Automatic Weight Redistribution Ensemble Model Based on Transfer Learning to Use in Leak Detection for the Power Industry

Creating an effective deep learning technique for accurately diagnosing leak signals across diverse environments is crucial for integrating artificial intelligence (AI) into the power plant industry. We propose an automatic weight redistribution ensemble model based on transfer learning (TL) for det...

Full description

Saved in:
Bibliographic Details
Published in:Sensors (Basel, Switzerland) Switzerland), 2024-08, Vol.24 (15), p.4999
Main Authors: Kwon, Sungsoo, Jeon, Seoyoung, Park, Tae-Jin, Bae, Ji-Hoon
Format: Article
Language:English
Subjects:
Accuracy
Acoustics
Aging
Artificial intelligence
Datasets
Deep learning
Efficiency
Electric power-plants
ensemble model
ensemble weight automatic redistribution
Fourier transforms
Leak detection
Machine learning
Neural networks
Noise
Power plants
Sensors
Signal processing
Support vector machines
Time series
transfer learning
Wavelet transforms
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Creating an effective deep learning technique for accurately diagnosing leak signals across diverse environments is crucial for integrating artificial intelligence (AI) into the power plant industry. We propose an automatic weight redistribution ensemble model based on transfer learning (TL) for detecting leaks in diverse power plant environments, overcoming the challenges of site-specific AI methods. This innovative model processes time series acoustic data collected from multiple homogeneous sensors located at different positions into three-dimensional root-mean-square (RMS) and frequency volume features, enabling accurate leak detection. Utilizing a TL-driven, two-stage learning process, we first train residual-network-based models for each domain using these preprocessed features. Subsequently, these models are retrained in an ensemble for comprehensive leak detection across domains, with control weight ratios finely adjusted through a softmax score-based approach. The experiment results demonstrate that the proposed method effectively distinguishes low-level leaks and noise compared to existing techniques, even when the data available for model training are very limited.
ISSN:1424-8220
1424-8220
DOI:10.3390/s24154999