Research On Multi-Source Information Fusion Based On Ontology For RSO Recognition

Space is the commanding highland of information for battlefield,with the advantages of information acquisition and transmission.He who takes over space definitely owns the power of acquiring information.Space object recognition is a foundation for realizing space situational awareness.Due to the limitations of various observation conditions,it is difficult for a single sensor to observe the state of the space object in a real-time manner,which is a bottleneck in the identification of the space object.Therefore,fusing multi-source data for space object recognition is an important choice to solve this problem.On the basis of ontology theory,this dissertation focuses on four scientific issues in space object recognition using multi-source data under distributed conditions.By comprehensively applying the related technologies of knowledge engineering,machine learning,multi-agent system and argumentation theory,This study systematically and deeply investigated ontology developing methods for target recognition,ontologies for space object recognition,zero-shot learning techniques for object recognition of unseen classes,and distributed space object recognition methods based on argumentation.This study owns a good originality and flexibility.The main research work and innovation of this dissertation are as follows.(1)An application ontology development method named SPIRALS for object recognition is proposed.This method provides a whole workflow for constructing object recognition ontology.It includes the use of domain knowledge and machine learning knowledge.It adds order-independent machine learning rules as the machine learning knowledge to the object recognition ontology,in order to further improve the performance of the ontology-based object recognition method and enhance the intelligibility of the object recognition.It contains the data-driven application ontology assessment method which is proposed for evaluating and promoting the object recognition ontology.An ontology-based space object recognition system was constructed.Experiments showed that SPIRALS can effectively improve multiple indexes which evaluate space object recognition.(2)Under the guidance of SPIRALS,the ontology for space object recognition named OntoStar is constructed.The space object recognition system named Clairvoyant is developed based on OntoStar.This system uses two different reasoning mechanisms which are object recognition reasoning using general inference engine Pellet and object recognition reasoning using dedicated inference engine.Clairvoyant is capable of recognizing space object with high accuracy.It is also able to maintain high object recognition accuracy and owns good robustness in the absence of features due to limited observations.Clairvoyant provides an understandable and verifiable process for space object recognition,which helps to improve the interpretability and credibility of space object recognition.Clairvoyant is more efficient using the dedicated inference engine than using general inference engine.(3)With the deepening of space research and the continuous development of space technology,space objects of unseen classes emerge continuously.Data for these space objects are scarce.Hence,there is no large amount of observation data available for training and adjusting model parameters.Zero-shot classification technique uses the knowledge learned from historical data of objects and domain knowledge about objects of unseen classes to achieve automatic classification of objects for unssen classes.This paper proposes a zero-shot classification method named CORL which combines ontology and reinforcement learning.This method first learns the hierarchical classification rules from the attribute description of the object classes based on ontology.These rules only contain the discriminant attributes.Then an optimal policy is obtained using reinforcement learning adaptively,allowing selecting the most discriminative rules for zero-shot classification.CORL achieves efficient classification of objects for unseen classes.(4)There may be situations where the recognition results are inconsistent when fusing multi-source data for space object recognition.This paper proposes a multi-agent collaborative argumentation model named AABO and its argument construction method to achieve multi-sensor decision-level information fusion.Arguments are different from evidence.Different from evidences,arguments are justifications for conclusions using evidences.Therefore,there is a difference between the aggregation of arguments and the accumulation of evidence.Arguments represent the basis for decision-making more accurately than evidences.In AABO model,different sources of information generate different arguments through processing sensing information.These arguments from different sources aggregate through dialectical analysis in the process of multi-agent argumentation,resulting in refinement of supporting the original conclusions and obtaining the best decision results.AABO can make correct decisions when the space object recognitions from distributed and multi-sources are inconsistent.It can also provide interpretable global object recognition results.