Research On Multi-faults Diagnosis Technology Of Thruster And Sensor For AUV

Marine resources have become increasingly prominent for human survival with the decrease of non-renewable resources on land. And ocean exploitation requires advanced technology and equipments. Underwater vehicle is the only equipment that can operate in deep sea at present. With the acceleration of ocean exploitation, higher and more urgent requirements are proposed for underwater vehicle. Autonomous Underwater Vehicle (AUV) could complete the exploration and operation tasks autonomously in oceans, with the security and reliability being the key problem during the research and application of AUV. As condition monitoring and fault diagnosis is the foundation and core technology to ensure AUV safety, the research into AUV fault diagnosis theory has great scientific and practical significance to improve its safety.Aiming at the multi-faults diagnosis for AUV, three main aspects have been discussed in this dissertation, namely, the fault isolation and location of AUV thruster and sensor, the detection of multi-sensors simultaneous faults and the dynamic classification of multi-faults models.A quantitative/qualitative diagnosis method is proposed in this dissertation to solve the problem of fault isolation and location of single or simultaneous faults of AUV thruster and sensor. The neural network fault detection observer model of AUV system is built to direct map fault parameters of thruster and sensor at the output of the model, and the faults of thruster and sensor are isolated based on the fault parameters values. After the fault is isolated, according the characteristic that there are different change trends of controlled and state variables when thruster or sensor fault occurs, the real-time change trends of AUV controlled and state variables are extracted, recognized and merged by dynamic trend analysis theory, and the real-time trend primitives description of each signal are achieved. The real-time trends primitive description is matched with the sets of trend primitives for AUV controlled and state variables in expert knowledge base to realize fault location. The results of the experiments show that the proposed method is feasible and effective to fault isolation and location of AUV thruster and sensor.A fault feature extraction method is proposed by combining wavelet analysis technology with neural network technology, and a fault detection method of fuzzy weighted attribute information fusion is proposed to solve the problem of multi-sensors simultaneous faults detection for AUV. On fault features extraction, with the problem of mutual aliasing between noise interference and fault signals, considering the different transfer characteristics in different resolutions after wavelet decomposition for noise interference signals and sensor fault signals, the reconstructured detail coefficients are fused after denoising using threshold to obtain the overall high-frequency detail information, which is taken as a type of fault feature, namely detail residuals; at the same time, as the actual measurement state value and theory state value show differences when fault occurs, the difference between fault detection observer model output value and actual measurement value of sensors is taken as another fault feature, namely the observer residuals. Multi-fault features information is consisted of the detail residuals and observer residuals. The experiment results indicate that the methods of fault extraction in this dissertation are effective.On fault detection, to solve the problem of consistency from multi-faults features to fault description, in consideration of the redundancy and contradictory between fault reasons and fault features, a method of fuzzy weighted attribute information fusion is proposed. The importance degree and confidence degree of the two fault feature values are transformed by the method of fuzzy synthetic conversion. Based on the results of conversion, the weighted fusion of every fault feature is made, and the simultaneous faults location of multiple sensors for AUV is realized. The results of pool experiment for AUV show that the method proposed in this dissertation is feasible and effectiveA fuzzy weighted support vector domain describing algorithm based on sample modeling about positive-negative type is proposed on multi-faults classification of AUV in this dissertation, in view of the situation that traditional describing algorithm of support vector domain description is only for one type of sample, and the problem about higher classification error rate caused by not considering the imbalance of samples numbers and different degree of sample density distribution in the process of classification. This algorithm adds negative type sample into positive type to training in the modeling process. Meanwhile local density and category weight are added, which makes the hyper sphere decision boundary to encompass the object sample and reject the non-object sample. Contrast experiment indicates that the improved algorithm proposed in this dissertation shows higher precision than the traditional algorithm. A dynamic model of multi-faults classification of AUV is built based on improved support vector domain description algorithm which proposed in this dissertation. This dissertation has proposed a method of selectivity adjustment based on distance determination, aiming to solve the problem of dynamic regulation of online failure mode feature space in the process of AUV multi-faults classification. Moreover, with the problem of determination of fuzzy sample points classification, this dissertation has proposed a strategy determinant about fuzzy sample points, which is based on the improvement of the traditional classification rule and considering fully about different size of classified balls. Both theoretical analysis and simulation experiment proves the rationality of strategy determinant proposed in this dissertation. At last, a method based on grading classification is proposee in this dissertation to solve multi-faults classification of AUV. The experimental results show that the methods propped is effective.