Air Handling Unit Fault Detection And Diagnosis Based On Semi-supervised Learning And Generative Adversarial Networks

Operating the heating,ventilation and air-conditioning system in a fault condition results in a large amount of power wastes.Air handling unit(AHU)is an important part of the air conditioning system and automated maintenance of AHU is therefore increasingly demanding.At the meanwhile,accurate and efficient detection and diagnosis of AHU faults is important.Data-driven fault detection and diagnosis(FDD)techniques have shown good results in identifying various AHU failures.Most data-driven fault detection and diagnostic methods today employ supervised machine learning techniques that assume a sufficient number of fault training data samples for availability.However,most existing data-driven FDD approaches simply adopt supervised machine learning techniques that presume the availability of a sufficient number of faulty training data samples.In real-life scenarios,the number of labeled fault samples is usually insufficient to support supervised learning methods.In addition,manual labeling is difficult and costly,which poses a challenge for AHU FDD.Therefore,from the perspective of the actual scene,this paper has done the following research to improve the efficiency of maintenance personnel in detecting and diagnosing AHU faults:1)Considering the case that the number of labeled fault samples in real-world scenarios is insufficient to support supervised learning methods,and manual labeling is difficult and costly.This paper proposes a framework for fault detection and diagnosis based on semi-supervised learning air handling units.This framework can enrich the training set by iteratively adding reliably labeled test samples to achieve accurate and efficient detection and diagnosis of air handling unit faults.Besides,the proposed framework can be easily extended using various existing classifiers such as SVM,KNN,ELM,and the like.After experimental analysis,the model has achieved high fault detection and diagnosis rates in the 2007 winter and summer data provided in the ASHRAE project 1312-RP project.2)To further improve the methodology applied in 1),the second half of this dissertation proposes a FDD model based on GANs(Generative adversarial network).The model is a hybrid model CW-GANs that combines CGANs(conditional Generative adversarial network)and W-GANs(Wasserstein Generative adversarial network).It uses a small number of fault samples to train the CW-GAN model to generate a large number of fault samples and uses the generated fault samples with a small amount of real The sample trains a classification model for detecting and diagnosing AHU faults.Through experimental analysis,the model can achieve better results than the semisupervised learning framework.