Physics-informed neural network supported wiener process for degradation modeling and reliability prediction

Due to strong data-processing capabilities, machine learning haves been widely applied and combined with stochastic processes to quantify the inherent uncertainty in degradation modeling. These approaches typically first extract health index using machine learning methods, then model them using stochastic processes. While, the machine learning models and stochastic processes are independent of each other, making it difficult to ensure their mutual compatibility. Furthermore, actual available data is often limited, which restricts the accuracy of extracting health indexes through machine learning methods. Hence, this paper proposes a prediction method based on physics-informed neural network supported Wiener process, which includes offline modeling and online prediction stages. In the offline modeling phase, degradation path is fitted using a deep network framework, and degradation mechanics-related prior physical knowledge is embedded into the network along with the Wiener process through parametric expression. Accordingly, a compound loss function is designed to simultaneously train network parameters and process parameters. In the online prediction phase, real-time data is integrated using Bayesian inference methods to update the process parameters, ensuring the robustness of the model. The effectiveness of this method is confirmed using actual datasets, highlighting that the accuracy can be guaranteed even without path information and/or sufficient data.