Sensor placement for diagnosability

Abstract

The concept of diagnostic accuracy is examined and redefined to support specific criteria for sensor placement. If the correctness of diagnoser operation is assumed, then any failure to diagnose accurately must be attributable to an inadequacy of sensor data. Inaccuracy in diagnoses can be expressed solely in terms of additional candidates whose faults cannot be ruled out. With ambiguity as the determiner of the quality of a diagnosis, user-defined diagnosability requirements can be expressed in terms of the types and instances of components which are permissible exceptions to perfect diagnosis. This requires a working diagnoser and a simulator, along with sets of system configurations, component fault modes, potentially measurable parameters, and uniqueness requirements for fault isolation. From these, the diagnosability of individual components can be determined for a particular attached subset of the available sensors. An optimum sensor assignment is one that satisfies the most requirements with a fixed number of sensors or, conversely, that minimizes the sensor requirements to achieve a given threshold of diagnosability. The considerable complexity of this search is reduced by exploiting sensor set minimality, structural knowledge, and diagnosis-free extension to the system level. Global optimization and sensor allocation do not add to the number of diagnoses required for diagnosability analysis. Finally, corrective measures are discussed for use when residual costs remain too high, or when redundancy is too low.