Online Inference for Adaptive Diagnosis via Arithmetic Circuit Compilation of Bayesian Networks

Abstract : Considering technology and complexity evolution the design of fully reliable embedded systems will be prohibitively complex and costly. Onboard diagnosis is a first solution that can be achieved by means of Bayesian networks. An efficient compilation of Bayesian inference is proposed using Arithmetic Circuits (AC). ACs can be efficiently implemented in hardware to get very fast response time. This approach has been recently experimented in Software Health Management of aircrafts or UAVs. However, there are two kinds of obstacles that must be addressed. First, the tree complexity can lead to intractable solutions and second, an offline static analysis cannot capture the dynamic behaviour of a system that can have multiple configurations and applications. In this paper, we present our direction to solve these issues. Our approach relies on an adaptive version of the diagnosis computation for different kinds of applications/missions of UAVs. In particular, we consider an incremental generation of the AC structure. This adaptive diagnosis can be implemented using dynamic reconfiguration of FPGA circuits.