融合Attention和GRU的压缩机组故障预警技术研究

Fault warning technology for compressor unit based on integration of Attention and GRU

现有天然气压缩机组普遍采用事后维修、定时检修方式,这会造成压缩机组的失修或过修现象。针对这一问题,提出了一种融合Attention和GRU的压缩机组故障预警技术,先利用RF算法筛选影响压缩机组对应故障的参数,将形成的数据集代入GRU模型进行训练和预测,并在隐含层和全连接层中建立Attention机制,用于对单一时间步内的关键性数据赋权,最后基于现场实际值和模型预测值的残差均值,通过计算三角函数隶属度确定不同时刻的风险等级。结果表明,注意力机制对模型精度的影响最大,其次为GRU模型;风险隶属度可实现预测数据风险信息的可视化,进气过滤器压差故障的预警时间可提前133h,压缩机喘振故障的预警时间可提前204min。研究结果可为过程控制系统的大数据分析及压缩机组故障超前预警提供实际参考。

The existing natural gas compressor unit generally adopts maintenance after a fault and regular maintenance,which will cause disrepair or over-repair of the compressor unit.To solve this problem,a fault warning technology for compressor units based on the integration of Attention and gate recurrent unit(GRU)was proposed.The random forest(RF)algorithm was first used to screen parameters affecting the corresponding faults of the compressor unit,and the formed data set was input into the GRU model for training and prediction.An Attention mechanism was established in the hidden layer and the fully connected layer to assign weight to the key data in a single time step.Finally,based on the residual mean of the actual field value and the predicted value of the model,the risk level at different time was determined by calculating the membership degree of the trigonometric function.The results show that the Attention mechanism has the greatest influence on the accuracy of the model,followed by the GRU model.The risk membership degree can realize the visualization of the risk information of the predicted data.The early warning time of the pressure difference fault of the air intake filter can be 133 h in advance,and that of the surge fault of the compressor unit can be 204 min in advance.The research results can provide a practical reference for big data analysis of process control systems and early fault warnings of compressor units.