<p>Shield attitude control is a critical aspect that must be continuously monitored during shield tunneling. To achieve scientifically rational settings for shield tunneling parameters, this study constructed multiple machine learning prediction models, including shield attitude deviations and tunneling speed, and optimized the hyperparameters of these models using Bayesian algorithms. Subsequently, a constrained grey wolf optimization(GWO) algorithm was employed to establish a real-time safety control method for attitude that considers tunneling efficiency, by dynamically updating the upper and lower bounds for adjustable parameters. The results indicate that the k-nearest neighbors (KNN) model achieved the highest prediction accuracy; however, due to its specific algorithmic principles, KNN is unsuitable for optimization tasks. Embedding the extreme gradient boosting model into the GWO algorithm yielded the best attitude control performance: the absolute attitude deviations were reduced by an average of 45.1% compared to actual values, while the rate of change for adjustable parameters did not exceed 30%. This approach ensures safety and tunneling efficiency during attitude correction and exhibits universal applicability. Compared with other optimization algorithms, GWO demonstrated significant advantages in both optimization effectiveness and computational time.</p>