Statistical characterization of semi-supervised neural networks for fault detection and diagnosis of air handling units

Abstract

Air handling units have been widely adopted in modern buildings for indoor air regulation and circulation tasks. The accurate and reliable fault detection and diagnosis (FDD) of air handling units (AHU) are of great significance to maintain indoor environment while ensuring the energy efficiency in building operations. Data-driven FDD approaches have gained great popularity due to their excellence and flexibilities for practical applications. Given sufficient labeled data, existing studies have validated the value of various supervised learning algorithms for FDD tasks. However, it can be very challenging, expensive, time-consuming and labor-intensive to obtain data labels for faulty operations, making it impractical to fully realize the potential of advanced supervised learning algorithms. To tackle this problem, this study proposes a novel semi-supervised FDD method using neural networks. The method adopts the self-training strategy for semi-supervised learning and has been tested for two practical applications, i.e., fault diagnosis and unseen fault detection. A number of data experiments have been conducted to statistically characterize the influence of key learning parameters, including the labeled data availabilities, the maximum semi-supervised learning iterations, the threshold and learning rate for pseudo-label data utilization. The results indicate that the method can effectively enhance model generalization performance by utilizing large amounts of unlabeled data. The insights obtained are helpful for developing advanced data-driven tools for smart building system fault detection and diagnosis.