Fault Diagnosis Research Of Refrigeration And Air Conditioning System Based On SVM

For the global energy shortage and pollution problem is getting more serious, how to improve the energy using efficiency has become a major research direction. Building energy consumption accounts for more than 30% of total energy consumption, and air-conditioning system is the main energy-consuming devices. When the air conditioning system run abnormally, the system's energy consumption will increase greatly. So researches concerning fault diagnosis of air-conditioning system are necessary. In this paper, two different air-conditioning systems were chosen as study subjects to implement related work.Chiller is a highly nonlinear complex system. The fault of its system will lead to an abnormal operation, and this will not only cause a decline in air quality in the work space, but also cause an increase in chiller energy consumption. In this paper, a group of normal data from Research Project-1043 and seven sets of fault data were selected to jointly establish training data. And support vector machine(SVM) method was employed to classify the data to test its fault diagnosis performance for chillers, four different fault levels were introduced to analysis classification accuracy of SVM with varying fault degrees. After addition of wavelet de-noising, accuracy rate had a decent promotion.For the multi-split VRF(variable refrigerant flow) system, the key of efficiently operation is to achieve the appropriate refrigerant charge amount(RCA). But it is difficult to do this because of the complexity of VRF systems. Therefore, it is significant to determine the VRF status quickly and correctly in the process of automatic control. This paper proposed a hybrid model that combines SVM(support vector machine), wavelet de-noising and mRMR(minimal-redundancy-maximal-relevance). Wavelet de-noising was employed to improve the quality of obtained data and SVM classifier's generalization ability and an improved mRMR algorithm was used for feature selection. Accounting to this method, the most suitable feature sequence was chosen. After this, a correlation analysis of features was implemented for further feature selection. Finally, the seven-feature subset(1B) was singled out by the mRMR-WD-SVM model, and only a 2.14% decline occur in accuracy compared with the initial feature set(18-features). In the case of the minimum model performance degradation, the use of the data is greatly reduced.Some sensors are not needed any more to the keep the balance between economic benefit and model performance. The new mRMR-WD-SVM model's results also showed that the results of relevance between feature and class and the correlation analysis both should be taken into account in the process of feature selection seriously.In general, the proposed hybrid SVM model was adequate for the jobs concerning data de-noising, feature selection and fault diagnosis. This work has a certain application value and deserve a further research...