Research On Fault Diagnosis By PCA For Autonomous Underwater Vehicle

As a key technology for autonomous underwater vehicle (AUV), the technology of fault diagnosis has an important academic and practical value. The thesis researches domestic and overseas methods of AUV, and bring principal component analysis (PCA) method into AUV fault diagnosis with the purpose to investigate the method's feasibility and validity. PCA method does not need an accurate mathematical model of AUV. It gets useful information from the normal data and establishes the PCA model for AUV. Abnormal behaviors or faults will be found by checking whether new data deviates from the PCA model.The thesis researches AUV fault diagnosis by using traditional PCA method. The simulation experiments show that the method is feasible. Then in view of the peculiarity of AUV and the disadvantage of traditional PCA method, many deep research aspects are done as follows:A method with the name of cumulative percent variance by average eigenvalue is proposed to choose principal components. An improved PCA method called sub-block PCA is designed, which considers relativity of variables. All variables are divided into different groups in terms of their mutual correlation coefficient to make the number of principal components for PCA model below three.It is prolix to get principal component scores with non-linear iterative partial least square (NIPALS), so NIPALS isn't suitable for AUV. A simple projective solution is deduced by using the data correlation coefficient matrix to calculate the PCA load matrix.In view of the disadvantage that the "false alarm" caused by the T~2 statistics maybe happen, the thesis proposes a solution method. The new method separates variables by their correlation with principal component scores. A new statistics is composed of the separated variables, and it is used associated with T~2 statistics to detect and diagnose faults. The simulation experiments based on sea testing data verify the validity of the method.The statistical characteristics of the Q statistics and the T~2 statistics are deduced for the typical sensor faults, which are the basis of the sensor fault diagnosis. Using principal component scores and principal component loads to reconstruct the fault data will bring estimation error to the results. The thesis uses a new method named iteration reconstruction in order to overcome the shortcoming. Experiments carried on "Beaver" verify the feasibility and validity of the method.