Abstract:
Reliable machinery health monitoring using measured vibration signals requires a good readability of time-frequency (TF) images. However, conventional TF methods suffer from a limited time–frequency resolution and cross-term interferences, which limit their practical applicability in health monitoring. To address this issue, a TF image improvement method using deep learning is proposed in this paper. The proposed method employs a deep learning technique known as conditional generative adversarial network (cGAN) to convert a noisy low-resolution TF image of a bearing vibration signal into a noise-free high-resolution image such that the true frequency characteristics of measured signals may be revealed. In this paper, the cGAN model is trained using a simulation-based dataset generated from a bearing analytical model. The trained cGAN model is then utilized to improve TF images generated from real bearings under different fault and operating conditions. The results reveal that the proposed image improvement method generates high-resolution TF representations which are better than both the traditional TF images and those generated using TF reassignment methods