Reconfiguration process for neuronal classification models: Application to a quality monitoring problem
Abstract
In the context of smart industries, learning machines currently have various uses such as self-reconfiguration or self-quality improvement, which can be classification forecasting problems. In this case, learning machines are tools that facilitate the modeling of the physical system. Thus, it is obvious that the model must evolve with changes in the physical system, thereby leading to adaptability/reconfigurability problems. Among the various tools reported previously, real-time systems seem to be the best solution because they can evolve autonomously according to the behavior of the physical system. In the present study, we propose a method for using learning machines efficiently in an evolving context. This method is divided into two components: (1) model conception by defining the objective function and influential factors, setting up data collection, and learning using multilayer perceptrons; and (2) monitoring system conception with the aim of tracking the misclassification rate, determining whether the physical system is drifting, and reacting by model adaptation based on the control charts. This paper focuses on the model monitoring procedure because the model conception procedure is quite classical. The proposed method was applied to a benchmark derived from previous research and then to an industrial case of defect prevention on a robotic coating line for which other methods have proved unsuccessful.
Origin : Files produced by the author(s)
Loading...