Abstract: The G-M cryocooler-based cryopump was identified as an important device widely used in fields such as semiconductor manufacturing, playing a critical role in achieving and maintaining ultra-high vacuum. Due to its long-term continuous operation, it is prone to mechanical wear and other failures, leading to a decline in cooling capacity and pumping characteristics. Therefore, conducting effective fault diagnosis is particularly crucial. An improved genetic algorithm combined with a back propagation neural network was proposed for fault diagnosis, overcoming the issues associated with traditional BP neural networks, which relied on initial weight and threshold settings and exhibited low optimization efficiency. The results showed that this method achieved a fault diagnosis accuracy of 98.05%, providing a scientific basis for health monitoring and fault warning of cryopumps.