基于分布式降噪正交自编码器的工业过程故障检测

Industrial Process Fault Monitoring Based on Distributed Denoising Orthogonal Auto-encoder

针对基于过程知识进行模块划分的方法,存在知识不足造成模块划分不准确以及数据中的噪声导致的过拟合影响到故障检测率的问题,提出一种变量模块划分并采用降噪正交自编码器建立分布式故障检测的方法(MBIDOAE)。首先利用互信息和最短距离的层次聚类将相关关系强的过程变量聚类,获得多个模块。然后采用降噪正交自编码器提取各模块过程变量的非线性特征构建分布式模型,引入随机噪声来增强自编码器的抗噪性并利用正交矩阵降低特征的冗余性。采用T2和SPE统计量作为故障检测指标,通过对TE过程的仿真,与PCA、AE、MBIPCA和所提MBIDOAE进行对比,验证了该方法的有效性。

Considering the fact that the inaccurate module division caused by insufficient knowledge and the over-fitting caused by the noise in data which affecting fault detection rate in the method of module division based on process knowledge,a method of variable module division and using de-noising orthogonal auto-encoder(MBIDOAE)was proposed to inplement distributed fault detection.Firstly,the process variables with strong correlation were clustered by adopting hierarchical clustering of the mutual information and the shortest distance to obtain multiple modules;then,a de-noising orthogonal auto-encoder was used to extract nonlinear features of the process variables of each module to construct a distributed model,including having random noise introduced to enhance anti-noise ability of the auto-encoder and an orthogonal matrix employed to reduce redundancy of the features;finally,the T2 and SPE statistics were used as fault detection indicators,and through simulating TE process,comparing with PCA,AE,MBIPCA and MBIDOAE proposed,the effectiveness of this method was verified.