Turbine blades are characterized by their intricate geometries， specialized materials， and complex manufacturing processes. Three-dimensional models constructed through point cloud alignment serve as critical tools for the comparative analysis and evaluation of blade manufacturing accuracy and quality. However， accurately identifying overlapping regions within point clouds during the alignment process presents significant challenges， compounded by complex calculations. To address these issues， a novel fast point cloud alignment method for complex hollow turbine blades was developed. This method extracted multi-level features from both source and target point clouds using deep learning techniques and facilitates the exchange of information between these features. Consequently， the global characteristics of the two point clouds could be aligned to focus on corresponding regions without the requirement for an attention mechanism. Experimental results indicate that the root mean square error （RMSE） for both the rotation RMSE（r） and translation RMSE（t） components in the ModelNet40 dataset alignments is reduced by 34% and 15%， respectively， compared to the previously established deep learning network PANet. Furthermore， continued training on turbine blade point clouds derived from the ModelNet40 dataset yielded RMSE（r） and RMSE（t） reductions of 83% and 46%， respectively. This method holds substantial promise for enhancing the evaluation processes related to the accuracy of turbine blade manufacturing in future applications.