Ultra-precision thinning of hard and brittle materials like monocrystalline silicon demands high dynamic stability in thinning spindle. To address the challenge of accurately detecting subtle spindle chatter anomalies in industrial environments characterized by high noise and limited data, this paper proposes a physics-guided dual-stream attention fusion transfer network (PG-AFNet). First, a physics-guided signal preprocessing method was developed. Using variational mode decomposition (VMD) and continuous wavelet transform (CWT) masking, one-dimensional dynamic features and high-frequency regions of interest (ROIs) rich in transient impact features were extracted. Second, the PG-AFNet architecture was designed. By introducing an attention mechanism, it achieves deep integration of one-dimensional purely dynamic sequences with two-dimensional spatiotemporal visual textures to capture surface damage features caused by subtle vibrations. Finally, systematic validations were conducted using a real silicon wafer thinning dataset with 197 real samples. By overcoming small-sample limitations via physical augmentation, PG-AFNet achieved an 82.45% (86.64% after data augmentation) diagnostic accuracy, significantly outperforming traditional baselines. Furthermore, a large-scale cross-load validation on the diverse CWRU dataset yielded an exceptional 99.68% accuracy under mixed-load conditions, conclusively verifying the model’s robust domain generalization. Lastly, a rigorous ablation study explicitly quantified the indispensable contributions of the physics-guided dual-stream architecture and attention fusion. This research provides a feasible theoretical foundation for intelligent surface quality monitoring in semiconductor hard-brittle material processing.

Real-time monitoring of wire bow is critical for ensuring wafer quality and preventing wire breakage in diamond multi-wire sawing (MWS). However, the deployment physical sensors in industrial MWS environments is hindered by severe sludge contamination, limited installation space, and high maintenance costs. To address these challenges, this paper proposes a novel data-driven soft sensor framework utilizing machine learning methods to predict wire bow based on readily accessible process data. A feature engineering pipeline, combining variance thresholding and correlation analysis, is established to identify key process variables. Subsequently, six representative ML algorithms are systematically evaluated, with eXtreme Gradient Boosting (XGBoost) optimized via two-stage hyperparameter optimization emerging as the superior model. Experimental results from an industrial MWS machine demonstrate that the proposed model achieves a coefficient of determination ($R^2$) of 0.992 and a mean absolute error (MAE) of 0.116 mm. Furthermore, the prediction is also extended to spatially distributed positions (head, middle, and tail) of the wire web. Finally, SHAP (SHapley Additive exPlanations) is utilized to elucidate the mechanical dependencies. This work provides a reliable and low-cost solution for wire bow monitoring during the MWS process.