<p>Abrasive belts are cutting tools used in belt grinding, featuring multiple abrasive grains and cutting edges. Once worn, these belts cannot be repaired, making monitoring their wear status crucial to prevent workpiece burns and guide tool replacements. This study proposes a wear condition monitoring method based on symmetric point pattern (SDP) multi-sensor fusion and a convolutional neural network–long short-term memory (CNN-LSTM) hybrid network. First, grinding experiments were conducted throughout the life cycle of the abrasive belt, collecting images of the belt surface along with sound, force, and vibration data. The U²-net network segmented the wear areas of the abrasive particles on the belt surface image, yielding a statistical value for abrasive wear. This data categorizes wear state into three types and seven wear levels. Next, the SDP image conversion algorithm transformed multi-sensor sequences from different wear states into two-dimensional images for input into the CNN-LSTM hybrid network framework for model training. Finally, the structural framework and hyperparameters of the preliminary classification model were optimized, resulting in a model that misclassified only 4 of the 530 test samples, achieving an accuracy of 99%. This method accurately monitors abrasive belt wear, which is significant for further controlling parameters to improve grinding quality.</p>