基于经验耦合函数的动态运行参数异常检测

Empirical Copula-based Anomaly Detection of Dynamic Runtime Parameters

随着工业智能化的发展,工业生产系统中的工业设备都具备了智能管控系统,其中重要需求之一是智能异常检测。实现智能异常检测通常需要从动态运行参数入手,但动态运行参数的流式数据形式以及高维数据耦合给可靠、高效的异常检测带来了很大困难。为此,提出了一种基于联合分布的动态运行参数异常检测方法。该方法首先从实时检测和整体检测两个角度对动态运行数据进行采样,然后结合经验耦合函数对联合分布进行建模,最后根据模型得到异常分数来判断异常。通过在大渡河流域水电站排水系统的排水泵动态运行参数数据集上的实验验证表明,该方法相比传统的异常检测方法效率更高,并且在曲线下的面积(area under curve,AUC)和平均精确率上均有提升。同时,该方法的可解释性也为工作人员故障排除以及后续维护提供了可靠依据。

As a result of the development of industrial intelligence,intelligent control systems have been equipped in industrial production systems,and one of the important requirements is intelligent anomaly detection.However,achieving intelligent anomaly detection is often difficult due to the streaming data format of dynamic runtime parameters and the coupling of high-dimensional data,which make it challenging to perform reliable and efficient anomaly detection.To this end,this paper proposes an anomaly detection method based on joint distribution.Firstly,the dynamic operating data was sampled from both real-time detection and overall detection perspectives.Then,the joint distribution was modeled using empirical Copula.Finally,the anomaly score was determined based on the model.The dataset of the drainage system in a hydropower station on Dadu River was used to investigate the effectiveness of the proposed method,and the results show that this method is more efficient than traditional anomaly detection methods and improves area under curve(AUC)value and average precision.Meanwhile,the interpretability of this method also provides a reliable basis for subsequent troubleshooting and maintenance of the drainage system.