| The air handling unit(AHU)is an important subsystem in the HVAC system,and its efficient and stable operation is essential to maintain indoor comfort.Due to the influence of external environment,improper human operation or defects in the AHU’s own structure,many different types of failures may be caused.When AHU stays in malfunctioning condition for a long time,it will reduce the equipment life,affect the indoor comfort,and even cause a lot of energy waste.Therefore,it is necessary to research an accurate and efficient AHU fault diagnosis method.To this end,this thesis proposes deep learning-based AHU fault diagnosis models for the AHU spatio-temporal characteristics and noise sensitive characteristics.The TRNSYS simulation platform is constructed for AHU multi-condition data acquisition and performance verification of the fault diagnosis model.The main innovations and contributions of this thesis are as follows:(1)To address the effects of the temporal dynamic characteristics and spatial correlation characteristics(spatio-temporal characteristics)of the AHU on fault diagnosis,a fault diagnosis model based on kernel slow feature analysis and deep learning(KSFA-VGG)is proposed.The model first extracts slowly changing features from dynamic data using the kernel slow feature analysis method(KSFA)and ranks the features according to their slowly changing degree.Subsequently,the data after feature extraction and feature ranking are converted into two-dimensional grayscale images and image expansion is performed using the sliding window method.Finally,a visual geometry group(VGG)based image classifier is constructed for the AHU fault diagnosis using the generated slow feature image dataset.The experimental results show that the KSFA feature extraction method can effectively extract slowly changing features from dynamic data;the data conversion into images method can enhance the neighborhood information and spatial features between feature variables;VGG has higher fault diagnosis accuracy compared with other classification methods.(2)To simultaneously deal with the negative effects of noise and spatio-temporal characteristics on the AHU fault diagnosis,a fault diagnosis model based on probabilistic slow feature analysis and attention residual network(PSFA-ARes Net)is proposed.Firstly,feature extraction and preliminary feature ranking are performed on the data using probabilistic slow feature analysis(PSFA).Subsequently,the feature variables are rearranged into numerical matrices and converted into corresponding spatial grayscale images at each moment using the improved data-to-image conversion method.Finally,the attention block attention module(CBAM)is added to the deep residual network(Res Net)to construct an attention residual network(ARes Net)based fault diagnosis model.The experimental results show that the proposed PSFA-ARes Net model has the best fault diagnosis performance under the influence of three different noise levels compared to other comparison methods.(3)To obtain richer AHU operation data and further verify the performance of the proposed fault diagnosis models,the building model and the variable air volume(VAV)air conditioning system are built using the TRNSYS simulation software.Firstly,the parameters of the building model and room areas are set.Secondly,select the appropriate components and make the correct wiring connections according to the operating principle of the variable air volume air conditioning system.Subsequently,the established simulation model is used to collect the normal operation data of the VAV air conditioner and to obtain the fault operation data by modifying some device parameters to simulate the occurrence of faults.Finally,the fault diagnosis performance of the KSFA-VGG and the PSFA-ARes Net models is verified using the collected data.The experimental results show that the KSFA-VGG and the PSFA-ARes Net models have better fault diagnosis performance in the new working environment. |
