Anomaly Detection inCombustion Engines withSound Recognition

Abstract: As a global manufacturer of commercial vehicles, Scania is aiming for delivering high-quality products to its customers. Therefore, testing the produced components before assembling the final product and delivering it to the customer is key. As such, the engine testing process is the final process step within the production of internal combustion engines. One part of this testing process is analyzing the sound profile of the engine that is produced during the hot test of the engine. Embedded in this process, this thesis aims for developing and evaluating an algorithm that can detect a faulty engine sound profile with anomaly detection.Within anomaly detection, autoencoder Neural Networks is a type of Neural Network that is trained to detect deviations of the given input to this normal behavior.To evaluate the suitability of autoencoder Neural Networks for the given application, two autoencoder architectures were developed and several experiments with those architectures were conducted. The focus of those experiments is laid on determining the effects of learning rate and batch size on its performance. Generally, the thesis finds that autoencoders are a suitable solution foranomaly detection of sound profiles in internal combustion engines. Nevertheless, in order to meet the requirements for full implementation of the proposed models, further efforts need to be taken. Here, the focus should layon increasing the number of normal and especially abnormal files. Furthermore, the required recall of 100% by Scania could not be achieved consistentlyin any of the tested experiment set-ups. Therefore, further tests for alternative architectures as well as a more extensive hyper-parameter tuning, for both preprocessing and Neural Network training, are to be conducted