<p>Measurement-While-Drilling (MWD) technology has become a primary method for acquiring rock mass strength parameters due to its real-time capability and high efficiency. However, the high-noise and nonlinear characteristics of MWD signals limit the prediction accuracy of traditional single machine learning models (e.g., SVM, LSTM). Although existing hybrid models (e.g., CNN-Transformer) can improve performance, the fusion of multi-branch heterogeneous features remains a core challenge. To address this, we propose a guided heterogeneous feature fusion network (GHT-Net) for accurate rock strength prediction in open-pit mines. First, we design a Feature Interaction Module (FIM) to project heterogeneous features from different branches into a unified representation space and adopt an adaptive weighting strategy for fusion. Second, we integrate a ContraNorm (CN) module into the Transformer-based feature extraction branch, which introduces negative features as contrastive samples to widen the disparity between primary and negative features, thereby dynamically enhancing feature discriminability. Finally, an Efficient Additive Attention (EAA) mechanism is incorporated after the feature interaction module, enabling additive interaction between global context vectors and local features in a low-dimensional space to strengthen global perception and training stability. Experimental results demonstrate that GHT-Net achieves superior performance with RMSE = 1.9762, MAE = 1.3687 and R²=0.9989. Evaluated on datasets of granite, limestone and sandstone, the proposed model outperforms four mainstream benchmarks including SVM, LSTM, LSTM-Transformer and CNN-Transformer in prediction accuracy, generalization and robustness. This study provides a novel approach for real-time and accurate rock strength prediction.</p>