Health condition identification for rolling bearing using a multi-domain indicator-based optimized stacked denoising autoencoder

Abstract

Stacked denoising autoencoder is one of the most classic models of deep learning. However, there are two problems in the traditional stacked denoising autoencoder: (1) the parameter selection of stacked denoising autoencoder mainly depends on expert experience and (2) stacked denoising autoencoder is mainly restricted to learn automatically single-domain features from raw vibration signals while identifying the fault type, which implies that no linear mapping relationship located in other domains of vibration data is neglected, which may lead to the imperfect diagnostic results. Consequently, to address these issues, learn the well-rounded feature representation, and improve recognition accuracy, this article presents a novel approach called multi-domain indicator-based optimized stacked denoising autoencoder for automatic fault identification of rolling bearing. First, multi-domain indicator of the original vibration signal is constructed through calculating the expression of different domains (e.g. time frequency domain, and time frequency domain). Second, the constructed multi-domain indicator is regarded as the input dataset to train stacked denoising autoencoder architecture containing three hidden layers, and a recently reported nature-inspired algorithm named grasshopper optimization algorithm is employed to synchronously determine the model parameters of stacked denoising autoencoder, which is aimed at learning more robust and reliable feature representation. Finally, the feature representation learned in the testing set is fed into the trained stacked denoising autoencoder model containing softmax classifier for identifying bearing health conditions. The presented method is evaluated using two bearing vibration datasets. Experimental results indicate that our approach can provide high-accuracy recognition over 99% for bearing health condition, and it achieves more decent and precise classification results compared with some shallow learning model and standard deep learning architecture.