HYBRID METAHEURISTIC ALGORITHM BASED OPTIMIZATION OF ECHO STATE NETWORK FOR FAULT PREDICTION IN AIRCRAFT ENGINES

The pervasive availability of data is the bedrock of the fourth Industrial Revolution (IR 4.0). Engineering Prognostics and Health Management (PHM) is one of the big beneficiaries of this abundant data. A shift had been made from the obsolete preventive maintenance to Predictive Maintenance (PdM). In PdM, sensor data is used to forecast the future state of equipment. Maintenance decisions are then based on these predictions. This has enhanced the maintenance of complex machines such as the aircraft. Hence, making them safer and cheaper to maintain. The tasks are mostly time-series predictions. Therefore, Recurrent Neural Networks (RNNs) are
excellent tools for such predictions. However, most traditional RNNs su↵er from the unstable gradient problem and computationally expensive training schemes. Echo State Networks (ESNs) and its families were developed to solve these issues. They do so by using a fixed large reservoir layer and a simplified linear regression training. However, the selection of ESN’s parameter and topology is difficult. Researchers formulate the issue as an optimization problem. Therefore, metaheuristics algorithms can be useful tools for solving these problems. In this research, a hybrid algorithm is developed to optimize the ESN.