Abstract 

This paper suggests an image gradient based method that determines the optimal image size for convolutional neural network (CNN)-based diagnosis of fluid-film rotorbearing systems. As distinct patterns improve the diagnosis performance, a criterion is defined to measure the intensity of patterns in an image. The proposed criterion is derived by segmenting an image by the size of the CNN filter and evaluating each segment through the use of image gradient analysis. Vibration signals from a testbed are used to demonstrate the proposed method. First, the signals are transformed into vibration images by using an omnidirectional regeneration technique. Then, vibration images of four different health states are analyzed using the suggested criterion. The analyzed results are compared to the performance of CNN based diagnosis. The results indicate that the proposed criterion can determine the optimal size range of the vibration image that gives the best performance for CNN-based diagnosis.