Time-to-Event Prediction Using Deep Learning Models: Application to GPU Failure Data

Neural network models gain significant popularity in recent years due to their ability to identify complex patterns. In the field of reliability research, efforts are made to develop neural network models for predictive reliability. However, research focused on utilizing neural networks to forecast graphics processing unit (GPU) failures remains limited. Analyzing the reliability of GPUs is crucial for effectively maintaining GPU systems and preventing issues related to GPU failures, such as safety concerns and interruptions in simulations. Most studies concentrate on predicting the remaining lifespan of GPUs, whereas our objective is to predict both lifespan and failure status. Furthermore, there is a growing trend of integrating statistical modeling with neural networks. By incorporating statistical concepts, we can create models that better represent reality and improve interpretability. Additionally, the model efficiently manages complex data structures like hierarchical clusters, which enhances its capacity to generalize. To improve our neural network model, we combine statistical concepts with neural network techniques. Building on these motivations, we outline our research approach as follows. We describe the development of deep learning models for GPU failure prediction, and explain how statistical concepts are integrated to improve model performance. Chapter~ref{cha:genintro} provides a general overview of the research presented in this dissertation. It begins by outlining the motivation behind the study and defining the research problem. Next, it offers a brief summary of the contributions made by this research. The chapter also explains the concept of neural network models, including their training processes. Furthermore, it covers embedding layers and the use of the GPU dataset. Finally, it introduces multitask output for neural networks, which will be utilized in Chapter~ref{cha:NNspatialRE}. Chapter~ref{cha:DL} introduces a deep learning approach for predicting GPU failure time and status. We propose two distinct neural network architectures to address these prediction tasks: TypeEmbedNet for failure type prediction and TimeEmbedNet for failure time prediction. Since our predictors are categorical variables, we employ embedding layers to transform them into lower-dimensional vectors. We utilize the Cross-Entropy Loss function to optimize TypeEmbedNet and the Mean Squared Error (MSE) for TimeEmbedNet. We develop evaluation metrics to comprehensively assess the models' performance in predicting both failure time and status, taking into account the characteristics of an imbalanced dataset. We integrate data frequency, F1 score, recall, and precision into the MSE by applying penalties based on classification outcomes. In Chapter~ref{cha:NNspatialRE}, we introduce a deep learning model for competing risks. We develop a custom loss function for competing risks that incorporates survival theory and is specifically adapted for time prediction. We compare this model to other machine learning approaches and benchmark it against a previously introduced parametric model. Additionally, we test the integration of spatial random-effect embeddings to model GPU failure outcomes. To achieve this, we compute correlation matrices based on two different spatial structures—physical and logical distances—and apply them using Cholesky decomposition. We describe how we assign a learnable embedding to each GPU location, incorporating these correlation matrices. The neural network outperforms both the machine learning models and the parametric model across various MSE metrics. However, adding spatial random effects to the neural network does not result in a significant improvement in predictive performance. Chapter~ref{cha:conclusion} summarizes the key findings of this study and explores several potential avenues for future research.

Name thesis:degree_name

Doctor of Philosophy

Level thesis:degree_level

doctoral

Discipline thesis:degree_discipline

Statistics

Department dc:contributor.department

Statistics

Grantor dc:publisher

Virginia Tech

Year dc:date.issued

2026

Author dc:creator

• Lee, Lina

Chair dc:contributor.committeechair

• Hong, Yili

Committee members dc:contributor.committeemember

• Kim, Inyoung
• Deng, Xinwei
• Du, Pang

• Competing risk
• Deep learning
• GPU reliability
• Random effect models
• Spatial embedding
• Time to failure prediction

Statement dc:rights

• In Copyright

Licence dc:rights.uri

• http://rightsstatements.org/vocab/InC/1.0/

Language dc:language.iso

en

Dc Identifier Other

vt_gsexam:45214

OAI identifier oai:identifier

oai:vtechworks.lib.vt.edu:10919/141029