عیب یابی آماری و روش تشخیص برای چیلرهای گریز از مرکز بر اساس نمایی با وزن میانگین متحرک نمودارهای کنترل و پسرفت بردار پشتیبانی

This paper presents a new fault detection and diagnosis (FDD) method for centrifugal chillers of building air-conditioning systems. Firstly, the Support Vector Regression (SVR) is adopted to develop the reference PI models. A new PI, namely the heat transfer efficiency of the sub-cooling section (ɛsc), is proposed to improve the FDD performance. Secondly, the Exponentially-Weighted Moving Average (EWMA) control charts are introduced to detect faults in a statistical way to improve the ratios of correctly detected points. Thirdly, when faults are detected, diagnosis follows which is based on a proposed FDD rule table. Six typical chiller component faults are concerned in this paper. This method is validated using the real-time experimental data from the ASHRAE RP-1043. Test results show that the combined use of SVR and EWMA can achieve the best performance. Results also show that significant improvements are achieved compared with a commonly used method using Multiple Linear Regression (MLR) and t-statistic.

Chiller plants contribute to a large portion of the energy consumption of HVAC&R system of buildings [1]. It typically accounts for 35–40% of the total building electricity consumption in commercial buildings in Hong Kong [2]. Significant amount of energy is wasted in the case of chiller performance degradation and malfunctions. The chiller component faults account for around 42% of the service resources and around 26% of repair costs [3]. It is obvious that maintaining the chiller system in healthy conditions and identifying faults in time are beneficial to both building energy saving and operating costs saving. Over the last decades, the development of automated fault detection and diagnosis (FDD) methods for chillers has been an area of active research. Detailed literature reviews on this subject can be found in the papers of Comstock et al. [3], Reddy et al. [4], Dexter and Pakanen [5], Katipamula and Brambley [6], Xiao et al. [7] and [8]. Among all of the methods, the methods, which use black-box approach to develop reference models and use t-statistic approach to detect faults, are efficient and widely used. Typical works can be found at Xiao et al. [7], Grimmelius et al. [9], McIntosh et al. [10], Chen and Braun [11] and Li and Braun [12], Jia and Reddy [13], Wang and Cui [14], Reddy [15]. Generally, these methods are developed as the following steps: 1) Reference PI models development. It is generally on the basis of physical chiller models [16] and [17] or black-box MLR (Multiple Linear Regression) models. 2) Confidence intervals (thresholds) calculation. They are usually gained through the t-statistic approach at a certain confidence level, e.g. 4.6% false alarm. 3) Fault detection. The residuals between benchmark PIs and current ones are calculated. A fault is detected as soon as the residuals are out of the confidence intervals. 4) Fault diagnosis. The fault is diagnosed according to the fault pattern rule table. The lessons learned from the success of these methods can be concluded as follows. Firstly, the performance indexes (also named characteristic quantities or characteristic parameters in some studies) can account for existing chiller faults efficiently. Secondly, the statistical approaches are introduced to detect faults, which can properly handle the uncertainties of both model-fitting errors and measurement errors. In the end, the expert knowledge is used to diagnose fault using rule tables. The low severity level is defined in this study as the level at which chiller performance is affected slightly. The available methods failed to detect and diagnose faults at low severity levels (SLs). For instance, the ratios of correctly diagnosed points were 0%, 0%, 0% at SL1 and 25%, 0%, 0% at SL2 respectively for the refrigerant leakage, condenser fouling and excess oil in Cui and Wang [14]. The MLR and t-statistic-based method in AHSRAE RP-1275 was reported that the ratios of correctly diagnosed points were 3.7%, 0%, 0% at SL1 and 7.4%, 0%, 0% at SL2 the refrigerant leakage, condenser fouling and refrigerant overcharge respectively [15]. It means that the operators have to wait for the fault alarms until the fault archives at a serious severity level. Obviously, significant amount of energy might have been wasted. The chiller might have been damaged. Hence, removing faults at low severity levels is crucial for chiller maintenance and costs saving. Three major reasons are found in this study, which contribute to the poor FDD performances at low severity levels using available methods: 1) The Type II errors in the t-statistic-based approaches. Generally, a fault is detected when the residuals of PIs are outside of the confidence intervals. However, the residuals are still within the confidence interval at the low severity levels. This is the Type II error (more details can be found in Section 2.4). 2) The accuracy of MLR-based reference models. The MLR-based models have better accuracy than physical models in most of conditions. It is widely used, e.g., RP-1043 [18], Reddy [15], Cui and Wang [14], etc. However, MLR is still a linear regression approach. The chiller is a typical non-linearly system. The accuracy can be improved using non-linear regression approaches. The accuracy is important to the width of confidence interval. The lower the accuracy is, the wider the confidence interval will be, and then the FDD performance will be poor. In this way, the confidence interval can be narrowed obviously. 3) The improper performance index for the condenser fouling. The condenser fouling has similar fault patterns with the refrigerant overcharge and non-condensable gas. To distinguish it with other faults, it is generally assumed that the deviation of sub-cooling temperature is within conference intervals or a positive value. Actually, the deviation is sometimes positive (mostly in low severity levels) and sometimes negative (mostly in serious severity levels), as shown in RP-1043 data. Such assumptions always lead to higher false diagnosis ratios. The objective of this study is to propose a robust statistical FDD method suitable for the detection and diagnosis of faults at low severity levels. Three innovations are adopted to overcome the shortcomings. Firstly, the Exponentially-Weighted Moving Average (EWMA) control charts are used to reduce the Type II errors. Secondly, Support Vector Regression (SVR) is implemented to develop reference models. Thirdly, a new performance index is introduced, which is named the heat transfer efficiency of the sub-cooling section (ɛsc). In the end, a new rule table is proposed. Both SVR and EWMA are statistical theories-based approaches and can efficiently handle uncertainties in FDD process. The proposed method is validated using the ASHRAE RP-1043 experimental data. Comparison is made with a typical FDD method from AHSRAE RP-1275.

A new statistical fault detection and diagnosis method is proposed in this study, which adopts three main innovations including a new performance index (ɛsc), the EWMA control charts and the SVR based reference models. Two comprehensive comparison studies were conducted using the experimental data from the ASHRAE RP-1043. It is found that the proposed method improved the FDD performances significantly, especially at low severity levels. For example, in the case of condenser fouling, the proposed method achieved the ratios of correctly diagnosed points of 7.7%, 45.2%, 60.7% and 100.0% at four severity levels (SL1 to SL4) respectively at the confidence level of 99.73%. Using the other methods including a commonly used method, this fault could not be correctly diagnosed at level SL1, SL2 and SL3. Significant improvements can also be found in the cases of other two typical faults (refrigerant leakage and refrigerant overcharge). Other three faults (i.e. non-condensable gas, reduced evaporator water flow rate and reduced condenser water flow rate) can be easily diagnosed similarly as other methods. The EWMA control charts contribute the most to the improvement on FDD performance. It reduces the Type II errors through taking into account the time series information using the weighting factor. Therefore, the EWMA-based methods achieved much higher correctly diagnosis ratios compared with the t-statistic-based methods. The heat transfer efficiency of the sub-cooling section (ɛsc) is an effective performance index to represent the effects of faults on the thermal performance of sub-cooling section. It is especially useful in improving the FDD performance for condenser fouling. As a new machine learning algorithm based on structural risk minimization from statistical learning theory, SVR could be more efficient than MLR approach using the actual site data of higher uncertainty level.