Abstract: This study proposes a rapid generalization method for spindle thermal deformation models based on transfer learning, aiming to improve the efficiency of batch implementation of machine tool thermal error compensation. First, spindle thermal characteristic experiments were designed, and a deep neural network model for spindle thermal deformation was established using the GRU (Gated Recurrent Unit) neural network. This network was trained with spindle speed and temperature data, achieving good prediction accuracy for spindle thermal deformation. Then, to address the generalization problem of traditional data-driven models, a rapid generalization method based on transfer learning was proposed. By utilizing transfer learning, common features of the spindle thermal deformation model were retained, and only the individual differences were corrected, thus reducing the amount of condition data required for remodelling. The experiments showed that, compared to traditional remodelling, transfer learning improved prediction accuracy by 38.6% with the same training dataset and reduced the training data by 66.7%, significantly lowering the time cost of thermal error compensation. This method has strong engineering application potential, as it effectively shortens the machine tool thermal characteristic testing period and enhances the generalization ability of the model.