基于多维状态参量和时间序列对电机的异常检测

针对传统对电机的异常检测常常出现误判或滞后的情况,采用基于时间序列对电机单一状态参量用AR拟合,同时利用SOM神经网络无监督的方式量化电机数据。然后,利用得到的量化序列结合AR曲线得到序列的转移概率,及早发现某种状态参量的异常变化；之后,DBSCAN算法挖掘多维参数之间特征关系来确定电机是否出现异常。最后结合实例说明该方法的检测过程,并对比验证了该方法的优越性。

Anomaly detection of electric machine based on multi-dimension state parameter and time series

In view of the traditional abnormal detection of the motor often appear in the case of miscarriage of justice or lag,this paper using AR fitting and SOM neural network unsupervised way to quantify the motor data based on the time series of the motor single state parameters.Then, the transition probability of the sequence is obtained by combining the quantitative sequence with the AR curve, and the abnormal changes of some state parameters are discovered as soon as possible.After that, DBSCAN algorithm mining multi-dimensional parameters of the relationship between the characteristics of the motor to determine whether there is abnormal.Finally, the detection process of the method is illustrated with an example, and the advantages of this method are verified by comparison.