One-shot learning for acoustic diagnosis of industrial machines

• Published in: Expert systems with applications 2021-09, Vol.178, p.114984, Article 114984
• Main Author: Ntalampiras, Stavros
• Format: Article
• Language: English
• Subjects: Deep learning; Domains; Drill bits; Fault diagnosis; Hand tools; Machine acoustics; Machine health condition monitoring; Machine learning; Malfunctions; Neural networks; One-shot learning; Online learning; Spectrograms; Tool wear
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.114984

•A system for automatic acoustic monitoring of machine health is presented.•We introduce the one-shot learning paradigm into the specific domain.•The system classifies machine states, detects novel ones, and incorporates online.•We achieve state of the art results under poor data and evolving environments.•Obtained predictions are interpretable by examining layer-wise activation maps. Automatic acoustic monitoring of machine health comprises a relevant field as, unfortunately, such equipment often suffers from faults, malfunctions, aging effects, etc. However, it is still an unexplored domain of research where the majority of existing works relies on traditional machine learning based approaches. After providing a critical survey of the available methods, this work highlights the most relevant limitations and designs a solution specifically addressing them. We introduce the one-shot learning paradigm into the specific domain and suitably extent it to (a) classify machine states, (b) detect novel ones, and (c) incorporate them in the class dictionary online. The backbone of the present system is a Siamese Neural Network (SNN) composed of convolutional layers. Conveniently, every processing stage depends on a standardized feature set free of domain knowledge, i.e. spectrograms. Interestingly, we enhance SNN’s classification ability by an appropriately designed data selection scheme. The proposed solution is applied on a publicly available dataset of vibration signals representing four states of a drill bit, i.e. healthy state, chisel wear, flank wear, and outer corner wear. After extensive experiments thoroughly examining every aspect of the proposed solution, it is shown to achieve state of the art results while using limited amount of training data. Importantly, at the same time it is able to operate under evolving environments. Last but not least, we show that the obtained predictions are interpretable, a property which is rapidly becoming a requirement in modern machine learning based technologies.