In model-based parametric fault diagnosis,due to the uncertainty of parameters,the nominal values of components are offset,and the frequent occurrence of error alarms will reduce the robustness of the residual evaluation method of conventional adaptive threshold.Aiming at the above problems,a method to enhance the performance of residual decision in fault detection stage was introduced.Based on the bond graph (BG) theory,the global analytic redundancy relation was used for fault detection and isolation in grid-connected inverters,and the hybrid diagnosis bond graph of linear fractional transformation was built for parametric fault diagnosis.The uncertain part could be decoupled from the model.The error alarm phenomenon in residual evaluation was reduced and its robustness was compared with incremental bond graph methods.In 20-sim software,the effectiveness of this strategy is verified by using NPC inverter as an example.