Adaptive Multimode Monitoring Based on Dictionary Learning with Online-Updating Window

In actual industrial processes, factors such as changes in production targets or related parameters can lead to the creation of multiple modes. A method has been proposed to overcome the problem of “catastrophic forgetting” of historical models due to the incompatibility of models learned from historical data and the updating of models with new modalities, with good results. However, this method is suitable for the case where a large amount of normal data with different modes are obtained in advance during the offline modeling phase, and if a new mode is encountered during the online process, the mode update cannot be performed in time. At the same time, the online process cannot process the data in advance, and it is often disturbed by outlier points and so on. To address this problem, this paper extends it to the online process based on the original method, and proposes a window update strategy to overcome the influence of outlier points. The effectiveness is demonstrated by a numerical example.