Classification of Time Series Measurement Data for Shift Control of Automatic Transmission of Vehicles Using Machine Learning Techniques 2020-01-0260
An efficient approach to classify time series physical measurement data of shift control for automatic transmission of vehicles is presented. Comfortable acceleration is the essential factor of today’s vehicle. Shift control of automatic transmission of vehicles directly contributes to the comfortable acceleration. Since calibration of automatic transmission of vehicles is time consuming task for expert engineers, the development of autonomous calibration is desired to reduce product development time in today’s competitive automobile market. In the stage of product development, it is difficult to obtain a large amount of physical measurement data. Therefore, we need to develop machine learning method for limited amount of data. For this purpose, we develop the method to classify time series measurement data of shift control for automatic transmission of vehicles. We use support vector machine (SVM) as a machine learning technique. Features, used by SVM, of time series measurement data of shift control of automatic transmission of vehicles is selected by expert engineers. In addition, the computation is too heavy to explore the optimal value of the high dimensional parameter space for our classification problem with grid search. To remedy this problem, we employ Bayesian optimization. Bayesian optimization is known to be much more efficient than brute force grid search, because it is a sequential parameter search strategy for global optimization. Combining SVM and Bayesian optimization, we successfully built a high accuracy classification scheme. As the consequence, our proposal method enables highly efficient calibration of automatic transmission of vehicles in the stage of the product development. We demonstrate the performance of proposed method for classification problem of shift control of AISIN AW’s automatic transmission of vehicles. The results of our experimentation show the expected average accuracy of 0.944 for up shift and 0.941 for down shift, that are promising enough in an actual use.