Improving Machine Learning Modeling of Nonlinear Processes Under Noisy Data Via Co-teaching Method

In practical implementation of machine learning modeling, training data often involve noise which can cause performance degradation as the machine learning model may overfit the noisy pattern. This work focuses on a novel machine learning training approach, termed co-teaching method, that utilizes both noisy data from industrial datasets and noise-free data from first-principles model solutions to improve model accuracy. Specifically, we consider an ASPEN dataset with non-Gaussian noise data, and develop two long short-term memory (LSTM) networks following the standard training process, and the co-teaching training method, respectively. A chemical process example is used to demonstrate the improved model accuracy under co-teaching method in both open-loop operation and closed-loop operation under model predictive control.

Automated summary

This work introduces a co-teaching method to improve model accuracy by using both noisy and noise-free data, and demonstrates its application in a chemical process through a case study.