Precision bone grinding constitutes a critical challenge for the safe operation of orthopedic surgical robots. A multi-modal sensing platform integrating force, displacement, acceleration, and laser displacement sensors was developed for comprehensive signal acquisition and depth prediction during bone grinding. A LSTM (long short-term memory) model was proposed to predict bone grinding depth. Experiments were conducted on cancellous bone samples with three densities (20, 30, and 40 pcf), and the collected experimental data were processed and mixed together for training, which showed that the mixed training strategy significantly improved the prediction accuracy and generalization ability of the LSTM model. Further testing on untrained 20 pcf cancellous bone samples showed that the model has strong robustness and high prediction accuracy. This study confirmed that the LSTM model can adapt to cancellous bone grinding scenarios with different characteristics, providing technical support for robotic applications in complex surgical scenarios.