Anomaly detection algorithm for rolling time series signal based on time series prediction
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
万方数据
维普
轧钢厂传统的时序生产异常检测通常依靠操作工经验判断或通过一些离线数据分析方法进行事后判断.实际生产环境中,依靠这类方法不能确保实时、准确的异常检测.为了应对这些挑战,提出了一种基于时序预测和序列相似性比对的轧钢时序信号异常检测算法.首先利用基于果蝇优化算法(fruit fly optimization algorithm,FOA)优化的Holt-Winters模型对当前轧制钢材的时序数据进行预测,然后利用动态时间规整算法(dynamic time warping,DTW)计算预测值与真实值的距离,最后根据距离值判断是否发生异常.利用某轧钢厂的轧机电流数据对所提算法进行了测试,结果表明,该算法预测部分的平均绝对百分比误差EMAP(mean ab-solute percentage error,MAPE)为1.73％,低于对比算法;此外,在对2天的生产数据进行检测时,本文所提方法成功检测出了电流数据中包含的2次片段异常.该方法有助于轧钢企业及时发现钢铁质量问题和设备故障问题,对规避批量产品质量问题和尽早发现设备异常起到重要作用.
Traditional detecting abnormalities in time series production in rolling mills usually rely on operator's experience or post judgment through some off-line data analysis methods.In the actual pro-duction environment,relying on such methods can not ensure real-time and accurate anomaly detec-tion.To meet these challenges,an algorithm for detecting abnormalities in rolling time series signals based on time series prediction and sequence similarity comparison was presented.The proposed al-gorithm firstly uses Holt-Winters model optimized by fruit fly optimization algorithm to predict the time series data of the current rolled steel,then calculates the distance between the predicted value and the true value by using the dynamic time warping algorithm,and finally judges whether an anom-aly occurs based on the distance value.The proposed algorithm was tested by using the current data of the rolling mill in a iron and steel rolling plant.The results show that the EMAP(mean absolute per-centage error,MAPE)of the prediction part of the algorithm is 1.73％,which is lower than the com-parison algorithm.In addition,the method proposed in this paper successfully detected two fragment anomalies contained in current data when detecting 2-day production data.This method is helpful for steel rolling enterprises to find steel quality problems and equipment failures in time,and plays an important role in avoiding batch product quality problems and finding equipment abnormalities as soon as possible.
steel rollingfruit fly optimization algorithmHolt-Winters modeldynamic time warpingtime series anomaly detection
NETL
NSTL
万方数据
维普
