A Bayesian Adaptive Resize-Residual Deep Learning Network for Fault Diagnosis of Rotating Machinery

Due to the high accuracy achieved in data-driven fault diagnosis, time frequency images generated by ContinuousWavelet Transform (CWT) are widely used as the input of deep learning methods. However, the image data require huge amount of data memories. An adaptive resize technique provides a reliable way for deducing the scale of image and achieving good effect. In this study, a novel Bayesian adaptive resize-residual network was proposed to resize the input data scale and extract the image feature for mechanical fault diagnosis. The CWT and Histogram Equalization (HE) algorithm were used to generate enhanced timefrequency images. The newly developed adaptive resize-residual network was applied for feature extraction, in which the adaptive resize block can adaptively resize input image by self-learning, and the residual block was used for classification. The Bayesian optimization was introduced to optimize the model hyper parameters and obtain an effective model. A testbeds of rolling element bearings are introduced to support the experiments. The experimental results indicate that the proposed Bayesian adaptive resize-residual network obtains superior recognition accuracy and outperforms many state-of-the-art methods. This method is conducive to improving the capabilities of rotating machinery fault diagnosis, and reduces the repair time of fault.