Reasearach On Automatic Transmission Shift Strategy Based On Deep Recurrent Neural Network

With the continuous increase of the market share of AT vehicles,people have put forward new requirements for the intelligence and automation of automatic transmission,and more and more scholars were attracted to carry out in-depth and comprehensive research on the intelligent shift strategy of automatic transmission.At present,in the research of automatic transmission intelligent shift,the selected neural network has no feedback connection and lacks associative memory function.Hence,it is hard to capture the internal relation between continuous data,which makes the shift model less intelligent in data processing.The above problems can be solved well when deep recurrent neural network(DRNN)was applied to automatic transmission shift strategy.Therefore,there are notable engineering significance to study shift strategy based on DRNN.Firstly,reviewing the history of automatic transmission at home and abroad,and analyzing the advantages and disadvantages of domestic and foreign research strategies,the appropriate solution was gotten.The DRNN is introduced into the shift strategy to make full use of the advantages of DRNN in processing time-series data.Before the neural network(NN)was trained,the vehicle transmission model based on Simulink was initially constructed,from which the shift strategy logic was designed.Then,the original data used for training the NN was generated,and the data would preprocessed so that it could be applied to the subsequent NN directly.Secondly,before the DRNN was built,the possible influence of its framework and various hyper-parameters on the training of DRNN is described respectively.Then,the theoretical principles of formulas in the process of DRNN forward calculation and back propagation were given.What is more,the corresponding NN framework would be designed and appropriate optimizer would be selected based on traditional model.Moreover,in order to verify that the DRNN has a unique adaptability,compared with the traditional NN in gear prediction,a gear prediction model based on back-propagation NN is added,and the hyper-parameters of the gear prediction model based on the different NN were same.Then,the preprocessed training samples were used to modify the weights of the two models respectively.After the training was completed,the simulation were carried out with the two gear prediction models respectively.Finally,the original hyper-parameters would be replaced by new hyper-parameters,so as to obtain a prediction model with higher prediction accuracy.Then,the simulation results were analyzed and the conclusions were drawn,After that,the innovation points were summarized and the deficiencies of this study were supplemented in the future study.Simulation and comparison results showed that the unique feedback framework of DRNN would associate the output of the current moment with the input of previous moments,which made it obtain sufficient prediction accuracy when the number of iterations increased significantly in processing time-series data(sensor data)and the defect of insufficient prediction accuracy of back-propagation NN was effectively improved.Therefore,the prediction accuracy of the gear prediction method based on DRNN can meet the requirements of engineering application,which lays the theoretical and technical foundation for the intelligent control of automatic transmission.