一种抑制残差污染的补偿型自联想神经网络  

A compensation auto-associative neural network for overcoming smearing effects

作　　者：任少君[1,2] 肖晋飞 司风琪[1,2] 曹越 陈家乐 Ren Shaojun;Xiao Jinfei;Si Fengqi;Cao Yue;Chen Jiale(Key Laboratory of Energy Thermal Conversion and Control of Ministry of Education, Southeast University, Nanjing 210096, China;School of Energy and Environment,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学能源热转换及其过程测控教育部重点实验室,南京210096 [2]东南大学能源与环境学院,南京210096

出　　处：《东南大学学报（自然科学版）》2020年第4期712-720,共9页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(51976031);中央高校基本科研业务费专项资金资助项目(2242019K40010);江苏高校“青蓝工程”资助项目;江苏高校优势学科建设工程资助项目。

摘　　要：针对常规自联想神经网络容易陷入残差污染而难以准确重构的问题,提出了一种新的补偿型自联想神经网络算法.该算法通过在常规自联想神经网络的测试过程中引入网络补偿层,建立了网络输入层与残差空间之间的调整机制,并采用梯度下降法快速获取目标变量的残差补偿幅值.给出了单变量和多变量残差补偿量计算流程,通过比较补偿后平方预测误差(SPE)统计量的大小来确定最佳补偿方向和补偿幅值,从而定位到异常点位置,并计算出模型重构值.通过仿真算例和工程算例验证了所提算法的有效性,算例结果表明,该算法能够在未知异常点位置的情况下,有效克服大幅度异常和多点并行异常造成的残差污染影响,其诊断和重构性能明显优于常规自联想神经网络和主成分分析算法.To overcome the smearing effect problem of the basic auto-associative neural network(AANN)that may reduce the reconstruction accuracy,a new compensation auto-associative neural network(CAANN)is proposed.In CAANN,a new compensation layer is introduced into the testing process of the basic AANN in order to develop an adjustment mechanism between the input layer and the residual space,and the residual compensation magnitude is calculated by the gradient descent algorithm.Then,a strategy of residual compensation for both single and multiple variable(s)is implemented to calculate the optimized compensation magnitude and direction by comparing the squared prediction error(SPE)after compensation,and further the source of anomalies is pinpointed and the reconstructed values are obtained.The effectiveness of the proposed method is evaluated on a validation example and an industrial example.The results demonstrate that the proposed CAANN can effectively inhibit the drawbacks of“smearing effects”in case of both large amplitude anomalies and multipoint concurrent anomalies without prior knowledge,showing a better performance than AANN and principal component analysis(PCA)methods for fault diagnosis and data reconstruction.

关 键 词：自联想神经网络 残差污染 残差补偿 故障诊断 

分 类 号：TP206.3[自动化与计算机技术—检测技术与自动化装置]