With the continued expansion of seismic observation networks and the rapid accumulation of continuous waveform data, intelligent seismic monitoring has placed increasingly stringent demands on real-time model inference and efficient edge deployment. In recent years, the DiTing seismic foundation model, with approximately one hundred million parameters, has demonstrated strong feature-representation capabilities in tasks such as earthquake detection and seismic phase picking. However, its large parameter count and high computational cost make direct deployment difficult in resource-constrained environments, including field monitoring terminals and domestically produced edge-computing devices. To address this problem, this study proposes a multi-level structured pruning method for the DiTing seismic foundation model that requires no post-pruning fine-tuning and is designed for deployment on the domestically developed RK3588 edge-computing platform.

The proposed method identifies the encoder as the primary compression target and performs importance evaluation and structured pruning at three structural levels: attention heads, feed-forward network （FFN） channels, and encoder layers. At the attention-head level, first-order Taylor importance, changes in P- and S-wave F₁ scores after masking individual heads, and attention-distribution entropy are integrated to evaluate each head in terms of task-loss sensitivity, actual influence on task performance, and attention-distribution characteristics. At the FFN-channel level, first-order Taylor importance is combined with activation magnitudes under actual seismic waveform inputs and weight norms to jointly characterize task-loss sensitivity, data-driven response intensity, and parameter-connection strength. At the encoder-layer level, the relative change between the layer input and output is combined with the mean importance of the attention heads within that layer, thereby evaluating each layer from the perspectives of overall feature-transformation capability and internal attention structure. The different importance indicators are normalized and fused through weighted summation to rank the candidate structures, and masking experiments are then used to determine appropriate pruning ratios. Based on this analysis, the number of attention heads in each encoder layer is reduced from eight to four, the FFN hidden dimension is reduced from 2048 to 1024, and the encoder depth is reduced from 24 to 16 layers, jointly compressing both the width and depth of the model.

The effects of structured pruning on model complexity and phase-picking performance are first evaluated on a GPU platform. The parameter count of the final pruned model decreases from 124.20 M to 56.89 M, corresponding to a reduction of 54.12%; its computational cost decreases from 59.03 GFLOPs to 30.77 GFLOPs, a reduction of 47.87%; and its weight-file size decreases from 475.50 MiB to 217.68 MiB, a reduction of 54.22%. Fixed-window evaluation is conducted on the complete DiTingV3 test set containing 78610 natural-earthquake records. The original model achieves P- and S-wave F₁ scores of 0.937 3 and 0.879 6, respectively, whereas the final pruned model achieves corresponding F₁ scores of 0.921 5 and 0.858 0. Thus, after more than half of the model parameters are removed, the P- and S-wave F₁ scores decrease by only 1.58% and 2.15%, respectively. These results demonstrate that, without post-pruning fine-tuning, the proposed method reduces the parameter count by more than half and the computational cost by nearly half while preserving most of the phase-picking capability of the original DiTing model.

The original and pruned models are further converted to the RKNN format and deployed on the RK3588 platform. In the fixed-window evaluation on the complete DiTingV3 test set, the original model achieves P- and S-wave F₁ scores of 0.936 9 and 0.879 6, respectively, while the final pruned model achieves corresponding scores of 0.920 0 and 0.856 2. These results are close to the GPU results of 0.921 5 and 0.858 0, indicating that model conversion and FP16 inference on the RK3588 platform introduce only minor additional performance degradation. In an inference-speed experiment using 5000 randomly selected long waveforms, the mean forward inference time per window decreases from 344.75 ms to 270.21 ms, a reduction of 74.54 ms or 21.62%, corresponding to a speedup of approximately 1.276 times. In terms of resource consumption, the RKNN model size decreases from 271.44 MiB to 136.41 MiB, a reduction of 49.75%, while the peak resident set size during sustained inference decreases from 535.40 MiB to 397.27 MiB, a reduction of 25.80%. These results demonstrate that the reductions in parameter count and FLOPs achieved by the proposed method can be translated into measurable inference acceleration, storage savings, and runtime memory reduction on the RK3588 platform.

In summary, the proposed multi-level structured pruning method jointly compresses attention heads, FFN channels, and encoder layers, thereby reducing both the width and depth of the DiTing seismic foundation model without requiring post-pruning fine-tuning. The method substantially reduces the parameter count, FLOPs, storage requirements, runtime memory consumption, and inference latency on the RK3588 platform while largely preserving P- and S-wave phase-picking performance on the complete DiTingV3 test set.