Deep learning based on image analysis for refrigerant charging and leakage detection in building heat pump

Abstract

In heat pump systems, refrigerant leakage and charging faults are the common issues. Diagnosing refrigerant leakage and charging faults of the heat pump systems are crucial for reducing system energy consumption and maintaining stable high-efficiency operation. With the iteration of computing technology, data-driven approaches play an important role in fault detection and diagnosis. This research introduces a novel algorithm that transforms one-dimensional data into image format using Gramian Angular Field (GAF) and optimizes hyperparameters through Triangular Topology Aggregation Optimization (TTAO) within a Parallel Convolutional Neural Network (PCNN). Additionally, the approach integrates a Multi-head Self-Attention Mechanism (MSA) and employs a Support Vector Machine (SVM) in lieu of a Softmax layer for enhanced fault detection efficiency. A dataset for refrigerant leakage and charging faults was created using a Water-to-water Heat Pump (WWHP) test bench, providing the basis for evaluation against innovative algorithm and three existing algorithms: SVM, CNN-SVM, and PCNN-MSA-SVM. The findings highlight that TTAO successfully optimized the solution, minimizing the adaptation value from 0.167 to 0.025. The iterative process consistently demonstrated low loss values and steady accuracy improvements, trending towards enhanced stability. The proposed algorithm significantly outperformed the compared methods, achieving an impressive 97.5% accuracy rate and enhancing fault detection by 34.2%, 9.2%, and 4.2% respectively. Moreover, it showed robust and uniform F1 Scores across different fault types, marking an average increase of 42.0% over traditional SVM. This methodology not only optimizes hyperparameters adaptively but also identifies the best parameter settings, improving algorithm performance substantially.