基于层级实时记忆算法的时间序列异常检测算法

时间序列异常检测是数据分析中一个重要的研究领域.传统的时间序列的异常检测方法主要通过比较检测数据和历史数据的差异程度，以判断被检测数据是否为奇异点（Surprise）、离群（Outlier）点等.然而序列和窗口的划分，状态的划分或者异常的定义和判定等问题，使得这类方法存在一定的局限性.本文针对传统时间序列检测算法不足，提出一种基于层级实时记忆算法的时间序列异常检测算法.该方法对时间序列内在模式关系进行学习，建立预测模型，通过比较预测值和真实值的偏离程度来判断数据是否异常.首先使用稀疏离散表征在保证保留数据相关性的同时又将数据离散化；然后输入到模型网络，预测下一时刻的数据值；最终根据预测值和真实值的差异为数据异常程度进行定量评分.在人造数据和真实数据上的实验表明，该方法能够准确、快速地发掘时间序列中的异常.

Time Series Anomaly Detection Model Based on Hierarchical Temporal Memory

Time series anomaly detection is an important area of data mining.Traditional methods of time series anomaly detection usually find the surprise,outlier,etc.,by comparing the data with the historical data.However,there are some limits with these methods,such as the inaccurate separation of the sequence,the false decision of the state and the window size or the incorrect definition and judgement of the anomaly.This paper proposes a time series anomaly detection model based on hierarchical temporal memory (HTM) to overcome the shortages of the traditional methods.This method can recognize and learn the intrinsic patterns in the time series and build a prediction model to determine an anomaly by comparing the real value with the predicted one.First,sparse distributed representation (SDR) is used to represent the raw data; then,the SDR is entered into the HTM model to make prediction;lastly,the proposed model evaluates the data by computing the difference of the actual value and the predicted one.The experiments on the artificial data and the real data show that HTM can detect anomalies accurately and quickly.