Fault diagnosis of control valves based on small-sample hybrid physics improved Resnet

Abstract

Pneumatic control valves, as vital components of industrial process automation, ensure the smooth operation of industrial production systems. However, they are susceptible to various malfunctions due to harsh working environments and complex transmission media, which can significantly impact production safety and efficiency. To address the challenge of obtaining fault data in actual operational settings, we constructed a fault test bench for pneumatic control valves and simulated a variety of fault conditions. We collected 421 fault data samples across 4 valve opening conditions, categorizing them into 27 distinct states with varying sample sizes, averaging 3-4 samples per state. To tackle the small-sample issue, we proposed a data augmentation method using periodic extension, validated through comparative analysis with other algorithms. Additionally, we innovatively analyze the data flow of pneumatic control valves and explore the relationships between different parameters. Based on these relationships, the input structure of the residual network is optimized. The above theoretical approach reduces the number of variables that need to be captured by the pneumatic control valve inspection system. Finally, through experiments under extreme conditions, our approach successfully diagnoses faults in 26 subclasses of pneumatic control valves, providing a reliable safeguard for industrial production safety and stability.