基于深度置信网络的多模态过程故障评估方法及应用

A Deep Belief Network-based Fault Evaluation Method for Multimode Processes and Its Applications

传统的多模态过程故障等级评估方法对模态之间的共性特征考虑较少,导致当被评估模态故障信息不充分时,评估的准确性较低.针对此问题,首先,提出一种共性–个性深度置信网络(Common and specific deep belief network,CS-DBN),该网络充分利用深度置信网络(Deep belief network,DBN)的深度分层特征提取能力,通过度量多模态数据间分布的相似性和差异性,进一步得到能够反映多模态过程共有信息的共性特征以及反映每个模态独有信息的个性特征;其次,基于CS-DBN,利用多模态过程的已知故障等级数据生成多模态共性–个性特征集,通过加权逻辑回归构建故障等级评估模型;最后,将所提方法应用于带钢热连轧生产过程的故障等级评估中.应用结果表明,随着多模态故障等级数据的增加,所提方法的评估准确率逐渐增加,当故障信息充足时,评估准确率可达98.75%;故障信息不足时,与传统方法相比,评估准确率提升近10%.

Traditional fault grade evaluation methods for multimode processes have not well consider the common features embedded in multimode process data,which led to the low evaluation accuracy for cases where there lacks of fault grade data for the operating mode under evaluation.To solve this problem,firstly,this paper proposes a common and specific deep belief network(CS-DBN),which fully utilizes the hierarchical feature extraction ability of deep belief network(DBN)to automatically obtain the common features that reflect the common information of multimode operating processes by measuring the similarity and difference in the distribution of multimode operating data,and obtain the specific features reflecting the unique information of each operating mode;Secondly,on the basis of CS-DBN model,the known fault grade data are gathered to formulate a multimode common and specific feature database,and the weighted logical regression method is used to develop a fault grade evaluation model;Finally,the proposed method is applied to the fault grade evaluation problem in a hot rolling mill process.The application results show that,with the increasing amount of multimode fault grade data,the evaluation accuracy of the proposed method gradually increases.For cases that the fault information is sufficient,the evaluation accuracy can reach up to 98.75%;For cases that the fault information is less sufficient,the evaluation accuracy by the proposed method improves nearly 10%compared with traditional methods.