A self-supervised learning approach for intelligent surface roughness monitoring in thin-walled component machining

Surface roughness monitoring is essential for process quality control in intelligent manufacturing, yet reliable prediction in thin-walled component milling remains challenging because roughness labels are sparse and multi-channel vibration responses are spatially and dynamically heterogeneous. This study proposes a spatially aware self-supervised contrastive learning framework using six-channel spindle–worktable vibration signals. The task is formulated as sparse-label anchor-bag roughness regression: only the bag corresponding to the central small region in each large region is assigned a measured roughness label, whereas the remaining small-region bags are used as unlabeled data for representation learning. The framework integrates a multi-channel vibration encoder, a physics-guided gated bag aggregation module, and a spatially aware weighted InfoNCE objective. Bag-level self-supervised pretraining enables the model to learn transferable representations while reducing excessive separation between physically related neighboring bags; the pretrained encoder and aggregation module are then fine-tuned using only labeled anchor bags. Experiments on two thin-walled helical milling datasets, Ti2AlNb and GH4169D, demonstrate robust performance under sparse annotation and material variation using a workpiece-level evaluation protocol. The proposed method achieves R2 = 0.9431 ± 0.0337 and RMSE = 0.0274 ± 0.0145 μm on Ti2AlNb, and maintains strong performance on GH4169D with R2 = 0.9246 ± 0.0424, RMSE = 0.0171 ± 0.0093 μm. Comparative and ablation analyses support the effectiveness of self-supervised pretraining, spatially aware contrastive learning, multi-channel fusion, and the anchor-bag supervision setting, providing a label-efficient route for surface quality monitoring in intelligent manufacturing.