CNN-BiLSTM-AM based thermal error modeling of motorized spindle for precision CNC cylindrical grinding machine

Abstract

The motorized spindle is the core component of a CNC grinding machine. During the machining process, the working conditions are complex and variable. The uneven temperature distribution caused by internal and external heat sources in the motorized spindle leads to nonlinear and time-varying thermal deformation, particularly, axial thermal elongation can exceed several tens of micrometers, significantly compromising the grinder’s machining accuracy. To effectively implement thermal error compensation with high accuracy and cost-efficiency, a CNN-BiLSTM-AM based thermal error prediction model for the motorized spindle is proposed, which combines CNN for spatial feature extraction, BiLSTM for capturing long temporal dependencies, and an attention mechanism to enhance feature relevance. The model’s prediction performance and robustness were validated through thermal characteristic experiments under various working conditions. In the predictions for Tests 2 and 3, the maximum residuals between the predicted and measured axial thermal deformation were 1.326 μm and 1.495 μm, with R² values as high as 0.991 and 0.992, MAEs as low as 0.524 μm and 0.589 μm, and RMSEs of 0.596 μm and 0.691 μm, respectively. All metrics outperformed those of comparative methods. Furthermore, the thermal compensation experiments were performed to verify the model’s practical applicability.