Abstract: Traditional drilling condition identification methods rely on expert experience and threshold judgment, neglect temporal dependencies, and support only a limited number of classes with low accuracy. To address these issues, an efficient drilling condition identification model based on the Transformer model was proposed. The model employed a multi-head self-attention mechanism and residual connections to capture long-range temporal dependencies in drilling data and added a feedforward fully connected classifier at the output end of the encoder to classify drilling conditions. Training used cross-entropy loss and the Adam optimizer to reach the optimal model parameters. Comparative experiments on a historical dataset of offshore drilling show that while all models’ loss curves converge, CNN model and CNN−LSTM model exhibit slow and unstable convergence within 100 iterations, whereas the Transformer model stabilizes rapidly. Adding the derived feature parameters has further improved the recognition accuracy and generalization ability of the model. These findings validate the model’s superiority in capturing multi-class long-term dependencies and offer an innovative solution for high-precision drilling condition identification. It has significant engineering value for enhancing the intelligence level of drilling supervision.