Avaliação de técnicas de Machine Learning na detecção e diagnóstico de falhas em válvulas de controle

Abstract

Process automation is one of the characteristics of Industry 4.0 that generates greater availability of historical data from extensive monitoring of equipment, processes and products. Thus, there is a need to develop means that enable the analysis of this information in search of solutions and opportunities to optimize production. The methodologies based on intelligent systems are being increasingly applied and studied for this purpose. Machine Learning techniques study the development of computational algorithms that allow the development of models based on the analysis of provided data and, therefore, the generalization of the information learned in the form of predictions. In the chemical industry, the Failure Detection and Diagnosis area uses this data to minimize failures and monitor the behavior of the process under these unusual conditions. Thus, the objective of this work is to evaluate different Machine Learning techniques in the Detection and Diagnosis of failures in control valves. The chemical process studied is an accumulation and separation vessel for mixtures, developed in the UniSim® Design software. The Python language will be used for data processing and algorithm development. The HAZOP methodology will be applied to determine the control valve that poses the greatest risk to the process. In this valve, two types of faults will be applied, gripping and pressure suppression. For Detection and Diagnosis of failures, models of Decision Trees, Neural Networks and Random Forests will be used. UniSim® Design provided the simulation and modeling of the chemical process, as well as the dimensioning of the equipment and instruments used in the system. The diameter obtained for the separation vessel was 0.76 m, and the length was 2.27 m. Through the HAZOP methodology, it was possible to identify the risk situations and their consequences in the chemical process. Thus, it was determined that the control valve that presents the greatest risk in case of failures is the VLV-001 valve. Exploratory data analysis was necessary to understand the simulated chemical process and the response of the variables to system modifications. The Random Forest technique was the model that presented the best results in the detection and diagnosis of failures in the control valve, with an accuracy of 82% in the tests. This result can be optimized after adjusting the hyperparameters, obtaining an accuracy of 84%. Based on the above, it is concluded that Machine Learning techniques are capable of detecting and diagnosing failures in control valves.