Training an electric signal-based generative model for anomaly detection in propeller condition diagnosis

In modern society, the necessity of Unmanned aerial vehicles (UAVs) is continually increasing. However, ensuring safety remains a critical challenge, particularly in detecting fault during flight. To address this issue, attempts have been made to predict fault during flight using data-based fault detection methods. However, due to the nature of aviation data, where there is an imbalance with significantly fewer fault data compared to normal data, it is difficult to apply general data-based fault detection methods in the field of UAV anomaly detection. Additionally, while using additional signals such as vibration or sound signals in data collection can improve the accuracy of predictions, it also increases system complexity, leading to challenges in collecting these additional signals. Therefore, this study utilized only normal current and voltage data, which can be collected from UAV motors without additional sensor attachments, as training data to address the aviation data imbalance problem and data collection challenges. The anomaly detection model was trained using a VAE-GAN (Variational Auto Encoder-Generative Adversarial Network)-based approach to perform binary classification of normal and abnormal conditions. The collected current and voltage data were preprocessed into image data through Markov transition matrices to extract their features. The generator of the GAN, trained with the preprocessed normal data, imitation the characteristics of the normal data during the training process, while the discriminator of the GAN learns to distinguish between the generated data and normal data, thereby learning the characteristics of the normal data. As a result of anomaly detection using the evaluation data model, the average accuracy of normal data was 74.05%; for Asymmetric fault, it was 75.10%; and for Symmetric fault, it was 94.05%. This demonstrates that anomaly detection for UAV motors, which are part of the propulsion system, can be effectively achieved using only normal electrical signals. It also shows that the VAE-GAN model is the most suitable approach for detecting anomalies in datasets with small inter-class variance, such as electrical signals.