| Summary: | Modern physical systems tend to have a high level of complexity that hinders the efficiency of human-based condition monitoring. Automatic and intelligent strategies, on the contrary, easily outperform the human expertise in terms of speed, accuracy and scalability. Focusing on faults, probably the most critical issue in condition monitoring, this paper presents a selected survey on the data-driven Fault Detection and Diagnosis (FDD) field analysed from a data mining perspective. Data pre-processing is identified as a fundamental step to reach satisfactory results in the FDD process. In this respect, Empirical Mode Decomposition, Wavelet and Walsh transforms are effective signal transformation tools. Principal Component Analysis and Fisher Discriminant Analysis are often used for feature reduction. Machine Learning techniques, such as Support Vector Machines and Neuro-Fuzzy are used to solve the core tasks of FDD, namely classification and novelty detection. Genetic Algorithms and Swarm Intelligence methods are usually applied for parameter optimization for the above mentioned techniques. It has also been observed that a particular approach, namely fault classification, is the most common FDD strategy. However, since it requires supervised learning, it is limited to applications where supervised data is available.
|
|---|
