|
| 基于混合专家模型的消防物联数据异常检测 |
| Mixture of experts model-based anomaly detection of firefighting IOT data in converter stations |
| 投稿时间：2025-05-13 修订日期：2025-06-20 |
| 查看/发表评论 下载PDF阅读器 |
| DOI： |
| 基金项目:国网安徽省电力有限公司科学技术项目（编号：B3120523000D） |
|
| 中文关键词: 异常检测 监督学习 异常生成 混合专家 |
| 英文关键词:anomaly detection, supervised learning, anomaly generation, Mixture of Experts (MoE) |
| 摘要点击次数: 258 |
| 全文下载次数: 0 |
| 中文摘要: |
| 消防物联设备在长期使用过程中会因环境干扰或自身故障产生异常时序。针对异常样本稀缺及传统模型泛用性不足的问题，本文提出基于异常生成算法与混合专家（MoE）的异常时序检测模型。首先，算法利用正常数据生成包含趋势异常、恒值异常等多类型异常样本，其次，构建融合CNN、LSTM、GRU和Transformer的MoE模型，使用门控网络获取子专家输出权重，集成时序局部特征提取、全局关系捕捉、长期依赖建模等能力，通过引入温度系数优化训练过程。基于换流站水池液位、室内温湿度及感烟探测器数据的实验结果表明，不同子专家对不同异常类型检测效果各异，而MoE模型综合性能超越或接近最优子专家，且仅使用生成异常训练的模型对显著或轻微异常均具备有效检测能力。 |
| 英文摘要: |
| Firefighting IOT devices will generate anomalous timings due to environmental disturbances or their own failures during long-term use. Aiming at the problems of scarcity of anomaly samples and insufficient generalizability of traditional models, this paper proposes an anomaly timing detection model based on anomaly generation algorithm and Mixture of Experts (MoE). First, the algorithm uses normal data to generate multi-type anomaly samples including trend anomalies, constant value anomalies, etc. Second, it constructs a MoE model that integrates CNN, LSTM, GRU and Transformer, and use gated networks to obtain output weights of sub-experts to integrate the abilities of timing local feature extraction, global relationship capture, and long term dependency modeling, etc., and optimizes the training process through the introduction of temperature coefficients. The experimental results based on the pool level, indoor temperature and humidity, and smoke detector data of the converter station show that different sub-experts have different detection effects on different types of anomalies, while the comprehensive performance of the MoE model exceeds or is close to the optimal sub-experts, and the model that is trained only with the generation of anomalies has the capability of detecting significant or minor anomalies effectively. |
| 关闭 |
|
|
|
