基于EKF-RNN算法的抗震棒材性能预报模型研究

Study on prediction models of anti-knock steel bar quality based on EKF-RNN algorithm

以抗震棒材的C、Si、Mn、P四种化学元素及冷却速度、终轧温度为输入;抗拉强度、断面收缩率作为输出建立了热轧抗震棒材力学性能预报模型。提出了一种基于扩展卡尔曼滤波的回归神经网络权值训练算法(EKF-RNN),对抗震棒材性能预报模型进行权值训练。并与随机梯度法训练回归神经网络权值的算法进行比较,仿真结果表明,随机梯度法存在局部极小值、收敛速度慢,而扩张卡尔曼滤波算法很好地解决了这些问题,并得到了比较满意的结果,更具优越性。

A prediction model for hot rolling anti-knock steel bar property has been established with 4 chemical elements（C,Si,Mn,P）,the cooling rate,the maximum temperature as input,and the tensile strength,the percentage of area reduction as output.Then an algorithm of recurrent neural network weight based on extended kalman filter trains（EKF-RNN）is proposed to carry out the weight training for model of predicting anti-knock steel bar property,which will be compared with the EKF-RNN algorithm trained by random gradient method.The simulation results show that the stochastic gradient method exists some problem such as local minima and slow convergence,while Extended Kalman filtering can solve it,which has obtained satisfied result.