Abstract: Discusses the architecture and functions required by the soft-sensing system for workpiece's microstructure and mechanical properties during hot rolling, especially the choice of auxiliary variables and the way to preprocess data both are integral to the soft-sensing system. Based on the mechanism model and artificial neural network (ANN), the architecture of the soft-sensing system is proposed via the technological mechanism analysis of hot rolling and system control requirements in site. The ANN input variables consist of the microstructure calculated by mechanism model and chemical composition of workpiece during rolling, thus normalizing the ANN architecture. During the application of the soft-sensing system, the short/long-term self-learning through correction model is used to meet the requirement of measuring accuracy for online detection.