HTPNet: A Hybrid Transformer-Based Dynamic Projection Network with Adversarial for Cross-Condition Bearing Fault Diagnosis

Intelligent fault diagnosis of rolling bearings under variable-speed operation is still difficult to achieve, since it is hard to learn speed-invariant fault representation and maintain the geometric structure of the fault discriminator under different operating conditions. To address this issue, this paper proposes a Hybrid Transformer-based Dynamic Projection Network (HTPNet) method for cross-condition bearing fault diagnosis. A Transformer-based feature extractor with learnable patch embedding captures long-range temporal dependencies from raw vibration signals to generate informative features. A class-specific hypernetwork learns dynamically the projection matrices, which project features into adaptive low-dimensional subspaces, and orthogonality regularization stabilizes the projections and decorrelates features. We propose the Gradient Reversal Layer for adversarial domain adaptation to learn condition-invariant representations and all-class projection scoring for making reliable predictions on unlabeled target domains. The purpose of this optimization is a unified objective that improves the feature discrimination, the consistency of the projections and the domain alignment. Experimental results on both HUST and CWRU datasets reveal that HTPNet consistently outperforms other state-of-the-art methods in terms of diagnostic accuracy, robustness to operating condition variations, additive noise, and feature interpretability.