摘要
火电厂分散控制系统(DCS)监测数据中往往夹杂着大量各类噪声信号,复杂工况的影响使得监测数据时间序列具有混沌特性。为满足后续数据应用时对其准确性和有效性的需求,需要对原始采集数据做降噪处理。混沌时间序列通常在相空间中处理,结合流形学习策略进行降维,使得位于高维空间中的噪声信号被剔除,仅保留低维空间上的有用信号,达到降噪目的。在改进局部保持投影基础上,采用余弦距离推导欧拉表示代替欧氏距离,并在投影时加入正交条件,旨在保留原始数据流中的非线性特性并解决邻域内投影过密集问题。对该算法进行仿真并在磨煤机状态监测数据清洗中进行应用,结果表明:相较于小波降噪及局部保持投影,该降噪方法能较好地修复相空间整体流形结构,使其更清晰平整光滑,在过滤掉高频噪声的同时更多地保留有用信号及其非线性特性;用该算法做数据清洗使得预测模型准确性更高,运算速度更快。
During the daily work of thermal power plants,DCS monitoring signal is usually mixed up with massive varieties of noise interference inevitably.Due to the complex working condition,time series signal has a feature of chaotic.In order to meet the requirements of accuracy and validity for data application,it is necessary to denoise the original collected data.Chaotic time series are generally processed in phase space,manifold learning is applied to carry out dimension reduction,thus to eliminate the noise in high dimension and preserves the desired signal in low dimension for denoising.Based on the research of LLP,the authors define cosine distance in Euler representation instead of Euclidean distance along with orthogonal constraint in process of projection to preserve the nonlinear characteristic of the original manifold and deal with concentrated projection of adjacent vector.The algorithm is simulated and applied in monitoring data cleaning of coal mill,and the results show that,compared with wavelet noise reduction and LLP,this noise reduction method can better repair the overall manifold structure of the phase space,make it clearer,more smooth,and retain more useful signals and their nonlinear characteristics while filtering out high-frequency noise.This algorithm makes the prediction model more accurate and faster in computation.
作者
陈宁
何新
吴智群
CHEN Ning;HE Xin;WU Zhiqun(Xi’an Thermal Research Institute Co.,Ltd.,Xi’an 710054,China)
出处
《热力发电》
CAS
CSCD
北大核心
2022年第7期95-102,共8页
Thermal Power Generation
基金
中国华能集团有限公司总部重大科技项目(HNKJ20-H74)。
关键词
相空间重构
局部保持投影
正交投影
欧拉表示
降噪
phase space reconstruction
locality preserving projection
orthogonal projection
Euler representation
noise reduction
