Abstract: Accurate discrimination between natural earthquakes (EQ) and mining-induced collapse events (SS) remains challenging because the two event types often exhibit similar waveforms in regional seismic networks. Manual identification is time-consuming and relies heavily on analyst experience, limiting its applicability to routine seismic monitoring. In this study, waveform records from the Huating mining area and its surrounding region, Gansu Province, China, were used to construct a single-trace classification dataset, with training and testing sets strictly separated at the event level to eliminate data leakage. Six representative deep learning models, including ResNet18-1D, VGG-1D, CRNN, ConvNeXt-1D, an enhanced ConvNeXt-1D, and a convolution-attention hybrid model (Transformer-1D), were evaluated under identical preprocessing, training, and evaluation protocols. Transformer-1D achieved the best overall performance among all models. At the trace level, it reached an accuracy of 0.9320, a macro-F1 score of 0.9315, and an ROC-AUC of 0.9875. After event-level probability aggregation, the accuracy and macro-F1 score both increased to 0.9846, with an ROC-AUC of 0.9991. Despite containing only 0.44 million parameters, approximately one-fifth of those in the enhanced ConvNeXt-1D model, Transformer-1D consistently outperformed all competing architectures. Additional experiments incorporating multi-scale convolutional embedding, attention pooling, and deeper Transformer encoders did not produce consistent performance gains, suggesting that a lightweight hybrid architecture provides a better balance between model capacity and the available training data. Model interpretation further showed that the conventional Grad-CAM method was ineffective for Transformer-based architectures because global self-attention weakens the spatial correspondence required by gradient-based localization. By computing relevance scores from encoder-output 词元s, meaningful temporal localization was recovered. The resulting attention maps consistently highlighted seismic phase arrivals and coda-wave segments, in agreement with the expected physical differences between natural earthquakes and collapse events. These results demonstrate that combining convolutional feature extraction with self-attention-based temporal modeling is an effective strategy for classifying long-duration one-dimensional seismic waveforms and provides a practical solution for the automatic identification of collapse events in regional seismic monitoring networks.