[Objective] To address the issues of false alarms and missed warnings in monitoring and early warning operations, an anomaly detection model is developed based on historical early warning experience and sensor monitoring data, aiming to provide support for reducing the risk of false alarm signals in geological disaster early warnings. [Methods] Using time-series data from landslide crack meters as an example, an anomaly detection model integrating irregular time feature encoding and a multilayer perceptron hybrid module (MLP-Mixer) is proposed. Through multi-task learning and knowledge distillation mechanisms, knowledge from anomaly detection tasks guided by expert judgment labels is distilled into precursor perception tasks, thereby fully utilizing historical expert experience and the implicit disaster dynamics information in irregular time-series data to improve anomaly detection accuracy. [Results] Experimental results show that the proposed method outperforms baseline models on the given dataset, achieving optimal performance in precision (80.36%), recall (95.41%), F1 score (87.24%), and area under the ROC curve (87.20%). [Conclusion] The model’s comprehensive advantages in recall and precision effectively reduce the risk of missed warnings and can be applied to automatically filter false alarm signals, thereby improving the efficiency and reliability of early warnings.