Agile knowledge management for situation awareness

Abstract

In the information age, access to data is easily achieved thanks to the development of new advanced sensors and information sources, able to measure all kind of features, to acquire several kind of information and to transfer those data fast and effectively all over the world. Taking into account newly developed sensors and networks, the access and communication of information are not unyielding and crucial as their analysis, aggregation and elaboration. It is within this context, that of Data Fusion finds its applicability. At the end of the 1990, Data Fusion doctrine is formalized as the ensemble of scientific techniques and algorithms, properly implemented in a single framework that is able to: (a) support human operators to gather huge quantities of heterogeneous data (some of which may not be synchronized) from sensors observing the scenario of interest; (b) detect and classify objects acting in the scenario; (c) understand the relationships among them, and the intent, and threats that they could cause; (d) foresee future evolutions of the scenario; and (e) take the best decisions in order to maximize human operator utility. The above mentioned process has been formalized in the Joint Directors of Laboratories (JDL) model, and progressively revised mainly to give greater emphasis to the human operator as an active sensor within the process, and to evaluate the quality of the whole inference process. Within the Data Fusion process, the goal of Situation Awareness (JDL Level 2) is to recognize relationships existing among objects observed in the scenario, in order to recognize situations of interest, and evaluate their threats. Hence, Situation Awareness should help human operators to be aware of the context they are observing, especially when the scenario is wide, or phenomena observed are complex and evolve fast. The focus of this PhD thesis is exactly Situation Awareness, and in particular knowledge management: in order to recognize situations and infer them from observations, knowledge models describing situations of interest must be effective, correct, and should be able to catch relevant and discriminant aspects. The definition of a good knowledge model is crucial for effective Situation Awareness, and it is usually hard because it requires experience in the domain, or the availability of huge quantities of data to be input to learning algorithms (that generates usually models difficult to interpretate). Moreover, once the model has been defined, the evaluation of its quality is difficult, especially in real-time, because the truth about the observed situation is not known. The goal of this thesis has been the investigation about effective knowledge management for correct inferences, and in particular the following aspects of knowledge management have been considered: • real-time knowledge model construction with regard to specific situations or events of interest, adopting Data Mining techniques; • real-time knowledge model refinement, according to metrics expressing the adequacy of the model to the observations gathered. Knowledge models employed in Situation Awareness usually differ from each depending on the mathematical approach adopted (Bayesian approach refers to Bayesian Networks, Hidden Markov Models requires a probabilistic inference algorithm, Evidence theory refers to cause-effects models). In this work, real-time model construction has been apply to Hidden Markov Models; while real-time knowledge refinement has been investigated with regard to Evidence Theory. Moreover, considerations derived from the implementation of Situation Awareness frameworks within the military context and critical infrastructure protection domain have been reported. Majour results of this research can be summarized in the characterization of the agility measure, able to quantify the capability of a model to revise itself by evaluating inconsistencies, contradictions and errors, and taking into account uncertainty of information employed. Model agility has be identified as a powerful feature in JDL Level 4 Process Refinement, because it can guide and improve the overall data collection process, eventually cueing the user or the system to search for lacking information. Main features identified for an agile model are the following: • an agile model does not require to be perfect since its construction: it can be obtained with imperfect knowledge of the whole system, because it is able to learn from its experience; • agility extends the model lifetime: agile models are able to manage a greater number of scenarios that maybe were not even included when the model was created; • an agile model is more resilient, more robust, and able to perform better and wider range of real life scenarios. Investigations about agility measure within Evidence Theory, have highlighted the inability of knowledge models and algorithms to recognize timedependent situations. In this regard, the trend of the empty set mass has been identified as an agility measure, able to identify the fitness of the model to the observed situations, and in particular model inadequacy to describe, and hence recognize, time-dependent patterns. It has been shown how to employ the measure for model review and correction, in order to allow in Evidence Theory dynamic pattern recognition, besides to static classification. Finally, research conducted for this PhD thesis have lead to the definition of a system architecture combining Data Mining and Data Fusion techniques in order to allow the construction of knowledge models able to recognize effectively situations of interest, that can be specified by the user in realtime. In the proposed framework Data Mining approach is employed to define correlations among data stored in databases, and events or objects of interest for the user; mined correlations are employed to build in real-time knowledge models to be adopted in the Situation Awareness process.