采用多变量耦合网络与变分图自编码器的机械设备异常检测方法

Anomaly Detection Method with Multivariable Coupling Network and Variational Graph Autoencoder

为了全面准确地检测出潜在的设备异常,针对汽轮机、风电转子等高关联耦合分布式系统多测点多源传感器产生的多维多态监测数据,提出了一种基于多变量耦合网络与变分图自编码器的异常检测方法。首先采用去趋势互相关分析(DCCA),定量分析多维变量间的耦合关系,构建复杂系统多变量耦合关系网络。在此基础上,建立基于无监督学习的变分图自编码模型,对系统多变量耦合关系网络进行特征提取,使用正常数据训练该模型,图卷积网络作为编码器学习输入数据的分布,采样获得其潜在表示,以实现耦合网络的重构,采用重建概率作为系统多维多态监测数据异常检测评价指标。最后以某火电厂汽轮机组转子系统监测数据为例开展异常检测分析,结果表明:考虑多维多态监测数据间的耦合关系,提高了系统异常检测的准确性和可靠性;引入基于变分图自编码器的无监督学习方法,降低了经验依赖性并克服了异常样本少的问题。

To comprehensively and accurately detect potential equipment abnormalities,the multi-dimensional and multi-state monitoring data generated by multi-point and multi-source sensors in highly correlated and coupled distributed systems,such as steam turbines,turbines,and wind power rotors are taken into account,and an anomaly detection method with multivariable coupling network and variational graph autoencoder is proposed.The detrended cross-correlation analysis(DCCA)method is used to quantitatively analyze the coupling relationship between multi-dimensional variables,and then a system multivariable coupling relationship network is constructed.Then a variational graph autoencoder based on unsupervised learning is established,and feature extraction on the system multivariable coupling relationship network is performed.The normal data are chosen to train the model,the graph convolutional network is taken as the encoder to learn the distribution of input data,and the potential representation is obtained via sampling to realize the reconstruction of the coupled network.The reconstruction probability is used as the system anomaly detection index.Taking an example,the anomaly detection analysis is carried out with the monitoring data of turbine rotor system in a thermal power plant.It reveals that considering the coupling relationship between multi-dimensional polymorphic data,the accuracy and reliability of the system anomaly detection are improved;the unsupervised learning method with variational graph autoencoder is able to reduce the empirical dependence and solve the problem of insufficient abnormal samples.