Abstract: When a pumping well indicator diagram is diagnosed by traditional artificial neural networks, the model is limited by the synchronous instantaneous input. It cannot reflect the cumulative time effect for continuous input signal and has low diagnostic accuracy. Aiming at solving this problem, we propose an extreme learning discrete process neural network. Three-spline numerical integration is applied to deal with the aggregation of discrete samples and weights in the time-domain. An extreme leaning algorithm is applied to the model's training and converts it to a least squares problem. The Moore-Penrose generalized inverse matrix and a hidden layer output matrix are used to compute the output weight. The training speed is enhanced. When the model is used to diagnose five common statuses in the indicator diagram, the discrete time sequence data on the displacement and load are taken as the model input. The experimental results show that the method has higher identification accuracy and faster learning speed than other process neural network.