Long short-term memory neural network based fault detection and isolation for electro-mechanical actuators

Abstract

In the new generation of aircraft, electro-mechanical actuators (EMA) have been replacing the conventional hydraulic versions. Despite the fact that a failure of this system can seriously affect the safety of vehicles and their operators, there are few studies that focus on fault diagnosis for the units. In this paper, we present an innovative fault detection and isolation method for the EMA. Our method is tested and verified against three types of failures. This novel fault diagnosis method works by creating a model for sensor data by using the known time series. It utilizes an advanced Long Short-term Memory (LSTM) neural network, which can effectively handle time series data in this domain. A modification to the LSTM network is applied in order to take advantage of the correlation between sensors. In addition, the algorithm uses a sliding window to improve performance of LSTM applied to fault isolation. Our research has revealed that the proposed algorithm is better able to detect faults when compared to traditional neural networks. We also compare our performance with the support vector machine algorithm and the typical LSTM algorithm. Ultimately, the proposed method performs superiorly for the task of fault isolation.