基于统计差分LPP的多模态间歇过程故障检测

Fault detection of multi-mode batch process based on statistics difference LPP

针对工业过程数据存在的非高斯和多模态特性,提出一种基于统计差分LPP的多模态间歇过程故障检测方法。该方法将统计模量分析的方法应用到间歇过程训练数据集中,计算统计过程变量的均值和方差,将不等长的批次变成等长的统计量,保证统计模量近似服从高斯分布;然后运用差分算法使多模态变为单模态;最后运用LPP算法进行降维和特征提取,计算样本的T^2统计量,并利用核密度估计确定控制限。对于新来的测试样本数据统计差分处理后,向LPP模型上进行投影,计算新数据的T^2统计量并与控制限比较进行故障检测。最后通过半导体过程数据的仿真结果表明,该算法的故障检测效果最好,验证了所提方法的有效性。

Aiming at non-Gaussian and multi-mode characteristics existed in industrial process data,this paper proposed a fault detection of multi-model batch process method based on statistics difference LPP. Firstly,it applied statistical pattern analysis to the batch process training data set to calculate the mean and variance of statistical process variables,and turned the uneven-length batches into equal-length statistics. It could ensure that the statistics pattern approximately obeyed the Gaussian distribution. Then it used the difference algorithm to transform the multi-mode into single mode. Finally,it used the LPP algorithm to reduce dimension and extract feature,and calculated the T^2 statistic of the sample. And it used the kernel density estimation to determine the control limit. This paper projected the new test sample data onto the LPP model after statistics difference processing,and calculated the T^2 statistics of the new data and compared them with the control limit for fault detection.Finally,the simulation results of the semiconductor process data show that this algorithm has the best fault detection effect,and demonstrates the effectiveness of the proposed algorithm.