HARFNet: Hierarchical Attention Residual Fusion Network for Industrial Chain and Supply Chain Risk Early Warning

In today's volatile global economy, industrial and supply chains are increasingly susceptible to disruptions caused by escalating geopolitical tensions, persistent logistics congestion, and sudden demand shocks. Establishing an effective early warning system for event severity is crucial for strategic buffer deployment and agile sourcing switching. However, existing warning indicators often manifest as heterogeneous mixed tabular data exhibiting weak separability, posing significant challenges for traditional prediction methods. To overcome these limitations, this paper introduces HARFNet, a novel Hierarchical Attention Residual Fusion Network. The architecture integrates three distinct components: Dual-Path Heterogeneous Encoding (DPHE), Hierarchical Dual Attention (HDA), and Gated Residual Cross-Path Fusion (GRCF). The study utilizes a robust dataset of 3,000 calibrated disruption events and 49 distinct early-warning features. HARFNet's performance is evaluated against ten classical learning algorithms through a rigorous one-round stratified experiment. The experimental results clearly demonstrate that HARFNet outperforms all compared models, achieving the highest macro-F1 score. Furthermore, the model provides interpretable insights by generating group-level attention weights. The findings confirm that categorizing supplier, logistics, demand, inventory, and visibility signals is particularly useful for accurate and actionable four-class severity warnings, thereby significantly enhancing supply chain risk management strategies.