Aplicação de técnica de análise multivariada em conjunção de técnicas de aprendizado de máquina e aprendizado profundo, empregados para detecção e diagnóstico de falhas na indústria

Abstract

The stability and safety of chemical industries depends on its monitoring capacity for Fault Detection and Diagnosis (FDD), faults are defined as unwanted behaviors of process parameters. The detection of failures and identification of their causes provide a basis for decision-making on equipment maintenance, which ensures the ability to operate chemical processes safely, guaranteeing protection of personnel and equipment. A modern production line can provide the necessary amount of data to solve this problem, however, given the complexity of industrial processes, the use of more traditional methods such as phenomenological modeling and signal analysisbecomeverydifficult,makingspacefortheuseofmultivariateanalysistechniques,such as Principal Component Analysis (PCA), in conjunction with machine learning techniques, or deep learning techniques, the latter two belonging to the context of the 4.0 industry, dependent on Big data. That said, this work aimed to develop, for an industrial application, strategies, in the form of computational routines, built with the programming language Python for fault detection with PCA and the fault diagnosis using machine learning through the Random Forest (RF) technique and deep learning through Recurrent Neural Networks (RNN) specifically the Gated Recurrent Unit (GRU), in addition, for the validation of the developed computational routines, The simulation of the chemical process Tennessee Eastman Process (TEP) was chosen, which is widely used as a benchmark in studies of control, modeling and data processing, as it behaves close to that of a complete real plant and is a complex problem. The PCA showed good results in detecting 11 of the 20 faults, the RF was able to efficiently diagnose 14 of the 20 faults, while the GRU far surpassed the RF performance for all failures, even though it took twice as much RAM and more than 10 times the execution time.