Gear Fault Diagnosis Method Based on Transfer Function and Vision Transformer

As a core transmission component of rotating machinery, gearboxes require effective health monitoring and fault diagnosis. To obtain a feature that directly characterizes the dynamic relationship between the two shafts while modeling correlations over the entire frequency band, this paper proposes a gear fault diagnosis method based on the shaft-to-shaft transmissibility and Vision Transformer (TF-ViT). An on-shaft vibration test rig was constructed, where sensors were mounted directly on the input and output shafts, with conductive slip rings ensuring signal transmission under rotating conditions. Modeling the gearbox as a dynamic system whose two measured responses are the input- and output-shaft vibrations, the shaft-to-shaft transmissibility T(f) is estimated from the responses via Welch’s method (the H1 cross-spectral ratio) to characterize the inter-shaft dynamic relationship. The transmissibility magnitude sequence was partitioned into equal-length patches and fed into a Transformer encoder after linear embedding and positional encoding; multi-head self-attention models dependencies between arbitrary positions across the entire frequency band, extracting global frequency-domain fault features. Experimental results over ten paired random-seed evaluations show that TF-ViT achieves the highest mean test accuracy among the compared dual-channel CNN, TCN, and ViT models.