Bayesian neural networks for calibration of air pollution sensors and fault detection in chemical processes.

Author(s)

Gustavo Ryuji Taira

• Document type: Master's Dissertation
• Press: São Paulo.
• Institution: Universidade de São Paulo (USP). Escola Politécnica (EP/BC)
• Defense date: 2022-09-02

Examining board members

Song Won Park; Olga Satomi Yoshida; Antônio Carlos Zanin

• Advisor: Song Won Park

Abstract

The development of machine learning techniques is considered one of the study areas that has revolutionized society in recent decades. Machine learning is a subarea of the artificial intelligence development area that aims to enable computers to perform tasks autonomously through learning processes based on data analysis. In this way, the development of machine learning techniques has provided great transformations in society by enabling computers to learn how to perform tasks and functions that until recently were only possible to be performed by humans. In addition, the use of machine learning techniques has become a trend in several scientific areas since it enables the training of data-driven models capable of making predictions for systems and phenomena that are difficult to model analytically. Among the various types of existing machine learning models, artificial neural networks are models that have recently gained great prominence in the scientific community. Artificial neural networks are machine learning models recognized for their ability to extract abstract and deep patterns present in data sets, and that recently, under the context of the development of deep learning techniques, have gained notoriety for their success in the solution of complex problems from different scientific areas. Despite this success, the models of artificial neural networks still have some problems that need to be solved. Commonly held criticisms of artificial neural networks are the tendency of these models to overfit and the inability to provide their predictions along with their respective uncertainty values. These problems impair the generalizability of these models and can generate security problems when using their predictions. In this context, the development of Bayesian artificial neural network models has established itself as a promising solution to these problems. Bayesian artificial neural network models are stochastic neural network models that have natural mechanisms of model regularization that avoid the problem of overfitting and that can directly represent the uncertainty values of their predictions. Studies on applications of Bayesian neural networks for solving real problems, however, are still very scarce in the literature compared to the number of studies on applications of traditional neural networks. Therefore, this work aims to prove the applicability of Bayesian neural network models to solve real problems addressed in two case studies. The first case is a study of the application of a Bayesian neural network model for the solution of a problem of calibration of low-cost air pollution sensors. The second case is a study of the application of a Bayesian neural network model for the solution of a problem of fault detection in a real chemical process. For each case study, the performance of the solution based on a Bayesian neural network is evaluated, as well as an analysis of the advantages and limitations of using a Bayesian neural network to solve the problem is performed. (AU)

• FAPESP's process: 19/08280-9 - Application of bayesian deep learning for Smart City monitoring and chemical industry processes operation
• Grantee: Gustavo Ryuji Taira
• Support Opportunities: Scholarships in Brazil - Master