Robust fault diagnosis of lifeboat davit hydraulic systems via cross-representation semantic alignment and sparse attention under low-SNR conditions

The hydraulic system of a lifeboat davit is a critical component for emergency lifesaving operations, where reliable condition monitoring is required under harsh marine environments. During long-term service, davits are exposed to fluctuating loads and complex operating conditions, causing fault signatures to become highly nonstationary and easily obscured by severe noise. This significantly limits the effectiveness of conventional feature extraction methods and unimodal learning strategies. To address this challenge, a cross-representation semantic alignment-driven multimodal fault diagnosis framework, termed the bidirectional gated recurrent unit (GRU) with attention encoder (Bi-GAE) and CLIP-Informer network (BCI-Net), is proposed for operation under low signal-to-noise ratio (SNR) conditions. Specifically, a Bi-GAE extracts fault evolution representations from raw time-series signals, while a CLIP image encoder captures weak-texture-aware spectral cues from continuous wavelet transform-based time–frequency maps, which remain informative when local fault textures are subtle. The temporal and time–frequency representations are first mapped into a shared representation space and fused via alignment-guided interactions to suppress representation-specific noise while retaining consistent fault-discriminative patterns. The fused representation is then modeled by an Informer classifier with sparse attention to capture long-range dependencies under nonstationary conditions. Extensive evaluations on a self-constructed lifeboat davit hydraulic system dataset and an auxiliary road machinery hydraulic system dataset demonstrate that BCI-Net effectively captures subtle fault characteristics under low-SNR conditions and achieves superior robustness compared with other models.