Abstract: To improve the prediction accuracy of grinding surface roughness of Al2O3-based ceramic insulating coating on bearing surface, a BP neural network prediction model was established which was consistent with the actual machining process. A method for measuring grinding wheel surface and quantifying abrasive particle characteristic parameters was proposed based on the principle of spectral confocal. A neural network prediction model for workpiece surface roughness was established, which took characteristic parameter K of grinding wheel surface, grinding wheel speed ω, workpiece feed speed υ, cutting depth ρ and normal grinding force F as input parameters. The model could directly reflect the time-varying state of grinding wheel surface. Finally, the prediction performance of the network was verified by the known grinding samples and the four groups of unknown test samples after grinding wheel passivation. For the known samples, the roughness predicted by BP network is consistent with the actual roughness, and the network output error is less than ±0.04μm. Further using the network for the grinding wheel after passivation to predict the unknown grinding test samples, the accuracy of the network prediction decreases, and the maximum error percentage is less than 20%. The neural network, which includes the characteristic parameters of abrasive particles on grinding wheel surface, can be used to predict the workpiece roughness of Al2O3-based ceramic insulation coating on bearing surface under the transient state of abrasive wear of grinding wheel, and the network has a certain generalization ability for unknown samples.