Unsupervised Domain Adaptation Algorithms in Industrial Environments: Recent Advances and Current Applications

Abstract

In the complex industrial environment, the original physical properties, such as rotating speed and various machines, which change with time, will inevitably lead to unrecoverable feature changes and the decline in the extensive use of models. Therefore, unsupervised domain adaptation (UDA) is the basic structure to bridge the gap between intelligent diagnosis and practical use. However, most of the existing UDA summaries only illustrate the algorithms one by one, and there is no systematic investigation based on the fundamental principles of industrial constraints. Because of the fragmentation of this method, a huge gap has been formed between the most advanced theoretical discovery and its application in practical engineering. In order to break this deadlock, this summary completely abandons the past algorithm-centered classification method, establishes a new orthogonal separation framework, and directly deduces the solution method of abstract algorithm from specific physical constraints. Within this big framework, this paper arrangement illustrates the following four mechanisms. First, the dynamic structure fixing mechanism in the ever-changing state; Second, the feature remodeling without using the original features while protecting the privacy of cross-enterprise; Third, the complementarity and diversity generated in the cold start state with very few samples; Finally, the review advocates that the following three are the established evolution paths to promote the new generation of industrial intelligence to realize large-scale automated operation: third, The structure based on physics is added to the center of the algorithm, and the engineering structure is transferred to light edge computing, and Industrial Foundation Models are used to close the macro data and physics.