Thermal induced error accounts for the most proportion of the total ma chining error.In order to improve the machining accuracy of machine tools,thermal error compensation is considered as the most economic way to reduce the effect of thermal error.However,as lacking of robustness evaluation methods for thermal error compensation models,it is difficult for a user to choose a suitable model for machine tools.In the actual use of models,abnor mal temperature data,which is inevitable for a long time monitoring,would result in the invalidation of the compensation models.In this paper,an esti mation mechanism as well as the performance indexes are designed for com pensation models evaluation and two classic compensation models based on multiple linear regression and artificial neural network are analyzed based on the estimation mechanism.Then the auto-regression based method is proposed to detective the abnormal temperature data and enhances the robustness of the compensation model.Finally,a set of experiments are carried out on a heavy duty machine tool to verify the feasibility and effectiveness of such methods.Research architecture is shown in Figure 1.