A Dynamic Domain\nAdaptation Regression Method for\nMultiple Working Conditions Based on Continual Learning

The process industry is the pillar of the national economy\nand\na key force in promoting social development. The complexity of its\nproduction processes and the diversity of working conditions lead\nto problems such as variable coupling, dynamic time variation, and\ndata distribution mismatches. Especially when faced with massive amounts\nof data, it is crucial to effectively learn new knowledge while retaining\nexisting knowledge. This helps avoid the dilemma of catastrophic forgetting.\nTo address the above problems, this paper proposes a dynamic domain\nadaptation regression method for multiple working conditions based\non continual learning. Construct a latent variable space to describe\nthe dynamic regression relationship between process variables and\nquality variables. Introduce domain adaptation regularization terms\nfor latent variable reconstruction and a dynamic continual learning\nregularization term. It is demonstrated on industrial data sets that\nthis method can maintain high prediction accuracy and low forgetting\ndegrees under multiple working conditions.