融合IMR-WGAN的时序数据修复方法  

Time Series Data Repairing Approach by Combing Iterative Minimum Repairing and Wasserstein Generating Adversarial Networks

作　　者：孟祥福[1] 马荣国 MENG Xiangfu;MA Rongguo(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105

出　　处：《小型微型计算机系统》2024年第3期641-650,共10页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61772249)资助;2021年度辽宁省教育厅科学研究经费项目(面上项目)(LJKZ0355)资助.

摘　　要：工业数据由于技术故障和人为因素通常导致数据异常,现有基于约束的方法因约束阈值设置的过于宽松或严格会导致修复错误,基于统计的方法因平滑修复机制导致对时间步长较远的异常值修复准确度较低.针对上述问题,提出了基于奖励机制的最小迭代修复和改进WGAN混合模型的时序数据修复方法.首先,在预处理阶段,保留异常数据,进行信息标注等处理,从而充分挖掘异常值与真实值之间的特征约束.其次,在噪声模块提出了近邻参数裁剪规则,用于修正最小迭代修复公式生成的噪声向量.将其传递至模拟分布模块的生成器中,同时设计了一个动态时间注意力网络层,用于提取时序特征权重并与门控循环单元串联组合捕捉不同步长的特征依赖,并引入递归多步预测原理共同提升模型的表达能力;在判别器中设计了Abnormal and Truth奖励机制和Weighted Mean Square Error损失函数共同反向优化生成器修复数据的细节和质量.最后,在公开数据集和真实数据集上的实验结果表明,该方法的修复准确度与模型稳定性显著优于现有方法.Industrial data is often subject to data anomalies due to technical failures and human factors.Existing constraint-based methods can cause repair errors because the setting of constraint thresholds is excessively loose or strict,and statistical-based methods have low repair accuracy for outliers with long time steps due to the smooth repair mechanisms.To address these problems,a time series data repair method combining minimum iterative repair and improved Wasserstein generative adversarial network(called for IMR_WGAN-GP)hybrid model with a reward punishment mechanism is proposed.Firstly,in the pre-processing stage,the anomalous data is retained and information annotation is carried out to fully mine the feature constraints between the anomalous value and the real value.Secondly,a nearest neighbour parameter clipping rule is proposed in the noise module to modify the noise vector generated by the minimum iterative repair formula.It is transferred to the generator of the simulated distribution module,while a dynamic temporal attention network layer is designed to extract temporal feature weights and combine with gated recurrent units in series to capture feature dependencies of different steps,and the recursive multi-step prediction principle is introduced to jointly enhance the expressive ability of the model;the Abnormal and Truth reward mechanism and Weighted Mean Square Error loss function is designed in the discriminator to jointly reverse optimize the detail and quality of the generator repair data.Finally,experimental results on both public and real datasets show that the repair accuracy and model stability of the method are significantly better than those of existing methods.

关 键 词：数据修复 改进Wasserstein生成对抗网络 Abnormal and Truth奖励机制 动态时间注意力机制 Weighted Mean Square Error损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]