| 초록 |
Contemporary chemical processes in diverse industries typically require multiple units which are connected via complex transport pathways. Due to the intrinsic complexity of the multi-unit process, fault in certain unit can propagate rapidly over the entire system. Although there have been several methods of detecting process anomaly signals, traditional principle component analysis (PCA)-based methodologies are limited in detecting specific error signals due to high frequency of false alarms as well as low true accuracy, resulting in time and profit losses. In this study, we developed a novel machine-learning based algorithms which can quantitatively measure disorder in multivariate time-series signals. The newly developed covariance-based methodology considerably reduces the false alarms while sustaining high accuracy for manifold error signals. Compared to conventional metrics, the newly developed measure exhibited much lower false alarm rates, while achieving enhanced true accuracy. The present algorithm is expected to work effectively in detecting various fault signals among multivariate time-series data, which should be valuable for maintaining the complex chemical processes. |
