A New Deep Stacked Architecture for Multi-Fault Machinery Identification With Imbalanced Samples

Effective intelligent fault diagnosis of rotating machinery using its vibrational signals has a considerable influence on certain analysis factors such as the reliability, performance, and productivity of a variety of modern manufacturing machines. Traditional intelligent approaches lack generalization schemes and add the burden of extracting features from data-driven cases. On the other hand, the Deep Learning (DL) studies have reported capabilities higher than the expectations of the researchers' objectives. In this context, this paper proposes a new deep architecture based on Stacked Variant Autoencoders for multi-fault machinery identification with imbalanced samples. The proposed model starts with a Variational Autoencoder (VAE) for facilitating data augmentation of small and imbalanced data samples using Gaussian distribution. After the preparation of suitable samples based on quality and size, the preprocessed vibration signals obtained are injected into the deep framework. The proposed deep architecture contains two subsequent unsupervised Sparse Autoencoders (SAE) with a penalty term that helps in acquiring more abstract and essential features as well as avoiding redundancy. The output of the second SAE is integrated on a supervised Logistic Regression (LR) with 10 classes. This is utilized for the proposed classifier training to achieve accurate fault identification. Experimental results show the efficiency of the proposed model which achieved an accuracy of 93.2%. In addition, for extensive comparative analysis issue, the Generative Adversarial Network (GAN) and triNetwork Generative Adversarial Network (tnGAN) were both implemented on the vibrational signal data, where the proposed method reported better results in terms of training and testing time and overall accuracy.