Abstract: To mitigate the impact of geological structural variations on reservoir prediction, this study proposes a novel seismic phase clustering analysis technique based on multi-granularity ensemble learning. The technique first extracts features at three scales: coarse-grained, fine-grained, and micro-grained. Coarse-grained features are derived using the Spearman correlation coefficient to reflect the macroscopic relationships between strata. Fine-grained features are extracted via long short-term memory (LSTM) networks to capture detailed characteristics among waveforms. Micro-grained features are obtained through dynamic time warping (DTW) distances to capture the microscopic differences within individual waveforms. Subsequently, through self-organizing map methods, clustering results are obtained for each granularity. A soft alignment-based ensemble learning technique is then applied to integrate the clustering results from different granularities, effectively addressing the limitations of single-granularity approaches influenced by geological structural variations. Experimental results demonstrate that the proposed multi-granularity ensemble learning algorithm significantly enhances seismic clustering accuracy and provides a valuable reference for reservoir prediction across different regions.

Key words: seismic phase clustering analysis, multi-granularity, ensemble learning, dynamic time warping (DTW), self-organizing maps (SOM)