Soft sensing is a promising technique for enhancing prediction accuracy and quality control in industrial processes, particularly in harsh environments. With the advance of machine learning recently, deep learning-based soft sensing becomes a new trend in intelligent manufacturing and industry. However, with the inherent noise and fluctuations in the process, the weak representational power from the recurrent networks falls short to model the spatial-temporal dependency between different process variables. In this work, we propose a dual-stage attention framework that integrates both self-attention and cross-attention mechanisms to capture the relations between endogenous (inherent to the process) and exogenous (external to the process) variables. Furthermore, it incorporates a patching strategy to construct a coarse-grained representation of the industrial data, providing a more holistic view of the system dynamics. The extensive experiments of a case study on the silicon content prediction task demonstrate superior performance compared to the benchmarks. We also provide ablation studies and experimental insights of the integral parts of the framework.