Autoencoders for Anomaly Detection in Commercial Truck Gearshifts

Abstract

Abstract The following study revolves around implementing a machine learning model that supports engineers in analyzing vehicle sensor data at Volvo Trucks. The aim is to classify different gear shift types and to automate the detection of abnormal gear shifts making decisions more efficient and data-driven. By applying clustering- and classification algorithms, the model demonstrates how unlabeled field test data can be classified in diverse gear shift types. Furthermore, through evaluation and comparision of dense, convolutional and long short-term memory (LSTM) autoencoder (AE) neural networks, this study exhibits how abnormal sensor data within specific gear shift types can be detected. Obtained results indicate adequate gear shift classification with an overall accuracy of 98%. Furthermore, a comparison of the autoencoders in regard to the performance metrics accuracy, precision and recall concluded that the convolutional autoencoder outperformed the other architectures with scores above 95%. Although further fine-tuning of the implemented model is possible, findings indicate that it is feasible to develop and use machine learning models for classification and anomaly detection of gear shift data within heavy-duty trucks.