Acoustic Emission-based Fault Diagnosis for Drones with Heterogeneous Multiverse Recurrent Expansion: Avoiding Representation Glitch

In drone technology, ensuring dependable fault diagnosis is of paramount significance. Drones encounter an array of operational challenges, encompassing motor failures and propeller malfunctions, necessitating the implementation of resilient and efficient diagnostic techniques. This paper therefore addresses this problem by exploiting the potential of representation learning, an effective strategy for managing complexities and concept drift within data. This methodology is built on a three-fold approach. First, it uses a dedicated data processing framework tailored to the unique attributes of acoustic emission-based datasets associated with drones. This framework encompasses tasks such as feature extraction, denoising, filtration, scaling, and balancing, thereby optimizing data for subsequent analysis. Second, this study introduces an innovative representation learning algorithm, denoted as Heterogeneous Multiverse Recurrent Expansion with Multiple Repeats (HMV-REMR). Within HMV-REMR, a spectrum of Recurrent Neural Network (RNN) variants is incorporated, including Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU). This selection ensures adaptive learning, delves deeper into the feature space, and alleviates the occurrence of so-called Representation Glitches frequently observed in basic REMR variants. In contrast to conventional RNN variants, basic MV-REMR, and existing state-of-the-art methodologies, the findings of this study demonstrate superior diagnosis capabilities for detecting faults in drones.