Improving accuracy of cavitation severity recognition in axial piston pumps by denoising time–frequency images

Abstract

The vibration signal is a good indicator of cavitation in axial piston pumps. Some vibration-based machine learning methods have been developed for recognizing pump cavitation. However, their fault diagnostic performance is often unsatisfactory in industrial applications due to the sensitivity of the vibration signal to noise. In this paper, we present an intelligent method for recognizing the cavitation severity of an axial piston pump in a noisy environment. First, we adopt short-time Fourier transformation to convert the raw vibration data into spectrograms that act as input images of a modified LeNet-5 convolutional neural network (CNN). Second, we propose a denoising method for the converted spectrograms based on frequency spectrum characteristics. Finally, we verify the proposed method on the dataset from a test rig of a high-speed axial piston pump. The experimental results indicate that the denoising method significantly improves the diagnostic performance of the CNN model in a noisy environment. For example, using the denoising method, the accuracy rate forcavitation recognition increases from 0.52 to 0.92 at a signal-to-noise ratio of 4 dB.