Wavelet Neural Network-Based Fault Diagnosis in Air-Handling Units

Abstract

Various sensor faults in long-term used heating, ventilation, and air-conditioning systems usually lead to a waste of energy consumption or decrease in indoor air quality. A wavelet neural network, which combines wavelet analysis with a neural network, is presented to diagnose the fixed and drifting biases of sensors in an air-handling unit. Wavelet analysis is employed to process measurement data, and a neural network is used to diagnose sensor faults. First, an abundance of historical data, under both normal and faulty operation conditions, are selected from the building automation systems as the training data. Wavelet analysis is employed to decompose these original data, and the eigenvector matrix representing the operation characteristic information of the air-handling unit is obtained. Then, the neural network is trained to learn various conditions using the eigenvectors and to distinguish between normal and faulty conditions. When new measurements are obtained, they are decomposed similarly using wavelet analysis. The new characteristic information after decomposition is selected as the input of the well-trained network. Finally, the outputs corresponding to the new condition are calculated out of the neural network. The diagnosis or identification is carried out by comparing these outputs with the objective vectors. Simulation tests in this paper show that wavelet neural networks can successfully diagnose fixed and drifting biases of sensors in air-handling units.