Data-based methods for fault diagnosis and prognosis – State of the art

For operators of industrial processes, monitoring the health of their equipment is one of the major preoccupations in preventing breakdowns. Over the last few decades, numerous methods and tools have been developed to detect the onset of degradation, perform diagnostics and estimate the remaining service life before failure (DEFAD, RUL). They are based on data, physical models or a combination of both. This article presents the state of the art of methods and tools that exploit only data collected on equipment or contained in databases, to carry out failure diagnosis and prognosis. To avoid any ambiguity in the terms used in this article, the first section presents the main definitions and terminologies proposed by ISO and NF-EN international standards (ISO 13372:2012, ISO 13379-1 ISO 13381-1, ISO 16079-1 and NF EN 13306). It proposes a classification of diagnosis and diagnostic methods into three families, data-driven, model-driven and hydrid. Following a description of the challenges and specific features of data-driven methods, the main steps required to diagnose and prognose the state of health with a view to making decisions on equipment maintenance strategies are described. These stages are divided into two phases: CBM (condition monitoring) and PHM (prognostics and health management) – RUL (remaining useful life). The CBM is dedicated to monitoring, detecting and diagnosing the state of health of the equipment, while the PHM – RUL provides a prognosis of the remaining useful life. The principles of data collection and storage, data processing, anomaly detection, residual life diagnosis and prognosis, and maintenance decision-making are all covered. Given the risks associated with errors in detection, diagnosis and prognosis, this section briefly reviews decision theory and proposes metrics for the confidence to be placed in diagnosis and prognosis. As many tools and methods are common to the resolution of diagnostic and prognostic problems, the second section is devoted to them. After a description of the principles of supervised and unsupervised learning techniques, it presents the two main categories of methods: statistical and artificial intelligence-based. Statistical methods include least-squares regression, Gaussian process regression and factorial methods (PCA). Following a description of machine learning concepts linked to the development of Big Data, cloud computing and data mining, an inventory of intelligence-based methods is presented. To illustrate the implementation of these methods and highlight their limitations, the third section presents the performance of their experimentation on three pieces of equipment: an aircraft engine, a lithium-ion battery and a wind turbine speed multiplier. For the aircraft engine, six methods for DEFAD prediction are compared: multiple linear regression, Ridge...