Scholarlink

<p>Monitoring and analyzing electricity consumption is essential for understanding how appliances are used. This insight enables intelligent energy feedback and demand-side management and helps both users and utility providers make smarter data-driven decisions. However, accurate classification of appliances remains challenging due to a lack of adequate labeled data and domain shifts caused by varying household characteristics, user behavior, usage patterns, and sampling rates. Additionally, existing models often rely on extensive labeled data for training, which can be costly and impractical. In this paper, we present a domain adaptation framework for household appliance classification that integrates depthwise separable convolution and window-based attention to effectively capture temporal features from resampled 1D power consumption data. The model is adversarially trained on source and target domains using negative log-likelihood loss with correlation alignment, preventing numerical instability. We also use the Wasserstein distance (WD) to assess the changes across different households to build more robust appliance classification in varying installations. The proposed approach offers a task-specific integration of lightweight feature extraction, adversarial adaptation, correlation alignment, and WD-based feature augmentation for appliance classification under domain shift. We evaluated the model in three scenarios: Source and target domains from different geographic regions, same region but different datasets, and same dataset but different households. Experiments on three public datasets show significant improvements in cross-dataset generalization. Furthermore, extensive experiments reveal that our model can offer a viable solution for appliance classification with minimal dataset requirements and 18.84% less MFLOPs, 8.79% less parameters, and 5.13% less storage, and noticeable performance of up to 51.23% improvement in macro F1.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Domain-aware appliance classification with hybrid temporal features and correlation-aligned adversarial training

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Mohammed Ayub,
  • El-Sayed M. El-Alfy

摘要

Monitoring and analyzing electricity consumption is essential for understanding how appliances are used. This insight enables intelligent energy feedback and demand-side management and helps both users and utility providers make smarter data-driven decisions. However, accurate classification of appliances remains challenging due to a lack of adequate labeled data and domain shifts caused by varying household characteristics, user behavior, usage patterns, and sampling rates. Additionally, existing models often rely on extensive labeled data for training, which can be costly and impractical. In this paper, we present a domain adaptation framework for household appliance classification that integrates depthwise separable convolution and window-based attention to effectively capture temporal features from resampled 1D power consumption data. The model is adversarially trained on source and target domains using negative log-likelihood loss with correlation alignment, preventing numerical instability. We also use the Wasserstein distance (WD) to assess the changes across different households to build more robust appliance classification in varying installations. The proposed approach offers a task-specific integration of lightweight feature extraction, adversarial adaptation, correlation alignment, and WD-based feature augmentation for appliance classification under domain shift. We evaluated the model in three scenarios: Source and target domains from different geographic regions, same region but different datasets, and same dataset but different households. Experiments on three public datasets show significant improvements in cross-dataset generalization. Furthermore, extensive experiments reveal that our model can offer a viable solution for appliance classification with minimal dataset requirements and 18.84% less MFLOPs, 8.79% less parameters, and 5.13% less storage, and noticeable performance of up to 51.23% improvement in macro F1.