Fault diagnosis of HVAC AHUs based on a BP-MTN classifier

• Published in: Building and environment Vol. 227; p. 109779
• Main Authors: Yan, Ying, Cai, Jun, Tang, Yun, Chen, Liang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.01.2023
• Subjects: AHU; BP algorithm; HVAC; Multi-dimensional Taylor network; Multi-dimensional Taylor network; HVAC; BP algorithm; AHU
• ISSN: 0360-1323; 1873-684X
• DOI: 10.1016/j.buildenv.2022.109779

HVAC Air Conditioning Units (AHU) adjust and deliver air to rooms through fans and ducts to meet human comfort needs. Fault diagnosis of AHUs helps to reduce energy consumption and meet human comfort needs, and thus is significant. As a network, the Multi-dimensional Taylor Network (MTN) approximates a nonlinear function with a polynomial network. It is suitable for embedding in a control system since it has a much simpler structure than a neural network while having high accuracy. However, the traditional MTN is usually used for model fitting but not for classification. To solve this problem, a Back Propagation Multi-dimensional Taylor Network (BP-MTN) classifier is proposed in this paper to diagnose the faults of AHUs. This BP-MTN classifier has three main features: 1) a fully connected layer is added after the output layer of the traditional MTN to solve the mismatch between the dimensionality of fault features and the number of categories; 2) the softmax layer is added in the traditional MTN to realize the classification; 3) ReLU function is added in the traditional MTN to improve the classification accuracy and reduce the model complexity; 4) the Back-Propagation (BP) algorithm based on the small batch gradient descent algorithm is used to train the BP-MTN classifier rather than the nonlinear least square used in the traditional MTN. Additionally, this paper explores the selection of polynomial orders and activation functions of BP-MTN through extensive experiments. The experimental results show that the BP-MTN can achieve the accurate classification of AHU faults effectively. •This paper provides a perspective for improving and applying MTNs to fault diagnosis while rare work applying MTNs for fault diagnosis.•The structure of a traditional MTN is modified to obtain an accurate BP-MTN classifier by adding a fully connected layer, a softmax layer, and multiple ReLU activation functions.•Unlike traditional MTNs which are trained by the nonlinear least square method, the BP-MTN classifier is trained using a Back-Propagation (BP) algorithm based on the small batch gradient descent algorithm.•The selection of hyperparameters for the BP-MTN and the selection of activation functions are explored through intensive experiments.