This thesis examines the application of deep learning for automatic machining feature recognition (MFR) from both clean and noisy CAD models, a crucial step in bridging the gap between design and manufacturing within an intelligent production pipeline. The traditional methods were limited and struggled when handling noisy, real-world data, creating a gap between idealised CAD environments and actual industrial applications. Also, recent deep learning approaches such as FeatureNet and Inception-based 3D CNN have demonstrated high accuracy on synthetic CAD data, but they rarely test robustness on noisy or scanned meshes, and this leaves a crucial gap. To address this, three voxel-based 3D convolutional neural networks were investigated: a FeatureNet-inspired model, a Pre-activated ResNet, and a lightweight InceptionLite architecture.