A High-Accuracy Deep Learning Framework for Fault Detection in Wound-Rotor Rotational Resolver

Resolvers are widely used in automotive, aerospace, and industrial systems where reliable position sensing is essential. Their durability ensures operation in environments where optical encoders and other sensors are prone to failure, but their susceptibility to electrical and mechanical faults can compromise safety, reliability, and performance. This study proposes an automated deep learning-based fault detection approach to identify and classify common faults such as eccentricity and short circuits in excitation or signal windings, which distort signals and impair position estimation. This is particularly important given the scarcity of methods that avoid manual feature engineering and perform fault classification automatically. Various fault scenarios were simulated to train a generalizable network, and several state-of-the-art architectures for temporal data classification were evaluated. While some models achieved acceptable performance, their micro F1 scores did not exceed 87%, motivating a hybrid architecture combining residual convolutional blocks, an LSTM layer, and an attention mechanism. The proposed model achieved a micro F1 score of 97% on unseen simulated scenarios and showed strong generalization in a laboratory experiment by correctly identifying the fault. These results confirm the effectiveness of the proposed hybrid architecture for accurate resolver fault detection and classification.