SSH-Net: A Deep Neural Network for Predicting Failure Time Distribution Functions under Competing Risks with Application to GPU Data

2020-01-01 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Hardware Reliability · Depth: Expert, extended

Summary

SSH-Net introduces a novel deep neural network (DNN) designed to predict failure time distribution functions under competing risks, specifically addressing challenges posed by complex application scenarios and hierarchical data structures. This approach associates the neural network structure with data structures, enabling different covariate groups to influence failure prediction through dedicated sub-networks. SSH-Net outputs cause-specific hazard functions and employs a penalized log-likelihood as its loss function. Validation through simulation studies demonstrated SSH-Net's superior prediction accuracy, consistently outperforming Neural Fine Gray (NFG) and DeepHit across Brier score, AUC, and RMSE metrics. The model was successfully applied to the Titan GPU failure time data, comprising 19,319 GPU units with 1127 OTB and 3093 DBE failures over nearly seven years, effectively predicting Cumulative Incident Functions (CIFs) and identifying critical factors like cage and cabinet locations influencing GPU reliability.

Key takeaway

For Machine Learning Engineers or Reliability Engineers modeling time-to-event data with competing risks, especially in systems with hierarchical covariates, you should consider SSH-Net. Its structured neural network design and penalized log-likelihood loss consistently deliver superior prediction accuracy (lower Brier score, higher AUC) over models like NFG and DeepHit. This approach helps you better understand and predict failure distributions, such as GPU failures influenced by physical location, enabling more informed maintenance and design decisions.

Key insights

SSH-Net improves competing risks survival analysis by structuring neural networks to data hierarchies, enhancing prediction and interpretability.

Principles

• Neural network structure should align with data hierarchies for better tuning.
• Penalized log-likelihood loss improves continuous time prediction accuracy.
• Segmented sub-networks enhance covariate group impact on predictions.

Method

SSH-Net constructs cause-specific hazard functions as piecewise constant segments. It processes hierarchical covariates via separate sub-networks, using a penalized log-likelihood loss with a smoothness term to predict failure time distributions.

In practice

• Predict GPU failure types (DBE, OTB) using hierarchical covariates.
• Analyze relative risk among failure types in complex engineering systems.

Topics

• Competing Risks
• Deep Neural Networks
• Survival Analysis
• GPU Reliability
• Failure Time Prediction
• Hierarchical Covariates

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• PHINN: Persistent Homology Inspired Neural Network for Rare-Event Time Series Generation — Rare events in time series possess stable topological fingerprints, enabling more effective generative modeling than statistical moments.
• Asymmetric-Loss-Guided Hybrid CNN-BiLSTM-Attention Model for Industrial RUL Prediction with Interpretable Failure Heatmaps — A hybrid CNN-BiLSTM-Attention model with asymmetric loss improves turbofan RUL prediction and interpretability.
• Stochastic Gradient Descent in the Saddle-to-Saddle Regime of Deep Linear Networks — SGD noise provides information about feature learning progression in deep linear networks without altering saddle-to-saddle dynamics.
• Bullet Trains: Parallelizing Training of Temporally Precise Spiking Neural Networks — Parallel associative scans and differentiable numerical solvers enable efficient, precise training of event-based SNNs.
• DIY #21 - Step-by-Step Guide to Time Series Forecasting with RNN — Vanilla RNNs effectively forecast cyclical time series data by learning temporal patterns through sequential processing.
• When Smaller Wins: Dual-Stage Distillation and Pareto-Guided Compression of Liquid Neural Networks for Edge Battery Prognostics — Dual-stage distillation and Pareto-guided compression enable accurate, compact LNNs for edge battery prognostics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.