People

Research activity in 2001

Model based reasoning has been the main area of investigation of our group in recent years. Several aspects have been explored.

From a methodological point of view, the recent contributions can be grouped as follows:

• Analysis of the problems arising when modeling and diagnosing complex dynamic systems. In particular, we focused on the study of the problem that have to be faced when modeling complex real world system with controlled dynamic behavior.

• Modelling with Process algebras. We investigated the use of the formalism of Process algebras (which is widely used for modeling in other areas of computer science) which proved to be a very flexible and powerful language for model-based reasoning.

• Integration of diagnosis in the design process of physical devices. If one analyses the design process of any significant complex systems, it is very common to see that diagnostic issues are usually taken into account only at the end of the process and are not integrated with the rest of the process. In particular, during the critical phases of the design process, when the actual architecture of the system is conceptualized, the control strategies are defined and models or prototypes of the system are simulated, diagnostic issues are not taken into account. Not only does this mean that the diagnostic software is not developed together with the control software, but, more critically, that issues such as the diagnosability of the system being designed or the analysis of the FMEA (Failure Modes Effect Analysis, which is very useful to discover safety critical faults or failures) are seldom and only partially considered. The goal of our work is to define defining a new process in which these issues are integrated within the design of a system and of its control strategies. The project also aims at defining and implementing a software toolkit supporting the new process. The toolkit integrates applications for design and simulation (e.g., Matlab Simulink) and model-based reasoning systems for diagnosis-related tasks.

• Embedding model-based diagnosis into automotive systems. The goal is to study how MBR technologies can be used to develop the diagnostic software that is emebdeed into real complex systems, especially in the automotive field.

• Multi-modal reasoning for diagnostic problem solving. In the last few years we have investigated the computational complexity of model based approach to diagnostic problem solving and we have studied under which condition the integration of case-based reasoning with model-based reasoning is useful for reducing the computational effort. In particular we have shown that instead of a fixed strategy (first CBR and then MBR is CBR fails in solving a case), the adoption of opportunistic strategies allows significant increases in performance. We have shown that significant gains both in terms of computation effort and in term of competence can be gained when it is possible to decide to activate the CBR or the MBR component on the basis of the difficulty of the diagnostic problem at hand. Since competence of the Case-based reasoning module depends on the content of the case memory, We have also investigated which is the impact of the content of the case memory on the quality of the solution of a diagnostic problem as well as on the performance of the multi-modal reasoning system. In order to deal with the "utility problem" we have developed learning techniques able to decide when to add a new case to the case memory and when cases have to be forgotten . In particular we have shown that significant results can be obtained both with learning (and forgetting) techniques based on notion of competence and with learning strategies based on the notion of usefulness of the cases in the case memory. These results have been obtained by a large experimental effort using various versions of the ADAPtER system for solving thousands of diagnostic problems of two different domains.

• Concise representation of diagnoses. In some domains the MBR approach to diagnosis produces too many diagnoses to be shown to the human supervisor: this is often the case in on-board diagnosis where the measurements available via sensors may not be sufficient for discriminating among competing diagnoses. For this reason, we developed a representation based on the notion of "scenario" able to capture in a unique structure a large number of tentative diagnoses and we have exploited such scenarios to develop efficient diagnostic strategies. These strategies has a significant impact in reducing the computational time and in reducing the information overload to the human supervisor since only the most plausible diagnoses are generated. Further work is currently being done for developing a library of abstract models, each model representing just the faults of the components which can be discriminated by taking into account a given class of observations.

The main area of application for the MBR techniques described above is the automotive one. In 2001, in fact we participated actively in a European Project: The IDD (Integrating Diagnosis and Design).

The project is supported by the EU under the Substainable Growth Programme (V Framework) and the partners are: Centro Ricerche Fiat, Daimler Chrysler, Renault, PSA - Peugeout Citroen, Magneti Marelli, Occ’m, Universitè Paris XIII, Technical University Munchen, Università di Torino.

The aim of the project is to integrate model-based diagnostic techniques in the tools used during the design of a system (CAD tools, simulation tools such as Matlab/Simulink). In this way the diagnosability of a system can be verified (and improved) during he design process and the diagnostic rules to be implemented in the control of the system can be produced automatically at the end of the design.

Another relevant domain for diagnostic problem solving concerns autonomous robots for space applications. In a project supported by ASI (Agenzia Spaziale Italiana) we have developed the diagnostic agent by taking as a test bed the SPIDER robotic arm. In this framework we have tested the techniques developed for a concise representation of the diagnosis and we have exploited the formulation of diagnosis in terms of constraints and variable assignment.. In particular, the diagnostic agent is able to deal with both detailed and abstract models and generates only the most probable diagnoses to be presented to the user via a graphical interface.