Herein, a phase-sensitive optical time domain reflectometer (φ-OTDR) pattern recognition method based on cross-model knowledge distillation is introduced to meet the demands for high precision and efficiency in distributed fiber optic pattern recognition. This method employed hierarchical token-semantic audio transformer as the teacher model and broadcasting-residual network as the student model. This setup enables the student model to achieve recognition performance comparable to transformer-like networks with disparate architectures using less parameters. For practical engineering experiments, a φ-OTDR was used as the signal acquisition device. The dataset used in practical engineering scenarios included signals from four categories, such as climbing nets, background noises, striking nets, and wind noises. Compared to typical deep learning algorithms, this improved algorithm demonstrates superior accuracy and faster convergence, resulting in higher recognition efficiency and offering considerable potential for engineering applications.