Abstract: A quadratic diagonal recurrent neural network(QD- RNN) is presented. The hidden layer of QDRNN is comprised of quadratic self-recurrent neuron, so the QDRNN is dynamic mapping. Two QDRNNs are utilized in a ship-gun control system, one is an identifier called quadratic diagonal recurrent neuroidentifier(QDRNI) and the other as a controller called quadratic diagonal recurrent neurocontroller(QDRNC). The hydraulic servo system of ship-gun is identified by the QDRNI which then provides the sensitivity information of the hydraulic servo system to the QDRNC. A generalized dynamic backprop- agation algorithm is developed and used to train both QDRNC and QDRNI. Due to the quadratic recurrence, the QDRNN can capture the dynamic behavior of the servo system. The test results for the hydraulic servo system indicate that the control precision of QDRNN is high and the adaptability is strong.

Key words: Recurrent neural network, Test study, Warship-guns