Análise de desempenho de método baseado em rede LSTM para classificação de falhas em um processo de controle de nível

Abstract

Due to the increasing demands in the operation monitoring of industrial plants, methodologies for fault detect and diagnose in the operation of these processes are gaining more and more importance, because they can contribute to more assertive and even predictive repairs in the components that generated such disturbances to the proper functioning of the system. With the growth of data-oriented approaches, Artificial Neural Networks have become considerable allies in solving these problems, and Recurrent Neural Networks, in particular, has gained strength due to their affinity in dealing with series that have temporal links between their samples, which is the case of industrial process variables monitoring. Due to this relevance, this dissertation analyzes the performance of Long Short-Term Memory (LSTM) recurrent neural network for the detection and classification of faults in a pilot-scaled level control process. For the performance evaluation, a methodology based on Monte Carlo statistical tests was used, in which the influence of the LSTM network hyperparameters, such as the number of layers and size of the input and regressors, was analyzed. The accuracy was the metric chosen to quantify the fault classification performance. The data set obtained from the operation of the pilot plant contained 23 situations of disturbances in this process, which resulted from disturbances applied to components such as sensors, valves, and the water tank itself. The adopted methodology proved to be quite efficient to examine both the performance and the robustness of these neural networks for the fault classification activity, in addition to indicating the best network architecture configurations.