Outlier Detection And Refrigerant Charge Amount Fault Diagnosis Of Variable Refrigerant Flow System Based On Hybrid Deep Forest Approach

The data-driven fault diagnosis approach,which is conducive to strengthening the building support system and realizing the building information construction,has gradually become the research hotspot of the heating,ventilating and air conditioning(HVAC)system.With high degree of automation data collection and transmission capabilities,the variable refrigerant flow(VRF)system is an important way to explore the data-driven approach in air conditioning system.Therefore,a hybrid deep forest approach of fault diagnosis is proposed and the refrigerant charge amounts experiment of VRF system is designed to verify the feasibility of the proposed approach in this dissertation.In view of the existing fault diagnosis researches,which focused on feature extraction during the data preprocessing,there is a lack of in-depth analysis of outliers in air conditioning system.Based on feature selection twice,the Isolation Forest(IF)algorithm is used to detect outliers.The performance of five diagnostic models before and after removing outliers is compared to verify the reliability of IF algorithm.The key features of air conditioning are visualized and analyzed by the box plot,and the practical significance of outliers in VRF system is explained.The results show that the diagnostic accuracy of the five improved models is improved compared with the original model,and the outliers of VRF system are mainly defrosting data.Compared with boxplot method,the IF algorithm builds detection model through subsample set,which can avoid masking problem caused by periodic defrosting of air conditioning system under long-term sampling,and is more suitable for outlier detection task of air conditioning system.Since the mainstream models rely on a large number of sensors to achieve highprecision diagnosis while the deep neural network has a complex and time-consuming hyper-parameters optimization strategy,the Cascade Forest(CF),a deep learning model based on non-neural network,is introduced into fault diagnosis field of air conditioning system.The results of Cascade Forest and four main diagnostic models,namely Back Propagation Neural Network(BPNN),Support Vector Machine(SVM),Multi-Layer Perceptron(MLP)and Long Short-Term Memory(LSTM),are analyzed from the perspective of low-dimensional diagnostic performance and hyper-parameters robustness.The results show that the proposed IF-CF model has better diagnostic performance with low-dimensional input features,which is more suitable for fault diagnosis tasks of VRF system with fewer sensors and incomplete input information.With 6-dimensional input features,the accuracy of IF-CF model is 94.16%,which is 10.02%,5.87%,5.26% and 3.34%higher than that of IF-BPNN,IF-SVM,IF-MLP and IF-LSTM models,respectively.In addition,the IF-CF model has stronger robustness with the brief hyper-parametric optimization strategy.With different configuration of hyper-parameters,the maximum accuracy difference of IF-CF model is only 2.04%,while the running time of IF-CF model is lower than that of IF-BPNN,IF-MLP and IF-LSTM models.In conclusion,the proposed hybrid deep forest approach can effectively eliminate the outliers of VRF system and explain the practical significance of outliers,which provides theoretical support for the outlier detection task of air conditioning system.Meanwhile,the proposed approach is fast,stable,suitable for fault diagnosis tasks with fewer sensors and incomplete input information,which has a certain application value in the construction of the actual diagnosis system of building air conditioning.