Trend-Aware Patch Feature Fusion Transformer for Fault Diagnosis of Vibration Signals

The vision transformer (ViT) has demonstrated reliable performance in industrial equipment fault diagnosis by capturing global features and long-range dependencies in vibration signals. Nonetheless, when processing vibration signals, ViT primarily focuses on the correlations between patches, while neglecting the trend variation information hidden between signal units. This limitation hinders the model’s ability to extract latent fault information. To address this issue, we propose a trend-aware patch feature fusion Transformer (TAPformer) for fault diagnosis of vibration signals. Specifically, we design both patch-level and signal-level trend-aware self-attention blocks and adopt a weighted approach to integrate coarse-grained patch-level trend features with fine-grained signal-level trend features at the output of each self-attention head. This approach enables the model to capture fault information hidden within trend variations, effectively enhancing its fault detection capabilities. TAPformer establishes optimal weighted fusion hyperparameters through ablation experiments and is tested on four public datasets. Compared to eight advanced models, TAPformer achieves the highest accuracy across all four datasets, demonstrating the effectiveness and generalization capability of the proposed method.