改进因果推断方法在发电机组辅机状态监测中的应用研究

Research of Improved Causal Inference Method Applied in Condition Monitoring System of Generator Set Auxiliary Equipment

数据分析技术是建设智慧电厂的核心要素,目前的大多数数据分析技术是利用数据间的强相关关系,因果推断技术的发展为利用两个强相关变量中呈现的因果不对称性发现可靠因果关系提供了有力工具。针对因果推断技术用于实际设备中模型难以选择、模型有效性和准确率无法判断的问题,本文提出一种基于因果函数的改进因果推断方法,并以一种新的角度使用因果效应强度变化监测系统运行状态。首先根据专家知识设置强、非因果关系,引入因果关系保持率CRR、因果信息不对称度DDA,改进因果关系模型的选择和评估过程。最后通过滑动窗口结合倾向得分匹配计算各方向因果效应强度用以监测系统状态。在某一次风机的系统算例中验证表明,改进因果推断方法能有效验证不同因果模型的有效性和准确度,优化因果发现过程,结合所选LiNGAM模型构建了合理可信的因果关系网络。根据因果关系网络通过因果效应强度的变化成功跟踪表征了系统状态,并发现了原始数据中无法发现的系统状态恶化节点。

As the core element of building smart power plant.At present,most data analysis technologies use the strong correlation between data.The development of causal inference technology provides a powerful tool to find reliable causal relationship by using the causal asymmetry presented in two strongly correlated variables.Aiming at the problems that it is difficult to select the model and judge the validity and accuracy of the model when the causal inference technology is used in practical equipment.An improved causal inference method based on causal function is proposed in this paper,and the operation state of the system is monitored by the change of causal effect intensity from a new perspective.Firstly,strong and non-causal relationships are set according to expert knowledge,and the causal relationship retention rate CRR and causal information asymmetry DDA are introduced to improve the selection and evaluation process of causal relationship model.Finally,the strength of causal effect in each direction is calculated by sliding window combined with propensity score matching to monitor the state of the system.In a system example of a primary air fan,the verification shows that the improved causal inference method can effectively verify the effectiveness and accuracy of different causal models,optimize the causal discovery process,build a reasonable and reliable causal network combined with the selected LiNGAM model,successfully track the system state through the change of causal effect intensity according to the causal network,and find the system state deterioration nodes that cannot be found in the original data.