Self-Modified Dynamic Domain Adaptation for Industrial Soft Sensing

Data-driven soft sensors have been widely applied to estimating important yet difficult-to-measure quality-relevant variables in industrial processes. The complex industrial data exhibits nonlinearity and dynamics due to changes in operating conditions. Soft sensors developed based on the assumption of identical and independent distributions often struggle to adapt to target domain data with significant distribution discrepancy, which poses a great challenge to traditional soft sensor approaches. Additionally, the dynamic features embedded in the practical industrial processes are of great importance for an accurate soft sensor. Existing soft sensor approaches pay little attention to the combined challenge of distribution discrepancy and dynamic feature transfer, referred to as the dynamic domain adaptation challenge. In this work, we propose a Self-modified Dynamic Domain Adaptation (SDDA) soft sensor approach to solve this problem. We develop a novel sequential optimization framework for dynamic domain adaptation, where the target samples are progressively incorporated and pseudo-labels are iteratively refined to preserve the underlying temporal dependency and enable efficient dynamic feature transfer. Also, we propose a feature alignment with the transfer component analysis (TCA) to avoid potential significant distribution discrepancy and guarantee a stable prediction. We demonstrate the superiority of the proposed method via two real-world industrial cases. Note to Practitioners - In many industrial processes, key quality variables are difficult to measure directly, while data-driven soft sensors provide an efficient alternative through indirect estimation. However, practitioners often face two major challenges: the mismatch between training and testing data distributions caused by time-varying operating conditions, and the dynamic characteristics inherent in industrial processes. Traditional soft sensor approaches rarely consider both issues simultaneously. This work presents a self-modified dynamic domain adaptation (SDDA) soft sensor to address these challenges. The proposed method adopts a sequential optimization framework to preserve intrinsic process dynamics and enable effective feature transfer. In addition, a transfer component analysis (TCA)-based feature alignment is introduced to mitigate distribution discrepancies and enhance prediction stability. For practitioners, the SDDA approach can be applied to a wide range of industrial systems, such as chemical and metallurgical processes, where process dynamics and data drift are inevitable. It provides a practical solution for improving the robustness and adaptability of soft sensors in real-world production environments.