Adaptive Densely Connected Convolutional Auto-Encoder-Based Feature Learning of Gearbox Vibration Signals

Abstract

Vibration signals are widely utilized for machinery fault diagnosis. Typical deep neural networks (DNNs), e.g., convolutional neural networks (CNNs), perform well in feature learning from vibration signals and have been applied in gearbox fault diagnosis. However, the supervised-learning-based DNNs depend on a large amount of labeled data, and the extracted features consist of much noise and redundant information. In this article, a new DNN, adaptive densely connected convolutional auto-encoder (ADCAE), is proposed for feature extraction from 1-D vibration signals directly in an unsupervised-learning way. First, adaptive attention mechanism (AAM) is proposed for feature filtering. Second, a multiscale convolution based on AAM is proposed for fusion of multiscale information. Moreover, a new unsupervised-learning network, densely connected convolutional auto-encoder (DCAE), is further developed to improve the information flow between encoder and decoder, which significantly improves the performance of feature learning and fault classification. The results on two cases show that ADCAE has good feature extraction performance on gearbox vibration signals. It performs quite better on gearbox fault diagnosis than typical DNNs, e.g., residual CNN (ResNet), densely connected CNN (DenseNet).