Graph-based convolutional neural networks for the classification of induction motor rotor bar faults using stator current & stray flux

Fault detection for induction motor diagnostics is crucial to ensure system reliability in electric drives. In recent years, research has shifted to combining signal processing and machine learning techniques to improve diagnostic accuracy under various machine states, notably the working conditions of induction machines with different speeds or load profiles. Modern approaches in diagnostic methods such as artificial neural networks rely on manual feature extraction. On the other hand, deep learning such as convolutional neural networks (CNNs) can automatically extract features from raw data in the learning process but are limited to grid-structured data. To this end, graph neural networks (GNNs), and particularly graph convolutional networks (GCNs), have become emerging solutions. This paper proposes a fault detection framework that integrates graph-theory-based learning and time-series analysis. The time segments of stator current and stray magnetic flux signals are constructed as graph nodes, and connections are established based on similarity metrics. Additionally, spatial features are extracted using GCNs, and time-series dynamics are modeled in combination with long short-term memory networks (LSTMs). The method is initially demonstrated using an extensive data sets of transient electromagnetic 2D finite element simulations of two induction motors of the same power rating and different rotor bar number. Then, it is verified experimentally via a range of laboratory measurements at several levels of loading, thus achieving non-invasive and highly accurate rotor fault detection across a range of rotor fault scenarios in various loading levels.