In recent years the growing amount of data generated in industrial processes has enabled the development of data-driven decision-making systems, placing strong interest on artificial intelligence, particularly when the world of Internet-of-Things is considered. Managing and monitoring this data flow is crucial in machining processes, where the health of the system is assessed through the analysis of different sources, such as vibration, temperature, electric and acoustic signals. Currently, the main theme in tackling this task is anomaly detection: identifying an anomalous state is equivalent to having adequate control of the process and being able to make decisions accordingly, which, if achieved in the early stages, gives an incredible return, avoiding failures and production interruptions. In this paper an approach combining Wavelet Packet Decomposition and Autoencoders for anomaly detection of CNC machining processes is presented. To this end, acoustic emission signals of a real-world use case are considered. To prove the effectiveness of the proposed system, a comparison with an Isolation Forest algorithm, a well-known benchmark in this field, is made. The results show an improvement of nearly 10% in terms of F1-score and accuracy, as well as the advantages of the encoding procedure. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.