To address the challenge of predicting fracture-cave structures in coal mine goafs, this paper proposes a novel identification technique based on seismic attribute fusion. First, extraction algorithms for multiple sensitive seismic attributes were implemented, and corresponding modules were developed on a self-developed software platform. Second, to overcome the lack of supervised information and background noise in goaf areas, unsupervised machine learning algorithms—including Fuzzy C-Means Clustering (FCM) and Principal Component Analysis (PCA)—were integrated to analyze and fuse the extracted seismic attributes. Then, the unsupervised learning results were innovatively constrained using multi-source information data to extract supervised samples. A deep learning model based on the Backpropagation (BP) neural network was constructed and trained to achieve seismic attribute fusion and fracture-cave structure identification in goafs. Finally, the proposed method was applied to a goaf area in the Modi Gully, Hejin, Shanxi Province. The results demonstrated effective fusion of multiple seismic attributes, with predicted fracture locations and goaf boundaries showing strong agreement with drilling data, This study provides a reliable technical approach for fracture-cave identification in goaf areas.