<p>Accurate prediction of milling tool wear enables timely assessment of machining conditions, improves tool-life utilization, enhances machining quality, and reduces production costs. Although multi-sensor-based tool wear prediction has received considerable attention, accurately integrating heterogeneous sensor signals and incorporating tool wear physics remain challenging. To address these issues, this study proposes a knowledge-enhanced physics-informed graph neural network (PIGNN) for tool wear prediction. First, multivariate time-series signals collected from multiple sensors are represented as a sensor graph, whose topology is adaptively learned from the monitoring data. A spatial–temporal graph convolutional network is then developed to capture the spatial dependencies among sensors and the temporal evolution of wear-related features. Finally, the monotonic evolution of tool wear and the underlying wear-rate mechanism are incorporated into the model as physical constraints. By embedding physically meaningful prior knowledge into the learning process, the predicted wear trajectories are guided toward a physically plausible solution space. The effectiveness of the proposed PIGNN is validated on two milling tool wear datasets, demonstrating its potential for online tool condition monitoring.</p>