Study On Fault Detection, Diagnosis And Prediction Of Refrigeration System

With the increase of the complexity of refrigeration system, the research of FDD (fault detection and diagnosis, FDD) and performance prediction have become a hot point in the field of the refrigerant. This paper focuses on the research of the fault diagnosis, fault detection and performance prediction of refrigerant system. The main works are summarized as follows:First, four abrupt faults and eight gradual faults of the most common refrigeration systems faults are selected to be simulated based on the survey results and the availability test rig. During the simulation, the required samples for the study are recorded and stored, then the faults results are analyzed and validated by the thermodynamics theory, finally the interrelationship between faults and their symptoms are summarized which provide the base for the following study.According to the different-degree damage by the two kind faults to refrigerant system and the actual demand for detection of them, this paper presents a particular fault detection strategy which is utilization of wavelet package transformation of the character signal to identify the abrupt faults and principle component analyze (PCA) method to detection the initial gradual faults. The faults detection strategy is validated by the measured experimental data, and shows satisfying results.Due to the multicollinearity existing among the measured experimental variables, direct utilization of them may badly reduce the FD (fault diagnosis, FD) model's performance. In order to solve this problem, this paper firstly uses the PCA method to preprocess the database, then the extraction components by PCA are used as the input features to developed the FD models. Based on the comparision of the several classical artificial intelligent methods, the SVM (support vector machine, SVM) algorithm is selected to developed FD models. Therefore, the combined PCA-SVM FD system is built for the refrigeration system. Several popular multi-SVM algorithms are analyzed and compared to build the FD model. For the results of PCA-SVM model, the fuzzy c-means cluster method and Euclidean distance are utilized to evaluate the degree of the same kind faults. The presented schemes are validateed and compared with ANN, and the results are satisfying: PCA-SVM combined diagnosis system presents excellent performance that is high diagnosis accuracy and less time to train model and test; the scheme of PCA-FCM and Euclidean distance can well class the same type faults.For the incompletely described samples could not be used by the existing previously trained refrigerant system FD models, this paper presents a novel FD strategy based on similar feature transformation method and SVM algorithm. Firstly, the unknown features in the incompletely described samples are transformed to the highly related known features by the similarity transform matrix, then the values of the unknown features are assigned by the corresponding features of the optimal case which is retrieved by measuring and comparing similarities between the retrieval feature vector and the completely-described samples in the historical database, and finally the regenerated completely-described samples are inputted to the precisely FD system to diagnose the refrigerant system. The presented FD strategy is validated, and achieved satisfying results.The refrigerant system is characterized by the strong nonlinear and time-varying process, and the experimental data is usually interfered by noise, so it is hardly to predict its performance precisely by traditional method under fault condition. To solve this problem, this paper presents two kind models that are SVM model for cause and effect problem and ARIMA-SVM hybrid model for time sequence problem. The SVM model for cause and effect is applied to predict heat flux of a frosting evaporator, and is compared to multivariable nonlinear regression model, then its'anti-noise ability and sensitivity are analyzed lucubrately. The ARIMA-SVM hybrid model is applied to predict the heat transfer of a frosting evaporator, and compared to the individual SVM and ARIMA model in aspects such as the performance index and the ability of prediction tendency and turn point of time sequence.