Inferência estatística no reator de Van de Vusse por meio de machine learning

Abstract

Currently, technical applications related to artificial intelligence, such as Machine Learning, are applied in several areas, but in chemical engineering its biggest of Soft Sensor and automatic failure detectors. These are algorithms that have an extremely malleable basic structure, great simplicity and adaptability to the data being worked on. The operation of a Soft Sensor in general is a signal estimation, interpolation and prediction (inference) that has the objective of generating a virtual measurement as a function of easy-to-measure variables in order to replace a variable experimentally monitored by analysis. The present work addressed a typical problem in the chemical industry, which is statistical inference, where the Van de Vusse reactor was used, which is a well-known process in the literature for having many nonlinearities and consequent complexity that makes monitoring and control difficult, which will be modeled with the help of Artificial Neural Networks (ANN's), which are computational models based on the functioning of the human brain, which, through the connection of artificial neurons, are able to learn complex relationships between data. Data for the evaluation of the efficiency of Machine Learning as a virtual sensor or Soft Sensor, were generated by modeling the Van de Vusse reactor mathematically represented from its mass and energy balances with systems of ordinary differential equations that were numerically solved by Euler's method. It was observed that the definition of the sampling time is an important step in the process, as it benefits the Soft Sensor performance metrics. For the established approach, the Artificial Neural Network performed a satisfactory result in the construction of the virtual sensor, proving to be viable for application in industrial plants. Both obtaining and using the data in the ANNs relied on the help of Python software, which provides free libraries for implementing the algorithms and has been showing great interest from the industry for the application of engineering 4.0.