Recognizing Similarities in Automatic Transmissions of Vehicles by Using Time Series Data and Autoencorders

In recent years, the development time of vehicles has further accelerated, and automation of the development is an urgent task. One example of time wasting tasks is gear-shift calibration. For this purpose, Kawakami et al. have studied OK/NG classification of shift quality by using neural networks. However, their classifiers have a problem in versatility over different AT hardwares. In this paper, we develop autoencoders to realize similar/not-similar classification on three AT hardwares of vehicles. These hardwares have different lock-up multi/single-plate clutch structures. Experimental results show that the performance of similar/not-similar classification is high in terms of AUC.