GPT-Assisted Multi-View Multi-Label Learning for Non-Intrusive Load Monitoring With Incomplete Labels and Data

The state-of-the-art research in non-intrusive load monitoring (NILM) has emphasized the difficulty of obtaining a considerable volume of labeled data and proposed weak learning algorithms for NILM. However, the availability of aggregate load measurement can be compromised due to temporary sensor failures, leading to missing data and incomplete labels. Most existing methods fail to address these issues simultaneously or extract comprehensive features. To overcome this, we propose GPT-iMVML, a generative pre-trained transformer (GPT)-assisted multi-view multi-label learning framework designed for NILM under incomplete data and label conditions. The model incorporates attention-based bi-directional long short-term memory (LSTM) networks and GPT to extract short-term and long-term features from diverse measurements, enhancing representation learning. A view availability matrix and attention mechanism are applied to fuse multi-view features effectively, while an LSTM-based decoder is used to validate and correct extracted features. Furthermore, an improved weighted binary cross-entropy loss is introduced to mitigate the adverse effects of incomplete labels on model training. Experimental results on public datasets demonstrate that GPT-iMVML significantly improves classification performance and robustness, outperforming typical baseline methods. The proposed framework effectively alleviates the impact of simultaneous data and label loss, offering enhanced feature extraction, generalization, and adaptability for NILM applications under realistic constraints.