异常值的检测及其对棉纱强力预测精度的影响

Outlier Detecting and Its Impact on Accuracy of Cotton Yarn Strength Prediction

异常值的存在会不同程度地影响BP神经网络对棉纱强力指标的预测精度,因此对原数据进行异常值检测及对检出异常值的处理是非常关键的.采用系统聚类中的k最近邻密度估计方法对数据样本中存在的异常值进行检测,使用多元回归分析方法对检出的异常值进行修正,对修正前后共4组棉纱强力数据样本进行预测,得到各自的预测精度,经过对比分析,发现修正样本的相对误差和均方误差均明显低于其他3组包含异常值的样本,且异常值的数量越多则预测精度越低.

With the existence of outlier, the accuracy of cotton yarn strength predicted by BP neural network was affected in different extent. Therefore, it was very important to find out outliers from raw data and decide which method for treatment to be used properly, k-nearest neighbor density estimation method with pertaining to hierarchy clustering was adopted for outlier detection, multiple regression analysis was recommended to revise the outlier. And then, based on all 4 data sets revised before and after, cotton yarn strength was predicted respectively. By means of comparative analysis, it is found that the relative error and mean square error of revised data set are all less than the other 3 data sets with outlier inside, furthermore, the more the quantity of outlier is, the lower the accuracy will be.