Proposta de arquitetura de autoencoder convolucional assimétrico para a análise de imagens de vibrações

Abstract

In the recent years, several industry segments have benefited from the advantages offered by health monitoring of rotating machines. Given the pace at which technological advances have occurred and the operational and economic gains provided by modern computational methods, there is an evident trend to apply artificial intelligence techniques in the maintenance of such equipment. This work addresses the study of artificial neural networks methods applied to the condition monitoring of rotating machines under different learning paradigms. A denoising convolutional autoencoder neural network configuration for the diagnosis of mechanical components is proposed, aiming the main advantages of different supervised and unsupervised architectures to optimize performance in the analysis of images generated from normalized vibration signals. The training of the model is carried out in two distinct phases, a pre-training by an asymmetric autoencoder and a fine-tuning by a classifier network. Through the selection of hyperparameters, experiments based on data extracted from bearings help to determine the structure that offers the most satisfactory results and to validate the generality of the developed method. A comparison between the proposed architecture and other architectures applied for the same purposes, including models proposed in other published works, allows to evaluate its performance in terms of accuracy and robustness to noise. The objective is to assess the effectiveness of the method through the use of two public experimental datasets produced under different machine operating conditions. The results obtained in both case studies justify the options adopted in the definition of the diagnostic system and prove that the proposed configuration offers greater faults classification accuracy and greater robustness in comparison to the other addressed architectures.