Acoustic Features and Autoencoders for Fault Detection in Rotating Machines: A Case Study

• Leonardo Afonso Ferreira Bortoni UFSC
• Pablo Andretta Jaskowiak UFSC

Resumo

Traditional Machine Fault Detection (MFD) techniques usually rely on multiple sensor data sources, such as vibration, temperature, force, and audio/acoustic signals. Acoustic signals, in particular, are quite appealing in the context of MFD, as they are often among the first manifestations of machine failure. Furthermore, they are associated with high sensitivity, environmental resilience, and do not require machine interference. Given these compelling characteristics, MFD based exclusively on acoustic signals can be highly beneficial. In this work, we evaluate an unsupervised MFD approach based on Autoencoders (AE) trained exclusively on features extracted from acoustic signals of a rotating machine. The data employed in this work comes from the Machine Fault Database (MaFaulDa), which includes information from vibration and velocity sensors, besides the acoustic measurements. This allows us to compare the performance of the AE models to that of supervised models (such as MLPs) trained on the same acoustic-based feature set, as well as feature sets that incorporate all sensors from MaFaulDa. Our results support that unsupervised MFD based on Autoencoders and acoustic signals is particularly appealing, as it requires only normal machine operation for training. Indeed, we obtained AUC values of 0.86 for the task.