Intelligent Bearing Fault Diagnosis Using Adaptive Time-Frequency Convolutional Features with Multi-task Learning

Abstract

Deep learning based fault diagnosis methods can extract representative and distinguish features automatically from raw signals and have achieved promising results in recent years. However, deep models rely heavily on massive labeled data and do not generalize well under varying load conditions. Besides, both temporal features and frequency features of raw vibration signals provide useful information in fault diagnosis and should be fully exploit to acquire better performance. To address these issue, a dual-input model based on a convolutional neural network with attention mechanism and multi-task learning strategy is proposed in this paper. The model uses both temporal and frequency adaptive features to achieve end-to-end fault diagnosis. First, a two-branch one-dimensional CNN is designed to learn rich and complementary features from both temporal and frequency data in parallel. Moreover, attention mechanism is adopted in the two-branch CNN to further increase the adaptability of the proposed model under different working loads. Finally, to obtain robust and discriminative features as well as enhance recognition accuracy, a multi-task learning strategy is also employed to better optimize the whole network. Experimental results over the standard Case Western Reserve University (CWRU) bearing fault diagnosis benchmark dataset demonstrate the superiority of our proposed model and can achieve a high fault recognition rate under variable load conditions with fewer training samples.