The pump card diagnostic recognition based on neural networks

摘要

Beam pumping is the most frequently used artificial lift technique. Down hole pump cards are used to evaluate performance of the pumping unit. Pump cards can be generated from surface dynamometer cards using one-dimensional wave equation which viscous damping. This paper describes an optimized multilayer feedforward network specialized in pattern recognition and expert reasoning. The network is trained to perform pump cards pattern recognition and crafted to simulate the expert decision-making. The system has been able to correctly identify problems in over 100 different training and test pump cards. The neural network requires total of 128 data points as input. 100 significant points are colected from the pump card perimeter itself and 26 data points represent the slope at selected points on the pump card perimeter and the remaining two data points describe the maximum and minimum load values.