A fault detection and diagnosis strategy of VAV air-conditioning systems for improved energy and control performances

Energy and Buildings 37 (2005) 1035–1048 www.elsevier.com/locate/enbuild A fault detection and diagnosis strategy of VAV air-conditioning systems for...
Author: Suzanna Bailey
9 downloads 0 Views 916KB Size
Energy and Buildings 37 (2005) 1035–1048 www.elsevier.com/locate/enbuild

A fault detection and diagnosis strategy of VAV air-conditioning systems for improved energy and control performances Jianying Qin, Shengwei Wang * Department of Building Services Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong Received 19 October 2004; received in revised form 8 December 2004; accepted 24 December 2004

Abstract This paper presents the results of a site survey study on the faults in variable air volume (VAV) terminals and an automatic fault detection and diagnosis (FDD) strategy for VAVair-conditioning systems using a hybrid approach. The site survey study was conducted in a commercial building. 20.9% VAV terminals were ineffective and 10 main faults were identified in the VAV air-conditioning systems. The FDD strategy adopts a hybrid approach utilizing expert rules, performance indexes and statistical process control models to address these faults. Supported by a pattern recognition method, expert rules and performance indexes based on system physical characteristics are adopted to detect 9 of the 10 faults. Two pattern recognition indexes are introduced for fault isolation to overcome the difficulty in differentiating damper sticking and hysteresis from improper controller tuning. A principal component analysis (PCA)-based method is developed to detect VAV terminal flow sensor biases and to reconstruct the faulty sensors. The FDD strategy is tested and validated on typical VAVair-conditioning systems involving multiple faults both in simulation and in situ tests. # 2005 Elsevier B.V. All rights reserved. Keywords: Variable air volume system; VAV terminal; Fault detection and diagnosis; Commissioning; Principal component analysis

1. Introduction Variable air volume (VAV) air-conditioning system, which is deemed more economical than other alternative systems, has been widely adopted in the buildings to maintain the cooling and heating demands. However, in complex VAV systems, faults at system level, subsystem level, component level, control and sensor level would not only reduce the economic benefits of the system but also lead to occupant discomfort. Though the benefits for fault detection and system improvement are difficult to quantify, the potential savings out of faulty and non-optimal operation of HVAC system alone in commercial buildings were estimated to be 20–30% [1]. Faults typically found in air-conditioning systems are due to improper design, application or operation of the systems [2]. A survey was conducted to sort out the top 10 faults of * Corresponding author. Tel.: +852 27665858; fax: +852 27746146. E-mail address: [email protected] (S. Wang). 0378-7788/$ – see front matter # 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.enbuild.2004.12.011

air-conditioning systems by collecting information from professionals [3]: (1) poor air quality, (2) water leakage, (3) room air temperature deviation due to excessive heat generation, (4) room air temperature deviation due to inadequate air-flow rate, (5) too much or less air volume of VAV unit, (6) excessive pressure difference across an air filter, (7) abnormal noise or vibration, (8) room air temperature deviation due to inadequate positions of diffusers, (9) false opening signal to a VAV unit control and (10) room air deviation due to insufficient water flow rate. Further investigation revealed that mechanical faults (such as coil and damper malfunction) were common also [3]. Fault detection and diagnosis (FDD) has been approved to be an essential and efficient supporting tool in fixing faults timely and reducing the impacts of them in building HVAC applications. However, previous FDD studies focused on the major equipment of the systems, such as air handling units (AHUs), fans and local feed water pumps. Various systems and methods were studied and developed. Lee et al. [4]

1036

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

Nomenclature A B e F DF G H P P DP R SPEa S t T Ta2 TF X y DZ

network relationship matrix (for nodes) basic circuit matrix (for circuits) process error air flow rate air flow variation vector of flow rate (G1, G2, . . ., Gn) vector of pressure increments caused by fan in the branch (H1, H2, . . ., Hn) pressure principal loading matrix vector of pressure difference in the branch (DP1, DP2, . . ., DPn) reversal threshold of SPE vector of resistance coefficient (S1, S2, . . ., Sn) time temperature threshold of T2 statistic total flow variable matrix process variable vector of pressure difference caused by altitude difference in the branch (DZ1, DZ2, . . ., DZn)

Greek letters a weighting factor L matrix of eigenvalues m control signal to terminal damper s standard deviation Subscripts and superscripts a number of PCs delt permissive period F air flow rate k kth variable max maximum min minimum new new samples set set-point st static th threshold ^ estimated output on the score space – estimated output on EWMA

generally described 11 faults of a system from fan failure to sensor failure and the use of a two-stage artificial neural network for fault diagnosis in a simulated AHU. Results demonstrated that the recovered estimate of the supply air temperature could be used in a feed-back control loop to bring the supply air temperature back to the set-point value. Dexter et al. [5–8] concentrated on coil heat exchange

process of AHUs and analyzed five faulty modes: fouled coil, valve leak, valve stuck closed, valve stuck midway and valve stuck open. They designed a robust fuzzy model for AHU fault diagnosis accounting the temperature sensor error. The model-based approach was successfully applied in a fault diagnosis scheme to remotely commissioning AHUs in a commercial building. Norford et al. [9,10] investigated both abrupt and degradation faults on several sections of three existing AHUs and proposed two methods for FDD, i.e. physical model-based method and grey box method. Both methods detected nearly all of the faults in the two matched AHUs. House et al. [11] studied several faulty cases of an AHU. Five classifiers, i.e. ANN classifier, nearest neighbour classifier, nearest prototype classifier, rule-based classifier and Bayes classifier, were tested for detecting and diagnosing seven faults of a simulated AHU system. They also proposed an expert rule set with 28 simple rules for fault detection in AHUs [12]. Field trials of the expert rule set successfully identified two occurrences of faults with mixing box dampers while they pointed out that the effort devoted to developing diagnostic capabilities for VAV boxes had been limited in comparison to AHUs and other types of HVAC equipment [13]. Researchers also paid more attention on subsystems in the recent years. Katipamula et al. [14,15] noticed that a failure of the economizer might go completely unnoticed. They designed an ‘outdoor-air/economizer diagnostician’ to monitor the performance of AHUs and automatically detect problems with economizer operation or ventilation problems for systems without economizers using decision tree method. Dodier et al. [16] particularly studied on fanpowered mixing box for both damper failure and power failure. The probabilistic inference methods were adopted in real-time diagnostic system (RTDS). The results of applying the RTDS to HVAC laboratory data were presented. The tests indicated that application of this inference system to the diagnosis of mixing box failures yielded encouraging results. Wang and Chen [17] paid particular attention to the air flow sensor failure of sensor-based demand control ventilation systems. They pointed out that fault-tolerant control for outdoor ventilation air flow rate based on neural network was applicable. As overall system reliable control relies on proper works of every component, few researchers began to particularly focus on VAV terminals and valves. Seem et al. [18] looked into VAV terminal on-line control recently. Two indexes were calculated from building management system (BMS) driven data for VAV box on-line monitoring and fault detection. A diagnostic method using a small number of cumulative sum statistical quality control charts to assess the performance of VAV boxes was proposed by Schein and House [19] afterwards. It had been embedded in commercial HVAC controllers and successfully tested using emulation and laboratory data. In the study, the faults detected are focused on damper stuck and oscillation. Multiple conflicting faults had not been identified. Yoshida et al. [3,20,21]

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

intensively worked on VAV damper failure recently and tested their approach on both sudden and consecutive faults. RARX modeling was used taking commercially available variables as the input. By making artificial faults in the tested system, it was demonstrated that the RARX method was robust and could even detected different faults including damper getting stuck at half open position. McGhee et al. [22] summarized typical failures in valves and actuators and classified faults in process valves and actuators. ANN method applied to valve process diagnosis was validated by experimentation and simulation test. Han et al. [23] presented an overall model-based FDD system to attack the problems on many levels of abstraction: from the signal level, controller programming level, system component level, all the way up to the information and knowledge processing level. This FDD strategy was under the assumption that the sensors in the system are reliable and fault free. However, the sensor faults, which could be undetected by the simple pre-test, would make the FDD rules malfunctioning. As VAV terminals serve the end users, their performance have significant effects on the environmental quality provided by HVAC system and the energy efficiency of buildings. Literature survey shows that study on the faults of VAV terminals is far from sufficient, particularly concerning the system VAV integrating a large number of VAV terminals. Most significant technical problem perceived in VAV systems is interaction among VAV units equipped with a control loop, where information exchange takes place between several control strategies [20]. This interaction must be carefully analyzed and measured for achieving optimal control, and therefore, in development of any FDD techniques. In this study, a full scale site survey on all the VAV terminals in a large commercial building was conducted summarizing 12 faults/symptoms, and further investigation on the suspected terminals summarized 10 root faults in the VAV terminal systems. On this basis, a FDD strategy is developed using a hybrid approach to deal with multiple VAV faults as an online commissioning tool. Expert rules and performance indices are adopted in the strategy. Principal component analysis (PCA) method is used for VAV terminal flow sensor bias detection and reconstruction. This paper presents the typical root faults in VAV air distribution systems as discovered in the site study, the FDD strategy for VAV air distribution systems and the validation of the strategy in simulation and in situ tests.

2. Site study—faults of VAV terminals and their root causes

1037

out in a commercial building located in Hong Kong, which is a 39-storey building completed in 1995. Pressure-independent VAV terminals (under cascade control) are used in the building. Totally, there are 1251 VAV terminal in the building. All the VAV terminals were re-commissioned and investigated in 2002. The building employs a fully automated BMS, which performs the environmental control of the indoor spaces. The AHUs and VAV systems provide ventilation, cooling or heating, as needed, throughout the year. Four variables are involved in the control of each pressure-independent VAV terminal: (i) zone temperature, (ii) zone temperature set-point, (iii) required air flow rate (or air flow set-point), (iv) measured air flow rate. Due to the limitation of BMS hardware, these four measurement and internal variables for the VAV terminal of three floors were logged at 5-min interval for 3 days simultaneously. The recommissioning and investigation of each VAV terminal was based on the trend data of 3 days. By investigating the operation of the VAV terminals, it was found that very often the measured VAV airflow could not approach the flow set-point and the space temperature could not approach the set-point. General screening got 261 ‘suspected’ VAV terminal out of 1251, which was 20.9% of the total terminals in the building. All ‘suspected’ VAV terminals were further examined and verified by the technical staff over 14 days. Such detailed checking found 12 faults/symptoms among the ‘suspected’ VAV terminal identified as listed in Table 1. The consequences of faults could be classified into four categories related to: poor environment, waste of energy, unreachable design value and physical damage [3]. The observations of this survey study match with the conclusion of the survey on AHU system faults conducted by Yoshida, which covered a wider range of faults from design faults to user-level faults in AHU and VAV systems. Concerning VAV terminals, their survey revealed that zone air temperature deviation and local DDC error were common, which was in line with our investigation results. Table 1 Summary of VAV terminal faults/symptoms No. 1 2 3 4 5 6 7 8 9

2.1. Faults identified in VAV terminals

10

The site study, along with the VAV system recommissioning involving all VAV terminals, was carried

11 12

Faults and symptoms

No.

Percentage (%)

Temperature sensor error DDC error Diffuser damper closed as requested by tenants Design flow too large VAV boxes dismantled by tenant Damper actuator failure Part of diffuser being wrapped by adhesive tape Temperature set-point too low Abnormal space temperature requested by tenants Temperature sensor located too close to VAV diffuser outlet Too many people in a room VAV box not accessible (cause unknown)

66 46 41

25.3 17.6 15.7

28 14 10 9

10.7 5.4 3.8 3.5

7 5

2.7 1.9

4

1.5

1 30

0.4 11.5

1038

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

2.2. Root faults of VAV air distribution systems In most cases, faults and their symptoms are mixed up. One fault may have different symptoms and one symptom may be associated with a few faults. The above 12 faults/ symptoms discovered in the site study could actually be either causes or symptoms. If investigating each case deeply, we found that the causes could be divided into two categories: mechanical failures (damper failure; diffuser wrapped, VAV terminal under/over capacity) and sensor/ controller failures (temperature/flow sensor failure; PI controller failure). Detailed site investigation found that there were many different causes related to various faults in VAV terminal. However, for the convenience and efficiency of commissioning and fault diagnosis, it is desirable to classify the causes of faults into a few general categories. Based on the actual physical cause of the faults discovered on the VAV terminal systems investigated, 10 root faults are summarized for the pressure-independent VAV terminal systems as listed below.          

Fault 1: Zone temperature sensor reading frozen Fault 2: VAV terminal under/over capacity Fault 3: VAV damper stuck Fault 4: VAV flow sensor reading frozen Fault 5: VAV flow sensor reading deviation to minimum/ maximum Fault 6: Poor tuning of static pressure control loop Fault 7: VAV controller failure Fault 8: VAV damper sticking Fault 9: VAV damper hysteresis Fault 10: VAV flow sensor bias

Zone temperature sensor bias would not be detected by the system characteristics and might be offset by adjusting the zone temperature setting. Thus, this fault is not included in this study. The same serial numbers of faults are used in the FDD study presented below.

3. FDD strategy for multiple faults in VAV distribution systems 3.1. Basic system and hierarchy of FDD strategy The study focuses on the VAV air distribution systems consisting of the typical pressure-independent VAV terminals employing cascade control loops as shown in Fig. 1. The pressure-independent flow controller is widely used in large complex VAV air-conditioning systems due to better control stability and faster response in load changes. It is achieved by employing two control loops separating the flow set-point reset based on the deviation of the measured air temperature in the occupied space from its set-point and flow control based on the deviation of the measured VAV air flow rate from its set-point. It allows the effect of fluctuations in the pressure supply, as the results of disturbances from other parts of the systems on the space temperature control, is eliminated as the inner loop response quickly to these fluctuations before they affect the space temperature control. Fig. 2 illustrates the hierarchy of the FDD strategy. Ten faults are dealt with by relevant FDD schemes in serial sequence at progressive steps presented in the figure, while Faults 2 and 3 are treated in parallel in Step 2 and Faults 7–9 are treated in parallel in Step 6. In this strategy, sequential arrangement is employed considering the interaction amongst the faults according to rule-based reasoning. The seven progressive steps are therefore arranged in an order that the FDD ability at the preceding steps would not be affected by the faults to be dealt at the later steps. Once the system is indicted to be fault free at preceding steps, the following steps begin their FDD process. The FDD schemes, relevant to each step in the FDD strategy, are briefed in the next session. 3.2. FDD schemes In this study, expert rules and performance indexes are selected to build the FDD schemes for Steps 1–6. For fault

Fig. 1. Control loops of pressure-independent VAV terminals.

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

1039

Fig. 2. Hierarchy of VAV system FDD strategy.

detection of VAV terminal flow sensor bias (Step 7), PCA method is applied to model the VAV systems by capturing the correlations between the measured variables. The VAV terminal damper openness is an important basis to build up the expert indexes, which are key elements for FDD. However, in normal pressure-independent VAV systems, the signal of damper openness is not available. For position algorithm controller, the control signal to damper (m) typically represents the position of an actuator and therefore the openness of the VAV damper [24]. Therefore, m is used to represent the damper openness in the FDD strategy developed in this study. To eliminate the effects of system dynamics and ensure the reliability of the measurements used, the measurements have to go through a filter before they are used for fault detection and diagnosis. The filter is constructed on the basis of exponential weighted moving average (EWMA) techniques for the calculation of indexes as shown in Eq. (1). y¯ k ¼ a¯yk1 þ ð1  aÞyk

(1)

The sensitivity of the filter is determined by the weighting factors a. A high value provides secure output, but reduce the tool’s capability to detect when the system is in transient state. Montgomery [25] describes the use of EWMA control charts and claims that the weighting factor between 75 and 95% works well in practice. A factor of 90% is used in this strategy. However, to increase the sensitivity, we choose 50% for the reversal counts in case of oscillation. Reversal counts are used in the strategy as one of the basis performance indexes, which is illustrated in a univariate statistical control chart (Fig. 3), where s is the standard deviation of the noise inherent in a process variable. The reversal counting starts when the process variable exceeds the threshold (R = 1) and one more reversal is counted (R = R + 1) once the variable exceeds the threshold at the opposite direction. In most circumstances, the maximum tolerable number of reversals is chosen to be 4 [26].

Fig. 3. Illustration of reversal counts.

Besides, the sensor reading frozen could be completely frozen at a fixed figure (Case 1) or floating within a certain range (Case 2) as shown in Fig. 4. The frozen of Case 2 is confirmed by the cumulative sum (CUSUM) control method that the CUSUM of the variable is within a certain limit (5s, 5s) [25]. In both applications of statistical control for reversal counting and sensor frozen detection, the control limits should be determined carefully. 3.2.1. Step 1—Fault 1 When the zone temperature sensor readings are frozen, the zone temperature could not be measured correctly and the VAV subsystem would be totally out of control eventually. Without resetting the zone temperature setpoint, the flow set-point calculated by the controller would be fixed within certain range. For instance, the flow set-point is usually fixed at the maximum or minimum if the fixed temperature reading is not equal to the set-point. The zone temperature sensor fault is detected when both zone

Fig. 4. Illustration of sensor reading frozen.

1040

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

temperature reading and the flow set-point are confirmed frozen over certain period (i.e. preset period, a threshold in time scale) by the above-mentioned 3s and CUSUM statistical control methods. For common VAV terminals, 3s for temperature was set at 0.3 8C and 3s for flow was set at 10 l/s. As the cooling load of a zone is seldom kept unchanged for half a day in practice, the preset period (tdelt) was set as 5 h. The space temperature changes significantly during morning pull-down period after switch on the system. The violation of frozen limit indicates fault free in this step and trigger the FDD process of the next step. 3.2.2. Step 2—Fault 2/3 VAV terminal under/over size (Fault 2) or damper stuck (Fault 3) would prevent the subsystem from providing appropriate air flow rate into the zone, therefore, the zone temperature cannot meet its set-point. A rule as shown in formula (2) is used to detect Faults 2 and 3, which is based on checking the deficiency of the temperature control loop and flow control loop as well as no variation of the measured flow rate over a preset period (tdelt). The temperature error threshold (Te) was set at 1.5 8C and the preset period was set as 30 min.     T¯  T set > T e   T¯  T set <  T e      m ¼ mmax Dtime > tdelt And (2)  m ¼ mmin     DF¯ And  0:05  F design

Faults 2 and 3 cannot be differentiated by the strategy and it does not cause much inconvenience in application as the fault is already focused on flow control devices. The hardware failure of damper stuck (mechanically stuck or actuator failure) and terminal under/over capacity are recommended to be investigated manually by adjusting the zone temperature set-point since both faults need manual rectification eventually. Furthermore, the site survey results indicated that most cases of ‘box under capacity’ were due to illegal local adjustments, i.e. diffuser dampers closed or diffusers wrapped with adhesive tapes, which also require maintenance works on site. Moreover, the improper location of zone temperature sensor is common after renovation. The fault of terminal under capacity may relate to this fault, which should be confirmed on site as well. Besides, the alteration of set-points to out of range values or significant temperature sensor bias would lead to the same symptoms. Those faults should be isolated from Fault 2 by further site investigation. 3.2.3. Step 3—Fault 4 If the reading of a terminal flow sensor is frozen within the normal range, the zone control loops would oscillate. The zone temperature oscillates around the set-point within

a small range, which usually will not be sensed by occupants, but the terminal damper oscillation could cause the actuator to wear out prematurely. To detect this fault, the number of flow set-point reversal (R F,set) is used as the fault indicator. Fault 4 is detected when the counted number of the reversal within a preset period (30 min) is over certain limit (selected to be four times) while the flow reading is detected frozen using the abovementioned univariate statistical control method. For common VAV terminals, 3s for flow was set at 10 l/s. 3.2.4. Step 4—Fault 5 The cascade control of VAV terminal would be totally ruined when the flow sensor reading deviates to the minimum/maximum flow of the terminal. The symptoms are that both flow set-point and measured flow are at the minimum while the zone temperature is relatively low, or on the contrary, both required flow and measured flow are at the maximum while the zone temperature is relatively high. As a VAV system consists of components and control loops, this fault might mix with other faults. Technical staff may focus on mechanical faults, such as terminal under/over capacity, and neglect the flow sensor problem in this case. Fault 5 is detected by the rule as shown in formula (3), which is based on checking the deficiency of the temperature control loop and flow control loop over the preset period. The temperature error threshold (Te) was set at 1.5 8C and the preset period (tdelt) was set as 30 min.    T¯  T set <  T e    T¯  T set > T e    m ¼ mmax   And  m ¼ mmin (3)   Dtime > tdelt F ¼ F min  And   F max    F set ¼ F min  And  F max 3.2.5. Step 5—Fault 6 The poor tuning of static pressure control loop would lead to unstable control of static pressure and therefore VAV damper premature wear and tear. Fault 6 is detected after counting excessive consecutive reversals (selected to be 20 times) of the static pressure over a preset period (10 min); 3s for static pressure was set as 10 Pa. 3.2.6. Step 6—Fault 7/8/9 Controller failure would lead to three types of unsatisfactory control performance [26]. They are unresponsive control process related to hardware failure, sluggish response related to lower gain and oscillatory behavior related to higher gain. Choiniere and Beaudoin [27] adapted the performance indexes method and demonstrated their application to detect and diagnose four main faults of the controllers: instability of set-point (i.e. flow set-point),

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

instability of output (i.e. the measured flow), temperature set-point not satisfied and airflow set-point not satisfied. Nevertheless, the root causes for the four main faults were not identified. Moreover, damper sticking or hysteresis would cause the above-mentioned faults as well [28]. A sticky damper results mainly from the increment of the static friction, which hinders the opening of the damper and causes damper oscillation. Hysteresis results mainly from slack in the actuator linkage mechanism, which gives rise to the deficiency of flow control (sluggish response and oscillation). Further rules should be added to perform a diagnosis. To identify the root cause, the secondary control loop (flow controller) is analyzed first. The reason is that the response may be misjudged as other faults if the air flow set-point is not reached because of faulty flow control, sticking or hysteresis. The control loop diagnostic method is shown in Fig. 5. As illustrated in Fig. 5, the fault of flow control saturation is detected when the flow set-point is not satisfied within the maximum permissive settling time (2 min) under the condition that m reaches its maximum or minimum. The sluggish response can be identified if the flow set-point is not satisfied within the preset period (2 min). Oscillation could be caused by both control loops due to the nature of cascade control, thus the oscillation of both control loops should be analyzed simultaneously. The oscillation in temperature control loop is certified by counting the excessive flow setpoint reversals (R 15) within the preset period (30 min). The oscillation in flow control loop could only be detected by counting the excessive flow reversals (R 20) within the preset period (30 min) under the condition that no excessive flow set-point oscillation is detected. For common VAV systems, the flow error threshold was set as 10 l/s, the flow error greater than the threshold means that the flow set-point is not satisfied. The temperature control (the primary control loop) faults can be analyzed similarly (Fig. 5). The fault of controller saturation is detected as the zone temperature set-point is not satisfied within the maximum permissive settling time (30 min) with the flow set-point reaches the minimum or maximum limit. The sluggish response can be identified if the temperature set-point is not satisfied within the preset

1041

period (30 min). For common VAV systems, the temperature error threshold was chosen as 1.5 8C. Damper sticking or hysteresis would cause oscillation or sluggish response similar to that the faulty flow controller may cause. To distinguish the mechanical faults from the improper flow controller setting, pattern recognition indexes are adopted to characterize the different response patterns. Previous studies [29,30] employed two dimensionless parameters (oscillation ratio and the close-loop response time) to characterize the close-loop response pattern. In this study, two pattern recognition indexes are designed to characterize the pattern of sluggish response and the pattern of oscillation, respectively. The pattern of sluggish response caused by the mechanical reason of hysteresis is distinguishable from the improper flow controller setting using the pattern ¯ After the recognition index designed as CUSUM (F). disturbance, the sluggish response caused by hysteresis is ¯ < 5s) identified by the control index (5s < CUSUM (F) taking the air flow rate before the disturbance as the expectation within the preset period (30 s). The pattern of oscillation caused by mechanical reasons of sticking and hysteresis is distinguished by the pattern recognition index designed as (F¯ k  F¯ k1 ). Oscillation caused by mechanical reasons is identified by the dominated points (more than 50%) of the recognition index qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  2k within the EWMA control limits 6s 1a 1þa ð1  a Þ; qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2k 6s 1a 1þa ð1  a ÞÞ [25]. 3.2.7. Step 7—Fault 10 VAV terminal flow sensor bias would not affect the normal control process if the readings are within the range as it can be compensated by resetting the air flow set-point. However, sensor soft fault would lead the bias to a certain level, which makes the reading reach the minimum/ maximum flow of the VAV terminal and ruins the control process. Early bias detection and data recovery of VAV terminal flow sensors are of great importance. As the network characteristics are related to a large number of

Fig. 5. Control loop diagnostic method.

1042

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

process variables in the system, a multivariate analysis technique, PCA, is selected as the suitable technique for sensor fault detection and data reconstruction. In FDD applications, the new observations (vector of measurements Xnew) are projected to a principal component (PC) subspace, which is determined based on the minimum VRE [31] from the training matrix, to get their PCA estimation (Eq. (4)). Both T2 statistic (Eq. (5)) and Q statistic (Eq. (6)), which is called square prediction error (SPE) as well, are used for fault detection. Generally speaking, T2 relates to process upset and SPE relates to sensor faults [32]. When faults exist, one or both thresholds would be exceeded. Contribution plot is used for multiple fault isolation. After flow sensor fault detection and isolation, sensor reconstruction is conducted to get the recovered data using the iterative approach [33]. Xˆ new ¼ X new PPT

(4)

2 T T T 2 ¼ Xnew PL1 a P X new  T a

(5)

Q ¼ SPE ¼ jjX new  Xˆ new jj2  SPEa

(6)

For VAV terminal flow sensor fault detection and diagnosis, PCA models at two levels are developed and used in serial, i.e. system level and terminal level models (Fig. 6). The system level model reflects that, in a network, the hydraulic characteristics are related to the static pressure (Pst), the damper position (m) and flow rate ( F) of all VAV terminals. As all VAV terminals are involved in the system level model, the reliability and sensitivity of fault detection and isolation may be affected by the process stability and multiple faults in the system. Therefore, terminal level PCA models are designed to further monitor on the suspicious terminal(s), which are isolated by the system level FDD. The faulty sensor is also reconstructed at the terminal level. The FDD scheme (Fig. 7) is strengthened by the recovered data and iteration of the FDD process. The process terminates until no further fault could be detected.

Fig. 6. PCA models at system level and terminal level.

4. Validation using simulation results An office building was simulated as the test facility for strategy validation. Dynamic simulation of the system with different combinations of faults provides a convenient and low cost tool in testing and evaluating the FDD startegy and schemes. 4.1. System description The building is a 46-storey commercial building located in Hong Kong. The floor under study is half a floor consisting an open plan office of about 1166 m2 usable floor area served by a central AHU with VAV systems serve the floor. There are 40 VAV dampers and over a hundred air diffusers associated with one AHU. The VAV terminals are pressure-independent VAV boxes under cascade control. The design air flow rates of the VAV system is 6 m3/s and the design VAV supply fan pressure at the location of pressure sensor is 650 Pa. Two variable blade angle fans are equipped as VAV supply and return fan, respectively. The pitch angle of the VAV supply (axial) fan is moderated to control the supply air static pressure. The return (axial) fan is used to control the ex-filtration flow rate in order to maintain positive pressure in the building. It is achieved by controlling the difference between the total supply and return air flow rates within the upper and lower limits by moderating the pitch angle of the return fan. In the simulation study, the floor area was simulated by dividing it into eight zones. 4.2. System simulation TRNSYS was used as the platform for the VAV system dynamic simulation. The models used were those developed by Wang [34] in this study except the system pressure-flow balance model. Those models include simplified building model, duct model, fan model, cooling coil model, DDC controller, sensor and actuator models. The ‘fluid flow rate and pressure calculation’ model developed by Dr. Y.X. Zhu of Tsing Hua University was modified as the system pressure-flow balance model for this simulation study. The model was designed for simulating the flow rate and pressure balance of a closed fluid network. Reference pressure is specified for calculating pressure of each node. The relationship of branches and nodes are specified as parameters. The flow resistance, fan/pump work characteristic and altitude difference of all branches are used as inputs to determine the flow rate and pressure distribution in the network as shown in Fig. 8. The parameters of VAV models to be used for simulation could be determined according to the VAV component characteristics given in manufacturer catalogues and/or empirical correlations given in handbooks. In this study, the VAV system performance data needed for determining the parameters of the VAV component models were obtained by monitoring the VAV system on site [34].

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

1043

Fig. 7. Structure of PCA-based flow sensor FDD scheme.

4.3. Simulation test results 4.3.1. Faults 1–9 Different combinations of faults were introduced in the simulations tests. As some of the results appeared straightforward, the detailed results are not presented if the patterns of the records described below are selfexplanatory. By fixing the temperature sensor output of Zone 6 at 24 8C (Fault 1), the FDD robustness of Step 1 was verified by the indexes of both fixed temperature and fixed flow setpoint over the preset period with or without other faults.

Another group of tests were conducted by setting VAV Terminal 6 damper stuck at 50% openness (Tset = 24 8C) to verify the strategy when Fault 3 was involved. Damper stuck (Fault 3) was detected after half an hour operation while all the three indexes in the FDD scheme of Step 2 were over the thresholds. Tests also confirmed that the ability of Step 2 would not be affected by the faults to be dealt at later steps. The third group of tests were carried out by fixing the flow rate reading of VAV Terminal 6. Frozen sensor reading (Fault 4) at 0.5 kg/s was detected after counting five reversals of the flow set-point (Step 3) as illustrated in Fig. 9. When the flow rate reading of Terminal 6 was fixed at the

Fig. 8. Schematic of system pressure-flow balance model.

1044

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

Fig. 9. VAV flow sensor reading frozen at 0.5 kg/s.

minimum, 0.22 kg/s, Fault 5 was detected after 30 min while strategy detected more than 1.5 8C temperature error under the conditions of m = mmax and F set = F min (Step 4). In fourth group of tests, Fault 6 (poor tuning of static pressure control) was detected after counting 20 consecutive reversals of the static pressure within 30 min (Step 5). This faulty pattern was also verified using laboratory tests by Seem et al. [18]. Fifth group of tests were conducted to verify the strategy when flow control faults exist, which were introduced by using too large and too small proportional gains for the flow control loop, adding damper sticking and damper hysteresis errors, respectively. The data were collected at the interval of 2 s as sampling frequency is an important issue when trying to unravel the true behavior of an unstable control loop [13]. The faults of excessive periods of controller saturation, sluggish response and oscillation were detected according to the control loop diagnostic method as shown in Fig. 5. The ¯ for sluggish response pattern recognition index, CUSUM (F), and the pattern recognition index, (F¯ k  F¯ k1 ), for oscillation to characterize the patterns of poor controller tuning, hysteresis and sticking are shown in Fig. 10. Due to set-point ¯ adjustment and/or disturbance occurrence, the CUSUM (F) kept almost unchanged for a certain period then deviated when the root cause of sluggish response was hysteresis. The oscillation caused by hysteresis or sticking was distinguished from controller poor tuning based on the dominated zeroes (F¯ k  F¯ k1 0) of the pattern recognition index. 4.3.2. Fault 10 Simulation of one operating day with fault free sensors was conducted for the training of PCA models. Both system

level training matrix (33 17) and terminal level training matrix (4 17) were constructed from the simulation results under normal operation. Three PCs were retained in both models based on the minimum VRE. In validation tests, developing sensor fault was introduced into Terminal 6 at 11:06 a.m. The FDD results of system level are shown in Fig. 11. Further study on suspected terminal (Terminal 6) based on the terminal level model confirmed the sensor bias and reconstructed the sensor. No further faults could be detected after replacing the faulty measurement with the recovered data. The serial use of two PCA models enhances the robustness of the FDD scheme.

5. In situ validation of FDD strategy 5.1. Faults 1–9 The in situ validation of the strategy was conducted partially by the data obtained from the above site survey study and partially by particular FDD tests at the same building when necessary. The performance of the FDD strategy is summarized below while presenting a few typical site test results. According to the performance pattern of the VAV 35 at the 31st floor, the zone temperature sensor frozen (Fault 1) was confirmed by the strategy, which verified the requirements of the performance indexes of Fault 1 and the FDD ability of strategy at Step 1. According to the site survey, Fault 2/3 was common. However, the database obtained during the site survey did not give the full data set of performance indexes for the FDD scheme as the control signal m was not logged. Since m cannot be logged by the existing BMS in this building, an in situ test on damper stuck (Fault 3) was carried out on an suspected stuck VAV damper (VAV Box 2 at 38th floor) by manual recording the damper control signal (0–10 V DC) using a portable digital voltage meter. Based on the data record, the strategy confirmed the damper stuck of VAV box by the scheme of Step 2 after analyzing 30 min data record. For VAV flow sensor reading frozen (Fault 4), another in situ test was carried out by replacing the flow sensor signal of VAV Box 30 at 18th floor with an emulated control signal of 4 V DC, which represented 200 l/s of the reading. Under

Fig. 10. Pattern recognition index for sluggish response and oscillation Fault 10.

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

1045

Fig. 11. System level fault detection and isolation.

the temperature set-point of 21.5 8C, the trend data the flow set-point were recorded at 1-min intervals for an hour afterwards. The strategy detected the fault after counting five consecutive reversals of the flow set-point by the scheme of Step 3 as shown Fig. 12. It was observed that VAV flow sensor reading fixed at minimum/maximum (Fault 5) happened on many of the 33 VAV terminals at 26th floor after roughly checking on the BMS logged data. The manual measurement on control signal, m, of five suspected VAV terminals were conducted besides BMS logging of the other variables required by the FDD strategy. Using the data record, the scheme of Step 4 of the strategy successfully diagnosed the sensor faults (fixed at minimum/maximum), which were further confirmed by manual checking. Two in situ tests involving VAV damper mechanical faults (sticking and hysteresis) were conducted. The fault of damper sticking was introduced by mechanically tightening up the actuator mechanism of a VAV box at 18th floor in the first test and the fault of hysteresis was introduced by slacking off the damper actuator connection of the same VAV box in the second test. The temperature set-point was kept fixed (Tset = 21 8C) for the sticking test and was adjusted to 18.5 8C at the beginning of the hysteresis test, respectively. The data trends of the flow set-point, the measured flow and the zone temperature were logged at 1min intervals for both tests as the existing BMS could not log the trend more frequently, which means that the logged

measurements could not reflect the full-scale faulty patterns as some essential data between two samples were missing from the records. Nevertheless, the pattern of flow oscillation for sticking and the pattern of flow sluggish response for hysteresis were obvious (Fig. 13). The logged flow measurements also showed that the flow measurement kept unchanged at the beginning of the hysteresis test and it was fixed most time during sticking test. The pattern recognition indexes for fault isolation among Faults 7–9 were supported by the in situ test results as well. 5.2. Fault 10 The PCA-based VAV terminal flow sensor FDD scheme was also tested using site measurements from another existing high-rise commercial building as its VAV damper position signals were logged by BMS. The air-conditioning system under study is a pressure dependent VAV system serving half floor open office. The system operates 12 h (8:00–20:00) in working days. The control and performance monitoring are handled by a BMS. Data trend at 30-min intervals of a week (five working days) were recorded by the BMS in data logging. After an initial checking on the data of VAV terminals, Box 34, 35, 37, 42 and 44 were excluded from the training matrix as the damper openness and the flow measured observed to be abnormal. The data of the first three working days are used to construct the training model. Applying a filter based on T2 statistic, outliers in the measurements were eliminated before applying the

Fig. 12. Flow set-point when VAV flow sensor reading frozen at 200 l/s (in situ test).

1046

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

Fig. 13. In situ test results of damper sticking and hysteresis.

PCA-based scheme. Three tests were conducted using the measurements of fourth and fifth days. In the first test, the measured flow of Box 38 was added with fixed bias of +100 l/s ( F38e1 = F38 + 100, Error I). In the second test, it was added with a developing bias of 4 l/s per hour ( F38e2 = F38 + 2 + 4t, Error II). In the third test, it was added with another fixed bias of +50 l/s ( F38e3 = F38 + 50, Error III). Fig. 14 presents the T2 statistic and SPE plot of the tests with Error I, Error II and Error III, respectively. The process was under control as indicated by T2 statistic plot, which was obviously below the threshold lines. The flow sensor bias was detected by SPE when it was +100 l/s (Error I). However, the developing bias (Error II) could be detected only when the reading deviation was significant (above 50 l/ s). It was also demonstrated by the last graph that the bias of +50 l/s (Error III) was marginally detectable. In the tests, the airflow rate of the concerned VAV box was around 150 l/s. The tests indicate that the sensor biases could only be

detected when those exceed a certain level in practical applications. As small sensor reading deviations would not affect the normal control process due to the compensation effect of the flow set-point reset in the cascade control loop, the sensitivity of the FDD strategy is acceptable. SPE contribution plot approach was used to isolate the faulty sensor in the test. In Fig. 15, for the fixed error test (Errors I and III), the average SPE contribution of each variable is compared (the first graph and the last graph). For the developing bias test (Error II), the SPE contribution of each variable was compared at 3-h intervals (the middle graph). The flow sensor of Box 38 in the test data matrix was identified to be faulty as it had a major SPE contribution. Investigating Box 38 with terminal level PCA model confirms the same diagnosis output. Experiences in situ tests show, when applying PCA-based flow sensor FDD strategy to real buildings, the training matrix should be carefully constructed. In the case of the studied system, if only the measurements of the first working

Fig. 14. T2 statistic and SPE plot of the tests in the real building (fourth and fifth days).

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

1047

Fig. 15. SPE contribution plot of the test in the real building (fourth and fifth days).

day are used to construct the training matrix, the faults could not be isolated as the training matrix does not cover the characteristics of the system at sufficient operating conditions, which actually leaded to extreme high values of both T2 and SPE for fault detection. As the operating condition varied according to the weather conditions, occupancy and internal loads, the training matrix selected should represent sufficient coverage of system operation conditions.

6. Conclusion Ten major root faults for VAVair distribution system were identified in the site survey study. Tests shows that the FDD strategy developed can detect faults in VAV systems effectively. This method is computationally efficient as the calculation of the performance indexes is straightforward without need in solving iteratively large equation systems. PCA-based method using models at both system level and component level is an effective approach for VAV terminal flow sensor FDD. The automatic FDD strategy can provide a simple and effective tool for automatic commissioning of VAV systems.

Acknowledgments The search presented in this paper was financially supported by a grant of the Research Grants Council (RGC) of the Hong Kong SAR. The authors would like to express their thanks to Swire Properties Ltd. for essential support to the site survey and the in situ tests.

References [1] J. Hyvarinen, S. Karki, Building optimization and fault diagnosis source book, IEA Annex (1996) 25. [2] R. Linder, C.B. Dorgan, VAV systems work despite some design and application problems, ASHRAE Transactions 103 (Part 2) (1997) 807–813. [3] H. Yoshida, T. Iwami, H. Yuzawa, M. Suzuki, Typical faults of airconditioning systems and fault detection by ARX model and extended Kalman Filter, ASHRAE Transactions 102 (Part 1) (1996) 557–564. [4] W.Y. Lee, J.M. House, D.R. Shin, Fault diagnosis and temperature sensor recovery for an air-handling unit, ASHRAE Transactions 103 (Part 1) (1997) 621–633. [5] A.L. Dexter, M. Benouatets, A generic approach to identifying faults in HVAC plants, ASHRAE Transactions 102 (Part 1) (1996) 550–556. [6] D. Ngo, A.L. Dexter, A robust model-based approach to diagnosing faults in air-handling units, ASHRAE Transactions 105 (Part 1) (1999) 1078–1086. [7] A.L. Dexter, D. Ngo, Fault diagnosis in air-conditioning systems: a multi-step fuzzy model-based approach, HVAC & R Research 7 (1) (2001) 83–102. [8] X.F. Liu, A. Dexter, Fault-tolerant supervisory control of VAV airconditioning systems, Energy and Buildings 33 (2001) 379–389. [9] S.R. Shaw, L.K. Norford, D. Luo, S.B. Leeb, Detection and diagnosis of HVAC faults via electrical load monitoring, HVAC & R Research 8 (1) (2002) 13–40. [10] L.K. Norford, J.A. Wright, R.A. Buswell, D. Luo, C.J. Klaassen, A. Suby, Demonstration of fault detection and diagnosis methods for airhandling units (ASHRAE 1020-RP), HVAC & R Research 8 (1) (2002) 41–71. [11] J.M. House, W.Y. Lee, D.R. Shin, Classification techniques for fault detection and diagnosis of an air-handling unit, ASHRAE Transactions 105 (Part 1) (1999) 1087–1100. [12] J.M. House, H. Vaezi-Nejad, J.M. Whitcomb, An expert rule set for fault detection in air-handling units, ASHRAE Transactions 107 (Part 1) (2001) 858–871. [13] J.M. House, K.D. Lee, L.K. Norford, Controls and diagnostics for air distribution systems, Journal of Solar Energy Engineering, Transactions of the ASME 125 (3) (2003) 310–317.

1048

J. Qin, S. Wang / Energy and Buildings 37 (2005) 1035–1048

[14] M. Brambley, R. Pratt, D. Chassin, S. Katipamula, D. Hatley, Diagnosis for outdoor air ventilation and economizers, ASHRAE Journal (October) (1998) 49–55. [15] S. Katipamula, R.G. Pratt, D.P. Chassin, Z.T. Taylor, K. Gowri, M.R. Brambley, Automated fault detection and diagnosis for outdoor-air ventilation system and economizers: methodology and results from field testing, ASHRAE Transactions 105 (Part 1) (1999) 555–567. [16] R.H. Dodier, P.S. Curtiss, J.F. Kreider, Small-scale on-line diagnosis for an HVAC system, ASHRAE Transactions 104 (Part 1) (1998) 530–539. [17] S.W. Wang, Y.M. Chen, Fault-tolerant control for outdoor ventilation air flow rate in buildings based on neural network, Building and Environment 37 (2001) 691–704. [18] J.E. Seem, J.M. House, R.H. Monroe, On-line monitoring and fault detection, ASHRAE Journal (July) (1999) 21–26. [19] J. Schein, J.M. House, Application of control chart for detecting faults in variable-air-volume box, ASHRAE Transactions 109 (Part 2) (2003) 671–682. [20] H. Yoshida, S. Kumar, ARX and AFMM model-based on-line realtime data base diagnosis of sudden fault in AHU of VAV system, Energy Conversion and Management 40 (1999) 1191–1206. [21] H. Yoshida, S. Kumar, Y. Morita, Online fault detection and diagnosis in VAVair handling unit by RARX modeling, Energy and Buildings 33 (4) (2001) 391–401. [22] J. McGhee, I.A. Henderson, A. Baird, Neural networks applied for the identification and fault diagnosis of process valves and actuators, Measurement 20 (4) (1997) 267–275. [23] C.Y. Han, Y.F. Xiao, C.J. Ruther, Fault detection and diagnosis of HVAC systems, ASHRAE Transactions 105 (Part 1) (1999) 568–578.

[24] K. Kamimura, A. Yamada, T. Matsuba, A. Kimbara, S. Kurosu, M. Kasahara, CAT (computer-aided tuning) software for PID controllers, ASHRAE Transactions 100 (Part1) (1994) 180–190. [25] D.C. Montgomery, Introduction to Statistical Quality Control, John Wiley & Sons Inc., 2001. [26] T.I. Salsbury, A practical algorithm for diagnosing control loop problems, Energy and Building 29 (1999) 217–227. [27] D. Choiniere, S. Beaudoin, Fault detection and diagnosis tool for VAV boxes, in: Proceedings of the IEA Annex 34 Meeting, 12–14 April 2000, Liege, Belgium. [28] K. Astrom, T. Hagglund, PID controllers: theory, design and tuning, Instrument Society of America, 1995. [29] J.E. Seem, Implementation of a new pattern recognition adaptive controller developed through optimization, ASHRAE Transactions 103 (Part 1) (1997) 494–506. [30] J.E. Seem, A new pattern recognition adaptive controller with application to HVAC system, Automatica 34 (8) (1998) 969–982. [31] S.J. Qin, R. Dunia, Determining the number of principal components for best reconstruction, Journal of Process Control 10 (2000) 245–250. [32] F. Doymaz, J.A. Romagnoli, A. Palazoglu, A strategy for detection and isolation of sensor failures and process upsets, Chemometrics and Intelligent Laboratory Systems 55 (2001) 109–123. [33] R. Dunia, S.J. Qin, T.F. Edgar, T.J. McAvoy, Identification of faulty sensors using principal component analysis, AIChE Journal 42 (10) (1996) 2797–2812. [34] S.W. Wang, Dynamic simulation of building VAV air-conditioning system and evaluation of EMCS on-line control strategies, Building and Environment 34 (1999) 681–705.

Suggest Documents