A multi-hypothesis estimation approach to diagnosis and prognosis of degrading systems

Complex aerospace systems are generally organized in a hierarchical manner, using a system-of-systems approach. A single capability area in such a system has numerous fault modes. Further, there may be several damage sensor measurements, each with its strengths and weaknesses in terms of accuracy. We propose a multi-hypothesis estimation (MHE) approach for degrading systems to address the problem of computing diagnostic likelihoods of damage state. In this, rather than storing marginal probabilities at the LRU level at each time step, which unnecessarily loses information, an adjustable number of possibilities of fault (i.e. fault hypotheses) are maintained. We only marginalize the likelihood distribution after sufficient information has been collected to utilize both fast and slow timescale patterns. The current damage estimation approach provides a unified framework in which to perform both sensor and temporal fusion of information to inform the fault likelihood estimates. Further, the current formulation is sufficiently general to handle coupled degradation dynamics, and a variety of differing failure models including expert system models and dynamical systems models. Our algorithm is based on standard Bayesian and Markov assumptions of the fault dynamics, and can be adjusted to trade off computational requirements with estimation fidelity (by adjusting the number of hypotheses kept). We demonstrate the approach and the algorithm using an example, patterned after a degrading mechanical system such as a bearing