Sucker rod pumping is a dominated artificial lift method for oil production engineering. In the production process, diagnosing the condition of pump using dynamometer card is vitally significant to monitor and manage the pumping system. With the ability to reflect arbitrary non-linear mappings, the BP neural network can be used in pattern recognition of the pump dynamometer card. In this paper, some key technologies of establishing reasonable neural network are introduced. The number of neurons in input layer depends on the selection of characteristic value. The number of neurons in hidden layer can be obtained by some models, optimum value will be chosen out. The number of neurons in output layer depends on the recognized behavior of pump. After the construction of neural network, the more effective and practical BP neural network will be obtained by suitable samples and appropriate training strategies.