基于归纳学习的结构损伤识别方法研究

Detection of Structural Damage by Inductive Learning Methods

采用归纳学习方法来识别结构损伤. 首先,通过对经典的决策树算法和序列覆盖算法进行结合与改进,得到一种高效且代价又小的归纳学习算法(RAC),同时引入装袋算法产生多个分类法,并用它们进行类预测,而且使用选票策略得出最佳类预测.其次,用正交最小二乘迭代算法作为径向基函数(RBF)神经网络的学习方法,通过“信息 贡献”准则进行正交变换来优选中心.最后,对上述归纳学习方法用于梁结构损伤定位的效果进行了实验评估.结果表明,对于 RAC算法和生成分类法的数目分别为 10 和 50 情况下的装袋算法,当损伤样本被噪声污染程度在100% 时,识别精度均可达到 90% 以上,而对于 RBF神经网络算法, 只有当损伤样本被噪声污染程度小于70% 时,识别精度才可达到90%以上.

Inductive learning methods were used for the detection of structural damage. Firstly, a inductive learning method (RAC) that is more efficient and effective, was developed by combining the basic decision tree learning algorithm with sequential covering algorithm and then improved; bagging method was used to generate several classification methods for predicting. Then the radial basis function (RBF) neural network was trained by orthogonal least squares (OLS) algorithm which selects the radial basis function centers by using information-contribution rule. At the end, the inductive learning methods mentioned above were used for the location detection of a beam structural damage. The detecting results show that the precision can be more than 90% for both RAC and bagging methods with the ensemble sizes of 10 or 50 when the noise level of test samples is within 100%. And the same identify precision can be kept for RBF neural network only if noise level of test samples is within 70%.