, and maintenance costs. However, due to the significant differences in combustion speed and lubricity among different fuels, existing injection quantity prediction models lack generalization ability.

[0003] Existing technologies have proposed solutions to the problems existing in multi-purpose fuel injection systems. For example, patent CN118940663A discloses "a method for predicting jet quantity," which addresses the problem of predicting the cyclic jet quantity during the main fuel injection process caused by pilot fuel injection in gas-liquid coupled injection. This includes sub-algorithms for calculating the fluctuation period, underdamped oscillation dissipation coefficient, and phase difference and jet quantity amplitude parameters. Regression analysis between predicted and measured values ​​reveals that the data converges on the regression line, with RMSE = 2.013 mg and R... 2 =0.9118, which shows that the accuracy of the prediction algorithm is in good agreement with the measured value.

[0004] However, the methods described in the patents still have certain shortcomings in practical applications: First, the working conditions and requirements of different fields (such as automotive and aviation) vary greatly in practical applications, and using a single detection method cannot guarantee high accuracy in various scenarios; Second, in the automotive field, the complexity and time delay of the system structure under transient conditions such as acceleration and deceleration make it difficult to accurately predict the gas injection quality, resulting in non-real-time results. Summary of the Invention

[0005] To address the above problems, this invention proposes a gas injection quality prediction method based on data fusion, aiming to improve the accuracy of gas injection quality prediction, optimize the real-time response capability of the prediction process, and enhance the adaptability to dynamic operating conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a gas injection quality prediction method based on data fusion, the steps of which are as follows: S1: Construct a diesel branch pressure fluctuation model, including an electrical simulation model, a fitting model, an online sensing iterative model, and a fitting iterative model for the hydraulic system; S2: Dataset acquisition and processing. The data collection process covers the normal operating conditions of the engine. The LM function is used to normalize the data, which is beneficial for data fitting. S3: Construct a neural network fitting model, which consists of an input layer, two hidden layers, and an output layer; S4: Construct a real-time injection quantity data fusion algorithm, simultaneously run a neural network fitting model and a diesel branch pressure fluctuation model, and perform linear weighted fusion according to the weights to generate the optimal value.

[0007] The data acquisition conditions in step S2 include: The data acquisition process covered 78 independent operating points, with the injection pressure ranging from 22MPa to 32MPa and the injection pulse width ranging from 0.5ms to 4ms. For each of the above operating points, dense sampling was performed at 0.1ms intervals within the short pulse width time interval corresponding to the injection pulse width to obtain the gas injection pressure and injection pulse width. The training set accounted for 70% of all samples, while the validation set and test set each accounted for 15% of the samples.

[0008] The parameters in the neural network fitting model in step S3 include: The parameter in the input layer is: gas injection pressure. Pressure difference before and after injection The injection pulse width is GET, the function in the hidden layer is a radial basis function, and the output layer performs a linear weighted sum of the outputs of the hidden layer to obtain the injection mass. .

[0009] The steps of the real-time injection volume data fusion algorithm in step S4 include: The first step is to calculate the initial weights, which are inversely proportional to the square of the RMSE. The second step is to obtain the predicted values ​​from the diesel branch pressure fluctuation model and the neural network fitting model. and ; The third step involves fusion calculation to obtain the optimal injection volume value, and then using a linear weighted fusion strategy to optimize the predicted values. and predicted value Optimize and integrate to obtain the optimal injection volume. ; Among them, when the sudden change in operating conditions satisfies When the conditions are met, a weight reset mechanism should be triggered: the weight parameters should be automatically reset to the initial weight parameters, which are determined using the initial weight calculation method described in the first step.

[0010] The fourth step is to combine the optimal values ​​generated in the third step using linear weighted fusion based on preset weights. Set the baseline injection value, and then derive the optimal value for the next injection process.

[0011] Specifically, during the initial injection, the predicted values ​​obtained from the diesel branch pressure fluctuation model are applied based on the initial weights. Predictions from neural network fitting models Linear weighting is applied to obtain the optimal injection quantity for that injection cycle. This optimal value is then used as the baseline injection quantity for the next injection cycle, and a new weight is calculated. Based on this new weight, the predicted value obtained from the diesel branch pressure fluctuation model is then applied. Predictions from neural network fitting models Linear weighting is performed to obtain the optimal spray volume for that spraying process. This process is repeated to ensure that the optimal spray volume is obtained after each spraying.

[0012] Beneficial effects: This invention combines a diesel branch pressure fluctuation model and a neural network fitting model using data fusion technology. This retains the neural network's ability to capture nonlinear relationships while utilizing the physical consistency of the mechanistic model, providing a more reliable injection quantity prediction scheme for engine control systems.

[0013] On the one hand, the dynamic weighting mechanism can quickly respond to changes in operating conditions. Compared with the predicted value obtained from the pressure fluctuation of a single diesel branch, the prediction error is reduced by 35% under the small pulse width condition of 0.5ms. Moreover, the data fusion process only needs to perform scalar multiplication and addition operations, and the prediction time for a single injection quantity is less than 1ms, which can meet the real-time requirements of the engine control cycle.

[0014] On the other hand, it exhibits excellent robustness and accuracy: by introducing a sudden change in operating conditions detection module, the interference of historical data on the current prediction can be avoided; combined with the "offline calibration + online optimization" control method, the robustness problem of injection quantity prediction is solved. Under the set prediction environment, the root mean square error (RMSE) of the three test results are 1.018mg, 0.937mg, and 0.980mg, respectively, indicating that the algorithm has better convergence than the comparison models (neural network fitting model and diesel branch pressure fluctuation model) when running continuously, the predicted value is closer to the measured value, and it remains stable under different operating conditions, confirming that it has good parameter adaptability and model generalization ability. Attached Figure Description

[0015] Figure 1 This is a flowchart of an algorithm for predicting gas injection quality based on data fusion; Figure 2 This is a diagram of the diesel branch pressure fluctuation model; Figure 3 This is a diagram of a neural network fitting model; Figure 4 This is a diagram of the real-time injection volume data fusion algorithm; Figure 5 , Figure 6 and Figure 7 This is a comparison chart of the injection results from the real-time injection quantity data fusion algorithm under different conditions, the injection results from the diesel branch pressure fluctuation model, and the injection results from the neural network fitting model. Detailed Implementation

[0016] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described in this application are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0017] This invention provides the following technical solution: a gas injection quality prediction method based on data fusion, the steps of which are as follows: S1: Construct a diesel branch pressure fluctuation model as follows Figure 2 This includes electrical simulation models, fitting models, online sensing iterative models, and fitting iterative models for hydraulic systems; The electrical simulation model diagram of the hydraulic system shows that the fuel system is abstracted into a one-dimensional hydroelectric model, and the frequency parameters are obtained. ; The fitting model uses injection pressure and pulse width as input conditions. The amplitude parameter is obtained by fitting the model graph and combining it with a function that changes with the current injection pressure. ; The online sensing iterative model takes the length of the high-pressure oil pipe and the length from the nozzle to the pressure detection point as inputs, and combines the diesel sound velocity, time, and delay time to obtain the phase difference. ; The fitting iterative model uses the pressure change process of fuel and gas at the injector inlet to sense the attenuation coefficient in real time. Combined with the previous pressure oscillation attenuation coefficient, the current injection environment is calibrated. The dissipation parameters are obtained through iterative and rolling optimization. .

[0018] S2: Dataset acquisition and processing.

[0019] The data acquisition process covered 78 independent operating points, basically covering the normal operating conditions of the engine. The injection pressure ranged from 22MPa to 32MPa, and the injection pulse width ranged from 0.5ms to 4ms. For each of the above operating points, dense sampling was performed at a time interval of 0.1ms within the small pulse width time interval corresponding to the injection pulse width to obtain the gas injection pressure and injection pulse width.

[0020] The training set comprises 70% of all samples, while the validation and test sets each comprise 15% of the samples.

[0021] Using the LM (Levenberg-Marquardt) function to normalize the data is beneficial for data fitting.

[0022] S3: Construct a neural network fitting model as follows Figure 3 .

[0023] The neural network fitting model consists of an input layer, two hidden layers, and an output layer. The parameter in the input layer is: gas injection pressure. Pressure difference before and after injection The injection pulse width is GET, the function in the hidden layer is the radial basis function (RBF), and the output layer performs a linear weighted sum of the outputs of the hidden layer to obtain the injection mass. ; The neural network fitting model is trained using a pre-set training set to learn the feature mapping relationship in the training dataset, thus obtaining the basic neural network fitting model. The validation dataset is input into the base neural network model, and the parameters in the base neural network fitting model are adjusted according to the performance metric until the neural network fitting model with the best performance on the validation dataset is obtained. Performance metrics are calculated by comparing the actual values ​​on the test set with the model's predicted values. The performance index formula is as follows: RMSE (Root Mean Square Error): Measures the difference between the predicted value and the actual value;

[0024] Where n represents the number of samples. This represents the actual measured mass of the gas jet. This indicates the jet quality output by the model.

[0025] R² (coefficient of determination): measures the goodness of fit of the regression model;

[0026] Where n represents the number of samples. This represents the actual measured mass of the gas jet. This indicates the jet mass output by the model. This represents the average of the true values.

[0027] S4: Construct a real-time injection volume data fusion algorithm to obtain the final predicted value, as shown below. Figure 4 .

[0028] The neural network fitting model and the diesel branch pressure fluctuation model are run simultaneously, and the fusion value is calculated through weights.

[0029] The detailed steps of the real-time injection volume data fusion algorithm are as follows: The first step is to calculate the initial weights.

[0030] The initial weights are calculated based on the inverse relationship between the weights and the square of the RMSE, i.e.

[0031] Based on the relationship between weights and RMSE, the calculation method for the initial weights can be obtained as follows.

[0032]

[0033] Based on the performance index formula, the RMSE1 of the diesel branch pressure fluctuation model is 2.013 mg, and the RMSE2 of the neural network fitting model is 1.591 mg.

[0034] Based on the relationship between weights and RMSE, the weights of the initial diesel branch pressure fluctuation model can be calculated. Neural network fitting model weights .

[0035] The second step is to obtain the predicted values ​​from the diesel branch pressure fluctuation model and the neural network fitting model.

[0036] After a single injection operation is completed, the host computer analyzes the collected timing signals. Predicted values ​​were obtained using a diesel branch pressure fluctuation model. Predicted values ​​are obtained by fitting a model using a neural network. .

[0037] The third step is to calculate the optimal injection volume.

[0038] A linear weighted fusion strategy is used to calculate the predicted values. and predicted value Optimize and integrate the process to obtain the optimal injection volume with extremely fast response speed. .

[0039]

[0040] in This represents the optimal value obtained after data fusion. This represents the predicted value obtained from the diesel branch pressure fluctuation model. This represents the predicted value obtained by fitting the model using a neural network. and The weights assigned to the two predicted values.

[0041] The formula for calculating the weight is:

[0042] This represents the standard deviation of the predicted value obtained from the diesel branch pressure fluctuation model relative to the previous best fusion value. It represents the standard deviation of the predicted value obtained by the neural network fitting model relative to the previous optimal fusion value.

[0043] When the sudden change in operating conditions satisfies When the conditions are met, a weight reset mechanism should be triggered. The weight parameters should be automatically reset to the initial weight parameters, which are determined using the initial weight calculation method described in the first step.

[0044] The fourth step is to combine the optimal values ​​generated in the third step using linear weighted fusion based on preset weights. Set the baseline injection value, obtain the algorithm weight of this injection according to the weight calculation formula, and then perform weighted fusion to obtain the optimal value of this injection amount.

[0045] Specifically, during the initial injection, the predicted values ​​obtained from the diesel branch pressure fluctuation model are applied based on the initial weights. Predictions from neural network fitting models Linear weighting is applied to obtain the optimal injection quantity for that injection cycle. This optimal value is then used as the baseline injection quantity for the next injection cycle, and a new weight is calculated. Based on this new weight, the predicted value obtained from the diesel branch pressure fluctuation model is then applied. Predictions from neural network fitting models Linear weighting is performed to obtain the optimal spray volume for that spraying process. This process is repeated to ensure that the optimal spray volume is obtained after each spraying.

[0046] The specific prediction environment is as follows:

[0047] Figure 5 , Figure 6 and Figure 7 This chart compares the injection results from the real-time injection quantity data fusion algorithm, the diesel branch pressure fluctuation model, and the neural network fitting model under different conditions. The MAP value in the chart refers to the value calibrated in offline testing, i.e., the theoretical value of the injection quantity under operating conditions. It can be observed that the injection quantity obtained by the real-time injection quantity data fusion algorithm is almost always closer to the theoretical value than that obtained by the single algorithm.

[0048] The final three performance metrics (RMSE) were 1.018 mg, 0.937 mg, and 0.980 mg, respectively. Experimental results show that the constructed data fusion algorithm converges more effectively and more closely approximates the measured values ​​compared to the two algorithms under continuous operation, and remains stable under different operating conditions. This confirms that the constructed data fusion-based gas injection quality prediction method has good parameter adaptability and model generalization ability.

Claims

1. A method for predicting gas injection quality based on data fusion, characterized in that, Includes the following steps: S1: Construct a diesel branch pressure fluctuation model, including an electrical simulation model, a fitting model, an online sensing iterative model, and a fitting iterative model for the hydraulic system; S2: Dataset acquisition and processing. The data collection process covers the normal operating conditions of the engine. The LM function is used to normalize the data, which is beneficial for data fitting. S3: Construct a neural network fitting model, which consists of an input layer, two hidden layers, and an output layer; S4: Construct a real-time injection quantity data fusion algorithm, simultaneously run a neural network fitting model and a diesel branch pressure fluctuation model, and perform linear weighted fusion according to the weights to generate the optimal value.

2. The gas jet quality prediction method based on data fusion as described in claim 1, characterized in that, The data acquisition conditions in step S2 include: The data acquisition process covered 78 independent operating points, with the injection pressure ranging from 22MPa to 32MPa and the injection pulse width ranging from 0.5ms to 4ms. For each of the above operating points, dense sampling was performed at 0.1ms intervals within the short pulse width time interval corresponding to the injection pulse width to obtain the gas injection pressure and injection pulse width. The training set accounted for 70% of all samples, while the validation set and test set each accounted for 15% of the samples.

3. The gas injection quality prediction method based on data fusion as described in claim 1, characterized in that, The parameters in the neural network fitting model in step S3 include: The parameter in the input layer is: gas injection pressure. Pressure difference before and after injection The injection pulse width is GET, the function in the hidden layer is a radial basis function, and the output layer performs a linear weighted sum of the outputs of the hidden layer to obtain the injection mass. .

4. The gas injection quality prediction method based on data fusion as described in claim 1, characterized in that, The steps of the real-time injection volume data fusion algorithm in step S4 include: The first step is to calculate the initial weights, which are inversely proportional to the square of the RMSE. The second step is to obtain the predicted values ​​from the diesel branch pressure fluctuation model and the neural network fitting model. and ; The third step involves fusion calculation to obtain the optimal injection volume value, and then using a linear weighted fusion strategy to optimize the predicted values. and predicted value Optimize and integrate to obtain the optimal injection volume. ; The fourth step is to combine the optimal values ​​generated in the third step using linear weighted fusion based on preset weights. Let the baseline injection value be used to determine the optimal value for the next injection process.

5. The gas injection quality prediction method based on data fusion as described in claim 4, characterized in that, When the sudden change in operating conditions satisfies When the conditions are met, a weight reset mechanism should be triggered: the weight parameters should be automatically reset to the initial weight parameters, which are determined using the initial weight calculation method described in the first step.

6. The gas jet quality prediction method based on data fusion as described in claim 4, characterized in that, During the initial injection, the predicted values ​​obtained from the diesel branch pressure fluctuation model are based on the initial weights. Predictions from neural network fitting models The injection quantity obtained by performing linear weighting is the optimal value for that injection. Using this optimal value as the baseline injection quantity for the next injection process, and calculating new weights, the predicted values ​​obtained from the diesel branch pressure fluctuation model are then applied based on these new weights. Predictions from neural network fitting models Linear weighting is performed to obtain the optimal spray volume for that spraying process. This process is repeated to ensure that the optimal spray volume is obtained after each spraying.

Images