基于深度学习的时间序列数据异常检测方法

针对类间分布不平衡的时间序列数据的异常检测问题，提出了一种基于深度卷积神经网络的检测方法.首先采用抽样法对不平衡时间序列数据进行预处理；其次，将处理后的时间序列数据转换为尺度一致、时长一致的片段；最后将数据送入具有4层隐藏层结构的卷积神经网络模型中进行异常检测.实验结果表明，所提方法弥补了现存的检测技术由于忽略数据分布的偏斜性而造成的少数类检测精度低的缺点，并通过与现有的时间序列分类方法的比较，验证了所提方法的高效性.

Time-series Data Anomaly Detection Method Based on Deep Learning

With regard to the anomaly detection problem of time-series data with a skewed between-class distribution,we propose a detection method based on deep convolutional neural network.First,we employ the sampling method to preprocess the unbalanced time-series data.Second,the original time-series data are converted into a series of continuous segments with a uniform scale and consistent duration.Finally,we feed the data into a convolutional neural network model with four hidden layers for anomaly detection.The experimental results show that the proposed method covers the shortage of existing detection technologies that ignore the skewness of data distributions and results in a low-detection precision.Compared with the existing time-series classification methods,the proposed method provided a satisfactory performance.