A fusion of attention and multi-task learning approach for surface quality prediction in turning

Surface quality is a critical indicator of mechanical part performance and service life, encompassing surface roughness, dimensional accuracy, and roundness. Traditional deep learning methods typically focus on predicting single surface quality parameters, failing to meet the industrial demand for real-time multi-parameter prediction. This paper proposes a novel multi-task learning approach for simultaneous prediction of surface roughness, dimensional accuracy, and roundness in CNC turning processes. A hybrid network architecture named MTL-GAMDenseNet-CA is developed, integrating global attention mechanism into the DenseNet encoder for enhanced feature extraction and coordinate attention into task-specific decoders for adaptive feature selection. To address the high computational cost of gradient-based loss balancing, a loss ratio normalized weight strategy is proposed. Multi-channel vibration signals from three sensor positions (tool, spindle bearing, and bearing end cap) are converted into relative position matrix images and fused using two-dimensional discrete wavelet transform. Experimental results based on 144 turning tests under 36 operating conditions demonstrate that the proposed method achieves a surface roughness recognition F1-score of 0.9907, dimensional accuracy prediction mean absolute error (MAE) of 0.0228 mm, Root Mean Square Error (RMSE) of 0.0279 mm, and roundness prediction MAE of 0.0323 mm, RMSE of 0.0420 mm. The multi-task model achieves computational efficiency equivalent to 34.12% of three single-task models, validating its effectiveness for intelligent manufacturing applications.