Gasoline engine exhaust system optimization method based on multi-objective improved artificial bee colony algorithm

Abstract

This invention relates to the field of gasoline engine exhaust system optimization technology, and discloses a gasoline engine exhaust system optimization method based on a multi-objective improved artificial bee colony algorithm. The method involves determining optimization variables and constraints; constructing a prediction model for multiple pollutant emissions in the gasoline engine exhaust system based on the XGBoost algorithm; initializing the bee population and related parameters; generating and updating new individuals using a search equation combined with a differential evolution algorithm during the hired bee phase; selecting individuals based on probability during the observation bee phase, generating and updating new individuals using a search equation combined with a particle swarm optimization algorithm; re-initializing individuals exceeding the threshold during the scout bee phase; and outputting the optimal parameter set upon iteration termination. This method aims to address the inability of existing gasoline engine exhaust system optimization methods to meet increasingly stringent environmental protection requirements.

Description

Technical Field

[0001] This invention relates to the field of gasoline engine exhaust system optimization technology, and specifically to a gasoline engine exhaust system optimization method based on a multi-objective improved artificial bee colony algorithm. Background Technology

[0002] With the rapid development of the automotive industry and increasingly stringent environmental protection requirements, the optimized design of gasoline engine exhaust systems has become a key aspect of improving engine performance and meeting emission regulations. The main function of a gasoline engine exhaust system is to expel the exhaust gases after combustion from the cylinders, while simultaneously purifying the harmful pollutants in the exhaust gases to reduce environmental pollution.

[0003] Currently, the optimization of gasoline engine exhaust systems mostly relies on empirical formulas or single-objective design methods.

[0004] In the actual design process, engineers usually optimize the exhaust back pressure or the purification efficiency of a certain type of pollutant by estimating core parameters such as exhaust manifold diameter, catalytic converter carrier volume or muffler expansion chamber size. They introduce electro-digital data processing technology and use computer-aided engineering to achieve physical field coupling analysis and target optimization.

[0005] For example, during optimization, the focus might be solely on reducing exhaust back pressure to increase engine power output, or only on improving the catalytic converter's ability to process nitrogen oxides. Purification efficiency.

[0006] However, this single-objective design optimization method has a series of serious problems. The various core components of the exhaust system, such as the exhaust manifold structure, catalytic converter parameters, and muffler dimensions, have extremely complex coupling relationships between flow resistance and purification efficiency. Taking the catalytic converter as an example, excessively increasing its carrier volume, while potentially improving efficiency to some extent... It improves purification efficiency, but at the same time it significantly increases exhaust back pressure.

[0007] Increased exhaust back pressure causes the engine to consume more energy to overcome resistance during exhaust, resulting in engine power loss and affecting the overall vehicle performance. This demonstrates that single-objective optimization often only solves local problems but can easily lead to a decline in other key performance indicators, failing to achieve optimal overall exhaust system performance. Summary of the Invention

[0008] The present invention aims to provide a method for optimizing gasoline engine exhaust systems based on a multi-objective improved artificial bee colony algorithm, in order to solve the technical problem that existing gasoline engine exhaust system optimization methods can no longer meet increasingly stringent environmental protection requirements.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a gasoline engine exhaust system optimization method based on a multi-objective improved artificial bee colony algorithm, comprising:

[0010] S1. Determine multiple optimization variables and the constraints on each optimization variable;

[0011] S2. Based on the goal of controlling the emissions of multiple pollutants in the exhaust system, an XGBoost prediction model is constructed to predict the emissions of multiple pollutants.

[0012] S3. Initialize the bee population. Randomly generate N bee colony individuals in the search space. Each bee colony individual corresponds to a set of optimization variable parameters. Set the artificial bee colony algorithm parameters. Use the XGBoost prediction model to calculate the pollutant emissions corresponding to each bee colony individual. Initialize the external archive set.

[0013] S4, the hired bee stage, uses a search equation improved by combining differential evolution algorithm to generate new candidate individuals, enhances local search ability, and updates individuals based on non-dominated ranking and crowding evaluation.

[0014] S5, Observation Bee Phase: Calculate the probability of each individual being selected based on non-dominated sorting and crowding. The observation bee selects individuals for searching based on this probability, generates new candidate individuals using a search equation improved by combining particle swarm optimization, and updates the individual and external archive set.

[0015] S6, Scout Bee Phase: For individuals whose Trial value exceeds the Scout Bee threshold, a random generation strategy is used to reinitialize the individuals to ensure the integrity of the non-dominated solution set.

[0016] S7. Iteration Termination: When the number of iterations reaches the maximum number of iterations, the iteration terminates, and the non-dominated solution set in the external archive is output as the final global optimal parameter set.

[0017] The principle and advantages of this scheme are: overcoming the limitations of existing single-objective optimization and empirical design of gasoline engine exhaust systems, by constructing an XGBoost prediction model with multiple core structural parameters as optimization variables, combining multiple pollutant emission optimization objectives, and using an improved artificial bee colony algorithm to achieve a balance between global search and local development, generating a Pareto optimal solution set covering different operating conditions, and achieving global optimal matching of exhaust system structural parameters, thereby simultaneously meeting the requirements of pollutant emission reduction and engine power performance protection under all operating conditions.

[0018] Preferably, as an improvement, in S1, the optimized parameters include the exhaust manifold diameter D, the catalytic converter carrier volume V, and the muffler expansion chamber volume. And the length L of the exhaust tailpipe; S2 contains multiple pollutants including hydrocarbons (HC), carbon monoxide (CO), and nitrogen oxides. And particulate matter (PM).

[0019] The beneficial effects of this improvement are: it clearly identifies key parameters such as exhaust manifold diameter and catalytic converter carrier volume as optimization targets, as well as HC, CO, and other parameters. Using PM2.5 as control targets makes optimization more targeted and comprehensive, enabling precise improvement of exhaust system performance and effective reduction of emissions of various pollutants.

[0020] Preferably, as an improvement, in S3, the parameters set for the artificial bee colony algorithm include the maximum number of iterations MaxGen, the number of hired bees, the number of observation bees, and the scout bee threshold Limit; each group corresponds to one set of optimization variables (D, V, ...). (L), calculate the pollutant emissions for each individual [ , , , And initialize the external archive set.

[0021] The beneficial effects of this improvement are: by reasonably setting parameters such as the maximum number of iterations and the number of bees, a clear framework and standard are provided for the operation of the algorithm, ensuring that the algorithm can efficiently search for the optimal solution within limited computing resources, while also guaranteeing the stability and reliability of the search process.

[0022] Preferably, as an improvement, in S4, during the hired bee phase, for each individual New candidate individuals are generated using a search equation improved by combining differential evolution algorithm. The search equation for the hired bee phase is:

[0023] ;

[0024] in, For the current individual The j-th dimension component; For newly generated candidate individuals The j-th dimension component; Let j be the j-th dimension component of the optimal non-dominated solution in the external archive set; The j-th dimension component of a randomly selected individual; For randomly selected individuals The j-th dimension component, Unlike the current individual ; is a random factor, and its value ranges from [-1, 1]; Let be another random factor, whose value ranges from [-1, 1];

[0025] Calculate the pollutant emissions of candidate individuals and update individuals based on non-dominant ranking and crowding evaluation: if a candidate individual is better than the current individual, replace the original individual and reset Trial; otherwise, Trial = Trial + 1.

[0026] The beneficial effects of this improvement are: by combining the differential evolution algorithm with the improved search equation, the local search capability is enhanced. By utilizing the optimal solution in the external archive set and the differential information of random individuals, it is possible to search for better solutions in the vicinity of the current individual more meticulously, thereby improving the search accuracy and efficiency of the algorithm and helping to quickly find high-quality non-dominated solutions.

[0027] Preferably, as an improvement, in S5, during the observation bee phase, the probability of each individual being selected by the observation bee is calculated based on the non-dominated ranking and crowding degree. The probability calculation formula is as follows:

[0028] ;

[0029] in, For individuals The probability of being selected by the observed bee; For individuals The fitness value is a comprehensive metric calculated based on non-dominated ranking and crowding. The sum of fitness values ​​of all individuals in the population; N is the total number of individuals in the population. It is the fitness value of the nth individual in the population; by iterating through n from 1 to N and summing the results, we obtain the sum of the fitness values ​​of the entire population.

[0030] Observe bees with probability Select individuals for searching, and use a combination of particle swarm optimization and other algorithms to improve the search equation and generate new candidate individuals. The search equation for the observation bee stage is as follows:

[0031] ;

[0032] in, For the current individual The j-th dimension component; For newly generated candidate individuals The j-th dimension component; Let j be the j-th dimension component of the optimal non-dominated solution in the external archive set; For randomly selected individuals The j-th dimension component, Unlike the current individual ; is a random factor, and its value ranges from [-1, 1]; Let be another random factor, whose value range is [-1, 1], which is related to... similar;

[0033] Calculate the pollutant emissions of candidate individuals. If a candidate individual is better than the current individual, replace it and reset the Trial. At the same time, update the external archive set to maintain the Pareto front solution.

[0034] The beneficial effects of this improvement are: by calculating the selection probability based on non-dominated sorting and crowding, and by combining it with the particle swarm optimization algorithm to improve the search equation, the observer bees can more effectively select individuals for searching. This probability-based selection mechanism and improved search method balance the algorithm's local exploitation and global exploration capabilities, further enhancing its search performance and enabling it to find solutions that better satisfy multi-objective optimization.

[0035] Preferably, as an improvement, in S6, during the scout bee phase, when an individual's Trial value exceeds the threshold Limit, the individual is reinitialized using a random generation strategy, and the external archive set is updated simultaneously; the formula for the random generation strategy is as follows:

[0036] ;

[0037] in, For the current individual The j-th dimension component; This is the lower bound of the j-th dimension variable; This is the upper bound of the j-th dimension variable; This is a random number generation function used to generate a random number that is uniformly distributed within the interval (0,1).

[0038] When the current iteration number Gen ≥ MaxGen, the iteration terminates, and the non-dominated solution set in the external archive is output as the final global optimal parameter set.

[0039] The beneficial effect of this improvement is that when an individual's Trial value exceeds a threshold, the individual is reinitialized using a random generation strategy, ensuring the integrity of the non-dominated solution set. This helps the algorithm escape from local optima in a timely manner, continue exploring the solution space, increases the algorithm's global search capability, avoids premature convergence, and thus finds a better global optimal parameter set.

[0040] Preferably, as an improvement, the constraints in S1 include:

[0041] The exhaust manifold diameter D ranges from 35 to 60 mm; the catalytic converter carrier volume V ranges from 1.2 to 2.5 L; and the muffler expansion chamber volume... The value range is 3.0~5.5L; the value range of the exhaust tailpipe length L is 400~800mm.

[0042] The beneficial effects of this improvement are: it clarifies the value range of each optimization variable, provides a reasonable boundary for the algorithm's search, avoids searching for unrealistic or unreasonable parameter combinations, improves the feasibility and practicality of the optimization results, and also reduces the algorithm's search space and improves search efficiency.

[0043] Preferably, as an improvement, in S2, the XGBoost model constructed for outputting pollutant emissions includes:

[0044] The input features of the XGBoost model are the exhaust manifold diameter D, the catalytic converter carrier volume V, and the muffler expansion chamber volume. The physical constraints for the four optimization parameters, including exhaust manifold diameter D ∈ [35, 60] mm, catalytic converter carrier volume V ∈ [1.2, 2.5] L, and muffler expansion chamber volume, are as follows: ∈[3.0,5.5]L, exhaust tailpipe length L∈[400,800]mm;

[0045] The output of the XGBoost model is the predicted pollutant emissions of a gasoline engine under different combinations of structural parameters. The target variables include the emissions of four pollutants: hydrocarbons, carbon monoxide, nitrogen oxides, and particulate matter.

[0046] The model training data comes from bench tests or high-fidelity engine simulation systems. The data on structural inputs and pollutant emission results corresponding to multiple parameter combinations are collected under typical operating conditions to construct a training sample set.

[0047] Model training includes: performing data preprocessing on the original dataset, dividing the preprocessed data into training and validation sets, using the XGBoost regression algorithm to construct four independent prediction sub-models for hydrocarbons, carbon monoxide, nitrogen oxides and particulate matter respectively, setting hyperparameters including learning rate, maximum tree depth, subsampling ratio and regularization parameter, and optimizing model performance and avoiding overfitting through cross-validation.

[0048] After the model is trained, it can be used to input any set of valid structural parameters (D, V, ...). Under the condition of L), the emissions of four types of pollutants, namely hydrocarbons, carbon monoxide, nitrogen oxides and particulate matter, are predicted, providing pollutant emissions as fitness assessment input for the subsequent optimization steps based on the multi-objective artificial bee colony algorithm.

[0049] The beneficial effects of this improvement are as follows: By constructing a pollutant emission prediction model based on XGBoost, unified modeling and output of emissions from four categories—hydrocarbons, carbon monoxide, nitrogen oxides, and particulate matter—are achieved. Compared to traditional explicit mathematical objective function expressions, this model better reflects the nonlinear coupling relationships and multi-objective influence mechanisms among exhaust system structural parameters. The model can automatically predict the emission levels of the four pollutants based on exhaust system parameter inputs, making the objective evaluation process quantifiable, rapid, and scalable, providing continuous, differentiable, and high-precision objective feedback for multi-objective optimization algorithms. With this data-driven model, the accuracy of individual performance evaluation and Pareto solution convergence efficiency during the optimization process can be significantly improved, facilitating the acquisition of better structural combinations within the design space and achieving synergistic suppression and global optimization of multi-pollutant emissions. Attached Figure Description

[0050] Figure 1 This is a flowchart of an embodiment of the present invention.

[0051] Figure 2 The fitness convergence curves of four multi-objective optimization algorithms during the iterative process are shown.

[0052] Figure 3 These are the predicted values ​​for the CO test set based on the XGBoost model. Detailed Implementation

[0053] The following detailed description illustrates the specific implementation method:

[0054] Example

[0055] The basics are as follows: Figure 1 As shown, the gasoline engine exhaust system optimization method based on the multi-objective improved artificial bee colony algorithm includes:

[0056] S1. Determine multiple optimization variables and the constraints on each optimization variable;

[0057] S2. Based on the goal of controlling the emissions of multiple pollutants in the exhaust system, construct an XGBoost model for outputting pollutant emissions.

[0058] S3. Initialize the bee colony. Randomly generate N bee colony individuals in the search space. Each bee colony individual corresponds to a set of optimization variable parameters. Set the artificial bee colony algorithm parameters, calculate the pollutant emission amount corresponding to each bee colony individual, and initialize the external archive set.

[0059] S4, the hired bee stage, uses a search equation improved by combining differential evolution algorithm to generate new candidate individuals, enhances local search ability, and updates individuals based on non-dominated ranking and crowding evaluation.

[0060] S5, Observation Bee Phase: Calculate the probability of each individual being selected based on non-dominated sorting and crowding. The observation bee selects individuals for searching based on this probability, generates new candidate individuals using a search equation improved by combining particle swarm optimization, and updates the individual and external archive set.

[0061] S6, Scout Bee Phase: For individuals whose Trial value exceeds the Scout Bee threshold, a random generation strategy is used to reinitialize the individuals to ensure the integrity of the non-dominated solution set.

[0062] S7. Iteration Termination: When the number of iterations reaches the maximum number of iterations, the iteration terminates, and the non-dominated solution set in the external archive is output as the final global optimal parameter set.

[0063] In S1, the optimized parameters need to cover key influencing factors of exhaust purification efficiency and system flow resistance. In this embodiment, the optimized parameters include exhaust manifold diameter D, catalytic converter carrier volume V, and muffler expansion chamber volume. The length L of the exhaust tailpipe. The constraints of each optimization parameter are determined based on the requirements of China VI emission regulations and indicators such as exhaust system performance.

[0064] Specifically:

[0065] To balance exhaust gas flow uniformity and flow resistance, the exhaust manifold diameter D is selected as the optimization variable. Constraints: The exhaust manifold diameter D ranges from 35 to 60 mm. The lower limit of the range corresponds to a small diameter, i.e., high flow velocity, which easily generates local turbulence and may reduce the intake uniformity of the catalytic converter; the upper limit of the range corresponds to a large diameter, i.e., low flow velocity and low flow resistance, but may increase the manifold volume and layout difficulty.

[0066] To ensure a balance between pollutant purification efficiency and exhaust back pressure, the catalytic converter carrier volume V is selected as the optimization variable. Constraints: The catalytic converter carrier volume V ranges from 1.2 to 2.5 L, meeting the China VI emission standards: hydrocarbon (HC) emissions ≤10 mg / km, carbon monoxide (CO) emissions ≤100 mg / km, and nitrogen oxide emissions ≤100 mg / km. Emissions ≤35mg / km, and exhaust back pressure ≤30kPa.

[0067] To balance noise reduction and flow resistance, the volume of the de-oscillator expansion cavity was selected. To optimize variables. Constraints: muffler expansion cavity volume. The value range is 3.0~5.5L, which ensures that the insertion loss is ≥25dB ​​and the local resistance coefficient of the expansion cavity is ≤1.8.

[0068] Based on the principle of exhaust airflow attenuation and back pressure matching, the exhaust tailpipe length L is selected as the optimization variable. Constraints: The exhaust tailpipe length L is in the range of 400~800mm, that is, to ensure that the airflow velocity at the tailpipe outlet is ≤30m / s, so as to avoid excessive aerodynamic noise generated by high-speed airflow.

[0069] In S2, various pollutants include hydrocarbons (HC), carbon monoxide (CO), and nitrogen oxides. and particulate matter (PM). In this embodiment, HC, CO, A pollutant emission prediction model was constructed using four pollutants: PM2.5, ... and PM2.5.

[0070] By optimizing the exhaust system's structural parameters, the emissions of various pollutants are minimized under typical engine operating conditions. The XGBoost model quantifies the "emission control target" into a computable optimization function, guiding the improved artificial bee colony algorithm to efficiently search for the optimal solution.

[0071] The key emission control objective for gasoline engine exhaust systems is to minimize pollutant emissions. The core function of the XGBoost model is to learn the nonlinear mapping relationship between "exhaust system parameters - operating conditions - emissions of various pollutants," serving as a "virtual emission test bench" to replace real bench tests and providing rapid emission predictions for optimization algorithms.

[0072] The XGBoost model is deeply coupled with the improved artificial bee colony algorithm, which transforms "multiple pollutant emission control" into the algorithm's fitness evaluation rules. The improved artificial bee colony algorithm's hired bee and observation bee stages require the XGBoost model to predict the emission amounts of candidate solutions and the Pareto non-dominated sorting to evaluate the quality of the solutions.

[0073] The XGBoost model uses a variety of pollutants, including hydrocarbons (HC), carbon monoxide (CO), and nitrogen oxides. With particulate matter (PM) as the prediction target, the input parameters are exhaust manifold diameter D, catalytic converter carrier volume V, and muffler expansion chamber volume. And the length L of the exhaust tailpipe.

[0074] A unified preprocessing was performed on 126 sets of experimental data: the Daubechies (db4) wavelet function was used for five-level decomposition and combined with hard threshold denoising, which can effectively suppress noise and preserve key information of BSFC signals.

[0075] For input parameters (D, V, ... The values ​​(L) are standardized using Z-Score standardization, as shown in the following formula: ,in The mean of the training set parameters, The standard deviation of the training set parameters is used to eliminate the influence of differences in parameter dimensions on model training.

[0076] The data split ratio is determined based on the sample size. Approximately 70% of the data is used as the training set and 30% as the test set. Therefore, the 126 test data sets are randomly divided into the training set and the test set in a 7:3 ratio, with the training set containing 88 data sets and the test set containing 38 data sets.

[0077] To avoid overfitting the training set, 5-fold cross-validation was used (the 88 sets of data in the training set were divided into 5 subsets, with 4 subsets used for training and 1 subset used for validation each time). The validation metrics were "mean absolute error (MAE)" and "coefficient of determination (R²)". 2 Root Mean Square Error (RMSE):

[0078]

[0079] in, The actual value of the target variable. These are the model's predicted values. The target variable is the average value, and n is the sample size. The evaluation metrics for this model are shown in Table 1 below:

[0080] Table 1 Model Evaluation Index Table

[0081]

[0082] Among them, due to the large magnitude of the output indicators, namely HC, CO, In ppm, its order of magnitude reaches Even if the relative difference between the predicted and actual values ​​is small and the model prediction is relatively accurate, the mean absolute error and root mean square error on the test set will still be relatively large.

[0083] The final hyperparameters for the XGBoost model were selected as follows: "Number of trees" = 44, "Maximum tree depth" = 3, and "Learning rate" = 0.16. An example model for predicting CO is shown below. Figure 3 As shown, the comparison between the prediction results and experimental values ​​of the established model on the CO test set demonstrates that the model can fully learn the complex nonlinear mapping between input and output within the known sample space.

[0084] In S3, the parameters set for the artificial bee colony algorithm include the maximum number of iterations (MaxGen), the number of hired bees, the number of observer bees, and the scout bee threshold (Limit). In this embodiment, N=60 bee colony individuals are randomly generated within the search space, with each group corresponding to one set of optimization variables (D, V, ...). L).

[0085] Set the maximum number of iterations MaxGen=100, the number of hired bees and observer bees to N / 2 each, and the scout bee threshold Limit=15; calculate the multi-objective optimization model value for each individual. , , , And initialize the external archive set.

[0086] In S4, during the hired bee phase, for each individual New candidate individuals are generated using a search equation improved by combining differential evolution algorithm. The search equation for the hired bee phase is:

[0087] ;

[0088] in, For the current individual The j-th component, which is part of the individuals being improved in the population, represents the location information of a solution in the current search process. For newly generated candidate individuals The j-th dimension component, during the optimization process, is a new individual component calculated through this equation, which is used for subsequent evaluation and selection operations to explore a better solution space.

[0089] Let be the j-th dimension component of the optimal non-dominated solution in the external archive set; the external archive is used to store high-quality solutions found during the search process. It is the j-th dimension component of the solution that performs best in the sense of multi-objective optimization among these high-quality solutions. By calculating the crowding degree, solutions with uniform distribution are preferentially selected to enter the external archive, and the component of the best solution is selected from them to participate in the generation of new individuals, which helps to guide the search direction towards a better region.

[0090] By randomly selecting the j-th dimension component of an individual and incorporating the information of that random individual into the generation of a new individual, we can increase the diversity of the population and prevent the algorithm from converging to a local optimum too early. For randomly selected individuals The j-th dimension component, Unlike the current individual . The random factor has a value range of [-1, 1], which expands the random search range and enhances the local exploitation capability of the algorithm. Let be another random factor, whose value range is [-1, 1], which is related to... Similarly, it is also used to expand the scope of random search and enhance local development capabilities.

[0091] Calculate the multi-objective optimization model value for candidate individuals and update the individuals based on non-dominated ranking and crowding evaluation: if a candidate individual is better than the current individual, replace the original individual and reset Trial; otherwise, Trial = Trial + 1. Trial is a counter used to record the current individual. No new candidate was found in multiple consecutive searches. The number of replacements. When the Trial value reaches a certain threshold, it may mean that the individual has fallen into a local optimum.

[0092] In S5, during the observation bee phase, the probability of each individual being selected by the observation bee is calculated based on non-dominated ordination and crowding. The probability calculation formula is as follows:

[0093] ;

[0094] in, For individuals The probability of being selected by the observation bee. During the observation bee phase, the observation bee uses this probability to decide which individual to select for subsequent search operations. The higher the probability, the greater the likelihood that the individual will be selected.

[0095] For individuals The fitness value is a comprehensive metric calculated based on non-dominated ranking and crowding.

[0096] Non-dominated ranking: In multi-objective optimization problems, individuals in the population are ranked hierarchically. Non-dominated solutions are placed in the first tier, and the remaining non-dominated solutions after removing the solutions from the first tier are placed in the second tier, and so on. Generally, individuals in better tiers, i.e., those in smaller tiers, have relatively higher fitness values.

[0097] Crowding: This measures the distribution of an individual in the solution space. A high crowding level indicates that the density of solutions around the individual is low, which helps maintain solution diversity. Crowding is also considered as a factor when calculating fitness values, resulting in sparser individuals having higher fitness values.

[0098] The group is stratified according to multi-objective non-dominance relationships. Individuals in the first stratum have a rank of 1, those in the second stratum have a rank of 2, and so on. Let the rank be denoted as _____. i The rank of an individual is ( ). i ).

[0099] Non-dominated solution: In a multi-objective optimization problem, if a solution is no worse than another solution in all objective functions, and is strictly better than another solution in at least one objective function, then the solution is said to be non-dominated by the other solution.

[0100] In other words, a non-dominated solution is one in the current solution set that no other solution can perform better on all objectives.

[0101] The specific process of non-dominated solution layering is as follows:

[0102] First non-dominated layer (Rank = 1): First, find all solutions in the entire solution group that are not dominated by any other individual. These solutions constitute the first non-dominated layer, and their rank is denoted as rank = 1.

[0103] These solutions are optimal in the current solution set because they perform no worse than other solutions on all objectives, and at least better on one objective.

[0104] Second non-dominated layer (Rank = 2): Next, all solutions from the first non-dominated layer are removed from the solution group, and then solutions that are not dominated by any other individual are selected from the remaining solutions. These solutions constitute the second non-dominated layer, and the rank is denoted as rank = 2.

[0105] These solutions are optimal among the remaining solutions.

[0106] And so on: Repeat the above process, each time removing the solution of the current best layer (i.e. the layer with the smallest rank value), and then selecting new non-dominated solutions from the remaining solutions to form the next layer, until all solutions are assigned to one layer.

[0107] For each objective m, sort the population in ascending order of the objective value, and maximize the contribution of the first and last individuals in the sorted sequence to that objective; for the middle individual i, the contribution value to objective m is:

[0108] ;

[0109] in Let m be the target value for the k-th individual. and These are the maximum and minimum values ​​of the objective in the current group, respectively. Then, the crowding degree of an individual is obtained by summing the contributions of that individual to all objectives:

[0110] ;

[0111] The fitness is obtained by normalizing and weighting the non-dominated hierarchy and crowding degree. :

[0112] ;

[0113] Among them, weight , ,satisfy .

[0114] The sum of fitness values ​​of all individuals in the population. N is the total number of individuals in the population. This is the fitness value of the nth individual in the population; by iterating through n from 1 to N and summing the results, we obtain the sum of the fitness values ​​of the entire population. This sum serves as a normalization function when calculating the probability of each individual being selected, ensuring that the sum of the probabilities of all individuals being selected is 1. The minimum crowding level among all individuals. This represents the maximum crowding level among all individuals.

[0115] Observe the bees using a roulette wheel method to determine probability. Select individuals for searching, and use a combination of particle swarm optimization and other algorithms to improve the search equation and generate new candidate individuals. The search equation for the observation bee stage is as follows:

[0116] ;

[0117] in, For the current individual The j-th dimension component; For newly generated candidate individuals The j-th dimension component; Let j be the j-th dimension component of the optimal non-dominated solution in the external archive set; For randomly selected individuals The j-th dimension component, Unlike the current individual ; is a random factor, and its value ranges from [-1, 1]; Let be another random factor, whose value range is [-1, 1], which is related to... similar.

[0118] Calculate the multi-objective optimization model value of the candidate individual. If the candidate individual is better than the current individual, replace it and reset the Trial. At the same time, update the external archive set to maintain the diversity of Pareto front solutions, i.e., non-dominated solutions.

[0119] In S6, during the scout bee phase, individuals whose Trial value exceeds the threshold Limit are reinitialized using a random generation strategy. Simultaneously, the external archive set is updated to ensure the integrity of the non-dominated solution set. The formula for the random generation strategy is as follows:

[0120] ;

[0121] in, For the current individual The j-th dimension component. This is the lower bound of the j-th dimension variable, and its value can be obtained from the constraints of each optimization variable in S1. Let the upper bound and lower bound of the j-th dimension variable be... Correspondingly. `<random_number>` is a random number generation function used to generate a random number uniformly distributed within the interval (0,1). This random number introduces randomness into the generation of individual components, making the value range of each generated individual uncertain, which helps to perform extensive searches in the solution space.

[0122] When the current iteration number Gen≥MaxGen, the iteration terminates and the non-dominated solution set in the external archive is output as the final global optimal parameter set. The non-dominated solution in the archive constitutes the "intake efficiency-power-torque optimal parameter set", which contains 15~25 sets of feasible parameters, covering the balance requirements under different operating conditions.

[0123] To verify the effectiveness of the multi-objective exhaust optimization strategy, this invention constructs exhaust manifold diameter D, catalytic converter carrier volume V, and muffler expansion chamber volume. And a multi-objective optimization model with four parameters, including exhaust tailpipe length L, and CO, A non-dominated ranking solution system is established with four emission indicators, namely HC, PM and PM, as the optimization objectives.

[0124] The baseline structure (control group) was set with D=46mm and V=1.3L. =3.8L, L=600mm, under these conditions the emission test results are: CO=4311.731ppm, =2253.865ppm, HC=2270.958ppm, PM=6.2 The pollutants were all at high levels, indicating that the structure failed to simultaneously achieve both oxidation efficiency and conversion efficiency, leaving room for further optimization.

[0125] Based on this, several typical non-dominated solutions were obtained through multi-objective optimization iteration, as shown in Table 2 below:

[0126] Table 2 Comparison of parameters and emission results of typical non-dominated solutions in multi-objective optimization

[0127]

[0128] in:

[0129] (1) The optimal solutions for CO and PM are consistent, with corresponding structures of D≈51.35mm and V≈1.51L. With a diameter of approximately 5.04L and a thickness of approximately 697mm, the emissions are approximately CO 4075ppm and PM 1.82, respectively. .

[0130] Analysis suggests that increased volume can reduce exhaust back pressure and slow down gas cooling, while a longer tailpipe allows for more thorough secondary oxidation reactions, resulting in a simultaneous and significant decrease in both CO and PM.

[0131] (2) The optimal solutions for NOx and HC are the same, corresponding to structures with D≈40.37mm and V≈1.67L. ≈4.75L, L≈495mm, the emission result at this time is The concentrations were approximately 592 ppm, and HC was approximately 1536 ppm, which were significantly lower than the control conditions.

[0132] It is inferred that due to the smaller pipe diameter and moderate... To moderately reduce the exhaust temperature, The formation is suppressed, while appropriately increasing V improves the catalytic reaction sufficiency, thereby... While decreasing, we should avoid a significant increase in HC to form a new equilibrium range.

[0133] As shown in Table 2, this invention avoids the "reduction" problem in traditional single-index parameter tuning. However, "HC and PM levels deteriorated sharply" or "reducing CO levels led to..." The problem is a one-way trade-off of "ascent"; the non-dominated solution set forms two distinct engineering optimal clusters:

[0134] (1) Big D + Big +L length→Suitable for CO / PM sensitive operating conditions;

[0135] (2) Small to medium D + moderate V / +Shorter L→Suitable for / HC preferred operating condition.

[0136] Therefore, the final output of 15 to 25 Pareto solutions constitutes a "structure-emission balance library" that can be directly used for exhaust development and design, and has strong engineering versatility.

[0137] As attached Figure 2 The figure shows the fitness convergence curves of four multi-objective optimization algorithms during the iterative process. (See attached figure.) Figure 2 As shown, the traditional NSGA-II converges slowly, exhibiting fitness step-down behavior in the first 150 generations; MOPSO converges quickly but suffers from significant oscillations in later stages; MOABC shows improved convergence performance but still exhibits global jumps. The improved MOIABC algorithm (black solid line) of this invention approaches a stable region within approximately 50 generations, demonstrating an overall convergence speed superior to the control algorithm, as shown below:

[0138] (1) The steep drop in the initial stage indicates that the global guided differential evolution search quickly exposes the optimal region;

[0139] (2) The mid-to-late stage tends to be smooth and there is no significant rebound, which proves that the PSO observation bee update mechanism with inertial weights effectively suppresses premature convergence.

[0140] (3) The lowest final fitness indicates the best approximation of the global optimum, and the Pareto solution has a more complete boundary and a more balanced distribution.

[0141] Therefore, the improved multi-objective artificial bee colony optimization framework proposed in this invention can obtain a globally optimal parameter set with fast convergence speed, complete Pareto front, and significant improvement in pollutants with fewer iterations, and finally output the optimal solution group for engineering deployment, providing a data foundation and strategy basis for subsequent exhaust structure matching, regulatory adaptation calibration and road durability verification.

[0142] This scheme avoids the problem of "one pollutant decreasing while other pollutants worsen" in traditional single-objective optimization. By improving the search equations of the hired bee and observer bee in the traditional ABC algorithm, a differential evolution search equation guided by the global optimum is designed in the hired bee stage to enhance local search accuracy. In the observer bee stage, a particle swarm search equation with inertial weights is used to balance global exploration capability. At the same time, by combining non-dominated sorting and congestion evaluation mechanisms, an external archive set is introduced to manage non-dominated solutions, realizing global optimization of key parameters of the exhaust system and construction of Pareto front solution set, ensuring that the solution set covers emission requirements under all operating conditions.

[0143] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for optimizing gasoline engine exhaust systems based on a multi-objective improved artificial bee colony algorithm, characterized in that, include: S1. Determine multiple optimization variables and the constraints on each optimization variable; S2. Based on the goal of controlling the emissions of multiple pollutants in the exhaust system, an XGBoost prediction model is constructed to predict the emissions of multiple pollutants. XGBoost prediction models include: The input features of the XGBoost prediction model are the exhaust manifold diameter D, the catalytic converter carrier volume V, and the muffler expansion chamber volume. The physical constraints for the four optimization parameters, including exhaust manifold diameter D ∈ [35, 60] mm, catalytic converter carrier volume V ∈ [1.2, 2.5] L, and muffler expansion chamber volume, are as follows: ∈[3.0,5.5]L, exhaust tailpipe length L∈[400,800]mm; The output of the XGBoost prediction model is the prediction results of pollutant emissions from gasoline engines under different combinations of structural parameters. The target variables include the emissions of four pollutants: hydrocarbons, carbon monoxide, nitrogen oxides, and particulate matter. The model training data comes from bench tests or high-fidelity engine simulation systems. The data on structural inputs and pollutant emission results corresponding to multiple parameter combinations are collected under typical operating conditions to construct a training sample set. Model training includes: performing data preprocessing on the original dataset, dividing the preprocessed data into training and validation sets, using the XGBoost regression algorithm to construct four independent prediction sub-models for hydrocarbons, carbon monoxide, nitrogen oxides and particulate matter respectively, setting hyperparameters including learning rate, maximum tree depth, subsampling ratio and regularization parameter, and optimizing model performance and avoiding overfitting through cross-validation. After the model is trained, it can be used to input any set of valid structural parameters (D, V, ...). Under the condition of L), the emissions of four types of pollutants, namely hydrocarbons, carbon monoxide, nitrogen oxides and particulate matter, are predicted, so as to provide pollutant emissions as fitness assessment input for the subsequent optimization steps based on multi-objective artificial bee colony algorithm. S3. Initialize the bee population. Randomly generate N bee colony individuals in the search space. Each bee colony individual corresponds to a set of optimization variable parameters. Set the artificial bee colony algorithm parameters. Use the XGBoost prediction model to calculate the pollutant emissions corresponding to each bee colony individual. Initialize the external archive set. S4, the hired bee stage, uses a search equation improved by combining differential evolution algorithm to generate new candidate individuals, enhances local search ability, and updates individuals based on non-dominated ranking and crowding evaluation. S5, Observation Bee Phase: Calculate the probability of each individual being selected based on non-dominated sorting and crowding. The observation bee selects individuals for searching based on this probability, generates new candidate individuals using a search equation improved by combining particle swarm optimization, and updates the individual and external archive set. S6, Scout Bee Phase: For individuals whose Trial value exceeds the Scout Bee threshold, a random generation strategy is used to reinitialize the individuals to ensure the integrity of the non-dominated solution set. S7. Iteration Termination: When the number of iterations reaches the maximum number of iterations, the iteration terminates and the non-dominated solution set in the external archive is output as the final global optimal parameter set. Specifically, in S4, during the hired bee phase, for each individual New candidate individuals are generated using a search equation improved by combining differential evolution algorithm. The search equation for the hired bee phase is: ； in, For the current individual The j-th dimension component; For newly generated candidate individuals The j-th dimension component; Let j be the j-th dimension component of the optimal non-dominated solution in the external archive set; The j-th dimension component of a randomly selected individual; For randomly selected individuals The j-th dimension component, Unlike the current individual ; is a random factor, and its value ranges from [-1, 1]; Let be another random factor, whose value ranges from [-1, 1]; The pollutant emissions of candidate individuals are calculated using the XGBoost prediction model, and individuals are updated based on non-dominant ranking and crowding evaluation: if a candidate individual is better than the current individual, the original individual is replaced and the Trial is reset; otherwise, Trial = Trial + 1. In S5, during the observation bee phase, the probability of each individual being selected by the observation bee is calculated based on non-dominated ordination and crowding. The probability calculation formula is as follows: ； in, For individuals The probability of being selected by the observed bee; For individuals The fitness value is a comprehensive metric calculated based on non-dominated ranking and crowding. The sum of fitness values ​​of all individuals in the population; N is the total number of individuals in the population. It is the fitness value of the nth individual in the population; by iterating through n from 1 to N and summing the results, we obtain the sum of the fitness values ​​of the entire population. Observe bees with probability Select individuals for searching, and use a combination of particle swarm optimization and other algorithms to improve the search equation and generate new candidate individuals. The search equation for the observation bee stage is as follows: ； in, For the current individual The j-th dimension component; For newly generated candidate individuals The j-th dimension component; Let j be the j-th dimension component of the optimal non-dominated solution in the external archive set; For randomly selected individuals The j-th dimension component, Unlike the current individual ; is a random factor, and its value ranges from [-1, 1]; Let be another random factor, whose value range is [-1, 1], which is related to... similar; The XGBoost prediction model is used to calculate the pollutant emissions of candidate individuals. If a candidate individual is better than the current individual, it is replaced and the Trial is reset. At the same time, the external archive set is updated to maintain the Pareto front solution.

2. The method for optimizing a gasoline engine exhaust system based on a multi-objective improved artificial bee colony algorithm according to claim 1, characterized in that: In S1, the optimized parameters include the exhaust manifold diameter D, the catalytic converter carrier volume V, and the muffler expansion chamber volume. And the length L of the exhaust tailpipe; S2 contains multiple pollutants including hydrocarbons (HC), carbon monoxide (CO), and nitrogen oxides. And particulate matter (PM).

3. The method for optimizing a gasoline engine exhaust system based on a multi-objective improved artificial bee colony algorithm according to claim 2, characterized in that: In S3, the parameters set for the artificial bee colony algorithm include the maximum number of iterations (MaxGen), the number of hired bees, the number of observer bees, and the scout bee threshold (Limit); each group corresponds to one set of optimization variables (D, V, ...). (L), calculate the pollutant emissions for each individual [ , , , And initialize the external archive set.

4. The method for optimizing a gasoline engine exhaust system based on a multi-objective improved artificial bee colony algorithm according to claim 3, characterized in that: In S6, during the scout bee phase, for individuals whose Trial value exceeds the threshold Limit, the individuals are reinitialized using a random generation strategy, and the external archive set is updated simultaneously. The formula for the random generation strategy is as follows: ； in, For the current individual The j-th dimension component; This is the lower bound of the j-th dimension variable; This is the upper bound of the j-th dimension variable; This is a random number generation function used to generate a random number that is uniformly distributed within the interval (0,1). When the current iteration number Gen ≥ MaxGen, the iteration terminates, and the non-dominated solution set in the external archive is output as the final global optimal parameter set.

5. The method for optimizing a gasoline engine exhaust system based on a multi-objective improved artificial bee colony algorithm according to claim 4, characterized in that, In S1, the constraints include: The exhaust manifold diameter D ranges from 35 to 60 mm; the catalytic converter carrier volume V ranges from 1.2 to 2.5 L; and the muffler expansion chamber volume... The value range is 3.0~5.5L; the value range of the exhaust tailpipe length L is 400~800mm.

Images