An integrated network architecture for data repair and degradation trend prediction

Abstract

This paper proposed a network framework, namely DR-DTPN, which integrates data repair and degradation trend prediction to address the serious deviation of equipment degradation trend prediction caused by the missing monitoring data and distribution changes. DR-DTPN considers the trend and periodic variations of the signal, and dynamically infer the latent vector of the spatial spectrum. The shared temporal dynamics of the training window are encoded as polynomial and trigonometric function, and the spatial spectral latent vector is inferred by minimizing the reconstruction error of the training window data through convex optimization. Dynamically inferred spatial spectral decomposition, which captures current temporal dynamics and correlations between features. DR-DTPN simultaneously interpolates past missing values and predicts future degradation trend values through optimal latent space spectral decomposition. Finally, this paper has validated the data interpolation performance and prediction performance of DR-DTPN on the PHM2012 bearing degradation data, and further verified its effectiveness through engineering applications on wind turbines and aero engines.