This paper presents an intelligent prediction method for rock mass classification based on multimodal drilling parameter fusion. The method is specifically applied to coal-bearing sandstone-mudstone formations, which are significantly influenced by tectonic activity and exhibit complex drilling data variations. We propose a novel decision level fusion approach-Adaptive Threshold Reclassification Decision-Level Fusion (ATRDF), which integrates the distinct physical characteristics of various drilling signals into a confidence-based decision-making process. By leveraging key drilling parameters, such as rotational speed (RPM), rate of penetration (ROP), Torque, weight on bit (WOB), and Vibration signals, the ATRDF method constructs a multimodal fusion model. This model uses key drilling parameters as features and rock mass classification as the target label. Experimental results demonstrate that the proposed method significantly enhances prediction accuracy, achieving an 89% classification accuracy under three complex geological conditions. Furthermore, we employ interpretable AI tools including SHAP and Grad-CAM to elucidate the decision-making process based on one-dimensional and two-dimensional signal features. This paper also investigates the optimization of confidence thresholds and decision logic within the ATRDF framework, providing valuable insights into its fusion process and underlying mechanisms.