Detecção de falhas em processos químicos contínuos : uma abordagem via ensemble learning e inferência bayesiana

Abstract

A fault is an unwanted event in any operation in a chemical industry. Its occurrence may cause safety, environmental and economic losses. Due to the complexity of chemical processes from one side, and the availability of very large data sets from the other side, the use of fault detection systems based on historical process data has increased. The Principal Component Analysis (PCA) is the main multivariate statistical method employed with this purpose. However, no particular technique is capable of describing any chemical process as a whole; thus, the combination of techniques constitutes in a potential alternative solution towards more efficient fault detection. The objective of this work is to investigate the combined use (ensemble learning) of three techniques, namely PCA, kernel PCA and dynamic PCA, through a Bayesian inference strategy. The case study is the Tennessee benchmark, which is the most usual in the Chemical Engineering community worldwide. The higher efficiency of the adopted approach was quantitatively verified by means of the monitoring metric called Missed Detection Rate (MDR). From the analysis of the results, it was observed that there is no loss of the detection capacity of the combined system in relation to the individual ones; the significant improvement in performance, from the metric Missed Detection Rate (MDR), mainly for five of six hard-todetect faults; with MDR values close to zero for four of these faults. In summary, it was possible to verify the significant gain when adopting a combined fault detection system in relation to the individual models. Greater certainty about the presence of a fault is a crucial factor in at least mitigating potential losses in industrial processes.