Deep learning approaches to problems in complex processes: early prediction of quadrotor failures and flight health monitoring models

Early detection of faults in advanced technological systems is crucial for ensuring reliability and safety. Despite the growing interest in artificial intelligence for fault detection, current methods often fall short in effectively utilizing extensive system information and sensor data. The presence of hidden faults in the collected data further directs the attentions to the need for advanced analytical techniques. This study presents two deep learning-based methods, fault prediction and flight health prediction, to early detect faults and monitor system health in real-time in complex systems. In the developed methods, feature dimensions were reduced using LSTM-based autoencoders, enabling the reduction of data into meaningful features. BiLSTM and LSTM models were used in the prediction and classification stages of the models. Both models were trained and validated with healthy and faulty data obtained from real quadrotor flights. Experimental results show that the fault prediction model detected fault indications with 0,9771 accuracy 30 seconds before a critical failure, while the flight health prediction model estimated the health status in real-time with an average MAE value of 0.1538 during flight. This demonstrates the significant potential of the proposed methods to enhance operational safety and reliability in complex systems, while also emphasizing the importance of comprehensive data and advanced analytical techniques in early fault detection and system health monitoring.