Objective An anomaly detection model was anstructed based on historical warning experience and sensor monitoring data to address the issues of false alarms and omissions in monitoring and early warning operations, in order to provide a scientific support for reducing the risk of false alarms 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 with a multilayer perceptron hybrid module (MLP-Mixer) was proposed. Through multi-task learning and knowledge distillation mechanisms, knowledge from anomaly detection tasks guided by expert judgment labels was distilled into precursor perception tasks, thereby fully utilizing historical early warning experience and the implicit disaster dynamics information in irregular time-series data to improve anomaly detection accuracy. Results The experimental results showed that the proposed method outperformed 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 alarms and can be used to automatically filter false alarm signals, thereby improving the efficiency and reliability of early warning.
文献参数: 曾振威, 欧阳淑冰, 李代超, 等.面向滑坡裂缝计时序数据异常检测的预警方法研究[J].水土保持通报,2026,46(1):228-235. Citation:Zeng Zhenwei, Ouyang Shubing, Li Daichao, et al. Early warning methods for anomaly detection in time-series data of landslide crack meters [J]. Bulletin of Soil and Water Conservation,2026,46(1):228-235.
表2与图4的消融试验结果反映了各组件对总体性能的贡献。移除时间嵌入(ours-without-time)以更大幅度的召回率下降换取模型精确率上升。这说明时间维的可学习表示通过对不规则时间信息的显式建模强化了模型对异常样本的覆盖度和敏感性,从而在低误报漏报要求的监测预警场景下具有必要性。将通道交互由MLP-Mixer替换为通道级Transformer(ours-transformer),两者指标表现相近,但前者(MLP-Mixer)召回率与AUC更加稳定,后者(transformer)精确率略高。这与两者的结构先验一致:MLP-Mixer的参数收缩与可分离映射在小样本场景下更加稳定,而注意力机制的冗余表达导致其覆盖不足,表现为召回劣化、分类性能下降;去除知识蒸馏(ours-without-KD)在精确率和AUC上表现相差不大,但召回率和F1分数小幅上升。这表明蒸馏主要作用于判别边界的校准与置信度分布的整形:在不显著牺牲精度的前提下提升对正例的覆盖,从而提高综合指标。此外,将两任务完全独立训练(anomaly detection/precursor of anomaly)发现,多任务学习在保持高召回的同时显著提高精确率,在F1分数上小幅上升,而AUC与最佳单任务接近,说明共享表征与蒸馏传递在两个任务中取得综合表现,能更加稳定地实现“高召回,较高精确率”的组合目标。综上所述,3项组件和多任务框架对模型性能的贡献具有显著的互补性和必要性。
Ministry of Natural Resources. The 14 th Five-Year Plan for National Geological Disaster Prevention and Control [R]. Beijing: Ministry of Natural Resources, 2022.
[3]
CookA A, MısırlıG, FanZhong. Anomaly detection for IoT time-series data: A survey [J]. IEEE Internet of Things Journal, 2020,7(7):6481-6494.
[4]
YangXue, QiXuejun, ZhouXiaobo. Deep learning technologies for time series anomaly detection in healthcare: A review [J]. IEEE Access, 2023,11:117788-117799.
LiuGang, YeLixin, ChenQiyu, et al. Abnormal event detection of city slope monitoring data based on multi-sensor information fusion [J]. Bulletin of Geological Science and Technology, 2022,41(2):13-25.
ZhangCongcong, WangXiuying. Research on detection method of abnormal data of precursory observation [J]. Technology for Earthquake Disaster Prevention, 2014,9(S1):615-621.
[9]
BarnettV, LewisT. Outliers in Statistical Data [M]. 3rd ed. Chichester, England; New York: Wiley, 1994.
LiQiaomei, ChengLibo. Prediction method of time series ARIMA model based on wavelet analysis [J]. Journal of Shenyang Normal University (Natural Science Edition), 2021,39(1):49-53.
XieLixia, WangJiamin, YangHongyu, et al. Review of anomaly detection methods for time series [J]. Journal of Civil Aviation University of China, 2024,42(3):1-12.
[14]
MalhotraP, VigL, ShroffG M, et al. Long short term memory networks for anomaly detection in time series [C]∥ Proceeding of the European Symposium on Artificial Neural Networks, Piscataway N J: IEEE, 2015
[15]
WuHaixu, XuJiehui, WangJianmin, et al. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting [J]. Advances in Neural Information Processing Systems, 2021,34: 22419-22430.
[16]
WuHaixu, HuTengge, LiuYong, et al. TimesNet: Temporal 2D-variation modeling for general time series analysis[C]∥ The Eleventh International Conference on Learning Representations (ICLR) 2023.
[17]
HuRongyao, YuanXinyu, QiaoYan, et al. Unsupervised anomaly detection for multivariate time series using diffusion model [C]∥ Seoul Korea: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024.
[18]
ZhongZhijie, YuZhiwen, XiXing, et al. SimAD: a simple dissimilarity-based approach for time-series anomaly detection [J]. IEEE Transactions on Neural Networks and Learning Systems, 2025,36(11):19669-19680.
ZhangLei, JuNengpan, HeChaoyang, et al. Real-time anomaly detection and analysis of time series data for crack gauge in landslides [J]. Chinese Journal of Rock Mechanics and Engineering, 2024,43(1):206-215.
GuoYizhen, YanChenxin, DuMiaozeng, et al. A Preprocessing Method for Landslide monitoring data based on LSTMAD algorithm and hermite interpolation: CN114970688A [P]. 2022-08-30.
[23]
WangW, MotaghM, PlankS, et al. Application of sar time-series and deep learning for estimating landslide occurrence time [J]. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2022,XLIII-B3-2022:1181-1187.
[24]
林经纬.福建省滑坡灾害特征及驱动因素分析[J].莆田学院学报,2015,22(5):83-88.
[25]
LinJingwei. Characteristics and driving factors of landslide hazard in Fujian Province [J]. Journal of Putian University, 2015,22(5):83-88.
[26]
ZhangWeihua. Geological disaster monitoring and early warning system based on big data analysis [J]. Arabian Journal of Geosciences, 2020,13(18):946.
[27]
吴辉隆.重庆市滑坡灾害监测预警优化研究[D].四川 成都:成都理工大学,2022.
[28]
WuHuilong. Research on optimization of landslide disaster monitoring and early warning in Chongqing [D]. Chengdu, Sichuan: Chengdu University of Technology, 2022.
ChenChunli, FangZhiwei. Research on an effective rainfall model for geological disaster early warning in Fujian Province, China [J]. Journal of Geomechanics, 2023,29(1):99-110.
[31]
ZengAiling, ChenMuxi, ZhangLei, et al. Are transformers effective for time series forecasting? [J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2023,37(9):11121-11128.
[32]
KlötergensC, YalavarthiV K, DerneddeT, et al. Mixing it up: Exploring mixer networks for irregular multivariate time series forecasting [J]. Journal of Time Series Analysis, 2024, 45(3): 100-120.
[33]
HintonG, VinyalsO, DeanJ. Distilling the knowledge in a neural network [EB/OL]. 2015:arXiv:1503.02531.
[34]
JhinS Y, LeeJ, ParkN. Precursor-of-anomaly detection for irregular time series [C]∥ Long Beach,CA,USA.ACM: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023.
[35]
LiuF T, TingKai ming, ZhouZhihua. Isolation forest [C]∥ Pisa, Italy: 2008 Eighth IEEE International Conference on Data Mining, 2009.
[36]
WangYuxuan, WuHaixu, DongJiaxiang, et al. Deep time series models: A comprehensive survey and benchmark [EB/OL]. arXiv:2407.13278, 2024.
[37]
LiuChen, HeShibo, LiShizhong, et al. Detecting both seen and unseen anomalies in time series [J]. ACM Transactions on Knowledge Discovery from Data, 2025,19(4):1-29.