Diagnosing Rolling-Element Bearing Faults in the Real World: Problem Solved?

An issue of great concern in the maintenance of critical infrastructures is the timely and accurate detection of mechanical faults, before these faults deteriorate into failures, causing major service disruptions. The ubiquity of rolling-element bearings (REBs) provides much incentives for online diagnosis of REB faults, spurring a crescendo of research activities that started decades ago. Near-perfect accuracies of state-of-the-art (SOTA) classifiers leveraging deep learning (DL) on the public-domain Case Western Reserve University (CWRU) dataset have been reported. However, the faults covered by this dataset were artificially induced in a controlled environment to bypass the class imbalance problem, and do not include cage faults. Among the open questions to be addressed are: How reproducible are SOTA performances? What are the associated resource requirements? How transferable is SOTA classification performance from CWRU data to actual field data covering naturally occurring faults? How well do SOTA classifiers perform when trained on this real-world dataset, without and with mitigation of the class imbalance problem? How generalisable is classifier performance in the real-world context? Set out to answer the preceding questions, the experimental study presented here provides (i) benchmark results on the CWRU dataset and a real-world dataset from South Australian Water Corporation (SA Water), and (ii) derived from the benchmark results, new insights regarding the state of the research field as well as an answer to the question in the title of this paper: is the problem of REB fault diagnosis a solved problem?