Intelligent optimization method for electronic injection fuel injection strategy based on deep learning

Abstract

The invention discloses an electronic injection fuel injection strategy intelligent optimization method based on deep learning, and the method comprises the following steps: S1, collecting and preprocessing multi-source working condition data of an engine, and generating a standard data set; s2, constructing a Bayesian neural network model, and outputting an integrated fuel injection parameter prediction value and an uncertainty estimation value; s3, optimizing model structure parameters and hyper-parameters by using an ant lion optimization algorithm to obtain an optimal structure and parameters; s4, training the model by using the optimal structure and parameters, and performing performance evaluation by using a standard data set; s5, collecting working condition data in real time, and outputting an optimal integrated fuel injection parameter prediction value and an uncertainty estimation value; and S6, periodically collecting feedback data, training the model in combination with the standard data set increment, and re-executing ant lion optimization in good time. According to the method, high-precision intelligent optimization of the oil injection strategy of the engine is achieved, the fuel economy is improved, emission is reduced, and the self-adaptive capacity under the complex working condition is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of engine intelligent control technology, and in particular to an intelligent optimization method for electronic fuel injection strategies based on deep learning. Background Art

[0002] With the rapid development of the automotive industry and the increasingly stringent emission regulations, the engine's electronic fuel injection system plays a vital role in improving fuel economy, reducing harmful emissions, and enhancing power performance. At present, engine injection strategies mainly rely on the electronic control unit to combine engine operating parameters for table lookup control or use traditional data-driven methods such as linear regression and neural networks to optimize injection parameters. The table lookup control method has limited response speed to changes in engine operating conditions and is difficult to adapt to complex and changing operating conditions in real time. Although traditional neural network methods have certain nonlinear modeling capabilities, in practical applications they often suffer from insufficient model generalization ability, susceptibility to noise interference, and poor robustness to unknown operating conditions. In addition, existing injection strategy optimization usually only focuses on a single model output and cannot effectively quantify the uncertainty of model predictions. This leads to a lack of safety redundancy in injection decisions under extreme operating conditions or data distribution deviations, affecting engine performance and system reliability.

[0003] With the development of artificial intelligence technology, Bayesian neural networks have been gradually applied to the field of complex system modeling due to their ability to output prediction results and their uncertainty, as well as their ability to resist overfitting and adapt themselves. However, single Bayesian neural networks still rely mainly on manual experience or simple search in terms of structure and parameter selection, making it difficult to achieve efficient global optimization for the high complexity and diversity of engine operating conditions. At the same time, existing optimization methods such as genetic algorithms and particle swarm optimization are prone to falling into local optimality in high-dimensional complex parameter spaces, lack dynamic diversity control and collaborative optimization mechanisms, and cannot fully utilize the advantages of Bayesian neural networks. It is difficult to meet the needs of real-time, precise and adaptive optimization of intelligent engine injection strategies.

[0004] How to organically integrate advanced intelligent optimization algorithms with Bayesian neural network structures to achieve multi-model integration and adaptive optimization of structural parameters, and introduce model uncertainty quantification and dynamic adjustment in injection decision-making has become a key technical problem that needs to be solved urgently in the field of intelligent injection strategy optimization.

[0005] Therefore, how to provide an intelligent optimization method for electronic fuel injection strategy based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based intelligent optimization method for electronic fuel injection (EFI) injection strategies. This method fully integrates the Ant Lion intelligent optimization mechanism with Bayesian neural network integrated modeling to model, train, and predict multi-source engine operating condition data. By dynamically optimizing model structure parameters and hyperparameters, it describes in detail an integrated algorithm for injection parameter prediction and uncertainty estimation, as well as an adaptive optimization process for the injection control strategy. This method offers advantages such as high injection control accuracy, strong model generalization and robustness, quantifiable uncertainty, continuous self-learning, and adaptability to complex operating conditions.

[0007] According to an embodiment of the present invention, a method for intelligent optimization of an electronic fuel injection strategy based on deep learning includes the following steps: S1. Collecting multi-source operating condition data of the engine and preprocessing the multi-source operating condition data to generate a standard data set; S2. Based on the standard data set, a Bayesian neural network model is constructed, which takes the standard data set as input and outputs the integrated injection parameter prediction value and the integrated uncertainty estimation value; S3. Based on the constructed Bayesian neural network model, the ant lion optimization algorithm is used to globally optimize the structural parameters and hyperparameters of the Bayesian neural network model to obtain the optimal structure and parameters of the Bayesian neural network model; S4. Based on the obtained optimal structure and parameters, train the Bayesian neural network model, use the standard data set for training, and evaluate the performance of the Bayesian neural network model; S5. During vehicle operation, the current engine operating condition data is collected in real time. Based on the trained Bayesian neural network model, the current operating condition data is input, and the optimal integrated injection parameter prediction value and integrated uncertainty estimate are output to dynamically adjust the injection control strategy. S6. Periodically collect engine operation feedback data after the injection control is executed, combine the standard data set with the engine operation feedback data, perform incremental training of the Bayesian neural network model, and re-execute the ant lion optimization algorithm in step S3 in a timely manner based on the operation results to achieve continuous adaptive optimization of the injection strategy.

[0008] Optionally, the multi-source operating condition data specifically includes engine speed, load, throttle opening, coolant temperature, intake pressure, air-fuel ratio, emission concentration and fuel injection amount.

[0009] Optionally, the preprocessing of the multi-source operating condition data specifically includes normalizing the multi-source operating condition data, eliminating outliers and performing feature engineering processing.

[0010] Optionally, the S2 specifically includes: S21. Define the standard dataset as an input vector , where each Indicates engine operating condition characteristics; S22. Initializing a Bayesian neural network model based on the input vector, the Bayesian neural network model comprising an input layer, m hidden layers, and an output layer, where m is a positive integer greater than or equal to 1; S23. Set the prior distribution type selection parameters for the weight parameters of the Bayesian neural network model , specify the prior distribution type of all weight parameters as Gaussian distribution; S24. Set the variational inference layer number parameters for the hidden layer structure of the Bayesian neural network model , determine the number or specific level of hidden layers involved in variational inference; S25. Based on the initialized Bayesian neural network model, construct k sub-models of the Bayesian neural network model with slightly different structures, and set the integrated weight parameters of the sub-models of the Bayesian neural network model , controls the weighted proportion of each Bayesian neural network model’s sub-model in the final output; S26, inputting the input vector into each sub-model of the Bayesian neural network model to obtain the injection parameter prediction value of each sub-model of the Bayesian neural network model , uncertainty estimates and confidence scores ; S27, based on the integrated weight parameters of the sub-model of the set Bayesian neural network model and confidence scores , using adaptive integration mechanism to calculate the integrated injection parameter prediction value and integrated uncertainty estimates : ; ; in, is the entropy adjustment coefficient, Represents the normalized entropy value of the sub-model confidence scores of all Bayesian neural network models, k is the number of sub-models of the Bayesian neural network model, is the weight corresponding to the sub-model of the j-th Bayesian neural network model, is the confidence score of the sub-model of the j-th Bayesian neural network model under the current working condition, is the weight corresponding to the sub-model of the l-th Bayesian neural network model, is the confidence score of the sub-model of the lth Bayesian neural network model under the current working condition; S28. Based on the standard data set, the posterior distribution of all weight parameters of the Bayesian neural network model is inferred by the Bayesian inference method, and the Markov chain Monte Carlo method is used to approximate the posterior distribution; S29, the obtained integrated injection parameter prediction value and integrated uncertainty estimates As the final output of the Bayesian neural network model.

[0011] Optionally, the S3 specifically includes: S31. The structural parameters and hyperparameters of the constructed Bayesian neural network model are set as the optimization objects of the Ant Lion Optimization Algorithm. The structural parameters include the number of network layers and the number of neurons in each layer. The hyperparameters include the prior distribution type selection parameters. , variational inference layer parameters , integrated weight parameters , where k is the number of sub-models of the Bayesian neural network model; S32. Initializing multiple antlion-ant populations based on the set optimization object, and independently generating initial population solution vectors in different subspaces of structural parameters and hyperparameters; S33. For each ant lion individual, set up an adaptive trap mechanism, and dynamically adjust the trap radius based on the global fitness distribution and the ant lion's historical predation success rate. and trap depth ; S34. For each ant individual, adaptively select the Levy flight path or the chaotic mapping path based on its historical fitness performance to form the current round of wandering strategy; S35. In each iteration, each ant randomly selects an antlion from the antlions as a leader according to the roulette wheel selection mechanism, updates its own position based on the antlion's trap radius and trap depth, and generates new structural parameters and hyperparameter solution vectors. S36. For each ant’s generated new structural parameters and hyperparameter solution vectors, a Bayesian neural network model is built and trained based on a standard data set to obtain an integrated injection parameter prediction. and integrated uncertainty estimates ; S37, based on the predicted value of integrated injection parameters and integrated uncertainty estimates , calculate the fitness function F, comprehensively consider the mean square error of the integrated injection parameter prediction value , the mean of the integrated uncertainty estimates and variance : ; in, 、 、 is the weight coefficient, are the real injection parameters of the standard data set; S38, sorting and evaluating the calculated fitness values, and performing trap updating, predation, individual position updating, and global optimal individual preservation operations in each antlion-ant population; S39. After a specified iteration cycle, each antlion-ant population exchanges and migrates excellent individuals to promote the search for the global optimal solution and maintain population diversity. S310: Determine whether the preset termination condition is met, and if so, output the optimal structural parameters and hyperparameter configurations of all current ant lion-ant populations. 、 、 , the optimal number of network layers and the number of neurons in each layer as the optimal structure and parameters of the Bayesian neural network model.

[0012] Optionally, the S4 specifically includes: S41. Build a Bayesian neural network model based on the obtained optimal structural parameters and hyperparameters; S42, dividing the standard data set into a training set and a validation set; S43, inputting the training set into the constructed Bayesian neural network model, performing parameter training, and updating the network weights and bias distribution; S44. After the training is completed, the validation set is input into the Bayesian neural network model to obtain an output result of the Bayesian neural network model on the validation set; S45. comparing the obtained output results of the Bayesian neural network model with the actual injection parameters of the validation set to evaluate the prediction performance and generalization ability of the Bayesian neural network model; S46. Determine whether the trained Bayesian neural network model meets the performance requirements based on the evaluation results. If so, save the Bayesian neural network model. If not, adjust the training parameters or return to step S43 for retraining.

[0013] Optionally, the S5 specifically includes: S51. During vehicle operation, continuously monitor the engine's operating status and collect real-time engine operating data, including engine speed, load, throttle opening, coolant temperature, intake pressure, air-fuel ratio, emission concentration, and fuel injection quantity. Preprocess the collected data to generate input vectors for intelligent optimization of the fuel injection strategy. S52, inputting the generated input vector into the trained Bayesian neural network model in real time, so that changes in the working conditions at each moment can be fed back to the Bayesian neural network model reasoning process in a timely manner; S53. Using the Bayesian neural network model, respectively obtain injection parameter prediction results and uncertainty estimation results corresponding to multiple sub-models of the Bayesian neural network model, as well as confidence scores of each sub-model of the Bayesian neural network model under the current operating conditions, and fully record the inference outputs of the sub-models of the Bayesian neural network model; S54. Based on the obtained sub-model outputs of all Bayesian neural network models, use an integration mechanism to perform weighted integration on the injection parameter prediction results and uncertainty estimation results of the sub-models of each Bayesian neural network model to obtain the optimal integrated injection parameter prediction value and integrated uncertainty estimation value under the current operating condition, and perform a validity check on the integration result. S55. Based on the obtained optimal integrated injection parameter prediction value and the integrated uncertainty estimate, and in combination with the actual needs of the engine control system, dynamically adjust the key injection control parameters of the injection pulse width and the injection timing.

[0014] Optionally, the S6 specifically includes: S61. After the engine injection control strategy is executed, periodically collect engine operation feedback data, including fuel consumption, emission indicators, power response, and various operating condition parameters; S62. Normalize the collected engine operation feedback data and merge it with the standard data set to construct an incremental data set for updating the Bayesian neural network model; S63, inputting the incremental data set into the trained Bayesian neural network model, and updating the parameters of the Bayesian neural network model using an incremental learning method; S64. After incremental training of the Bayesian neural network model, re-evaluate the performance of the Bayesian neural network model based on the latest incremental data set to test the prediction accuracy and generalization ability of the Bayesian neural network model under various working conditions; S65. Based on the evaluation result of the Bayesian neural network model, determine whether the structural parameters and hyperparameters of the Bayesian neural network model need to be re-optimized. If necessary, start the ant lion optimization algorithm in step S3 to perform global optimization on the structural parameters and hyperparameters of the Bayesian neural network model. S66. Reapply the updated and optimized Bayesian neural network model to the real-time optimization control of the engine injection strategy to achieve continuous adaptive optimization of the injection control strategy and enter the next cycle of data collection and Bayesian neural network model update process.

[0015] The beneficial effects of the present invention are: This invention achieves intelligent optimization of the engine's electronic fuel injection strategy by deeply integrating the Ant Lion Optimization Algorithm with a Bayesian neural network model. This overcomes technical bottlenecks such as slow response, difficulty adapting to complex operating conditions, and inability to quantify and predict uncertainty in traditional fuel injection control methods. By constructing an integrated Bayesian neural network model, it not only outputs the optimal predicted values ​​of the injection parameters, but also simultaneously provides uncertainty estimation results, providing a more scientific and secure decision-making basis for the fuel injection control strategy. By introducing multiple structural parameters and hyperparameters and performing global collaborative optimization through the Ant Lion Optimization Algorithm, it can effectively improve the adaptability of the model structure and parameters and the global search capability, avoid falling into local optimality, and enhance the prediction accuracy of the fuel injection model and the stability of the system.

[0016] This invention establishes a periodic model adaptive update and feedback optimization mechanism, continuously incrementally training and dynamically optimizing the model based on the latest engine operating feedback data. This ensures that the injection strategy maintains excellent self-learning and adaptive capabilities even under long-term engine operation and complex and variable operating conditions. Compared to existing technologies, this invention significantly reduces fuel consumption, improves emissions performance, enhances power response, and ensures the long-term reliable operation of the engine injection system, possessing broad engineering application value and widespread potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of an intelligent optimization method for electronic fuel injection strategy based on deep learning proposed by the present invention; Figure 2 This is a structural schematic diagram of a Bayesian neural network model for an intelligent optimization method for electronic fuel injection strategy based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0018] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0019] refer to Figure 1 and Figure 2 , an intelligent optimization method for electronic fuel injection strategy based on deep learning, comprising the following steps: S1. Collecting multi-source operating condition data of the engine and preprocessing the multi-source operating condition data to generate a standard data set; S2. Based on the standard data set, a Bayesian neural network model is constructed, which takes the standard data set as input and outputs the integrated injection parameter prediction value and the integrated uncertainty estimation value; S3. Based on the constructed Bayesian neural network model, the ant lion optimization algorithm is used to globally optimize the structural parameters and hyperparameters of the Bayesian neural network model to obtain the optimal structure and parameters of the Bayesian neural network model; S4. Based on the obtained optimal structure and parameters, train the Bayesian neural network model, use the standard data set for training, and evaluate the performance of the Bayesian neural network model; S5. During vehicle operation, the current engine operating condition data is collected in real time. Based on the trained Bayesian neural network model, the current operating condition data is input, and the optimal integrated injection parameter prediction value and integrated uncertainty estimate are output to dynamically adjust the injection control strategy. S6. Periodically collect engine operation feedback data after the injection control is executed, combine the standard data set with the engine operation feedback data, perform incremental training of the Bayesian neural network model, and re-execute the ant lion optimization algorithm in step S3 in a timely manner based on the operation results to achieve continuous adaptive optimization of the injection strategy.

[0020] In this embodiment, the multi-source operating condition data specifically includes engine speed, load, throttle opening, coolant temperature, intake pressure, air-fuel ratio, emission concentration and fuel injection amount.

[0021] In this embodiment, the preprocessing of the multi-source operating condition data specifically includes normalizing the multi-source operating condition data, eliminating outliers, and performing feature engineering processing.

[0022] In this embodiment, S2 specifically includes: S21. Define the standard dataset as an input vector , where each Indicates engine operating condition characteristics; S22. Initializing a Bayesian neural network model based on the input vector, the Bayesian neural network model comprising an input layer, m hidden layers, and an output layer, where m is a positive integer greater than or equal to 1; S23. Set the prior distribution type selection parameters for the weight parameters of the Bayesian neural network model , specify the prior distribution type of all weight parameters as Gaussian distribution; S24. Set the variational inference layer number parameters for the hidden layer structure of the Bayesian neural network model , determine the number or specific level of hidden layers involved in variational inference; S25. Based on the initialized Bayesian neural network model, construct k sub-models of the Bayesian neural network model with slightly different structures, and set the integrated weight parameters of the sub-models of the Bayesian neural network model , controlling the weighted proportion of each Bayesian neural network model's sub-model in the final output, wherein the integration weight parameter is determined by proportional distribution according to the prediction accuracy and error size of each Bayesian neural network model's sub-model on the validation set; S26, inputting the input vector into each sub-model of the Bayesian neural network model to obtain the injection parameter prediction value of each sub-model of the Bayesian neural network model , uncertainty estimates and confidence scores ; S27, based on the integrated weight parameters of the sub-model of the set Bayesian neural network model and confidence scores , using adaptive integration mechanism to calculate the integrated injection parameter prediction value and integrated uncertainty estimates : ; ; in, is the entropy adjustment coefficient, which is obtained by gradually adjusting different values ​​using the cross-validation method on the validation set to select the parameters that optimize the prediction performance and uncertainty measurement of the integrated model. Represents the normalized entropy value of the sub-model confidence scores of all Bayesian neural network models, k is the number of sub-models of the Bayesian neural network model, is the weight corresponding to the sub-model of the j-th Bayesian neural network model, is the confidence score of the sub-model of the j-th Bayesian neural network model under the current working condition, is the weight corresponding to the sub-model of the l-th Bayesian neural network model, is the confidence score of the sub-model of the lth Bayesian neural network model under the current working condition; S28. Based on the standard data set, the posterior distribution of all weight parameters of the Bayesian neural network model is inferred by the Bayesian inference method, and the Markov chain Monte Carlo method is used to approximate the posterior distribution; S29, the obtained integrated injection parameter prediction value and integrated uncertainty estimates As the final output of the Bayesian neural network model.

[0023] In this embodiment, S3 specifically includes: S31. The structural parameters and hyperparameters of the constructed Bayesian neural network model are set as the optimization objects of the Ant Lion Optimization Algorithm. The structural parameters include the number of network layers and the number of neurons in each layer. The hyperparameters include the prior distribution type selection parameters. , variational inference layer parameters , integrated weight parameters , where k is the number of sub-models of the Bayesian neural network model; S32. Initializing multiple antlion-ant populations based on the set optimization object, and independently generating initial population solution vectors in different subspaces of structural parameters and hyperparameters; S33. For each ant lion individual, set up an adaptive trap mechanism, and dynamically adjust the trap radius based on the global fitness distribution and the ant lion's historical predation success rate. and trap depth The dynamic adjustment of the trap radius and trap depth refers to the use of a nonlinear scaling function to calculate in real time based on the current ant lion individual fitness ranking in the population and the historical predation success rate. The higher the fitness and the higher the predation success rate, the smaller the trap radius and the deeper the trap.

[0024] S34. For each ant individual, adaptively select the Levy flight path or the chaotic mapping path based on its historical fitness performance to form the current round of wandering strategy; S35. In each iteration, each ant randomly selects an antlion from the antlions as a leader according to the roulette wheel selection mechanism, updates its own position based on the antlion's trap radius and trap depth, and generates new structural parameters and hyperparameter solution vectors. S36. For each ant’s generated new structural parameters and hyperparameter solution vectors, a Bayesian neural network model is built and trained based on a standard data set to obtain an integrated injection parameter prediction. and integrated uncertainty estimates ; S37, based on the predicted value of integrated injection parameters and integrated uncertainty estimates , calculate the fitness function F, comprehensively consider the mean square error of the integrated injection parameter prediction value , the mean of the integrated uncertainty estimates and variance : ; in, 、 、 is the weight coefficient, 、 、 By using the cross-validation method on the validation set, different value combinations are gradually adjusted, and the coefficient combination that makes the comprehensive performance index of the Bayesian neural network model optimal is selected as the final value. are the real injection parameters of the standard data set; S38, sorting and evaluating the calculated fitness values, and performing trap updating, predation, individual position updating, and global optimal individual preservation operations in each antlion-ant population; S39. After a specified iteration cycle, each antlion-ant population exchanges and migrates excellent individuals to promote the search for the global optimal solution and maintain population diversity. S310: Determine whether the preset termination condition is met, and if so, output the optimal structural parameters and hyperparameter configurations of all current ant lion-ant populations. 、 、 , the optimal number of network layers and the number of neurons in each layer as the optimal structure and parameters of the Bayesian neural network model, where the preset termination condition refers to reaching the set maximum number of iterations or the improvement of the optimal fitness of the population within several consecutive generations is less than the predetermined threshold.

[0025] In this embodiment, the S4 specifically includes: S41. Build a Bayesian neural network model based on the obtained optimal structural parameters and hyperparameters; S42, dividing the standard data set into a training set and a validation set; S43, inputting the training set into the constructed Bayesian neural network model, performing parameter training, and updating the network weights and bias distribution; S44. After the training is completed, the validation set is input into the Bayesian neural network model to obtain an output result of the Bayesian neural network model on the validation set; S45. comparing the obtained output results of the Bayesian neural network model with the actual injection parameters of the validation set to evaluate the prediction performance and generalization ability of the Bayesian neural network model; S46. Based on the evaluation results, determine whether the trained Bayesian neural network model meets the performance requirements. If so, save the Bayesian neural network model. If not, adjust the training parameters or return to step S43 for retraining. Whether the performance requirements are met is determined by setting indicator thresholds for the Bayesian neural network model's prediction error, uncertainty, and convergence speed.

[0026] In this embodiment, the S5 specifically includes: S51. During vehicle operation, continuously monitor the engine's operating status and collect real-time engine operating data, including engine speed, load, throttle opening, coolant temperature, intake pressure, air-fuel ratio, emission concentration, and fuel injection quantity. Preprocess the collected data to generate input vectors for intelligent optimization of the fuel injection strategy. S52, inputting the generated input vector into the trained Bayesian neural network model in real time, so that changes in the working conditions at each moment can be fed back to the Bayesian neural network model reasoning process in a timely manner; S53. Using the Bayesian neural network model, respectively obtain injection parameter prediction results and uncertainty estimation results corresponding to multiple sub-models of the Bayesian neural network model, as well as confidence scores of each sub-model of the Bayesian neural network model under the current operating conditions, and fully record the inference outputs of the sub-models of the Bayesian neural network model; S54. Based on the obtained sub-model outputs of all Bayesian neural network models, use an integration mechanism to perform weighted integration on the injection parameter prediction results and uncertainty estimation results of the sub-models of each Bayesian neural network model to obtain the optimal integrated injection parameter prediction value and integrated uncertainty estimation value under the current operating condition, and perform a validity check on the integration result. S55. Based on the obtained optimal integrated injection parameter prediction value and integrated uncertainty estimate value, combined with the actual needs of the engine control system, dynamically adjust the key injection control parameters of the injection pulse width and injection timing. The dynamic adjustment of the key injection control parameters of the injection pulse width and injection timing in combination with the actual needs of the engine control system specifically includes making a comprehensive judgment on the engine target output, emission requirements and fuel economy targets under the current operating conditions based on the integrated injection parameter prediction value and uncertainty estimate value, and correcting the injection pulse width and injection timing parameters in real time to optimize the engine operating indicators and meet safety and environmental protection constraints.

[0027] In this embodiment, S6 specifically includes: S61. After the engine injection control strategy is executed, periodically collect engine operation feedback data, including fuel consumption, emission indicators, power response, and various operating condition parameters; S62. Normalize the collected engine operation feedback data and merge it with the standard data set to construct an incremental data set for updating the Bayesian neural network model; S63, inputting the incremental data set into the trained Bayesian neural network model, and updating the parameters of the Bayesian neural network model using an incremental learning method; S64. After incremental training of the Bayesian neural network model, re-evaluate the performance of the Bayesian neural network model based on the latest incremental data set to test the prediction accuracy and generalization ability of the Bayesian neural network model under various working conditions; S65. Based on the evaluation result of the Bayesian neural network model, determine whether the structural parameters and hyperparameters of the Bayesian neural network model need to be re-optimized. If necessary, start the ant lion optimization algorithm in step S3 to perform global optimization on the structural parameters and hyperparameters of the Bayesian neural network model. S66. Reapply the updated and optimized Bayesian neural network model to the real-time optimization control of the engine injection strategy to achieve continuous adaptive optimization of the injection control strategy and enter the next cycle of data collection and Bayesian neural network model update process.

[0028] Example 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to a certain bus group to verify the actual engineering effect of the intelligent fuel injection strategy optimization method of the present invention. For a long time, the bus group has faced practical operational difficulties such as heavy pressure on urban road traffic, complex vehicle operating conditions, slow response of traditional fuel injection strategies, high fuel consumption and large emission fluctuations. Especially in special environments such as high temperature, heavy rain, morning and evening rush hours, some buses have experienced slow power response, large instantaneous fuel consumption fluctuations and occasional excessive exhaust gas. For this reason, the group's technical department decided to divide 30 buses of the same model into an experimental group and a control group in May 2024. The experimental group will all upgrade the intelligent fuel injection strategy optimization system of the present invention, and the control group will continue to use the traditional table lookup plus PID fuel injection adjustment method.

[0029] During the implementation process, the technical team first installed a real-time engine data collection and upload module on all experimental vehicles. This module collects key parameters such as engine speed, load, throttle opening, temperature, air-fuel ratio, injection volume, NOx, and CO every second, and performs local preprocessing and standardization. Throughout May, the experimental group of vehicles collected approximately 1,100 hours of engine operation data, generating a complete standard data set. This data set was then used to build a Bayesian neural network model, and the model structure parameters and hyperparameters were globally optimized using the Ant Lion Optimization Algorithm. Five differentiated Bayesian neural network models were introduced into the model integration structure. The prior distribution, number of inference layers, and other parameters of each sub-model were dynamically selected and optimized by the Ant Lion Optimization Algorithm. In the model integration structure, five Bayesian neural network models with different structural parameters or hyperparameters were set. The prior distribution type, number of layers involved in inference, and other parameters of each sub-model were dynamically selected and optimized by the Ant Lion Optimization Algorithm.

[0030] After the model was deployed, the buses in the experimental group automatically switched to the intelligent injection strategy on all routes and under various complex operating conditions. The vehicle's engine operating conditions were input into the integrated Bayesian neural network model in real time at every moment, and the optimal injection parameters and uncertainty estimates were output, which were used by the vehicle ECU to dynamically adjust the injection pulse width and injection timing. At the same time, feedback data after the injection control was executed was automatically uploaded to the server, and the model incremental data was automatically updated daily. Model self-learning and parameter re-optimization were performed during the off-peak hours of each month to ensure that the injection strategy always adapted to the latest operating environment. The entire test period was from May to August 2024. During this period, the routes, passenger capacity, and weather conditions of the experimental and control groups remained basically the same, ensuring the objectivity and scientific nature of the comparison results.

[0031] During four consecutive months of actual operation, the experimental group of vehicles demonstrated significant performance improvements. The experimental group's average fuel consumption per 100 kilometers was 29.0 liters, a 10.8% decrease compared to the control group's 32.5 liters. NOx emissions averaged 0.94 g / km, 21.7% lower than the control group's 1.20 g / km. CO emissions averaged 0.12 g / km, down from the control group's 0.18 g / km. Power response time was shortened from 0.55 seconds in the control group to 0.40 seconds in the experimental group, resulting in smoother acceleration and shifting. Over the four months, the experimental group experienced a 52% decrease in injection-related alarms, a 13% decrease in engine maintenance, and an 80% reduction in vehicle downtime due to injection issues. Furthermore, the average model confidence score for vehicles using the proposed method was significantly higher than that of the control group, demonstrating the model's enhanced adaptability to complex operating conditions and significantly reduced injection prediction uncertainty. Through continuous feedback and self-learning mechanisms, the experimental group's fuel consumption and emissions indicators were consistently maintained at optimal levels. The savings in fuel and maintenance costs generated significant economic benefits for the bus group.

[0032] Table 1 Comparative data of intelligent fuel injection optimization and traditional fuel injection strategies in actual bus operation ; The data in Table 1 demonstrates that the intelligent fuel injection optimization method of the present invention demonstrates significant advantages in actual bus operations. In terms of fuel economy, the experimental group's average fuel consumption per 100 kilometers was 29.0 liters, significantly lower than the control group's 32.5 liters, resulting in a fuel saving rate of 10.8%. This significantly reduces fuel costs for bus companies. Furthermore, regarding environmental emissions, the experimental group's average NOx emissions were only 0.94 g / km, and its average CO emissions were 0.12 g / km, both lower than the control group's (1.20 g / km and 0.18 g / km, respectively), effectively improving the urban atmospheric environment and complying with higher emission regulations.

[0033] In terms of power performance, the experimental group's vehicle power response time was 0.40 seconds, a significant improvement from the control group's 0.55 seconds. This resulted in smoother and more timely acceleration, enhancing the driving experience and ride comfort. Regarding vehicle health and maintenance, the number of injection-related alarms dropped to 12 in the experimental group, compared to 26 in the control group. This demonstrates that the intelligent injection strategy significantly reduced injection system anomalies and improved engine operating stability. The engine maintenance rate dropped from 12.4% to 10.8%, and downtime was reduced from 10 to 2 days, significantly improving vehicle availability and business operational efficiency.

[0034] The performance of the intelligent control model itself showed that the average confidence score of the experimental group model reached 0.92, surpassing the control group's 0.77. This demonstrates that the integrated Bayesian neural network model of the present invention has greater adaptability and prediction confidence for complex operating conditions, making injection decisions more scientific and reliable. The average injection prediction uncertainty also decreased from 0.37 to 0.18, further demonstrating the stability and robustness of the model's prediction results. By adopting the intelligent injection strategy of the present invention, public buses not only save fuel, reduce emissions and maintenance costs, but also improve power performance and system stability, demonstrating its high engineering value and potential for widespread adoption.

[0035] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent optimization method for electronic fuel injection strategy, characterized in that: The steps include: S1. Collecting multi-source operating condition data of the engine and preprocessing the multi-source operating condition data to generate a standard data set; S2. Based on the standard data set, a Bayesian neural network model is constructed, which takes the standard data set as input and outputs the integrated injection parameter prediction value and the integrated uncertainty estimation value; S3. Based on the constructed Bayesian neural network model, the ant lion optimization algorithm is used to globally optimize the structural parameters and hyperparameters of the Bayesian neural network model to obtain the optimal structure and parameters of the Bayesian neural network model; S4. Based on the obtained optimal structure and parameters, train the Bayesian neural network model, use the standard data set for training, and evaluate the performance of the Bayesian neural network model; S5. During vehicle operation, the current engine operating condition data is collected in real time. Based on the trained Bayesian neural network model, the current operating condition data is input, and the optimal integrated injection parameter prediction value and integrated uncertainty estimate are output to dynamically adjust the injection control strategy. S6. Periodically collect engine operation feedback data after the injection control is executed, combine the standard data set with the engine operation feedback data, perform incremental training of the Bayesian neural network model, and re-execute the ant lion optimization algorithm in step S3 in a timely manner based on the operation results to achieve continuous adaptive optimization of the injection strategy.

2. The method for intelligent optimization of electronic fuel injection strategy based on deep learning according to claim 1, characterized in that: The multi-source operating condition data specifically include engine speed, load, throttle opening, coolant temperature, intake pressure, air-fuel ratio, emission concentration and fuel injection amount.

3. The method for intelligent optimization of electronic fuel injection strategy based on deep learning according to claim 1, characterized in that: The preprocessing of the multi-source operating condition data specifically includes normalizing the multi-source operating condition data, eliminating outliers and performing feature engineering processing.

4. The method for intelligent optimization of electronic fuel injection strategy based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. Define the standard dataset as an input vector , where each Indicates engine operating condition characteristics; S22. Initializing a Bayesian neural network model based on the input vector, the Bayesian neural network model comprising an input layer, m hidden layers, and an output layer, where m is a positive integer greater than or equal to 1; S23. Set the prior distribution type selection parameters for the weight parameters of the Bayesian neural network model , specify the prior distribution type of all weight parameters as Gaussian distribution; S24. Set the variational inference layer number parameters for the hidden layer structure of the Bayesian neural network model , determine the number or specific level of hidden layers involved in variational inference; S25. Based on the initialized Bayesian neural network model, construct k sub-models of the Bayesian neural network model with slightly different structures, and set the integrated weight parameters of the sub-models of the Bayesian neural network model , controls the weighted proportion of each Bayesian neural network model’s sub-model in the final output; S26, inputting the input vector into each sub-model of the Bayesian neural network model to obtain the injection parameter prediction value of each sub-model of the Bayesian neural network model , uncertainty estimates and confidence scores ; S27, based on the integrated weight parameters of the sub-model of the set Bayesian neural network model and confidence scores , using adaptive integration mechanism to calculate the integrated injection parameter prediction value and integrated uncertainty estimates : ； ； in, is the entropy adjustment coefficient, Represents the normalized entropy value of the sub-model confidence scores of all Bayesian neural network models, k is the number of sub-models of the Bayesian neural network model, is the weight corresponding to the sub-model of the j-th Bayesian neural network model, is the confidence score of the sub-model of the j-th Bayesian neural network model under the current working condition, is the weight corresponding to the sub-model of the l-th Bayesian neural network model, is the confidence score of the sub-model of the lth Bayesian neural network model under the current working condition; S28. Based on the standard data set, the posterior distribution of all weight parameters of the Bayesian neural network model is inferred by the Bayesian inference method, and the Markov chain Monte Carlo method is used to approximate the posterior distribution; S29, the obtained integrated injection parameter prediction value and integrated uncertainty estimates As the final output of the Bayesian neural network model.

5. The method for intelligent optimization of electronic fuel injection strategy based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31. The structural parameters and hyperparameters of the constructed Bayesian neural network model are set as the optimization objects of the Ant Lion Optimization Algorithm. The structural parameters include the number of network layers and the number of neurons in each layer. The hyperparameters include the prior distribution type selection parameters. , variational inference layer parameters , integrated weight parameters , where k is the number of sub-models of the Bayesian neural network model; S32. Initializing multiple antlion-ant populations based on the set optimization object, and independently generating initial population solution vectors in different subspaces of structural parameters and hyperparameters; S33. For each ant lion individual, set up an adaptive trap mechanism, and dynamically adjust the trap radius based on the global fitness distribution and the ant lion's historical predation success rate. and trap depth ; S34. For each ant individual, adaptively select the Levy flight path or the chaotic mapping path based on its historical fitness performance to form the current round of wandering strategy; S35. In each iteration, each ant randomly selects an antlion from the antlions as a leader according to the roulette wheel selection mechanism, updates its own position based on the antlion's trap radius and trap depth, and generates new structural parameters and hyperparameter solution vectors. S36. For each ant’s generated new structural parameters and hyperparameter solution vectors, a Bayesian neural network model is built and trained based on a standard data set to obtain an integrated injection parameter prediction. and integrated uncertainty estimates ; S37, based on the predicted value of integrated injection parameters and integrated uncertainty estimates , calculate the fitness function F, comprehensively consider the mean square error of the integrated injection parameter prediction value , the mean of the integrated uncertainty estimates and variance : ; in, 、 、 is the weight coefficient, are the real injection parameters of the standard data set; S38, sorting and evaluating the calculated fitness values, and performing trap updating, predation, individual position updating, and global optimal individual preservation operations in each antlion-ant population; S39. After a specified iteration cycle, each antlion-ant population exchanges and migrates excellent individuals to promote the search for the global optimal solution and maintain population diversity. S310: Determine whether the preset termination condition is met, and if so, output the optimal structural parameters and hyperparameter configurations of all current ant lion-ant populations. 、 、 , the optimal number of network layers and the number of neurons in each layer as the optimal structure and parameters of the Bayesian neural network model.

6. The method for intelligent optimization of electronic fuel injection strategy based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Build a Bayesian neural network model based on the obtained optimal structural parameters and hyperparameters; S42, dividing the standard data set into a training set and a validation set; S43, inputting the training set into the constructed Bayesian neural network model, performing parameter training, and updating the network weights and bias distribution; S44. After the training is completed, the validation set is input into the Bayesian neural network model to obtain an output result of the Bayesian neural network model on the validation set; S45. comparing the obtained output results of the Bayesian neural network model with the actual injection parameters of the validation set to evaluate the prediction performance and generalization ability of the Bayesian neural network model; S46. Determine whether the trained Bayesian neural network model meets the performance requirements based on the evaluation results. If so, save the Bayesian neural network model. If not, adjust the training parameters or return to step S43 for retraining.

7. The method for intelligent optimization of electronic fuel injection strategy based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51. During vehicle operation, continuously monitor the engine's operating status and collect real-time engine operating data, including engine speed, load, throttle opening, coolant temperature, intake pressure, air-fuel ratio, emission concentration, and fuel injection quantity. Preprocess the collected data to generate input vectors for intelligent optimization of the fuel injection strategy. S52, inputting the generated input vector into the trained Bayesian neural network model in real time, so that changes in the working conditions at each moment can be fed back to the Bayesian neural network model reasoning process in a timely manner; S53. Using the Bayesian neural network model, respectively obtain injection parameter prediction results and uncertainty estimation results corresponding to multiple sub-models of the Bayesian neural network model, as well as confidence scores of each sub-model of the Bayesian neural network model under the current operating conditions, and fully record the inference outputs of the sub-models of the Bayesian neural network model; S54. Based on the obtained sub-model outputs of all Bayesian neural network models, use an integration mechanism to perform weighted integration on the injection parameter prediction results and uncertainty estimation results of the sub-models of each Bayesian neural network model to obtain the optimal integrated injection parameter prediction value and integrated uncertainty estimation value under the current operating condition, and perform a validity check on the integration result. S55. Based on the obtained optimal integrated injection parameter prediction value and the integrated uncertainty estimate, and in combination with the actual needs of the engine control system, dynamically adjust the key injection control parameters of the injection pulse width and the injection timing.

8. The method for intelligent optimization of electronic fuel injection strategy based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. After the engine injection control strategy is executed, periodically collect engine operation feedback data, including fuel consumption, emission indicators, power response, and various operating condition parameters; S62. Normalize the collected engine operation feedback data and merge it with the standard data set to construct an incremental data set for updating the Bayesian neural network model; S63, inputting the incremental data set into the trained Bayesian neural network model, and updating the parameters of the Bayesian neural network model using an incremental learning method; S64. After incremental training of the Bayesian neural network model, re-evaluate the performance of the Bayesian neural network model based on the latest incremental data set to test the prediction accuracy and generalization ability of the Bayesian neural network model under various working conditions; S65. Based on the evaluation result of the Bayesian neural network model, determine whether the structural parameters and hyperparameters of the Bayesian neural network model need to be re-optimized. If necessary, start the ant lion optimization algorithm in step S3 to perform global optimization on the structural parameters and hyperparameters of the Bayesian neural network model. S66. Reapply the updated and optimized Bayesian neural network model to the real-time optimization control of the engine injection strategy to achieve continuous adaptive optimization of the injection control strategy and enter the next cycle of data collection and Bayesian neural network model update process.

Images