TThe tunneling performance of shield machines is greatly influenced by varying geological conditions. This study investigates an electrically-driven earth pressure balance shield machine, analyzing 3,761,006 tunneling data points across 961 rings. The dataset includes six geological combinations, such as sandy cohesive soil layers, along with corresponding tunneling parameters. Through correlation analysis, key features strongly related to tunneling speed, including total thrust, sync grouting volume, and foam pressure, were identified. To address the issue of uneven data distribution in practical tunneling projects, Gaussian resampling was applied, resulting in a dataset with 19 950 valid samples. A tunneling speed prediction method for shield machines based on the Kolmogorov-Arnold Network (KAN) was then proposed. The KAN model approximates nonlinear relationships by combining multi-level composite functions, breaking down the complex nonlinear interactions into simpler univariate function combinations. This approach ensures high prediction accuracy while significantly improving computational efficiency. Using the Shenzhen-to-Daya Bay Metro Shield Tunneling Project as a case study, experiments showed that the KAN model outperforms CNN and LSTM models in handling high-dimensional data and nonlinear coupling relationships. The prediction results align closely with measured data, with prediction errors ranging from 5.12% to 7.02% in simpler geological conditions (such as completely weathered mixed granite and strongly weathered mixed granite). In mixed geological layers, the prediction errors are higher, but the overall average error remains below 15%. This method offers strong decision support for optimizing shield machine operations under complex geological conditions. In the future, geological spatial distribution data will be incorporated into sequential modeling, and cutterhead wear prediction will be added to provide a more comprehensive intelligent management solution for shield tunneling.