To solve the problem of thermal error of heavy-duty CNC machine tools, a thermal error separating and modeling approach based on linear regression theory is proposed. Firstly, the thermal errors are separated from geometric error to get thermal error parameters. Secondly, based on temperature sensors optimizing distribution strategies, like main factor strategy and irrelevance strategy, the key points of thermal errors are selected. Finally, a multiple linear regression model is constructed for thermal errors, and an experiment is carried out on a TK6920 heavy-duty CNC floor-type boring and milling machine tool, and the results indicate that the thermal error of column is reduced from 0.45 mm and 0.25 mm to 0.13 mm and 0.09 mm in X and W directions, that is 71％ and 64％ of thermal error can be compensated with this model. Error compensation technology can reduce the thermal error of heavy-duty CNC machine tools, therefore the manufacturing precision can be improved.