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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Acoustic Signal Processing

Background:

• Integrating AI into the power plant industry requires robust leak detection systems.
• Site-specific AI methods face challenges in diverse operational environments.
• Accurate diagnosis of leak signals is crucial for plant safety and efficiency.

Purpose of the Study:

• To develop an effective deep learning technique for diagnosing leak signals across diverse power plant environments.
• To overcome limitations of site-specific AI methods using transfer learning.
• To improve the accuracy and reliability of AI-based leak detection.

Main Methods:

• Proposed an automatic weight redistribution ensemble model based on transfer learning (TL).
• Processed time series acoustic data into 3D root-mean-square (RMS) and frequency volume features.
• Employed a two-stage TL process: initial domain-specific training followed by ensemble retraining with adjusted weights via softmax scores.

Main Results:

• The proposed method effectively distinguishes low-level leaks from noise.
• Achieved accurate leak detection even with very limited training data.
• Demonstrated superior performance compared to existing techniques in diverse environments.

Conclusions:

• The developed TL-based ensemble model offers a generalized solution for AI-driven leak detection in power plants.
• This approach enhances the applicability of AI in industrial settings by addressing data scarcity and environmental diversity.
• The technique provides a significant advancement in ensuring the operational integrity of power plants through reliable leak identification.