Abstract: Flash butt welding is the most critical process in mooring chain manufacturing. Its process simultaneously involves coupled interactions of electricity, heat, mechanics, and metallurgy, exhibiting strong non-stationarity, multi-stage process evolution, and strong multivariate coupling. Existing methods struggle to effectively identify anomalies in the flash butt welding process online. To address this issue, a multi-scale Transformer anomaly detection method based on HP filter decomposition and trend-branch Mamba enhancement (namely, TD-Mamba) was proposed. First, HP decomposition was used to decouple the raw multivariate time-series signals collected during the mooring chain flash butt welding process into a trend component that reflects the evolution laws of stages such as preheating cycles, continuous flashing, and upsetting response, and a residual component that characterizes local disturbances such as current spikes, position jitter, and sudden pressure changes. Second, a Mamba state-space module was introduced into the trend branch to enhance the capability of modeling long-range dependencies such as the stage evolution of flash butt welding and displacement trend changes, while a multi-scale Transformer was employed in the residual branch to capture short-term anomalies and cross-stage correlation features. Finally, an unsupervised anomaly score was constructed using the KL divergence of multi-view feature distributions. Experimental results show that the TD-Mamba method achieves a precision of 92.7%, a recall of 100.0%, and an F1-score of 96.2% on the self-built mooring chain flash butt welding dataset and outperforms existing mainstream methods on most public time-series datasets, demonstrating high detection sensitivity and stability especially for long-term trend anomalies.