Electric Motor Drive Anomaly Detection Using AutoGluon

This paper proposes an anomaly detection method for electric motor drive using state-of-the-art convolutional neural network (CNN) based deep learning methods. As a first step, the motor controller time-series data are preprocessed and transformed into images, and thus the anomaly detection problem is formulated into an image classification problem. AutoGluon is specialized in automating the optimization process of model selection, network topology, and hyperparameter tunning. Under-the-hood, AutoGluon Multimodal (AutoMM) uses PyTorch Image Models (TIMM), which provides a wide variety of pre-trained image models. Different anomalies based on real-life data are labeled as multiple classes. A technique to automate the process of preparing and labeling large-scale train/validate/test data is introduced. AutoMM demonstrates strong capability in handling unbalanced dataset. In the end, the prediction results of different models are compared and analyzed.