A geometric alignment approach to parameter-estimation-based fault diagnosis.

摘要

A method for achieving the detection, isolation, and estimation of faults is presented. The systems considered are Single-Input-Single-Output (SISO) systems represented by input-output models. Assuming that the model parameters are multilinear in the physical (diagnostic) parameters, the influence of each physical parameter on the model parameters (called the influence vector) can be interpreted as a fault template line associated with that physical parameter. The influence matrix is assumed to have been computed off-line at the nominal (fault-free) point, and stored for later use in on-line fault diagnosis. After a fault is detected, as indicated by an intolerable change in the identified models, the faulty physical parameter is first isolated and then estimated using the influence matrix, as proposed by Doraiswami and coworkers (1993). It is well known that all parametric-model identification techniques require rich signals, such as a pseudo random binary signal, as test inputs to the system to be monitored. Many industrial systems, however, may not allow the feeding of such (persistently exciting) signals as inputs. This fact may therefore be a major drawback for the above-mentioned parametric-model FD techniques, especially noting that the task of parametric model identification has to be repeated each time the system is to be checked for fault detection. To avoid the use of persistently exciting test inputs required by the identification process in this parametric-model-estimation scheme, a nonparametric-model-estimation method is developed based on the estimates of the impulse response of the system. Also, the generalized case of diagnosing successive faults will be discussed as a parameter tracking problem.