Fault detection and diagnosis for industrial processes based on clustering and autoencoders: a case of gas turbines

Abstract

Industrial machinery maintenance constitutes an important part of the manufacturing company’s budget. Fault Detection and Diagnosis (henceforth referenced as FDD) plays a key role on maintenance, since it allows for shorter maintenance times and, in the long run, to train predictive maintenance algorithms. The impact of proper maintenance is reflected on an especially costly type of industrial machine: gas turbines. These devices are complex, large pieces of machinery that cause considerable service disruption when downtime occurs. In an effort to shorten these service disruptions and establish the basis for the development of predictive maintenance, we present in this paper an approach to FDD of industrial machinery, such as gas turbines. Our approach exploits the data generated by industrial machinery to train a machine-learning based architecture, combining several algorithms with autoencoders and sliding windows. Our proposed solution helps to achieve early malfunctioning detection and has been tested using real data from real working environments. In order to build our solution, first, we analyze the behavior of the gas turbine from a mathematical point of view. Then, we develop an architecture that is capable of detecting when the gas turbine presents an abnormal behavior. The great advantage of our proposal is that (i) does not require existing disruption data, which can be difficult to obtain, (ii) is not limited to processes with specific time windows, and (iii) provides crucial information in real time to the monitoring staff, generating valuable data for further predictive maintenance. It is worth highlighting that although we exemplify our approach using gas turbines, our approach can be tailored to other FDD problems in complex industrial processes with variable duration that could benefit from the aforementioned advantages.