钢材性能与工艺参数的交叉预测模型研究

Research on cross prediction model of steel performance and technological parameters

为了克服BP神经网络算法的固有缺陷,增强其在钢材生产实践和事故调查中应用实效,提出了对BP神经网络算法的改进措施,并通过排列组合建立了钢材性能、温控工艺和化学成分添加工艺三组数据的6种交叉预测模型。该方法将遗传算法与小生境技术和禁忌算法相结合,优化了BP神经网络初始权值和阈值的选取,避免陷入局部最优解;利用Cauchy误差估计器代替传统均方差估计器,改进了BP神经网络误差统计方法,降低少数异常输入元素对结果的影响。实验结果表明,所有模型训练后检验精度均较高,总体平均相对误差较小,表明了该算法模型具备良好的预测能力和现实可行性。

To overcome the inherent defects of BP neural network algorithm, enhance its application of quality ior steel produc lion practice and accident investigation, an improvement actions to BP neural network algorithm is proposed. Through the permutation and combination, six cross prediction models of three sets of data including steel performance, temperature control technology and chemical composition addition technology, are built. This method combined niche technology and tabu search with genetic algorithm （OA）, the selection of BP neutral network initial weights and threshold values is optimized and local opti- mum is avoided. Cauchy error estimator instead of traditional mean-square deviation estimator is used to improve error statistical method of BP neutral network, and the influence of abnormal input elements on results is reduced. Results show that all the model test precision is higher after training, the overall average relative error is smaller, good prediction ability and practical feasibility of the presented method is demonstrated.