Home | Business News | Browse by Publication | H | HVAC & R Research

A novel strategy for the fault detection and diagnosis of centrifugal chiller systems.

Article Excerpt
INTRODUCTION

Fifteen to thirty percent of the energy waste in commercial buildings is due to the performance degradation, improper control strategy, and malfunction of the HVAC&R and chiller systems. To maintain the chiller system properly and save energy, equipment must be strurdy enough to last a long time, the indoor environment must be satisfactory, and a fault detection and diagnosis (FDD) strategy must be employed.

Stylianou and Nikanpour (1996) proposed an FDD method that included an off-cycle module, a start-up module, and a steady-state module for reciprocating chillers. In the steady-state module, the chiller performance was verified based on the universal chiller model developed by Gordon and Ng (1995). The residuals between the estimated variables by the linear regression models and the measured variables were matched to the known rule pattern to diagnose the chiller faults. Peitsman and Bakker (1996) compared two FDD methods based on two black-box models. One is a multiple-input/single-output autoregressive with exogenous inputs (ARX) model and the other is an artificial-neural-network-based (ANN-based) model. The black-box models contained 14 system models for fault detection and 16 component models for fault diagnosis. The selected inputs from the current and previous time steps and outputs from two previous time steps were used in the ANN models. Comstock et al. (2001) presented eight common faults tested on a 90-ton centrifugal chiller at four levels of severity. The eight faults could be detected using a rule table composed of measurement deviation. Both McIntosh et al. (2000) and Jia and Reddy (2003) implemented the FDD strategy using the characteristic quantities (CQs), which are sensitive to some faults but not to others. Wang and Cui (2006) developed a strategy that consisted of a model-based chiller FDD scheme and a sensor FDD and estimation (FDD&E) scheme. Six performance indices deduced from theoretical analysis were used to characterize the health condition of the centrifugal chiller and the principle component analysis (PCA) method was used to handle the sensor faults. The sensor FDD&E strategy for centrifugal chillers was enhanced using wavelet analysis discussed in Xu et al. (2008).

Artificial intelligence (AI) techniques such as fuzzy logic and neural network have also been applied in the FDD field. Although the application of AI in FDD for chiller systems is rare (Bailey 1998), several researchers have discussed its application in air-handling units. For example, the ANN application was referred to in Lee et al. (1996) and Lee et al. (1997), and the fuzzy logic application was referred to in Ngo and Dexter (1999) and Dexter and Ngo (2001).

Detailed literature reviews relating to FDD for chiller systems can be found in Comstock et al. (1999a) and Reddy et al. (2001). More general literature reviews relating to FDD for building systems can be found in Katipamula and Brambley (2005a, 2005b).

Traditional FDD methods for chiller systems require numerous measurements to constitute a rule table. Although this situation is improved by using the deduced performance indices such as CQs proposed by McIntosh et al. (2000), Jia and Reddy (2003), and Wang and Cui (2006), it is still difficult to decide how sensitive the selected performance indices are to the considered chiller faults. This becomes especially challenging when some faults have the same symbolic pattern of rules such as two typical chiller faults, refrigerant overcharge and noncondensables in the refrigerant. These faults are discussed in Comstock et al. (2001). Wang and Cui (2006) developed an adaptive threshold determination method to benchmark the performance indices (PI) residuals. It was found that the method is very effective when the fault severity is high but not very effective when the fault severity is low.

In this paper, a novel chiller FDD strategy is proposed that is based on the fuzzy modeling and ANN techniques. Based on the sensitivity analysis of PIs to the chiller faults, the PI residual between the estimation by the regression model of the normal test and fault tests is selected to form a qualitative rule table. The proposed fuzzy model is used to quantify the rule table in order to deduce a quantitative diagnostic classifier, which has been effective in diagnosing the investigated seven chiller faults. The fault identification is conducted by ANN using the quantitative diagnostic classifier. The factors that affect the diagnosis results are also discussed. The strategy is validated using the laboratory chiller data obtained from ASHRAE Research Project RP-1043 (Comstock and Braun 1999c).

STRUCTURE OF FDD STRATEGY AND MODELING METHOD

The structure of the proposed FDD strategy, which includes three major steps, is illustrated in Figure 1. The first step is the data preprocessing, which includes removing data outliers, disposal by chiller models (described in the section focusing on validation of regression models), and normalization of the PI residuals. The chiller model is trained off-line with the data from the normal test and fault tests. The chiller performance variables are the operating conditions that should fall into the range of the operating conditions of normal tests and fault tests. As the basis of comparison, the chiller performance variables should be the same in the training and validation processes. In the first step, a qualitative diagnostic classifier is deduced from the sensitivity analysis of various PIs to the faults. The analysis is processed considering both fault category and fault severity levels. The second step is to deduce the quantitative diagnostic classifier from the conventional qualitative classifier using the fuzzy models. The conventional linguistic IF-THEN rules, such as IF A increases and B decreases THEN C is abnormal, are changed into the fuzzy description, which is realized by the proposed fuzzy modeling method. The quantitative diagnostic classifier composed of standardized quantitative PIs (SQPs) is achieved during this step; how to determine the SQPs is covered in a later section. The third step is to identify faults using the quantitative diagnostic classifier. An ANN, which is very effective in pattern recognition, is used as the tool for fault identification. In Figure 1, most processes in the model training and FDD strategy are the same, except on three occasions. The first is when data used in the chiller model establishment of the training process and FDD strategy are different. The second is when the ANN model in the training process is used to determine the network connecting weights and biases for utilization of FDD strategy while it is applied in FDD strategy for fault identification. The third is when analysis of the sensitivity of PI residual percentages to the faults is needed in the training process to determine the parameters (centers and widths introduced in a later section) of the membership function (MF), and the parameters are used in FDD strategy for reconstruction of a new quantitative diagnostic classifier. It is worth noticing that fault detection and fault diagnosis are simultaneously performed by the FDD strategy. Therefore, there is no separation between the FDD processes in Figure 1.

[FIGURE 1 OMITTED]

Fuzzy Model--Quantitative Diagnostic Classifier

As the conventional qualitative or linguistic diagnostic classifier based on the trend pattern of PI residual...

View this article FREE - Now for a Limited Time, try Goliath Business News
Free for 3 Days!