استراتژی FDD مبتنی بر شبکه های بیزی برای پایانه های حجم هوای متغیر

This paper presents a diagnostic Bayesian network (DBN) for fault detection and diagnosis (FDD) of variable air volume (VAV) terminals. The structure of the DBN illustrates qualitatively the casual relationships between faults and symptoms. The parameters of the DBN describe quantitatively the probabilistic dependences between faults and evidence. The inputs of the DBN are the evidences which can be obtained from measurements in building management systems (BMSs) and manual tests. The outputs are the probabilities of faults concerned. Two rules are adopted to isolate the fault on the basis of the fault probabilities to improve the robustness of the method. Compared with conventional rule-based FDD methods, the proposed method can work well with uncertain and incomplete information, because the faults are reported with probabilities rather than in the Boolean format. Evaluations are made on a dynamic simulator of a VAV air-conditioning system serving an office space using TRNSYS. The results show that it can correctly diagnose ten typical VAV terminal faults.

Variable air volume (VAV) air conditioning systems are widely used in offices and commercial buildings nowadays. Building professionals usually consider that VAV systems have better performance in terms of thermal comfort and energy saving than fan coil unit systems and constant air volume systems. However, VAV terminals easily suffer from various faults which cause the performance of VAV systems to hardly meet the high expectations. Qin and Wang found that 20.9% of 1251 VAV terminals were ineffective in a site survey conducted in a commercial building in Hong Kong [1]. Preventive maintenance of VAV terminals is a difficult task since a large number of VAV terminals are installed above ceilings. Fault detection and diagnosis (FDD) tools for VAV terminals are essential for reliable indoor environment control, saving maintenance efforts, and eliminating the associated energy waste. There was little research conducted on FDD of VAV terminals in the last decades. Yoshida proposed an automatic regressive exogenous (RARX) model and an extended Kalman filter model to detect faults in a VAV unit and an air handling unit (AHU) cooling coil system [3] and [4]. Seem et al. described a set of indexes to assess the performance of control loops and to detect faults in VAV terminals and AHUs [5] and [6]. The performance indexes were embedded in commercial VAV terminal controllers to quickly identify terminals that were not operating correctly. Schein proposed VAV box performance assessment control charts (VPACC) to assess the performance of pressure independent VAV boxes with hydronic reheat coils. VPACC introduced a small number of CUSUM charts to accumulate the error between a process output and the expected value of the output [[7] and [8]]. These FDD methods for VAV terminals focused on fault detection and seldom considered fault diagnosis. Qin and Wang proposed a hybrid approach to diagnose ten typical faults in VAV systems utilizing expert rules, performance indexes and statistical process control models [1]. Principal component analysis (PCA) was used to detect flow sensor biases. Wang et al. designed a rule-based classifier consisting of twenty expert rules and fault isolation algorithms to diagnose fifteen faults in VAV terminals [9], which was able to diagnose faults using sensor measurements and control signals which are commonly available in building management systems (BMSs). The above FDD methods for VAV terminals can normally provide good results; however, they rarely considered the realistic situation where only uncertain and incomplete information is available for conducting FDD. Uncertainties widely exist in measurements, fault symptoms, fault–symptom relationships, expert knowledge, FDD results, etc. For instance, different faults may cause similar fault symptoms and a fault may exist with certain probability when a symptom is observed. Therefore, it is more reasonable to give the probabilities of faults at given symptoms in FDD results. However, most existing FDD methods report the FDD results in the Boolean format, i.e. Yes/Faulty and No/Normal. In addition, due to the limited number of instruments, incomplete records of system design and operation data, insufficient memory capacities of control stations and building automation systems, etc., the information available for conducting FDD is usually incomplete. Using incomplete information for FDD is also a big challenge for most existing FDD methods. On the other hand, some useful information, which is very helpful for FDD, was often overlooked. For instance, the prior probabilities of the temperature sensor fault and the damper actuator failure are 25.3% and 3.8% respectively, according to the survey [1]. When a VAV terminal is abnormal, the possibility of the temperature sensor fault is much higher than that of the damper actuator failure. Such prior experience or knowledge about faults has seldom been used by existing FDD methods. FDD experts have already recognized that FDD of VAV terminals is quite challenging. The major reasons are listed as follows. Firstly, there are generally very few sensors installed in VAV terminals. The information is extremely insufficient which makes it difficult to diagnose the faults [1]. Secondly, faults may propagate by control loops, which lead to complex relationships between faults and symptoms. Thirdly, limitations associated with controller memory and communication capabilities further complicate the task [8]. Fourthly, the number of different types of VAV boxes and lack of standardized control sequences add extra complexities [8]. Lastly, there is almost no preventive maintenance due to the large number of VAV terminals installed above ceilings [9]. Although FDD of VAV terminals is challenging, it is interesting to see that domain experts can always find the sources of faults. Experts diagnose faults using as much useful information as possible, e.g. fault symptoms (from BMS or on-site test), configurations (e.g., control strategy, set-points), performance of peer VAV terminals in the same zone or similar zones, cooling load, as well as the experts' knowledge/experience. Fault diagnosis of VAV terminals may be more efficient and effective if the FDD methods can work in a similar way as that used by FDD experts. In this study, an intelligent method for FDD of VAV terminals is proposed to simulate the diagnostic thinking of experts based on Bayesian belief network (BBN) theory.

This study develops a robust DBN-based FDD strategy based on probability analysis and graph theory for VAV terminals. The strategy is evaluated using simulation tests in which ten typical faults of VAV terminal are introduced. All faults are correctly diagnosed with high beliefs. The diagnostic Bayesian network developed is composed of three types of nodes, i.e. fault nodes, BMS evidence nodes and additional information nodes. Additional information nodes usually need manual inputs; however, they are only used when the BMS evidence nodes cannot isolate the fault. A recommended manual checklist will be produced in this circumstance. Generally speaking, the DBN is an open system, which means that new fault nodes and evidence nodes can be conveniently added into the network without significantly revising the current structure. The prior probabilities and the conditional probabilities need to be assigned by the experts, which may be subjective to some extent. This study also analyzes the sensitivities of the DBN-based FDD strategy to the pre-assigned probabilities. The results show that small variations in the probabilities won't change the FDD results so long as the qualitative probability relationships between the states of one node (prior probabilities) or between the faults and symptoms (conditional probabilities) are correct, i.e. assign a large value to the high probability event and assign a small value to the low probability event. Rules play important roles in the DBN-based strategy which makes it similar to the widely used rule-based FDD methods. However, faults are reported with probabilities in the DBN-based strategy rather than in the Boolean format (i.e. Normal or Faulty). The performance of the DBN-based strategy under the circumstance that only incomplete information is available for FDD is better than that of the rule-based methods. Meanwhile, the DBN-based strategy is tolerant with various uncertainties, such as measurement noises, considering that uncertainties are usually low probability events. The DBN developed in this study is applicable to the pressure-independent VAV terminals operating at the cooling mode. If the control strategy changes, some measurements may not be available which causes the rules to become inapplicable. For example, the supply air flow rates are rarely needed for the pressure-dependent VAV terminals, but they are used in the DBN developed in this study. However, the method for developing the DBN presented in this study is suitable for developing DBN for other types of VAV terminals and the heating modes. The DBN-based strategy can be used for on-line and off-line FDD of VAV terminals. Further validation in field application can improve its performance.