Small-sample fault diagnosis of the dual-cycle dehumidifier based on its operating mechanism

Abstract

The dual-cycle dehumidifier is an emerging system integrating refrigeration and dehumidification via an additional cooling coil and plate heat exchanger (PHE), with broad application prospects in data centers and commercial buildings. However, its complex structure and limited fault samples create bottlenecks for fault diagnosis. To address these challenges, this study investigates operational parameters under seven typical faults based on thermodynamic principles and system operating mechanisms. Six mechanism-driven diagnostic features are constructed. Four mainstream classifiers including random forest (RF), support vector machine (SVM), convolutional neural network (CNN), and deep neural network (DNN) are adopted to assess the features. The RF model yields the highest accuracy of 95.09 % and 97.68 % at training set proportions of 10 % and 20 %, respectively. The other three models all attain over 93 % accuracy at the 20 % training ratio, verifying feature compatibility across diverse classifiers. Feature ablation experiments show the two-feature subset (refrigerant mass flow rate and dehumidification amount) yields 94.04 % accuracy. When introducing single-sensor interference at a signal-to-noise ratio (SNR) of 20 dB, the average accuracy remains at 92.99 %, demonstrating the presented method maintains stable diagnostic capability. The proposed small-sample fault diagnosis method helps enhance equipment operational reliability and may facilitate wider application of dual-cycle dehumidification systems.