Research On Gear Shift Decision Based On SVM Learning Model

The shift decision problem is the core problem of the shift control of the automatic transmission of the vehicle,and is the basis for the improvement of the shift control quality.The goal is to select a reasonable shift timing or shift point in the complex driving conditions.When to shift is a prerequisite for shift control,directly affecting the power,economy and driving of the vehicle.The traditional shift decision method is based on the fixed MAP form.The two-parameter to three-parameter theoretical design method proves that introducing more effective control parameters will greatly improve the adaptability of the shift law,but according to the shifting law MAP design method,In addition to calibrating the data,the designer has to carry out a large number of strategy development work to solve various problems represented by adaptability in the practical process.The design workload and human design factors are very large.Moreover,this calibration correction is generally carried out by separating the driver's style and intention from the environmental factors,and the overlay correction is performed by identification one by one,but the style and intention of the driver while driving the vehicle are actually made according to the environment.Not a single style and intention,such as the modified superposition of the ramp will cause the superposition of power demand to increase.At the same time,with the comprehensive development of artificial intelligence technology,people's demand for intelligence is constantly improving,and the method of gear decision-making must be continuously developed intelligently to meet the needs of the times and markets.Using artificial intelligence technology to solve the above problems is an effective way and an attempt.The machine learning algorithm aiming at artificial intelligence is the main means to realize the artificial intelligence method at present.It is the key technology development of artificial intelligence to practicality.The representative algorithm SVM is a data mining technology that automatically performs simulation and implementation.A limited number of examples(data)to learn the knowledge of the system,free from the assumptions of traditional statistical optimization methods for the amount of data.SVM can automatically optimize the classification hyperplane in the space of multi-dimensional control parameters according to strict mathematical principles,improve the adaptability of shift decision,and can greatly reduce the workload and design human factors of designers,the black box characteristics of machine learning algorithms It is possible to learn the mapping knowledge of driving style and intention into driving parameters,and it is not necessary to separate them,but to learn as the response knowledge of other control parameters,which is more in line with the natural behavior of the driver's driving operation.The quantity learning technology solves the requirements of one-time sample training,and also provides the ability of online continuous learning,providing an interface for personalized learning and laying a foundation for intelligence.This paper combines the National Natural Science Foundation's research on SVM modeling learning shift decision-making.1)The basic research of shifting theory,the theoretical basis and basic method of shift decision-making problem,and the adaptability.The traditional shift decision-making method is too complicated and difficult to use when analyzing multi-parameters.Therefore,the adaptability is not good and a large number of corrections are used.And strategy formulation to improve adaptability creates a burden on the designer's workload,and has no continuous learning ability;while SVM has a strict theoretical foundation,superior performance and generalization ability when dealing with multi-dimensional problems,not like neural networks.Get into a local optimal solution and get rid of the designer's a priori limits.At the same time,the power transmission system simulation model of the automatic transmission vehicle is established,which provides support for the gear test of the subsequent shift decision model.The model is verified by the road test data of the target vehicle.2)Construction of SVM model.The structure of SVM model is selected,and the factors affecting the performance of SVM model are analyzed,including the selection of kernel function,feature parameters and model parameters.Based on the transformation theory of the kernel method,the kernel functions suitable for the shift decision-making problem are analyzed and selected;the candidate targets of feature selection are determined according to the physical meaning of the shift decision-making problem;in order to avoid the problem that the performance of the learning machine is difficult to grasp due to the interdependence of feature selection and model parameter optimization,an improved genetic algorithm is proposed to link feature selection with model parameters.By improving the genetic operator,the optimization efficiency and global convergence ability of the algorithm are improved.The model parameters suitable for the calibration data of gear shifting decision are given.At the same time,the characteristic parameters of gear shifting decision problem are determined.3)Construction of multi-classification SVM decision tree structure.Since shift decision-making is a multi-classification problem,the multi-classification structure of SVM is a key problem affecting the performance of the model.Based on the analysis of the existing multi-classification methods,the binary tree structure is selected as the multi-classification structure of the shift decision-making SVM model on the premise of improving the performance of the model and the number of classifiers.The number of each group of SVM classifiers is reduced to n-1,which can be solved simultaneously.An improved particle swarm optimization(PSO)algorithm is proposed to optimize the structure of binary tree.The optimal classification order is selected from the root node,which improves the stability of the whole model for the accuracy of gear shifting decision.4)Recognition of general environmental features and working conditions.Shifting decision-making is faced with many environmental characteristics and working conditions in the process of vehicle driving.The general environmental characteristics are numerically computed by load calculation and added into the list of input characteristic parameters of SVM model as characteristic parameters to improve the adaptability of the model.The ability to map high-dimensional spatial features is an important factor for the success of machine learning algorithm.However,for practical problems such as shift decision-making,in different working conditions,when environmental features are added to the feature,the driver itself is more affected by the working conditions.Therefore,there are a large number of data cross-samples,and it is necessary to train their respective SVM models to adapt to this distribution change.ART2 learning neural network is used to identify four typical working conditions as the label of classifier,which makes the whole model more suitable for working conditions.5)Incremental learning of SVM.In order to improve the disadvantage of the machine learning machine,it needs to train the whole data set at one time,conform to the working characteristics of the step-by-step gear decision-making calibration,andmake the model have the online learning ability beyond the production stage.The incremental learning SVM algorithm is introduced to solve the problem.Firstly,based on the physical meaning of shift decision-making problem,a strategy to determine incremental candidate sets is proposed to improve the performance of the model.LSVM is used to optimize the solution of the problem,and the algorithm based on inherited iteration results can greatly improve the operation speed of the incremental SVM algorithm and keep its performance unchanged.At the same time,in order to avoid the explosion of incremental training set,this paper proposes a new strategy based on the concept of shell vector.Sample elimination algorithm improves the performance of reduced learning through shell vector calculation and counting elimination strategy.6)Verification of vehicle test.Firstly,the rationality of model gear decision-making is verified by vehicle model;a real vehicle data acquisition platform and a test platform are built,and the SVM gear decision-making method is tested and verified by the self-developed software and hardware automatic transmission system.