摘要
提出了一种使用稀疏编码对机械频谱信号自学习并识别故障的方法。首先分别对每类频谱信号进行字典学习得到每类信号的字典,然后依次计算测试样本在各个类别的字典上的稀疏重构系数,利用稀疏重构系数与对应类别的字典重构测试样本。最后将重构残差作为识别依据,对机器状态进行判断。通过将振动信号从时域转化到频域,将复杂的移不变稀疏编码问题转化为普通的稀疏编码,并且得益于高效的K-SVD字典学习算法,计算效率得到了大幅提高。所提方案直接使用原始频谱信号作为训练集,不仅省去了特征提取过程,而且保留了更丰富的信息。经实验验证,该方案较基于时域的移不变稀疏编码具有更高的计算效率、准确率和稳定性。相对于常规诊断算法,除了有准确率的优势外,不易受负载变化的影响,所需人工干预较少。
An automatic learning and recognition scheme using sparse coding based on freqency domain signals was proposed here.Firstly,each dictionary for per class of frequency domain signals was obtained with a dictionary learning algorithm.Later,the test samples were sparselyly represented,respectively using the dictionaries of each class to calculate corresponding sparse reconstruction coefficients.Afterwards,the dictionaries with corresponding coefficients of the same class were applied to reconstruct the test samples.Finally,the reconstructed residual was taken as the criterion to determine machine states.Through converting vibration signals in time domain into those in frequency domain,a complex shift-invariant sparse coding problem,was simplified as an ordinary sparse coding one,and with the help of the effective K-SVD algorithm,the whole efficiency was further significantly improved.The original spectra singals were directly used as training samples in the proposed scheme,so that the complicated feature extraction was not needed,and more information was reserved.The test verification showed that the proposed technique improves greatly the efficiency and robustness compared to the shift invariant sparse coding in time-domain;compared with the traditional algorithms,besides the advantage of accuracy,this proposed scheme needs less cost and is affected less by load variation.
出处
《振动与冲击》
EI
CSCD
北大核心
2015年第21期59-64,共6页
Journal of Vibration and Shock
基金
国家自然科学基金项目(61472444
61472392)
关键词
故障诊断
特征提取
稀疏编码
K-SVD
字典学习
K-SVD
fault diagnosis
feature extraction
sparse coding
K-SVD
dictionary learning
