Fault Diagnosis of Pumps Based on Acoustic Images and an Improved Pre-Trained CNN Model

Fault diagnosis of mechanical equipment is a crucial research direction. It helps improve production efficiency, reduce costs, and ensure worker safety. This paper focuses on fault diagnosis based on the acoustic signals of pumps, which are collected from real factory working environments and contain obvious background noise. This paper proposes an improved Convolutional Neural Network (CNN) model to handle this task. Additionally, this paper selects the most suitable time-frequency transform method for feature extraction from the acoustic signals. Due to the significant imbalance in the dataset, two data balancing methods are employed, and the most effective one is chosen. Finally, the model's fault diagnosis capability is further enhanced through pre-training, which is conducted on a related rotating machinery dataset. The results show that the value of Area Under the Curve (AUC) reached 98.34%, indicating that the proposed model's fault diagnosis capability is superior to other models in the research.