DPMSAN: A Dual-Path Multiscale Attention Network With Diverse Local–Global Interactions for Industrial Quality Prediction

Soft sensing is pivotal for monitoring crucial quality variables in complex industrial processes, which inherently exhibit multiscale spatiotemporal dynamics driven by material flows and equipment topology. However, existing deep learning methods face a critical theoretical bottleneck: Standard architectures often struggle to model diverse local–global interactions, including multiscale local features, intrascale temporal dependencies, and interscale dynamic interactions, while emerging multiscale models typically rely on fixed convolutional kernels or rigid periodic assumptions. Consequently, these limitations hinder the joint modeling of such local–global interactions and time-varying industrial process behaviors, which are essential for characterizing complex process dynamics. To overcome these challenges, this article proposes a dual-path multiscale attention network (DPMSAN). First, parallel 1-D Deformable Convolutions are used to construct adaptive multiscale feature representations. Second, a dual-path encoder combines ProbSparse intrascale attention with a frequency-enhanced interscale attention mechanism to model long-range temporal dependencies within each scale and dynamic interactions across scales. In addition, a gated scale feature fusion module and a scale-aware weighted decoder are introduced to adaptively aggregate multiscale information for quality prediction. Experiments on four industrial datasets show that DPMSAN achieves competitive and generally improved performance overx several recent baselines on most evaluation metrics, indicating its effectiveness for industrial quality prediction under different sampling settings.