Técnicas de deep learning: autoencoder e long short-term memory (lstm) no desenvolvimento de sensores virtuais e na detecção e diagnóstico de falhas em processos reacionais complexos

Abstract

With the modernization and application of Artificial Intelligence in the context of Industry 4.0, deep learning techniques have been extensively explored as an alternative to optimize industrial processes, mitigating errors caused by human failures or physical equipment limitations in processing large amounts of data. Currently, the major applications of deep learning techniques include chemical processes, statistical inference with softsensors, and fault detection and diagnosis. This study aims to utilize current deep learning techniques, namely Autoencoder and Long-Short-Term Memory (LSTM), to develop softsensor and fault detection and diagnosis strategies in complex reactive systems. The models were created in two stages: the first stage focused on the creation, training, and evaluation of softsensors for different datasets, while the second stage aimed at developing fault detection and diagnosis strategies in two distinct scenarios: LSTM in the same stage, and the second scenario utilized Autoencoder for fault detection and LSTM for fault diagnosis. The scenarios were subsequently evaluated and compared. The results obtained revealed that softsensors developed using LSTM deep learning technique exhibited better performance using two datasets: one with a sampling time of 3 seconds and 2 times delays, yielding a coefficient of determination of 0.999998 and a mean squared error of 0.37x10-8; and another dataset with a sampling time of 9 seconds and 3 time delays, presenting a coefficient of determination of 0.999998 and a mean squared error of 0.35x10-8. Regarding fault detection and diagnosis, the combined use of Autoencoder and LSTM techniques in the second scenario achieved superior results with 99% accuracy and a 1.02% error rate, while scenario 1 showed 98% accuracy and a 1.18% error rate. This integrated approach in scenario 2 proved to be more effective in fault identification, highlighting its practical relevance in the context of complex industrial systems. Therefore, it can be concluded that the deep learning techniques analyzed in this study provide promising solutions for softsensor simulation and fault detection in complex reactive systems. Furthermore, these findings support the importance of this work in the development of new technologies for the education of chemical engineers, especially considering the challenges posed by Industry 4.0.