Audio Anomaly Detection in Cars

Sammanfattning: Audio anomaly detection in the context of car driving is a crucial task for ensuring vehicle safety and identifying potential faults. This paper aims to investigate and compare different methods for unsupervised audio anomaly detection using a data set consisting of recorded audio data from fault injections and normal "no fault" driving. The feature space used in the final modelling consisted of: CENS (Chroma energy normalized Statistic), LMFE (Log Mel Frequency Energy), and MFCC (Mel-frequency cepstral coefficients) features. These features exhibit promising capabilities in distinguishing between normal and abnormal classes. Notably, the CENS features which revealed specific pitch classes contribute to the distinguishing characteristics of abnormal sounds. Four Machine learning methods were tested to evaluate the performance of different models for audio anomaly detection: Isolation Forest , One-Class Support Vector Machines, Local Outlier Factor, and Long Short-Term Memory Autoencoder. These models are applied to the extracted feature space, and their respective performance was assessed using metrics such as ROC curves, AUC scores, PR curves, and AP scores. The final results demonstrate that all four models perform well in detecting audio anomalies in cars, where LOF and LSTM-AE achieve the highest AUC scores of 0.98, while OCSVM and IF exhibit AUC scores of 0.97. However, LSTM-AE displays a lower average precision score due to a significant drop in precision beyond a certain reconstruction error threshold, particularly for the normal class. This study demonstrates the effectiveness of Mel frequency and chroma features in modelling for audio anomaly detection in car and shows great potential for further research and development of effective anomaly detection systems in automotive applications.