DHCEN-DTC: an ensemble learning approach for small-feature recognition in machining process classification

The capability to automatically learn from the design and manufacturing data of discrete manufacturing processes is essential for the development of future data-driven Computer-Aided Process Planning systems. This investigation presents an integrated learning approach, merging principles of deep learning with those of machine learning, to address the issue of detecting minuscule features due to constrained voxel resolution. The proposed model effectively learns local and global features of voxelized workpieces and better understands how quality information impacts classification performance. The model architecture consists of a dual-branch network that combines hybrid convolutional neural networks with a batch-attention Transformer encoder. The hybrid convolution branch extracts shape features, while the Transformer encoder extracts machining information features. These features are fused to jointly learn shape and quality attributes. A decision tree is also trained on a newly constructed dataset to distinguish between finishing and roughing processes. Finally, an ensemble learning strategy is employed, where a learnable linear layer fuses the output probabilities of both models. The proposed approach was evaluated on the MDP and EMDP datasets, achieving accuracies of 100 and 99.72%, respectively, which represents an improvement of 7.59 and 6.66% over the baseline model, outperforming 13 state-of-the-art backbone network models.