Abstract: One-dimension vibration signals are often used for gearbox monitoring and fault diagnosis to perform maintenance timely and then reduce losses. Thus, the accuracy and reliability of the fault diagnosis model are determined by the key fault features extracted from one-dimension vibration signals. Typical deep neural networks (DNNs), e.g., convolutional neural network, have performed well in fault diagnosis and been widely used in machine fault diagnosis. However, supervised learning of CNN often requires a large amount of labeled data thus limits its wide application. So a new DNN, one-dimension residual convolutional auto-encoder (1DRCAE), is proposed for unsupervised learning and feature extraction from vibration signals, which significantly improves diagnosis accuracy of the gearbox fault. First, the effective integration of the one-dimension convolutional layer and the auto-encoder is realized, and a deep one-dimension convolutional auto-encoder is constructed in the 1DRCAE network. Secondly, the residual learning mechanism is introduced to train the one-dimension residual convolutional auto-encoder to learn effective features from one-dimension vibration signals. Finally, based on the features extracted by 1DRCAE, a classifier is constructed using a small amount of labeled data for fine-tuning and then is used for the gearbox fault diagnosis. The vibration signals from a gearbox test rig are used to verify effectiveness of 1DRCAE for gearbox fault diagnosis. The experimental results show that this unsupervised learning of 1DRCAE has good denoising and feature extraction ability. It performs much better on feature extraction than the typical DNNs, e.g., deep belief network (DBN) and stacked auto-encoders (SAE). 1DRCAE also shows better performance on fault diagnosis than that of 1-D CNN (1-DCNN), etc.

Key words: gearbox fault diagnosis, feature learning, deep learning, convolutional auto-encoder, residual learning