|
| 基于无监督深度自编码器的森林火险因子时序异常检测研究 |
| Research on Time Series Anomaly Detection of Wildfire Risk Factors Based on Unsupervised Deep Autoencoder |
| 投稿时间：2025-12-12 修订日期：2026-04-22 |
| 查看/发表评论 下载PDF阅读器 |
| DOI： |
| 基金项目:应急管理部重点科技计划项目(2024EMST010101) |
|
| 中文关键词: 森林火灾、时序异常检测、多尺度卷积、极值理论（POT） |
| 英文关键词:Wildfire, Time series anomaly detection, Multi-scale convolution, POT(Peaks Over Threshold) |
| 摘要点击次数: 98 |
| 全文下载次数: 0 |
| 中文摘要: |
| 现代森林火险预警系统依赖传感器监测火险因子，但传感器易受老化或环境干扰产生故障，削弱系统可靠性。针对火险因子时序数据异常样本稀缺、模式复杂的问题，本文提出一种自适应强、面向单维时序的无监督多尺度深度自编码器异常检测方法。方法通过多尺度卷积捕捉不同时间尺度变化模式，结合长短期记忆网络与注意力机制强化关键时间步特征表示。在重构基础上，利用极值理论中POT（Peaks Over Threshold）方法确定误差阈值实现异常判定。实验结果表明，在无先验条件下，模型在大气压力、地表温度和地表湿度异常检测中F1分别为0.9959、0.9606和0.9041，并在综合识别准确率、精度及F1上优于多种重构及预测基线模型，验证了其在复杂单维时序异常检测中的有效性。 |
| 英文摘要: |
| Wildfire is a typical natural disaster that threatens human life and property. Modern wildfire prevention systems rely on sensors to monitor fire risk factors, but sensors are prone to anomalies due to aging or environmental interference, which reduces system reliability. To address the challenges of scarce anomalous samples and complex patterns in time series data of fire risk factors, this paper proposes an adaptive unsupervised multi-scale deep autoencoder method called MCLAD for single-dimensional time series anomaly detection. The method employs multi-scale convolution to capture patterns at different temporal scales and combines LSTM with an attention mechanism to enhance feature representation of critical time steps. Based on reconstruction, the POT extreme value theory is used to determine error thresholds for anomaly detection. Experimental results show that, without prior knowledge, the model achieves F1 scores of 0.9959, 0.9606, and 0.9041 for anomalies in atmospheric pressure, surface temperature, and surface humidity respectively, outperforming multiple reconstruction and prediction baseline models in accuracy, precision, and F1 score, demonstrating its effectiveness for complex single-dimensional time series anomaly detection. |
| 关闭 |
