Unsupervised Anomalous Sound Detection Using Hybrid Machine Learning Techniques

The early detection of defects in mechanical equipment is of paramount importance in the industrial field. Research into data analysis methodologies for extracting valuable data from mechanical equipment has highlighted the significant value of these technologies. One such study is the early and accurate detection of anomalies in the acoustic data of machinery. In this paper, we propose an effective technique for the detection and classification of rare event anomalies in the data derived from the rotational noise of automotive motors. MFCC extraction and smoothing techniques were used to select minimal features for optimal performance, and Principal Component Analysis (PCA) was applied to extract salient features. These features are capable of distinguishing between normal and anomalous data. Additionally, an unsupervised learning algorithm was applied to the dataset to differentiate between normal and anomaly data. Experimental results showed that the proposed method can effectively detect sound anomalies with a high accuracy of 99.4% and is also capable of detailed classification of anomalous data.