The grinding wheel condition critically affects surface integrity and process stability in precision grinding. During continuous grinding, abrasive grains exhibit multiple coexisting physical phenomena, including chip removal, self-sharpening, stuffing, and passivation, which cannot be adequately represented by conventional single-state monitoring approaches. In this study, we present an acoustic emission (AE)-based multi-label deep learning framework for intelligent grinding wheel condition recognition and dressing timing determination. AE signals collected during continuous surface grinding were transformed into time–frequency representations and labeled through the direct optical observation of abrasive grain conditions. A Residual Neural Network (ResNet)-based multi-label classifier was trained to simultaneously identify the four grinding phenomena. The proposed model achieved a macro-averaged F1-score of approximately 0.80, with area-under-curve values ranging from 0.73 to 0.81. Consecutive prediction behavior was further analyzed under continuous grinding conditions, showing that persistent AE-predicted degradation phenomena correspond to irreversible surface roughness deterioration beyond Ra = 0.4 μm. On the basis of these findings, an experimentally validated dressing timing criterion was established using consecutive multi-label predictions and surface roughness verification. The proposed approach provides a physically interpretable and data-driven solution for intelligent grinding process monitoring and dressing decision-making.