It proposes a fault diagnosis method of multi-stage reciprocating compressor valve based on LM (Levenberg-Marquardt) algorithm to optimize the BP (Back Propagation) neural network. Six-level pressure differences and six-stage temperature differences of 6M25-185/314 hydrogen nitrogen compressor regarded as the input vector of the network, to establish the LM-BP neural network model which can be used in online monitoring and fault diagnosis of the one-to-six level valve fault of the reciprocating compressor. 100 groups of fault data as the network training samples and 30 sets of data as the network detection samples for fault diagnosis, the results show that, compared to the variable gradient BP neural network and RBF neural network, LM-BP neural network is more rapid and more stable and the accuracy rate of diagnosis reaches above 96%. Built by using the Matlab software platform, fault diagnosis model of LM-BP neural network is simple and can be easily used in engineering practice.