<p>In industrial scenarios, deep learning registration predominantly relies on the degree of inverse transformation between the predicted transformation matrix and the actual transformation matrix as the loss function. This approach is challenging to adapt to cross-source point cloud local positioning and global registration tasks, and it exhibits a lack of robustness when the features of actual workpieces change rapidly. This paper proposes a registration strategy based on the actual transformation matrix, which corrects the normalization effect and achieves reproducible results in multiple models. The ROPNet model has been augmented with a real-time adjustment module for the range of changes, and the network architecture has been lightened. The design experiment compares conventional methods with other networks. ROPNet-CS has been shown to demonstrate comparable performance to RANSAC when operating on cross-source point cloud datasets. The registration of 16 pairs of collected point clouds on the equipment utilized in this study is achieved with a processing time of 1.8906&#xa0;s. It has been demonstrated that the proposed model exhibits a range of advantages over conventional deep learning algorithms. For instance, it has been shown to achieve a 62.3% enhancement in registration accuracy and a 22.6% increase in iteration speed when compared to the FMR algorithm. This paper tested the robustness of the model in registering noisy point clouds under different point numbers and overlap conditions and showed good robustness on a small car model and cylinder head workpiece dataset with a structural loss ratio of 70%.</p>