Combining Similarity Measures and Left-Right Hidden Markov Models for Prognostics of Items Subjected to Perfect and Imperfect Maintenance

In the nuclear industry, high system reliability requirements are essential since in-service failure can result in undesirable consequences in terms of costs or safety. However, the current approach to maintaining systems and components is costly and known to involve overly conservative periodic maintenance activities. It is, therefore, appropriate to develop monitoring, detection, and predictive tools to enable operators to create optimal maintenance strategies. These strategies can vary from the substitution of an item to its repair, intending to avoid unexpected consequences. The repair can restore the item’s functionality to an as-good-as-new condition (perfect repair) or sometimes can keep some accumulated degradation and change the item’s degradation rate (imperfect or partial repair). Current techniques and models that can perform prognostics with extraordinary accuracy are often designed on the assumption that following maintenance, the item is restored to an as-good-as-new condition. When these models are used to predict items that follow imperfect repairs, the predictions are likely to be inaccurate. Therefore, the present work focuses on the condition-based prognostics of items, considering and handling the criticalities that arise after the items undergo different kinds of repairs. The proposed solution involves a data-driven framework that employs Left-Right Gaussian Hidden Markov Models (LR-GHMMs). These models can intrinsically manage accumulated degradation. The idea is to train different LR-GHMMs, each specific to a degradation path, and then combine them to cover possible intermediate paths. The effectiveness of the approach is tested in two case studies. In the first one, we consider simple artificial sequences that are useful to explain the method’s capabilities. In the second case study, we consider semi-simulated nuclear data describing the degradation transients of a condenser that undergoes fouling. The framework is trained with data collected from items that start without accumulated degradation. The test data represent either new items or items that undergo imperfect repairs. The results demonstrate an attractive elasticity of the framework in adapting to nonstandard degradation behaviors. In addition, the applications provide interpretable and highly accurate outputs.