Fault detection and diagnosis for chillers and AHUs of building ACMV systems
Date of Issue2017-06-12
School of Electrical and Electronic Engineering
Building worldwide contributes to 40\% of the global energy consumption, and most of that is due to Heating, Ventilation, and Air-conditioning (HVAC) systems. A large part of this energy is wasted because of poor maintenance, inevitable degradation, and improperly controlled equipment. Therefore, it is of practical relevance and significance to study Fault Detection and Diagnosis (FDD) techniques for smart buildings aiming at saving energy and offering more comfortable dwelling environment. Researchers have been tackling building FDD task with a wide variety of techniques, such as analytical model-based, signal-based and data-driven methods. Recently the data-driven method has shown its advantage in dealing with complex systems with random penetrations. Most of the existing works tend to formulate the data-driven FDD as a pure fault types identification task. Problems such as severity levels identification, inter-dependence information incorporation, and essential features selection have long been ignored. This dissertation addresses the aforementioned problems, and the details are summarized as follows. First of all, the building FDD task is directly formulated as a multiple classification problem. A Discriminant Analysis-based Fault Classification (DAFC) method is driven to conduct the detection and diagnosis. Linear Discriminant Analysis (LDA) is firstly adopted to project the high dimensional data into a lower dimensional space so as to achieve optimal class separation and maximum original information maintenance. Derived from the K-means Clustering, DAFC classification applies two criteria to make a decision. The testing data set is classified to a certain cluster if: 1), it is the closest to that cluster by Manhattan distance; 2), Manhattan distances between the testing data set and that cluster are within a certain range. By feeding the training and testing data to DAFC, fault type is diagnosed at the first stage, and the corresponding severity level is identified at the second stage. The proposed two-stage data-driven FDD strategy is validated by the experimental data collected by the ASHRAE Research Project 1043 (RP-1043). Results show that it can detect and diagnose chiller faults and the corresponding severity levels effectively. Although the two-stage FDD strategy generates satisfactory results, it only works well when the number of included classes is small. Formulating the FDD task as a pure multiple classification problem is not effective enough when the number of included classes becomes large. Thus, a Tree-structured Fault Dependence Kernel (TFDK) method is proposed to identify fault type as well as fault severity level in a unified large margin learning framework. TFDK adopts structured labeling to incorporate the inter-class dependence information and deals with the streaming data with a corresponding on-line learning algorithm. As an improvement of traditional classification methods, it encodes the dependence information in its feature mapping and takes regularized misclassification cost as the learning objective. Similarly, following the ASHRAE Research Project 1043 (RP- 1043), TFDK is applied to solve the FDD for a 90-ton centrifugal water-cooled chiller. Experimental results show that compared to conventional classification methods, TFDK can significantly improve the FDD performance and recognize the fault severity levels with high accuracy. Lastly, previous works have justified that buildings and their operation can greatly benefit from rich and relevant data sets. More specifically, data has been analyzed to detect and diagnose system and component failures that undermine energy efficiency. Among the vast amount of measured information, some features are more correlated with the failures than others. However, there has been little research to date focusing on determining the types of data that can optimally support FDD. Thus, a novel optimal feature selection method, the Information Greedy Feature Filter (IGFF) method, is proposed to select essential features. On the one hand, the selection results would serve as a reference for configuring sensors in the data collection stage, particularly when the measurement resource is limited. On the other hand, with the most informative features selected by IGFF, the performance of building FDD could be improved and theoretically justified. A case study on Air Handling Unit (AHU) FDD is conducted based on the ASHRAE Research Project 1312 (RP-1312). Numerical results show that compared with several baselines, the FDD performances of conventional classification methods are greatly enhanced by IGFF. In summary, this dissertation studies the data-driven techniques and proposes several effective strategies to solve the FDD problem for building chillers and AHUs. Compared with previous works, the proposed DAFC can identify the fault severity level at a second stage after fault types have been diagnosed. This dissertation also focuses on recognizing both fault types and the corresponding severity levels in a unified learning framework. Hence, the inter-class fault dependence information is included with tree-structured labeling by the proposed TFDK algorithm. Besides, by selecting essential subsets of variables that are more correlated to faults with the proposed IGFF algorithm, not only the FDD accuracy is improved, but also the FDD application becomes more convenient and practical.