Generally, dynamic model selection is implemented using algorithms that need a feedback from the system’s output; but, in many real-world applications this feedback is not available. For dealing with this challenge, a novel two-step machine learning approach is introduced by designing a dynamic model selection strategy where the selector only requires information about the system parametrization and not about its output. The first step of this approach is centred on a selection procedure that determines the most adequate model from a digital library with eight machine learning techniques using k-Nearest Neighbours. In the second step, the selected model is then used to make the prediction. The proposed approach is validated in a case study for predicting surface roughness of a micro-machining process which presents complex cutting phenomena, such as built-up edge and micro-burr formation. The experimental results corroborate the advantages of the proposed method increasing R2 from 0.892 to 0.915 and decreasing mean absolute percentage error from 19.79 % to 14.63 % when compared to the best individual models’ metrics.

his work was supported by the H2020 project “Platform enable KITs of Artificial Intelligence for an Easy Uptake of SMEs (KITT4SME)”, Switzerland, grant ID 952119; Ministerio de Ciencia e Innovación (MICINN) project "Self-reconfiguration for Industrial Cyber-Physical Systems based on digital twins and Artificial Intelligence. Methods and application in Industry 4.0 pilot line", Spain, provisional grant ID PID2021-127763OB-100” supported by MICINN.

Peer reviewed