Unsupervised and Supervised Learning Techniques for the Detection of Anomalous Sounds for Machine Condition Monitoring

Summary

The increasing demand for automated machine condition monitoring in industrial environments has led to extensive research into developing efficient fault detection techniques. In response to this need, this thesis explores both supervised and unsupervised learning methods for identifying anomalous sounds in machinery, which are critical indicators of potential mechanical failures. Focusing on the automated detection of machine faults—a crucial aspect of the Fourth Industrial Revolution, characterized by factory automation and the implementation of artificial intelligence—this study aims to enhance real-time anomaly detection capabilities. Analyzing the acoustic characteristics of machines to identify irregular operations offers a faster and more efficient approach to industrial automation, enabling the rapid detection of malfunctions. Given the challenges in acquiring and labeling abnormal machinery sound data, which are often rare, this study places a significant emphasis on exploring unsupervised approaches in addition to supervised ones, evaluating multiple one-class classification algorithms.

To comprehensively assess the performance of the classifiers, two distinct training strategies were employed: ID-dependent and ID-independent. The ID-dependent strategy used combined normal and anomalous sound segments from all machine IDs per machine type for training, allowing for detailed learning about specific machine types but limiting generalizability. The ID-independent approach involved training the classifiers by leaving out one machine ID at a time, using it as the test set in each iteration, which provided a more robust evaluation of the model’s ability to generalize to new, unseen machine IDs. The research utilizes the MIMII dataset, which includes audio recordings of various industrial machines—Fans, Pumps, Sliders, and Valves—under both normal and faulty operating conditions. Several machine learning models were explored, including supervised models like Support Vector Machines (SVM) and Logistic Regression (LR), as well as unsupervised approaches employing one-class classification algorithms such as One-Class SVM (OCSVM), Isolation Forest (IF), Local Outlier Factor (LOF), Elliptic Envelope (EE), and Autoencoders (AE). In the unsupervised context, models were trained exclusively on normal sounds to learn their distribution characteristics, with unseen sounds flagged as anomalous if they deviated from these patterns.

Results indicated that supervised models like SVM achieved high accuracy, with Area Under the Curve (AUC) values exceeding 97% and F1-scores above 0.88 in ID-dependent scenarios. However, their performance significantly declined in ID-independent scenarios, particularly for the Valve machine type. Focusing on unsupervised learning methods, Autoencoders (AE) generally outperformed others, particularly in ID-independent scenarios, with notable results for Fan (AUC 89%, F1 0.628) and Slider (AUC 89%, F1 0.572) machine types. While AE models showed strong overall performance, they occasionally struggled with precision-recall balance, especially in the Pump category. Other methods, such as One-Class SVM and Isolation Forest, also yielded promising results, with the former excelling in ID-independent scenarios and the latter in ID-dependent contexts. These findings highlight the importance of choosing the appropriate unsupervised method based on the specific application and data characteristics. The contributions of this thesis involve the application of efficient preprocessing methods and the extraction of features utilizing the pyAudioAnalysis library, which improved the model’s performance. The results indicate that unsupervised models are capable of effectively generalizing across various machine IDs for a given machine type, presenting a scalable and efficient solution for real-world predictive maintenance.