Open AccessReinforcement Learning for Predictive Maintenance of Industrial Plants
Author(s) -
Petia Koprinkova–Hristova
Publication year - 2013
Publication title -
information technologies and control
Language(s) - English
Resource type - Journals
eISSN - 2367-5357
pISSN - 1312-2622
DOI - 10.2478/itc-2013-0004
Subject(s) - reinforcement learning , predictive maintenance , computer science , predictive power , artificial intelligence , model predictive control , artificial neural network , element (criminal law) , key (lock) , machine learning , reinforcement , mill , control (management) , engineering , reliability engineering , mechanical engineering , philosophy , computer security , structural engineering , epistemology , law , political science
The reinforcement learning is a well-known approach for solving optimization problems having limited information about plant dynamics. Its key element, named “critic” is aimed at prediction of future “punish/reward” signals received as a result of undertaken control actions. The main idea in the present work is to use such a “critic” element for prediction of approaching alarm situations based on limited measurement information from the industrial plant. In order to train the critic network in real time it is proposed to use a special kind of a fast trainable recurrent neural network, called Echo State Network (ESN). The approach proposed is demonstrated on an example for predictive maintenance of a mill fan in Maritsa East 2 Thermal Power Plant