<p>The surface quality in a grinding process is influenced by machining parameters and the condition of the grinding wheel. As the wheel wears and dulls, fluctuations in surface roughness occur. The study leverages a deep learning model to predict surface roughness and adjust grinding parameters to mitigate such fluctuations. A probabilistic surface roughness prediction model based on a Bayesian Multilayer Perceptron (BMLP) was developed to achieve accurate surface roughness predictions during grinding. The BMLP treats the weights and biases of hidden layers as Gaussian distributions, enabling probabilistic predictions with corresponding confidence intervals. The model takes grinding sequence, grinding wheel speed, and feed rate as input parameters, with the grinding surface vectors as the output, establishing a nonlinear relationship between them. To enhance prediction accuracy, Bayesian optimization of hyperparameters was employed to search for the optimal solution in the hyperparameter space. To increase the model’s reliability, existing predictive equations for grinding surface roughness were used for validation. The prediction results demonstrate the accuracy of the BMLP, while the validation losses prove the effectiveness of Bayesian hyperparameter optimization. Additionally, the surface fitting results from existing surface roughness prediction equations further highlight the model’s exceptional ability to capture grinding surface features, expanding the application scenarios of deep learning in the field of machining.</p>