| 초록 |
Quality process variables that are hard to measure in real-time by conventional sensors are challenging in optimal control designs to make prudent and timely economic decisions toward safe and effective industrial operations. Due to offline measurement limitations, the untimely availability of quality-related variables affects the interconnected online process monitoring, optimization, and control systems. These limitations have inspired the development of linear and deep neural network (DNN) models, achieving varying degrees of success. While DNNs are known to provide high accuracy and handle nonlinearities of complex processes, they are extremely non-interpretable, making them hard to trust in making delicate decisions. Hence, this study proposes an interpretable inferential sensor composed of neural additive networks to infer quality variables from primary measurements to make sensitive decisions. Not only is the proposed model comparable to typical DNNs in accuracy, but how the model makes predictions can be understood. The results suggest that the proposed model is extensible to diagnosing processes that DNNs cannot offer. |
