基于DAE和TCN的复杂工业过程故障预测  

Fault prediction of complex industrial process based on DAE and TCN

作　　者：高学金 马东阳[1,2,3,4] 韩华云 高慧慧 Gao Xuejin;Ma Dongyang;Han Huayun;Gao Huihui(Department of Information,Beijing University of Technology,Beijing 100124,China;Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China;Beijing Laboratory for Urban Mass Transit,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]数字社区教育部工程研究中心,北京100124 [3]城市轨道交通北京实验室,北京100124 [4]计算智能与智能系统北京重点实验室,北京100124

出　　处：《仪器仪表学报》2021年第6期140-151,共12页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61803005,61640312,61763037);北京市自然科学基金(4192011,4172007);山东省重点研发计划(2018CXGC0608);北京市教育委员会项目资助。

摘　　要：为实时监测复杂工业过程的故障状态,精确预测故障趋势,提出基于降噪自编码和时间卷积网络的故障预测方法。首先,利用随机森林算法筛选故障相关特征。之后,利用堆栈降噪自编码网络提取非线性特征以及特征重构,并根据重构误差构造平方预测误差(SPE)统计量作为故障状态特征。最后,针对时间卷积网络残差模块中的ReLU激活函数在负区间内导数为零导致部分神经元无法被激活的问题,设计基于自门控激活函数(Swish)和滤波器响应(FRN)规范化的时间卷积网络(SFTCN)。将得到的SPE组成时间序列,利用SFTCN的预测模型实现其状态趋势预测。通过在TE仿真平台数据和美国密歇根大学智能维修中心实测的轴承全生命数据上的实验表明,与未改进的时间卷积网络对比,所提方法的预测平均绝对百分比误差至少降低20.9%,具有较高的应用价值。In order to monitor the state of complex industrial process in real time and predict the fault trend accurately,this paper presents a fault prediction method based on denoising auto encoder(DAE)and temporal convolutional network(TCN).Firstly,the random forest algorithm is used to filter out the features related to faults.Then,the nonlinear features of input data are extracted and the original features of input data are reconstructed,and the squared prediction error(SPE)statistics is established based on the reconstruction error to reflect the state characteristics of the faults.Finally,considering that the derivative of ReLU activation function in the residual module of TCN is zero in the negative interval,which may cause certain neurons to fail to activate,a Swish activation function and filter response normalization-based temporal convolutional network(SFTCN)is proposed.By constructing the obtained SPE into time series,the SPE prediction can be realized based on the SFTCN.Experiments are conducted with the data of Tennessee Eastman(TE)process and the life-cycle vibration data of rolling bearings measured by the center for intelligent maintenance systems of the University of Michigan.Results show that compared with the unmodified TCN,the average absolute percentage error of the proposed method is reduced by at least 20.9%,which has high application value.

关 键 词：故障预测 随机森林算法 降噪自动编码 时间卷积网络 复杂工业过程 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置] TH17[自动化与计算机技术—控制科学与工程]