Classification of Gear-shift data using machine learning

Abstract: Today, automatic transmissions are the industrial standard in heavy-duty vehicles. However, tolerances and component wear can cause factory calibrated gearshifts to have deviations that have a negative impact on clutch durability and driver comfort. An adaptive shift process could solve this problem by recognizing when pre-calibrated values are out-dated. The purpose of this thesis is to examine the classification of shift types using machine learning for the future goal of an adaptive gearshift process. Recent papers concerning machine learning on time-series are reviewed. Adata set is collected and validated using hand-engineered features and unsupervised learning. Four deep neural networks (DNN) models are trained on raw and normalized shift data. Three of the models show good generalization and perform with accuracies above 90%. An adaption of the fully convolutional network (FCN) used in [1] shows promise due to relative size and ability to learn the raw data sets. An adaptation of the multi-variate long short time memory fully convolutional network (MLSTMFCN) used in [2] is superior on normalized data sets. This thesis shows that DNN structures can be used to distinguish between time-series of shift data. However, much effort remains since a database for shift types is necessary for this work to continue.