基于双通道回声状态网络的时间序列补全及单步预测

Time Series Completion and One-step Prediction Based on Two-channel Echo State Network

随着物联网的发展,众多传感器采集到大量具有丰富数据相关性的时间序列,为各种数据挖掘应用提供强大的数据支持。然而,一些客观或主观原因(如设备故障、稀疏感知等)往往会造成采集到的数据出现不同程度的缺失。虽然已有很多方法被提出用于解决这一问题,但这些方法在数据相关性方面或考虑不够全面,或计算成本过高。而且,现有方法仅关注对缺失值的补全,未能兼顾下游应用。针对上述不足,设计了一种兼顾补全与预测任务的双通道回声状态网络。两个通道的网络虽共用输入层,但具有各自的储备池和输出层。两者最大的区别是左/右通道的输出层分别表示输入层前/后一个时刻对应的目标值或预补值。最后将两个通道的估计值进行融合,充分利用来自缺失时刻之前和之后的数据相关性以进一步提升性能。两种缺失现象下(随机缺失和分段缺失)不同缺失率的实验结果表明,所提模型无论是在补全精度还是预测精度上都优于目前流行的各类方法。

With the development of the Internet of Things,numerous sensors can collect a large number of time series with rich data correlation,providing powerful data support for various data mining applications.However,some objective or subjective reasons(such as equipment failure,sparse sensing)often lead to the loss of collected data to varying degrees.Although many approaches have been proposed to solve this problem,data correlation is either not fully considered or computationally expensive.In addition,existing methods only focus on the completion of missing values,and fail to take into account downstream applications.Aiming at the above shortcomings,this paper designs a two-channel echo state network to achieve both the completion task and the prediction task.Although the two channels share the input layer,they have their own reservoir and output layer.The biggest difference between them is that the output layer of the left/right channels respectively represents the target value or prefilled va-lue corresponding to the moment before/after the input layer.Finally,by fusing the estimates of the two channels,the data correlation from before and after the missing moments is fully utilized to further improve performance.Experimental results of diffe-rent missing rates with two missing mechanisms(random missing and piecewise missing)show that the proposed model is superior to the current methods in both completion accuracy and prediction accuracy.