Feature Selection for Remaining Useful Life Prediction in Hard Disk Drives with Missing Data

Gabriel L. S. Felix, Francisco Pereira, Francisco D. B. S. Praciano, João P. P. Gomes, Javam C. Machado2023 · DOI

Abstract

This paper proposes a two-stage feature selection approach for the problem of Remaining Useful Life (RUL) prediction in Hard Disk Drives (HDDs) with missing data. First, a wrapper method is employed, utilizing a regression estimator to identify the most informative features for RUL prediction. The selected feature set is then evaluated in the second stage using a neural network model, with a focus on assessing the imputation performance for missing data. The goal is to determine a feature subset that enhances RUL prediction accuracy and exhibits robustness in handling missing data scenarios. This approach addresses the challenges of missing data and provides insights into the most relevant features for accurate RUL prediction.

Introduction

Hard Disk Drive (HDD) failures represent a significant challenge for data centers, leading to data loss and operational downtime. Modern drives utilize Self-Monitoring, Analysis, and Reporting Technology (SMART) to track health metrics. However, applying machine learning to this data is complicated by high dimensionality—manufacturers report numerous attributes, many redundant—and frequent missing data due to irregular sampling or system gaps. This study aims to identify a compact, "imputable" subset of SMART features that maintains high RUL prediction accuracy even in the presence of data gaps.

The researchers utilized the Backblaze (2023) public dataset, specifically focusing on the Seagate ST4000DM000 model. The dataset comprised over 1.6 million daily records from nearly 5,000 drives, covering the final 360 days of their operational life.

Methods

To manage the computational complexity of performing feature selection directly on deep learning models, a two-stage framework was developed.

Stage 1: Feature Filtering. A sequential wrapper method was applied using 15 lightweight regression estimators (e.g., Lasso, Decision Trees, ElasticNet) to identify candidate feature subsets that best predict RUL in a simplified setting.

Stage 2: Neural Evaluation. The candidate subsets were evaluated using a specialized Recurrent Neural Network (RNN) with a Long Short-Term Memory (LSTM) layer. As shown in Figure 1, the architecture splits into two distinct branches: an RUL branch for predicting days until failure and an imputation branch for reconstructing missing SMART values. This dual-task approach ensures that the selected features are not only predictive but also robust to data loss.

Dual-branch RNN architecture for simultaneous RUL prediction and data imputation.

Results and Discussion

The evaluation revealed that non-linear feature selection methods, particularly Decision Trees (DT), outperformed linear models when integrated into the neural network. This suggests that the relationships between HDD health metrics and failure are fundamentally non-linear.

The study identified a "Golden Subset" of 12 SMART features (IDs: 1, 5, 9, 10, 184, 188, 191, 192, 193, 197, 240, and 242). The Decision Tree subset achieved the lowest Mean Squared Error (MSE) for RUL prediction at 0.01527. While the Extra Tree method showed slightly better imputation performance, the DT subset provided the best overall balance for the primary prediction task.

Comparison of regression and imputation losses across different feature selection estimators.

As illustrated in Figure 3, there is a clear trade-off between regression loss and imputation loss across different estimators. The Decision Tree and Extra Tree models occupy the lower-left region of the plot, indicating superior performance in both tasks compared to models like OMP or Poisson regression.

Conclusion

This research demonstrates that a two-stage feature selection process can effectively identify a robust subset of SMART attributes for HDD failure prediction. By integrating imputation directly into the model architecture, the framework handles the inherent "holey" nature of real-world monitoring data. Future work will investigate the generalizability of the identified 12-feature subset across different HDD models and manufacturers to determine if a universal health-monitoring standard can be established.

Summary updated

Abstract

This paper proposes a two-stage feature selection approach for the problem of Remaining Useful Life (RUL) prediction in Hard Disk Drives (HDDs) with missing data. First, a wrapper method is employed, utilizing a regression estimator to identify the most informative features for RUL prediction. The selected feature set is then evaluated in the second stage using a neural network model, with a focus on assessing the imputation performance for missing data. The goal is to determine a feature subset that enhances RUL prediction accuracy and exhibits robustness in handling missing data scenarios. This approach addresses the challenges of missing data and provides insights into the most relevant features for accurate RUL prediction.
BibTeX
@misc{felix2023feature,
  author = {Gabriel L. S. Felix and Francisco Pereira and Francisco D. B. S. Praciano and João P. P. Gomes and Javam C. Machado},
  title = {{Feature Selection for Remaining Useful Life Prediction in Hard Disk Drives with Missing Data}},
  year = {2023},
  doi = {10.5753/sbbd\_estendido.2023.233372},
}
Figure