I Simposio de Postgrado 2023. Ingeniería, ciencias e innovación

MÓDULO_ 05 Ingeniería Eléctrica 112 FAULT DETECTION AND CLASSIFICATION BASED ON DATA-DRIVEN MUTUAL INFORMATION ESTIMATOR IN THE TENNESSEE EASTMAN PROCESS ABSTRACT Effectively determining abnormal operating conditions and identifying their origin are among the main concerns of process engineers and plant operators. Efforts in this task will help to improve process operations and maintenance, increase plant throughput, decrease process downtime, and comply with ever- stricter environmental rules and safety regulations. These fault or abnormality detection and identification are included in the modern concept of Prognostics andHealthManagement (PHM). We propose a method for detecting failures in an unsupervised way, only requiring a model of the plant to be trained in a healthy condition; as a consequence, we not only get to know if the system is operating abnormally, but we also get signatures for the different fault modes that can help in fault classification. We achieve this by computing the mutual information between the input and a quantity we call residual which in this case is equal to the prediction error of the model and imposing a threshold on this quantity effectively giving us an hypothesis test for the operation of the plant. To show our method, we use a well-known benchmark, the Tennessee Eastman Process simulator, consisting of a complex chemical plant where several faulty modes can be simulated on its different subsystems. Our method not only proves to be accurate in detecting faulty modes but is also highly interpretative. Tomás Rojas 1 *, Camilo Ramírez 1 , Ferhat Tamssaouet 2 , Jorge F. Silva 1 , Marcos Orchard 1 1Departamento de Ingeniería Eléctrica, Universidad de Chile. 2PROMES-CNRS, University of Perpignan. *Email: tomas.rojas.c@ug.uchile.cl