SGST: A Novel Approach Based on Machine Learning for Cavitation Fault Diagnosis

Cavitation is a common fault phenomenon in hydraulic machinery, which can affect the performance of equipment, causing serious economic losses and safety accidents. Many methods for cavitation fault diagnosis have been proposed and validated to be effective, but due to the limited amount of cavitation data and the insufficient feature extraction ability of the models used, there are still serious issues with the accuracy of fault state judgment for these methods. To improve the effectiveness of diagnose cavitation faults, this paper proposes a method entitled SGST which consists of Short-time Fourier Transform (STFT), Generative Adversarial Network (GAN) and Swin Transformer. In this method, first, the vibration signals of hydraulic mechanical equipment are pre-emphasized to compensate for attenuation of high-frequency components. Then STFT is used to transform the signals into spectrograms to capture abundant features based on time-frequency domain. Next, GAN is adopted to perform data augmentation on the spectrograms, increasing the scale and diversity of the dataset. Ultimately, the augmented spectrograms are input into the Swin Transformer network for classification training. The experimental results show that this method achieves an accuracy of 98.6% on the validation set, which is better than the existing cavitation fault diagnosis methods, proving its feasibility and effectiveness.