Incremental learning approach for improved prediction

Prognostics is the key function in prognostics and health management (PHM), which can provide remaining useful life of systems in real-time so that timely maintenance plans can be scheduled to avoid system downtime and even catastrophic events. In system prognostics, fault degradation models are necessarily established to describe the fault evolution dynamics and used to extrapolate the future health conditions. However, it is very challenging to build an accurate fault degradation model considering the complex fault growth dynamics and numerous modeling uncertainties, such as unit to unit variation. Particularly, in data driven modeling methods, the variations of loading conditions, environments and usage patterns will influence greatly the fault modeling accuracy. Some research has been conducted to tackle this problem by utilizing real-time monitoring data to update the fault model in terms of model parameters and even model structures to accommodate these varying factors. But whenever new data are available, it becomes difficult to determine how to retain the prior learned model while also learning new fault degradation dynamics. That is, how to learn new knowledge without forgetting what was learned previously. In this paper, we develop a new model update and fusion method for prognostics by using incremental learning. A case study is given to validate the developed approach via the battery degradation data.