Transfer learning based on improved stacked autoencoder for bearing fault diagnosis

Deep transfer learning algorithm is regarded as a promising method to address the issue of rolling bearing fault diagnosis with limited labeled data. Stacked autoencoder (SAE) has been widely employed in deep transfer learning research since it is a semi-supervised algorithm. However, there are still some limitations for the transfer learning based on SAE, including the vanishing gradient problem caused by the sigmoid activation function in SAE, and low accuracy under the condition of cross-domain or limited labeled training data. In this work, an improved SAE based on convolutional shortcuts and domain fusion strategy (ISAE-CSDF) is proposed for fault diagnosis of rolling bearing. The sparse term Kullback-Leibler (KL) divergence in the original SAE is replaced with the convolutional shortcuts to prevent vanishing gradient problem and improve the feature extraction ability. The domain fusion strategy can transfer commonly shared feature information from various domains. The feasibility of ISAE-CSDF is validated on two publicly available bearing datasets and a custom-built experiment device. Results show that ISAE-CSDF outperforms the state-of-art methods in the context of different working conditions, cross-domain, and limited labeled data. (c) 2022 Elsevier B.V. All rights reserved.