Data-Driven Techniques for Mechanical Fault Classification

Data-driven techniques for machine condition monitoring are typically first evaluated on relatively simple experimental datasets, where they often achieve near-perfect classification accuracies. Such results do not establish real-world applicability or whether the associated computational demands are justified. This paper introduces an experimental dataset designed to better approximate selected real-world operating conditions. The dataset is used to evaluate both traditional and novel techniques for machinery fault classification and to assess their suitability for resource-constrained computing environments, such as edge-computing applications on mobile or remote equipment. Four distinct data-driven methods were evaluated using several metrics in addition to classification accuracy. The traditional machine learning techniques struggled to identify fault conditions under unseen speed profiles, whereas the deep learning models performed substantially better. Computational metrics showed that some methods required considerably more memory and processing resources than others. One deep learning model had memory and computational requirements that appear theoretically compatible with real-time implementation on platforms such as the Jetson Nano or Raspberry Pi 5. These results highlight both the potential of continued deep learning development for machine condition monitoring and the limitations of traditional approaches under more variable operating conditions.