In order to tackle the technical bottleneck of predicting jamming and empty drilling faults of rock drills in drill-and-blast tunnel construction, a method of fault classification and prediction of the rock drill based on multi-feature time-series labeling Transformer was proposed. By collecting the key high-frequency while-drilling parameters of the rock drill under various working conditions, and integrating the thresholds of these parameters in the faulty states, a labeled dataset of jamming and empty drillin was constructed. A multi-feature time-series labeling strategy was designed to convert raw data into sequences of embedding vectors with temporal relationships. Building upon this, a multi-head self-attention mechanism was employed to mine long-term dependencies among the multiple features. Combined with a feedforward neural network and a dynamic slicing optimization strategy, and enhanced by residual connections and layer normalization, a time-prospective Transformer model was constructed. This model ultimately achieved the dual functions of fault classification and prediction. The experimental results demonstrated that the proposed method achieved an accuracy of 93.233% in the classification and prediction of jamming and empty drilling faults of the rock drill, significantly outperforming comparative models such as CNN (convolutional neural network), LSTM (long short-term memory), CNN-LSTM, RNN (recurrent neural network), and iTransformer. Visualization results of features using t-SNE (t-distribution stochastic neighbour embedding) revealed superior intra-class clustering and inter-class separation characteristics for the proposed model. Furthermore, it exhibited the lowest training loss and an inference time of merely 0.014 6 s, meeting the real-time warning requirements. The research results provide a reliable technical approach for classifying and predicting the faults of rock drills under complex geological conditions.