过程控制时间序列中异常值的动态检测

针对传统小波异常值检测方法的不足以及控制调节系统在调节阶段采集的震荡数据所具有的特点, 提出了适用于调节系统震荡数据异常检测的自回归模型(auto-regression, AR)与小波相结合的在线异常值检测方法. 该方法通过引入改进的鲁棒AR模型, 克服了传统小波分析方法检测控制过程数据异常值时存在的不足; 为了避免传统异常值检测方法需要事先设定检测阈值的问题, 算法引入隐马尔科夫模型(hidden Markov model, HMM)来分析小波系数, 并在线更新HMM参数, 提高了算法的检测精度. 通过实验与应用证明了本文提出的异常数据检测方法更适合震荡的控制过程数据, 具有一定的实用性.

Dynamic outlier detection for process control time series

To improve the traditional outlier-detection by using wavelet analysis method and to deal with the instability characteristic of data from regulatory control process, we propose an improved outlier-detection method. This method combines an improved robust auto-regression (AR) model with the wavelet analysis method to eliminate the deficiency of the wavelet method in outlier-detection. To avoid the requirement of a pre-selected threshold value in the traditional method, we introduce the hidden Markov model (HMM) which analyzes the wavelet coefficients and updates online the coefficient values to improve the detection precision. Experiments and applications show that this method is especially suitable to oscillatory data in control processes.