Development And Application Of Fault Diagnosis System For Water Chiller In Office Building

Office building energy usage is rapidly rising in step with the economy’s rapid expansion.Due to their large construction volume,high individual energy demand,and huge energy-saving potential,office buildings have been the focus of building energy efficiency research.Due to chiller unit failures,HVAC systems,which are the biggest energy-consuming components of office buildings,are regularly subjected to abnormal operating conditions,resulting in considerable energy losses and catastrophic accidents.As a result,implementing fault diagnostic technology into chiller plants is critical for maintaining safe HVAC system operation and reducing energy consumption caused by defects.To that end,this thesis employs an office building chiller plant as a research object,investigating a variety of topics such as typical fault principle analysis and modeling,fault data simulation,experimental platform construction,machine learning method driven,and office building site application,among others.With the use of automated machine learning,joint platform simulation,fault diagnostic system development,and other technologies,intelligent fault diagnosis of office building chiller plants may be achieved,resulting in lower energy consumption.This thesis’s core research work consists of the following:（1）To address the problem that the fault data available for training machine learning methods for office building chillers is minimal and easily leads to overfitting of diagnostic models,a joint simulation platform for chillers based on typical fault modeling was established.Five types of typical faults were chosen as the major research objectives for fault principle analysis and process modeling based on the available experimental settings,and 20 fault trials were simulated utilizing the joint simulation platform to provide the data base necessary for the study.Finally,dynamic simulation data is utilized to test the suggested method’s efficacy.（2）While traditional machine learning techniques rely heavily on expert tuning and human intervention,automated machine learning can streamline the model development process by automating a series of procedures such as data processing,model generation,and parameter optimization for the research object’s task situation.As a consequence,an auto machine learning theory-based fault diagnosis approach is proposed to solve the chiller unit fault diagnostic process’s difficult to apply difficulties of high feature redundancy,problematic parameter optimization,excessive human interaction,and high algorithm model complexity.First,a Maximum Correlation and Minimum Redundancy algorithm（mRMR）is used to extract important feature information from the training data and reduce feature redundancy;then,using a long and short-term memory（LSTM）model,temporal association information between unit data is preserved.,and an Efficient Neural Architecture Search algorithm（ENAS）is used to train and optimize the model in the AutoKeras automatic machine learning framework.The ideal neural network topology and hyperparameter configuration are then automatically determined,allowing for the model’s automation and easy deployment.The experimental findings reveal that the LSTM-based automated machine learning approach improves performance significantly in both low and complicated fault instances,demonstrating the system’s efficacy and superiority.（3）This study developed a chiller plant fault diagnosis system,which includes personnel information management,equipment information management,historical data query,equipment fault diagnosis,and other functions,to address the necessity of the methodological model in terms of practical application.The system was used at an office building chiller plant in Jinan in an actual project to assess its practicality and reliability,and it yielded significant results in terms of successfully processing data and diagnosing faults.