Semi-supervised learning for early detection and diagnosis of various air handling unit faults

• Published in: Energy and buildings Vol. 181; pp. 75 - 83
• Main Authors: Yan, Ke, Zhong, Chaowen, Ji, Zhiwei, Huang, Jing
• Format: Journal Article
• Language: English
• Published: Lausanne Elsevier B.V 15.12.2018; Elsevier BV
• Subjects: Air handling unit; Artificial intelligence; Diagnosis; Diagnostic systems; Fault detection; Fault detection and diagnosis; Fault diagnosis; HVAC; Learning algorithms; Machine learning; Semi-supervised learning; Support vector machine; Training; Support vector machine; Semi-supervised learning; Fault detection and diagnosis; Air handling unit
• DOI: 10.1016/j.enbuild.2018.10.016
• ISSN: 0378-7788

•The work introduces a novel semi-supervised approach to detect and diagnose faults for AHUs.•80% accuracy rate is reached using a training set with 8000 normal samples and only around 30 samples for each fault type.•This work addresses the tradeoff between the initial number of faulty samples and the final classification accuracy.•This work addresses the tradeoff between the initial number of faulty samples and the computational cost.•This work addresses the tradeoff between the threshold of confidently levels and the final classification accuracy. Modern data-driven fault detection and diagnosis (FDD) techniques show impressive high diagnostic accuracy in recognizing various air handling units (AHUs) faults. Most existing data-driven FDD approaches simply adopt supervised machine learning techniques that presume the availability of a sufficient number of faulty training data samples. However, in real-world AHU FDD scenarios, the number of faulty training samples is not enough to support supervised learning methods, since faults are usually fixed within short periods of time. In this study, a semi-supervised learning FDD framework is proposed to deal with the above problem. By using the proposed framework, the training pool can be enriched by iteratively inserting confidently labeled testing samples, which mimics the scenario of detecting faults the earliest possible. Furthermore, the proposed framework can be easily extended with various kinds of state-of-art classifiers. Three important tradeoffs are observed through a series of experiments. With a reasonably small number of faulty training data samples available, the performance of the proposed semi-supervised learning technique is comparable to the classic supervised FDD methods.