Acoustic emission (AE) monitoring captures transient elastic waves generated by rock fracturing. Reliable identification of fracture-related AE signals supports rock-mass damage monitoring and subsequent early-warning analysis of mining-induced geohazards. However, AE signals are often corrupted by operational noise, hindering accurate detection. To improve AE signal identification under complex field conditions, raw binary AE data from a mine site were parsed, segmented, and manually labeled, and a binary dataset of fracture-related AE signals and industrial noise was constructed. Seven time-domain statistical features and three frequency-band energy-ratio features were extracted from each segment and combined into a 10-dimensional multi-domain feature vector. A hybrid model integrating a 1D convolutional neural network (1D-CNN) with a local-window self-attention mechanism was developed. The proposed model was compared with eight models: a standalone 1D-CNN, a 1D-CNN with global self-attention, a BPNN, an SVM, logistic regression, k-nearest neighbors, random forest, and XGBoost. On an independent test set comprising 400 samples, the proposed model achieved mean accuracy, precision, recall, and F1-score values of 96.75 ± 0.71%, 93.99 ± 1.20%, 99.90 ± 0.22%, and 96.85 ± 0.67%, respectively, across five repeated runs. It achieved the highest mean accuracy and F1-score among the nine models. Across five fixed feature permutations, recall remained stable, whereas precision, accuracy, and F1-score showed greater variation. These findings indicate the potential of the proposed model for front-end classification of fracture-related AE signals in mine-site monitoring.