<p>Accurate prediction of grinding wheel wear is essential for ensuring machining quality, improving production efficiency, and enabling intelligent manufacturing. However, the wear evolution of grinding wheels is influenced by machining conditions, material properties, and dynamic process characteristics, resulting in complex nonlinear behaviour that is difficult to measure directly. To address these challenges in practical industrial applications, this study first investigates the relationship between grinding wheel wear evolution and workpiece dimensional accuracy, and proposes an indirect wear characterization method based on vibration signals to enable rapid mapping of wear information. Furthermore, considering that vibration signals collected from different positions contribute unequally to wear characterization, a multi-scale convolutional neural network with adaptive-weighted ensemble learning (EWMCNN) is developed. The model extracts multi-band vibration features through a multi-scale convolutional architecture and employs a meta-learner to dynamically assess the importance of features from different positions, thereby enabling adaptive decision-level fusion through dynamic weight allocation. Finally, based on a grinding vibration data acquisition platform and incorporating discrete wavelet transform denoising and data enhancement, the proposed method enables dynamic prediction of grinding wheel wear. Ablation experiments demonstrate that the proposed EWMCNN achieves superior prediction accuracy and generalization capability compared with other baseline models, providing an efficient and reliable solution for online monitoring of grinding wheel wear.</p>