A degradation modeling method based on artificial neural network supported Tweedie exponential dispersion process

Degradation modeling is crucial for predicting product performance and reliability. Stochastic process-based methods are widely used due to their ability to incorporate uncertainties. These methods typically involve three main components: stochastic process models, degradation paths, and model parameters. However, traditional approaches often overlook the inherent uncertainties in both the model and degradation path, leading to potential modeling errors. This paper proposes a novel approach that combines artificial neural networks with the Tweedie exponential dispersion process framework to adaptively fit the stochastic process model that best reflects the actual degradation trend. The hybrid model is parameterized by process parameters and network parameters, and offline-trained using gradient-based algorithms. For predicting degradation in new individuals with incomplete data, the process parameters are regarded as random variables to account for individual heterogeneity and time-varying uncertainties. Bayesian inference is used to estimate process parameters based on the trained model, with real-time data used to update the parameters for improved accuracy. A simulation dataset based on a non-standard process validate the method’s effectiveness. Furthermore, a real degradation dataset is used to demonstrate its application in engineering scenario. Results show that the proposed approach better captures true degradation trends, offering higher prediction accuracy compared to conventional models.