Complementary Representations of Invariant in Domain Generalization for Industrial Data Drift
针对工业过程数据漂移问题，提出不变互补域泛化方法，通过协变表示补充不变表示，从信息论角度建立联合学习目标，理论证明优化目标等价于最小化未见域经验风险上界，在燃气轮机和聚酯酯化数据集上验证有效性。
It is well known that data drift may occur in complex industrial processes, which can cause changes in the distribution of data sampled by sensors. Therefore, the ability to generalize across unseen domains is essential for monitoring systems deployed in industrial processes. The invariant-complement domain generalization (ICDG) is proposed to alleviate data drift in industrial processes. This study reveals how the proposed covariant representation complements the invariant representation. Additionally, it derives novel theoretical error bounds characterizing the relationship between seen and unseen domains. Intuitively, the proposed invariant-complement representation method mitigates the influence of variant factors and encourages the learning of invariant and covariant representations. From the perspective of information theory, the boundaries for invariant and covariant representations are established and integrated as a joint learning objective with multiple information constraints. Theoretically, we elucidate that optimizing the ICDG objective function is equivalent to minimizing the upper bound of the empirical risk associated with unseen domains. This result helps ensure accurate prediction of quality variables under data drift. Case studies on the gas turbine (GT) dataset and the actual polyester esterification dataset validate the effectiveness of the proposed ICDG. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/heheding/ICDG</uri>