Inductive Situation Evolution Modeling

The effective monitoring and control of large-scale infrastructures in domains like traffic, energy or industry gets more and more challenging, not least due to the tremendous technological innovations over the last years. Game-changing technologies like the "Internet-of-Things" deliver a huge stream of sensor data about current states of the monitored environment like traffic flow, energy consumption or machine wear, providing the basis for an early detection and prediction of potential critical situations (e.g., wrong-way driver) and their evolutions (e.g., wrong-way driver moves towards tunnel), thus allowing for alleviation or even prevention of critical situations. Such vast amount of data, however, entails also the massive danger of information overload, thus making an adequate IT-support on basis of so-called "Situation Awareness Systems" indispensable. The crucial success factor for the effectiveness of these systems are suitable patterns of relevant critical situations in terms of so-called "Situation Evolution Models", serving as the pre-requisite for an automatic detection thereof within the sensor data stream at runtime. Existing approaches, however, either focus on a manual, and thus elaborate and error-prone, creation of these models through domain experts or on an automatic analysis of historical sensor data on basis of machine learning algorithms, which are, however, not able to recognize situations which are rare or never occurred in the past. Combined approaches considering both, human intelligence and sensor intelligence are still in their infancy. The goal of this project proposal is therefore to intertwine both approaches in an innovative and synergetic way to allow for an interleaved and iterative creation of precise models of relevant critical situation evolutions together with predictions of their possible future occurrence as well as to provide for an automatic deployment of these models to existing Situation Awareness Systems independent of a certain domain. On the one hand, driven by an appropriate automatic data analysis, already existing Situation Evolution Models should be validated, complemented and refined and, on the other hand, these models should be employed in order to target the automatic analysis in a more precise way. As underlying research methodology, as common in engineering disciplines, serves the constructive "Design Science Research" method, whereby a software prototype is realized allowing to evaluate the effectiveness of the developed approach on basis of real world sensor data. Finally, in a broader context, this project proposal is expected to provide a valuable contribution to the burning question of how to harness reliable, machine-processible knowledge out of the permanently growing unreliable big data stream, we are confronted with these days.

The research project "Inductive Situation Evolution Modeling" (inSiTUEVO) investigated how intelligent information systems can exploit existing situation logs to derive abstract models of these situations, which are a prerequisite for enabling automated situation detection and prediction. The proposed basic research on algorithms and machine learning would find practical application in control center monitoring software, as required in road traffic control, air traffic control, maritime monitoring, cyber security, power plants and power grids, industrial production etc. In such control rooms, human operators need to spot critical situations emerging in the monitored environment and project their further evolution, in order to undertake the appropriate (counter)actions. In road traffic control, such a situation could be the formation of a traffic jam following an accident. To prevent this situation from escalating, the human control center operator needs to take appropriate action, like displaying corresponding warnings on variable message signs to inform approaching motorists, until the traffic jam has dissolved following the clearance of the accident site. Intelligent software should support the operator in this time-critical situation assessment: By describing such situations abstractly through the spatio-temporal relations holding between their comprised events (e.g., that the traffic jam event appears after and is located at the rear of the accident event), these "situation models" can be compiled to corresponding pattern recognition and tracking rules. Given such rules, the vast real-time data streams collected from the monitored environment can be automatically filtered to detect corresponding events, fuse them to a coherent situational picture, and provide forecasts on its further development. But how can such situation models be obtained? The rules for such "expert systems", as the name implies, are usually provided by human domain experts, who "write down" their domain knowledge into corresponding formal descriptions or models that can be processed by software, representing a time-intensive and cumbersome process. To address this problem, inSiTUEVO investigated whether such situation models could be derived "inductively", by rather learning them from already recorded data. The devised "situation mining" approach represents a dedicated data mining algorithm for reconstructing situation models from recorded event logs. Thus, human domain experts can be proposed with automatically mined models and are hence relieved from the tedious work of manually specifying every detail of potential situations of interest, allowing to incorporate both human expertise and empirical data. The developed formal notation moreover paved the way for adapting established sequence prediction techniques from natural language processing to the task of situation evolution prediction. With these learned predictive models, human operators' decision making can be supported by providing empirically grounded forecasts. By leveraging Information Fusion and Artificial Intelligence techniques, the proposed research aims at supporting information systems' human users in maintaining situation awareness of their monitored environments.