Classification of time series using information granules for efficient detection of unmanned aerial vehicles faults

Classification of time series data is of significant importance across various domains. In particular, an important issue is to ensure the proper operation of vehicles such as drones. In this study, we propose a novel approach that utilizes the paradigm of information granules for time series classification. The proposed methodology transforms consecutive time windows of the original vibration signal into information granules, which are compact statistical summaries describing the signal dynamics within each segment. These granules serve as the input for machine learning classifiers and provide a more interpretable and noise-robust representation of the data. The proposed method is tested on data collected during the operation of unmanned aerial vehicles (UAVs), demonstrating its effectiveness. Experimental results indicate that the classification accuracy significantly improves when using information granules compared to raw data. In this study, the classification task involves distinguishing five classes corresponding to different rotor fault levels (0%, 25%, 50%, 75%, and 100% faulty rotors). This allows for precise identification of UAV operational health. In the case of the Tree Ensemble method, classification accuracy increased from 57.5% (raw data) to 100% (using information granules from four combined time windows), highlighting the substantial enhancement achieved through the proposed approach. Additionally, the study explores the impact of different time window lengths and training set sizes on classification performance, providing insights into optimizing the proposed method for practical applications.