Few-Shot Learning for Image-Based Nonintrusive Appliance Signal Recognition

Area of Science:

• Artificial Intelligence
• Computer Vision
• Energy Systems

Background:

• Deep learning models for nonintrusive load monitoring (NILM) require extensive data, posing challenges in acquisition, storage, and computation.
• Appliance signal data often suffers from imbalance due to varying activation periods.

Purpose of the Study:

• To develop and evaluate a computer vision deep learning approach for recognizing disaggregated appliance signals using a reduced dataset.
• To address the high data demands of deep learning in NILM applications.

Main Methods:

• Utilized Siamese and prototypical few-shot classification algorithms on a reduced dataset.
• Implemented a similarity test to ensure data quality before deep learning model input.
• Focused on few-shot learning techniques to handle limited data samples.

Main Results:

• Achieved acceptable performance in appliance signal recognition.
• Demonstrated the viability of Siamese networks for one-shot recognition and prototypical networks for handling data imbalance.
• Validated the effectiveness of the approach with limited data samples.

Conclusions:

• Few-shot learning is a promising strategy for nonintrusive load monitoring with reduced datasets.
• The developed computer vision approach offers an efficient alternative to traditional data-intensive deep learning methods in NILM.
• This method can significantly lower the barriers to implementing NILM systems.