(370k) Hierarchical Multi-Class Segmentation for Identification of MIMO Systems from Historical Data

In this era of large data, utilizing stored historical data for various applications like fault diagnosis, troubleshooting etc. have been explored. Utilizing historical data for model identification requires identification of regions of high quality data with minimal disturbance effects. Hierarchical classification based interval halving technique has been proposed for SISO systems in [1] for identifying regions of high information content from historical data. This method consists of two sequential binary classifications: (i) to identify data quality and (ii) to detect presence of disturbance. Extending this method for MIMO systems pose several challenges such as: collinearity of input variables, presence of correlated input moves, and regions of data where only a few of the inputs are persistently exciting. While performing step / similar tests in the plant, the user ensures that the data collected contains no correlated moves, thereby minimizing collinearity between inputs. Historical data has to be assessed for the presence of these phenomena. Regions where only a few of the inputs show significant variation can also be effectively utilized for model identification. In this work, the segmentation method proposed in [1], is extended to MIMO systems. Results on several case studies will be presented to highlight the efficacy of the proposed approach.