Industrial processes with multiple operating grades have become increasingly important in satisfying the requirements of agile manufacturing and a diversified market. However, because of the unknown distribution discrepancy of process data collected from different grades, the development of reliable quality prediction models is still intractable, especially for the grades with limited quality measurements. In this study, a novel framework of an adversarial transfer learning (ATL)-based soft sensing method was designed for the quality inferring of multigrade processes. Treating each grade as a domain, the concept of ATL was adopted to learn a suitable feature transformation between different domains, which reduces the data distribution discrepancy and enriches the information provided by the target domain containing limited labeled data. Subsequently, a domain adaptation-based soft sensor was built in a supervised manner, and it outperformed conventional prediction models in terms of the range of prediction domains and prediction accuracy. Through case studies, the feasibility of the developed method was illustrated via a simulated example and an industrial multigrade polymerization process. The benefits of the ATL-based soft sensor were discussed by visualizing the feature transformation.