Sensor fault detection and diagnosis of air handling units

Abstract

HVAC systems in modern buildings are heavily instrumented to realize automatic monitoring and control. Measurements from sensors not only indicate operation conditions, but also take part in control and optimization. Accurate measurements are essential for reliable monitoring, control and energy management. On the other hand, the performance of fault detection and diagnosis (FDD) methods applied in HVAC systems depends strongly on the quality and reliability of sensor measurements. However, sensors may suffer from various faults, including drift, bias, accuracy decrease and complete failure. Although a great deal of research has been carried on the FDD of HVAC components, not much has concerned sensor faults and sensor validation in HVAC systems. This thesis developed a robust strategy, which consists of a basic scheme and a condition-based adaptive scheme, for online detection and diagnosis of sensor faults in air handling units (AHUs). The basic scheme includes three major steps: data pre-process, sensor fault detection, and sensor fault diagnosis. Transient data and outliers are removed from the training data and new samples using a data pre-processor. The PCA method is adopted and improved for sensor FDD of AHUs. Correlation matrixes are used to depict correlations among variables in air handling processes. Statistics are used to measure the variance of the correlations, and their upper limits usually define the normal ranges of the variance. If the statistics exceed the normal ranges, it indicates that the correlations among variables are disturbed and something abnormal has happened. Two kinds of statistics are used in this FDD application, i.e. the Hotelling T2 and the Q-statistic. The Hotelling T2 is mainly used to preprocess the training data and measurement data to be analyzed to improve their quality. The Q-statistic is used to detect abnormalities. Independent heat balance and pressure-flow balance PCA models were developed to make variables in each PCA model more closely correlate, therefore enhance the stability of PCA models. The usage of two PCA models in parallel also helps to isolate faulty sensors. In order to enhance the ability of the PCA method in isolating faulty sensors, the Signed Diagraph (SDG) is adopted to supplement the Q-contribution plot which is commonly used for fault isolation in PCA-based FDD methods. The SDG is a graph qualitatively demonstrating causal interactions among variables due to basic principles and control strategies. Fault patterns are built according to the fault directions (negative or positive biases) and the changing directions of affected variables (increasing or decreasing magnitudes) by analyzing the simplified SDG of typical AHUs. By matching signs of residual elements with the fault patterns, faulty sensors could be successfully isolated. The fault isolation ability of the basic scheme is greatly enhanced by the SDG-based fault isolation technique. A robust sensor FDD strategy is developed for automatic online application in AHUs, in which sensor faults are detected and diagnosed using the basic scheme, while the PCA models are updated to follow the normal shifts of air-handling processes using so-called condition-based adaptive scheme. The condition-based adaptive scheme updates the PCA models to follow the change of operating condition using the outdoor climate instead of time. The conventional time-based adaptive method may result in that the PCA models adapt not only to the normal shift of a process, but also to slowly developing faults, such as sensor drifting. The condition-based adaptive scheme can improve sensitivities of the PCA models to drifting faults. A PCA model database is also built up during the adaptive process, sorted in descending order of the outdoor air temperature, and then according to the outdoor air humidity. This database is easy to be used to monitor the air-handling process in long-term. Software package is developed on the platform of FORTRAN PowerStation to implement the robust sensor FDD strategy. It is also integrated with a BMS management platform named IBmanager which is an open IBMS integration and management platform. Users can view the sensor FDD results via network by opening a web page. The robust strategy including both the basic scheme and the condition-based adaptive scheme is validated using dynamic simulation tests and sitedata retrieved from the BMS of a real building in Hong Kong. Various faults including fixed biases and drifting faults were introduced to the simulation tests and real air-handling unit. Results show that the PCA-based sensor FDD strategy is powerful and robust in detecting and diagnosing sensor faults in AHUs.