Abstract: The data cleaning process is an important aspect of data preprocessing for online monitoring data of boiler equipment.Aiming at the problems of cumbersome data cleaning steps that easily lead to the destruction of continuous data,a data cleaning method of stacked denoising auto-encoder based on hybrid Adaptive Moment Estimation (Adam) and Stochastic Gradient Descent (SGD) algorithm optimization was proposed.A hybrid algorithm of Adam and SGD was introduced to continuously adjust the network parameters of the stacked denoising auto-encoder model.The model was utilized to train the normal status data to obtain the hidden features,and the reconstruction error of the data was obtained.The model was used to detect the abnormal data,and its reconstruction error was applied to analyze the influence of various types of data on the model.Thus the “dirty data” and the abnormal data reflecting the equipment failure were classified and repaired quickly.The cleaning and repairing experiments of the boiler monitoring data of a power plant proved that the proposed method could accurately identify the “dirty data”,and the repaired data could also follow the distribution law of the data,meet the data cleaning requirements,laying a good foundation for subsequent data analysis and mining and equipment fault diagnosis.

Key words: boiler equipment, online monitoring, data cleaning, deep learning, stacked denoising auto-encoder, feature extraction

摘要： 数据清洗过程是对锅炉设备在线监测数据预处理的一个重要环节,针对数据清洗步骤繁琐,易导致连续性数据被破坏等问题,提出一种基于混合自适应性矩估计和随机梯度下降算法优化的堆栈降噪自编码器的数据清洗方法。首先,引入自适应性矩阵估计和随机梯度下降的混合算法,以不断调整堆栈降噪自编码器模型的网络参数。其次,利用模型训练正常状态数据,获取数据的隐藏特征,得到正常状态下的重构误差。再次,用该模型检测异常状态数据,根据其重构误差分析各种类型的数据对模型的影响,并对“脏数据”和反映设备故障的异常数据进行快速分类清洗修复。通过某电厂锅炉监测数据的清洗修补实验,证明了该方法能准确识别出“脏数据”,修补后的数据亦能遵循数据整体的分布规律,满足了数据的清洗要求,为后续数据分析挖掘和设备故障诊断工作奠定了良好的基础。

关键词: 锅炉设备, 在线监测数据, 数据清洗, 深度学习, 堆栈降噪自编码器, 特征提取