Semi-Supervised Fault Detection for Building Energy Systems

Fault labels for building energy systems are scarce because faults are rare and costly to annotate. This paper proposes a semi-supervised fault detection method that combines a physics-informed residual generator with a consistency-regularized classifier, enabling effective training with a small fraction of labelled samples.

问题

建筑能源系统的故障样本稀缺且标注成本高，纯监督学习方法难以适用。

方法

提出半监督故障检测框架：

1. 基于物理规则生成残差特征，降低对标注数据的依赖；
2. 采用一致性正则化的分类器，利用大量未标注样本提升泛化能力；
3. 引入故障样本增强策略，缓解类别不平衡。

结论

在公开数据集与某公共建筑实测数据上，在仅 10% 标注样本条件下，所提方法的 F1 值较纯监督基线提升约 14 个百分点。

@article{Z.Xu20242024-fau,
  title   = {Semi-Supervised Fault Detection for Building Energy Systems},
  author  = {Z. Xu and H. Zhao},
  journal = {Applied Energy},
  year    = {2024},
  doi     = {10.1016/j.apenergy.2024.100008},
}