Adaptive Multi-Model Fault Diagnosis of Dynamic Systems for Motion Tracking

For most real-world systems, the exact description of possible faults is unknown, making these faults difficult to detect, and even more difficult to identify. The most promising way is to use multiple hypotheses for faults to find the best fitting fault model by comparing system measurements with the predictions of the multi-model algorithm. However, this may lead to the need for infinite hypotheses. We propose a novel multi-model approach that considers a small number of different models with a known macro-structure and unknown parameters, combining system identification with simultaneous fault diagnosis. The unknown parameters in the models are estimated using a maximum likelihood approach. The fitted models are then used in an interacting multiple model algorithm to determine the most likely model that best describes the system behavior at any moment in time. An overfitting problem emerging from short data sequences is discussed, and two solutions are introduced. First, a regularization term in the probability estimation is suggested to penalize frequent parameter changes that signal possible overfitting. Second, an algorithm with a shifted data set is presented. The effectiveness of the algorithms is demonstrated on a motion tracking problem where the different fault hypotheses represent the macro-behavior of a moving object, and the real system switches between different modes. In a comparison, the proposed algorithms are the only ones that reliably identify the defined faults. They can be easily adapted to other fault diagnosis problems.