Real World Comparative Analysis of Unsupervised Machine Learning Techniques for Anomaly Detection in Washing Machine Production

In modern industrial production lines, ensuring product quality, customer satisfaction and minimizing production costs is of primary importance. Many learning techniques to address the issue of fault detection require training on labelled databases with a large number of anomalous audio samples that, however, are difficult or impossible to obtain. Furthermore, understanding which audio features are really crucial for anomaly detection is non-trivial. The article presents a comparative analysis of three unsupervised machine learning techniques based on the analysis of audio files/features, suitable to the case where a significant number of anomalies is not available; and a strategy for isolating audio features that are really important for anomaly detection. Experimental results show that the technique based on the isolation of the correct audio features is better than brute-force techniques.