近红外光谱煤质分析模型中异常样品的剔除方法

Rejecting Method of Abnormal Samples in Analysis Model of Coal Quality of Near-infrared Spectrum

针对建立近红外光谱煤质定量分析模型时训练集中的异常样品严重影响模型预测精度的问题,提出一种二次诊断法剔除异常样品:利用模糊C均值聚类法对样品进行聚类,得到可疑样品;将可疑样品作为验证集,通过PCA-GA-BP模型进行二次诊断,剔除异常样品。实验对比了训练集中异常样品剔除前后,模型对15组待测样品的预测能力,结果表明该方法能够准确剔除异常样品,并有效提高模型的预测精度。

In building of quantitative analysis model of coal quality of near-infrared spectrum,abnormal samples in training set seriously influence forecast precision of the model.So the paper proposed a new method of rejecting abnormal samples by twice diagnoses： get suspicious samples by clustering training samples based on fuzzy C-mean algorithm,and taking suspicious samples as validation set,reject abnormal samples by secondary diagnosis through PCA-GA-BP model.An experiment was done to compare predicting ability of the model with and without abnormal samples in training set by 15 groups of samples,and the result showed that the method can reject abnormal samples accurately,and improve forecasting precision effectively.