Desenvolvimento de estratégias utilizando estatística multivariada e técnicas da engenharia 4.0 no acompanhamento e monitoramento de processos químicos

Abstract

The concept of engineering 4.0 is already a reality within current industries. Therefore, it is necessary to develop strategies for using a set of techniques, such as artificial intelligence, connectivity, data science, internet of things. Just as, with the advent of new software, the same occurs with a more effective use of multivariate statistics. In chemical processes, an inherent problem is obtaining process control and efficiency parameters that are difficult to measure, for technical or economic reasons. Multivariate statistics and engineering 4.0 can collaborate to solve these problems, for example, in the construction of virtual analyzers. In a typical Vinyl Monochloride (MVC) industry, a problem that fits into this context is the measurement of 2-chlorobutadiene, necessary to control the production process, avoiding significant losses in process efficiency. In this sense, this work aims to develop computational algorithms in Python with a focus on creating virtual sensors for predicting the contaminant 2-chlorobutadiene, using Artificial Neural Networks (ANN) and multivariate statistics. Due to the complexity of the process and its numerous variables, multivariate analysis techniques also gain prominence in exploratory data analysis, the most common of which is Principal Component Analysis (PCA). The results indicate that the concentration of 2-chlorobutadiene has a greater correlation with the variables: furnace temperature, flow rate of 1,2-EDC, and pressure of vaporized 1,2-EDC. Therefore, these variables were chosen as input data for training the analyzer's ANN. The results indicate that, among the virtual sensors constructed, the ANN without PCA, with a hidden topology layer [3, 7, 1] presented the best performance and effectiveness in estimating the contaminant concentration, reaching an accuracy of approximately 99.4%. Furthermore, this study contributes to greater robustness in the MVC production process, providing a basis for implementing control and monitoring strategies in complex reaction processes. Additionally, the results obtained have potential applications in other areas of the chemical industry, allowing the development of virtual analyzers to improve process reliability and efficiency.