Denoising attention loss network for microseismic monitoring of highway slopes

Microseismic monitoring of highway slopes is one of the key technologies for the early warning of geological hazards such as unstable rockfalls of highway slopes. To address the problems of severe noise interference and highly imbalanced microseismic event samples in actual monitoring scenarios, a Denoising Attention Loss Network was proposed for identification of microseismic events of highway slopes. First, a deep time–frequency denoising module with dual time–frequency constraints was designed, in which an encoding-sampling module was constructed and combined with time–frequency loss optimization to enhance noise suppression capability. Second, a hybrid attention mechanism incorporating channel attention and temporal attention was introduced to improve the feature extraction ability of microseismic signals. Finally, a momentum-adaptive weighted class-balanced loss was proposed to dynamically adjust loss salience and address the data imbalance problem. Extensive experiments conducted on highway slope microseismic datasets demonstrated the robustness of the proposed method under strong noise interference and class imbalance, highlighting its practical value for highway slope rockfall monitoring.