Dead band detection for fiber optic gyroscopes based on Patch-Decomposed Transformer

Dead band detection is a critical step in the performance evaluation and quality assurance of high precision Fiber Optic Gyroscopes (FOGs). However, accurate detection remains challenging because dead bands usually appear as contextual anomalies with trend variations and temporal dependencies rather than isolated point anomalies. Therefore, this paper proposes a novel deep learning framework termed the Patch-Decomposed Transformer for dead band detection in FOG test data. The proposed framework employs a patching mechanism to preserve local temporal structures while reducing computational complexity. In addition, a deep series decomposition module is introduced to explicitly separate complex dynamic components and enhance temporal characteristics, thereby improving the detection performance. Experimental results on the FOG dead band dataset demonstrate that the proposed method significantly outperforms existing time frequency analysis methods and data-driven approaches. It achieves a F1-score of 97.15% with a boundary localization error of 18.28 samples. In addition, the impact of key hyperparameters on model performance is comprehensively investigated to achieve optimal results.