Computationally Efficient Tiered Inference for Multiple Fault Diagnosis

Abstract

Diagnosing multiple-component systems is difficult and computationally expensive, as the number of fault hypotheses grows exponentially with the number of components in the system. This paper describes an efficient computational framework for statistical diagnosis featuring two main ideas: (1) structuring fault hypotheses into tiers, starting from low cardinality fault assumptions (e.g., single fault) and gradually escalating to higher cardinality (e.g., double faults, triple faults) when necessary; (2) at each tier, dynamically partitioning the overall system into subsystems, within which there is likely to be a single fault. The partition is based on correlation between the system components and is dynamic: when a particular partition is ruled out, a new one is constructed based on the updated belief. When no viable partition remains, the search proceeds to the next tier. This approach enables the use of single-fault diagnosis, which has only linear complexity, to the subsystems avoiding exponential hypothesis explosion. We demonstrate the concepts and implementation via examples and simulation. We analyze the performance and show that for practical systems where most components are functioning properly, the proposed scheme achieves a desirable tradeoff between computational cost and diagnosis accuracy.