Open Access Embargo
Date of Award
Industrial and Manufacturing Engineering
The reliability estimation of products has crucial applications in various industries, particularly in current competitive markets, as it has high economic impacts. Hence, reliability analysis and failure prediction are receiving increasing attention. Reliability models based on lifetime data have been developed for different modern applications. These models are able to predict failure by incorporating the influence of covariates on time-to-failure. The covariates are factors that affect the subjects’ lifetime.
Modern technologies generate covariates which can be utilized to improve failure time prediction. However, there are several challenges to incorporate the covariates into reliability models. First, the covariates generally are high dimensional and topologically complex. Second, the existing reliability models are not efficient in modeling the effect on the complex covariates on failure time. Third, failure time information may not be available for all covariates, as collecting such information is a costly and time-consuming process.
To overcome the first challenge, we propose a statistical approach to model the complex data. The proposed model generalizes penalized logistic regression to capture the spatial properties of the data. An efficient parameter estimation method is developed to make the model practical in case of large sample sizes. To tackle the second challenge, a deep learning-based reliability model is proposed. The model can capture the complex effect of the data on failure time. A novel loss function based on the partial likelihood function is developed to train the deep learning model. Furthermore, to overcome the third difficulty, we proposed a transfer learning-based reliability model to estimate failure time based on the failure time of similar covariates. The proposed model is based on a two-level autoencoder to minimize the distribution distance of covariates. A new parameter estimation method is developed to estimate the parameter of the proposed two-level autoencoder model.
Various simulation studies are conducted to demonstrate the proposed models. The results show that the proposed models outperformed the traditional statistical and reliability models. Moreover, physical experiments on advanced high strength steel are designed to demonstrate the proposed model. As microstructure images of the steels affect the failure time of the steel, the images are considered as covariates. The results show that the proposed models predict the failure time and hazard function of the materials more accurately than existing reliability models.
Aminisharifabad, Mohammad, "Deep Learning Based Reliability Models For High Dimensional Data" (2019). Wayne State University Dissertations. 2213.
Available for download on Wednesday, August 25, 2021