Adapting to Changes: A Novel Framework for Continual Machine Learning in Industrial Applications

Abstract

This paper is dedicated to solving the problem of concept drift in industrial plants using artificial intelligence methods. For this purpose, methodological approaches and procedures are considered and analyzed. Based on the findings, reference architectures were developed at different abstraction levels that can be used in an industrial environment and enable continuous machine learning. Continuous machine learning offers the possibility of adapting to dynamic changes in the production environment, which are reflected in constantly changing data sets. Through a combination of machine learning techniques, a novel and practical framework for continuous learning, also known as lifelong learning, is presented. The integration of problem-focused machine learning methods is advancing in production, e.g., predictive maintenance, process optimization or fault detection. Thereby, fully or semi-automated adaptations to changing environments requiring continuous improvements are less often explored, although practical use cases often require adaptive capabilities as the physical data distribution may change over time. In this paper, the application was continuously improved based on case studies and empirical results, and finally validated with a quality assurance application. Various methods and approaches for detecting concept and data deviations, retraining, packaging and model updating had to be investigated, which led to the question of what a real industry-oriented implementation could look like. The result is a reference architecture that can run on cloud and edge computing resources. This reference architecture is validated in real-world application in the parquet production sector, proving its feasibility and efficiency.