Abstract: Based on the advantages of both the expert knowledge and deep learning methods, the reactor coolant pumps are taken as the pilot object to research and design the a state monitoring and diagnosis system for reactor coolant pump in M310 Unit. The data management and diagnosis modules was designed additionally in the system as well as retaining the original measurement system. In the system, a high-precision diagnostic model that does not rely on a large amount of fault data and can continuously learn and iterate was established by integrating the advantages of measurement point optimization technology, expert knowledge, and deep learning. Finally the identification and diagnosis of common fault defects in pilot pumps were achieved successfully