A Deep Learning-Based Reliability Model for Complex Survival Data

The reliability of products is a critical issue as it has high economic impacts, especially in current competitive markets. In modern applications, the complex and high-dimensional data of products are collected that can be used for reliability analysis and failure prediction. The existing reliability approaches, however, cannot efficiently model complex covariates and their effects on the time-to-failure of products. In this article, we propose a novel deep learning-based reliability approach to model the complex relationship between the covariates and product failure. To estimate model parameters, neither the traditional deep learning parameter estimation method nor the maximum likelihood estimation method is applicable. To overcome this difficulty, a new model parameter estimation method is developed based on the partial likelihood framework. Furthermore, as there are often only a limited number of samples for real-world reliability problems, a new penalized partial likelihood estimation method is developed for this special circumstance. The developed method is capable of estimating model parameters for censored reliability data. A simulation study is conducted to verify the developed methods. The proposed method is justified by a real-world case study of the reliability analysis of materials. The case study shows that the proposed model outperforms the existing ones.