A two-stage fault diagnosis strategy for air handling units via a backpropagation multidimensional Taylor network fitter and a novel statistical process control

Abstract

Faults in Heating, Ventilation, and Air Conditioning Air Handling Units can lead to increased energy consumption and failure to meet human comfort standards. Accurate and prompt fault diagnosis of Air Handling Units is therefore critical. Although widely-used deep learning methods have shown high precision in fault diagnosis, they present two main drawbacks: 1) deep learning approaches are often “black box” models lacking interpretability and rely on labeled data; 2) their complex structures entail substantial computational resources, which may not be conducive to timely fault diagnosis. To overcome these challenges, this paper introduces a method for diagnosing faults in air handling units that combines a novel statistical process control strategy, a backpropagation multidimensional Taylor network fitter, and a D-matrix approach. The backpropagation multidimensional Taylor network, using a polynomial network with only one hidden layer, approximates nonlinear functions. Compared to deep learning models, it maintains high accuracy with a simpler structure, enabling rapid fault diagnosis. Additionally, the novel statistical process control strategy employs the Markov property for fault detection, characterized by low false alarm rates, interpretability, robustness, minimal labeled data requirements, and suitability for real-time monitoring. Experimental results validate the effectiveness of our method in diagnosing faults in air handling units, achieving an average precision of 99.9%, surpassing several commonly used techniques.