A Decentralized Boltzmann-machine-based fault diagnosis method for sensors of Air Handling Units in HVACs

Abstract

As a key module in a Heating, Ventilation, and Air Conditioning (HVAC) system, an Air Handling Unit (AHU) is controlled based on information collected by sensors to satisfy human thermal comfort and air quality requirements. Fault diagnosis is critical since it allows maintenance crews to know which faults have occurred to improve system availability. However, fault diagnosis in AHUs is challenging because of the following reasons. First, widely used fault indicators are correlated with changing environments, e.g., weather dynamics or occupants, thus may not be enough to distinguish faults. Second, existing decentralized fault diagnosis methods developed for sensors require solving many optimization problems, leading to high computational requirements. To overcome these challenges, this paper develops a decentralized Boltzmann-machine-based method. To address the first issue, residuals between actual values of several sensor readings and their estimates are considered as fault indicators since they are not related to changing environments. To address the second issue, a novel decentralized voting mechanism is developed based on the convergence characteristic of the Boltzmann machine to locate sensor faults while avoiding solving many optimization problems. However, the established Boltzmann machine usually has an asymmetric weight matrix, and thus it does not converge to the state estimates of sensors. To guarantee convergence, a new symmetrization method is developed to symmetrize the Boltzmann machine by adding an extra unit into the Boltzmann machine to reset the weight matrix while retaining the original voting. Experimental results demonstrate that our method can effectively diagnose sensor faults with high diagnostic accuracy.