Statistical Properties and Design Criterions for AI-Based Fault Isolation

Fault diagnosis in the presence of noise and model errors is of fundamental importance. In the paper, the meaning of fault isolation performance is formalized by using the established notion of coverage and false coverage from the field of statistics. Then formal relations describing the relationship between fault isolation performance and the residual related design parameters are derived. For small faults, the measures coverage and false coverage are not applicable so therefore, a different performance criteria, called sub-coverage, is proposed. The performance of different AI-based fault isolation schemes is evaluated and it is shown that the well known principle of minimal cardinality diagnosis gives a very bad performance. Finally, some general design guidelines that guarantee and maximize the fault isolation performance are proposed.