Research On Technology Of UUV Data Mining Oriented To Marine Environment

Since the system design, sensor precision and control algorithms for UUV havebeen gradually improved, the UUV itself can usually generate large scale data. Aiming atimproving its ability of self-diagnosing and perceiving in the marine environments, thisdissertation is expected to discover useful knowledge in the UUV data throughtechniques in data mining.Three typical data mining methods, namely Association Rule, Naive BayesianClassifier (NBC) and Support Vector Machine (SVM), are utilized to analyze the voltagedata of the lithium-ion battery and dead reckoning data in UUV. All these methods arebased on the fundamental theories and algorithms in data mining and adapted accordingto the characteristic and analysis demand of the UUV data. The specific researchcontents and results are as follows:The features of UUV data in marine environments are studied and the data samplesunder different features are incluced, thus providing evidence for chosing proper datamining methods. Besides, we review some fundamental concepts and mechanisms indata mining. And the feasibility of applying data mining techniques to UUV data is alsoanalyzedBy virtue of the Association Rule based on the Apriori algorithm, a new method isproposed to extract the features describing the battery damage. When the UUV ismoving, there exists some feature difference that can distinguish the damaged batteriesfrom the normal batteries. Thus the feature selection problem in modeling the lithium-ion battery damaging can be sovled effectively.Naive Bayesian Classifier (NBC) is employed to evaluate the availability oflithium-ion battery in UUV. The state of the battery goes through a transitional processchanging between being available and unavailable, whose break point can be detected bythe NBC. The evaluation results accord with the ground truth, which indicates theeffectiveness of the adopted method.Support Vector Regression (SVR) is used to predict the error of dead reckoning.Instead of solving for the mathematical models directly, SVR can find the relationshipbetween the speed&pose of UUV and its dead reckoning error. Therefore this error canbe successfully predicted.