Sequential detection of soft failures in linear systems

On-line and off-line schemes for sequentially detecting significant system changes which can be represented by deviations in the statistical parameters of errors is considered. The model used does not specifically include the nature and time onset time of failure. The on-line scheme utilizes a Kalman filter corresponding to the design values of statistical system parameters and is based on a recursive relation for the nonwhite filter innovations. The scheme has modest computational requirements and involves monitoring the sampler number trajectory for sudden and nontransient changes. Following detection of a soft failure, the estimation schemes can be used to identify the failed parameters. The off-line scheme is a maximum likelihood estimate. Two examples are presented to illustrate the implementation of the algorithms and demonstrate the detection and robust estimation capabilities of the schemes. The problem of maneuvering targets is reformulated in the soft failure framework via sudden changes in the appropriate variables. The failure isolation capabilities of a bank of detectors scheme are demonstrated via the maneuvering target problem.