Developing a data science platform for pipeline applications Tennessee Eastman (Atena Editora)

Abstract

A data science platform in a smart chemical plant enables simultaneity of multiple applications. However, in each application there is a wide variety of algorithms that can be implemented. Therefore, the objective of this work is to evaluate the performance of the algorithms to be implemented in a data science platform for applications related to fault classification, fault detection and virtual sensor in the Tennessee Eastman (TE) process. To achieve this, a synthetic data set obtained by simulating the TE process was built, where each algorithm was trained, optimized and evaluated according to the type of data science task associated with each application. The results show that the artificial neural network only achieved the best performance in fault classification. In the virtual sensor, gradient boosting (GB) and k-nearest neighbor (k-NN) achieved the best performance. Meanwhile, in fault detection, most of the evaluated algorithms achieved a fault detection rate around 88%. In conclusion, algorithms based on artificial neural networks did not achieve the best performance in all implemented applications, being surpassed by other non-linear algorithms.