Abstract: The Spectral Residual(SR)algorithm is originally proposed as an algorithm for image saliency detection,which can also be used for unsupervised anomaly detection of time series.The improvement of SR algorithm from frequency domain transformation,smoothing algorithm,removing the seasonal influence,and adjusting the abnormal judgment threshold were studied,and a multivariate time series anomaly detection method based on SR algorithm was proposed.Experiments showed that the improvement proposed in this paper could improve the accuracy of anomaly detection,remove the seasonal influence caused by environmental factors,and detect anomalies in a better time than the existing algorithms.In addition,the proposed algorithm could adaptively adjust the abnormal judgment threshold according to actual needs.To adapt to the multivariate time series data that often appear in industrial systems,the Independent Component Correlation Algorithm(ICA)for processing multivariate data was combined on the basis of the SR algorithm,so that the algorithm was suitable for multivariate time series.Experiments showed that the algorithm combining spectral residual algorithm and independent component analysis could be applied to automatic detection of anomalies in industrial systems,and could ensure the accuracy and real-time performance required by the algorithm.

Key words: time series anomaly detection, spectral residual algorithm, unsupervised algorithm, independent component correlation algorithm

摘要： 谱残差算法是一种针对图像显著性检测的算法,也可用于无监督时间序列的异常检测。从频域变换、平滑算法、去除季节性影响、阈值自适应调节等多个环节研究针对谱残差算法的改进,提出一种基于谱残差算法的多变量时间序列异常检测方法。实验证明,所提出的改进可以提高异常检测的准确率,去除环境因素造成的季节性影响,且检测异常用时优于已有算法。另外,所提算法可以根据实际需要,自适应调节异常判定阈值。为了适应工业系统常出现的多变量时间序列数据,在谱残差算法的基础上结合用于处理多变量数据的独立成分分析算法,使算法适用于多变量时间序列。实验表明,谱残差算法与独立成分分析结合的算法能够应用于工业系统的异常自动检测,并且可以保证算法所需的准确性和实时性。

关键词: 时间序列异常检测, 谱残差算法, 无监督算法, 独立成分分析