Optimized LightGBM Power Fingerprint Identification Based on Entropy Features.

Area of Science:

• Electrical Engineering
• Computer Science
• Data Science

Background:

• Massive power fingerprint data presents challenges with unbalanced categories and limited data transmission rates for Internet of Things (IoT) communications.
• Efficient extraction and identification of power fingerprints are crucial for large-scale IoT systems.

Purpose of the Study:

• To propose an optimized LightGBM (Light Gradient Boosting Machine) method for power fingerprint extraction and identification.
• To address data imbalance and reduce data transmission volume in IoT environments.
• To improve the accuracy and efficiency of power fingerprint recognition.

Main Methods:

• Extracted voltage and current signals using time-domain and V-I trajectory features, constructing a 56-dimensional feature set with six entropy features.
• Employed the Boruta algorithm with LightGBM for feature selection, identifying a 23-dimensional optimal feature subset with five entropy features.
• Utilized the Optuna algorithm to optimize LightGBM hyperparameters and improve the loss function for imbalanced datasets.

Main Results:

• Successfully reduced the computational complexity of feature extraction.
• Significantly decreased the amount of power fingerprint data transmission.
• Achieved high recognition accuracy and efficiency, meeting the demands of massive power fingerprint identification systems.

Conclusions:

• The proposed optimized LightGBM method effectively handles imbalanced datasets and reduces data transmission in power fingerprint identification.
• The method demonstrates superior performance in terms of accuracy and efficiency for large-scale IoT applications.
• This approach provides a viable solution for practical power fingerprint recognition systems with limited resources.