[Paper Review] Remaining Useful Life Prediction Using Temporal Deep Degradation Network for Complex Machinery with Attention-based Feature Extraction

Yuwen Qin, Ningbo Cai+4 more|arXiv (Cornell University)|Feb 21, 2022

Reliability and Maintenance Optimization6 citations

TL;DR

This paper proposes the Temporal Deep Degradation Network (TDDN), an end-to-end deep learning framework that combines 1D CNN for temporal feature extraction and an attention mechanism to enhance RUL prediction in turbofan engines. The method achieves state-of-the-art accuracy on the C-MAPSS dataset by learning monotonic degradation trends and focusing on key fault characteristics through adaptive attention weights.

ABSTRACT

The precise estimate of remaining useful life (RUL) is vital for the prognostic analysis and predictive maintenance that can significantly reduce failure rate and maintenance costs. The degradation-related features extracted from the sensor streaming data with neural networks can dramatically improve the accuracy of the RUL prediction. The Temporal deep degradation network (TDDN) model is proposed to make the RUL prediction with the degradation-related features given by the one-dimensional convolutional neural network (1D CNN) feature extraction and attention mechanism. 1D CNN is used to extract the temporal features from the streaming sensor data. Temporal features have monotonic degradation trends from the fluctuating raw sensor streaming data. Attention mechanism can improve the RUL prediction performance by capturing the fault characteristics and the degradation development with the attention weights. The performance of the TDDN model is evaluated on the public C-MAPSS dataset and compared with the existing methods. The results show that the TDDN model can achieve the best RUL prediction accuracy in complex conditions compared to current machine learning models. The degradation-related features extracted from the high-dimension sensor streaming data demonstrate the clear degradation trajectories and degradation stages that enable TDDN to predict the turbofan-engine RUL accurately and efficiently.

Motivation & Objective

• To address the challenge of accurate RUL prediction in complex machinery under noisy, high-dimensional sensor data.
• To improve feature extraction from raw sensor streaming data by capturing degradation trends and fault characteristics.
• To develop a deep learning model that dynamically weights critical degradation stages for better prediction performance.
• To enable robust RUL prediction in complex operating conditions with minimal reliance on expert-defined health indicators.
• To support privacy-preserving deployment via integration with federated learning for distributed industrial applications.

Proposed method

• A 1D CNN is applied to raw sensor data to extract temporal features that reflect monotonic degradation trends.
• A moving window strategy is used to process sequential sensor data, enabling localized feature learning across degradation stages.
• An attention mechanism computes dynamic weights for abstract features, emphasizing those most relevant to degradation development.
• The attention weights are learned end-to-end, allowing the model to focus on critical fault patterns during prediction.
• The TDDN framework integrates 1D CNN and attention modules into a unified architecture for joint optimization.
• The model is trained end-to-end on the C-MAPSS dataset, with hyperparameters like window size tuned for optimal performance.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively extract degradation-related features from high-dimensional, noisy sensor streaming data?
• RQ2 How does attention mechanism improve RUL prediction by focusing on key fault characteristics in time-series degradation?
• RQ3 What is the impact of moving window size on capturing degradation evolution in complex machinery?
• RQ4 How does the TDDN model compare to existing deep learning methods in terms of RUL prediction accuracy on benchmark datasets?
• RQ5 Can the TDDN framework be extended to federated learning settings to preserve data privacy in distributed industrial environments?

Key findings

• The TDDN model achieves the best RUL prediction accuracy among existing deep learning methods on the C-MAPSS dataset.
• Attention weights evolve meaningfully across degradation stages, with higher values consistently assigned to features reflecting critical fault development.
• The 1D CNN effectively extracts monotonic degradation trends from fluctuating raw sensor data, improving feature representation.
• The attention mechanism successfully identifies and emphasizes key fault characteristics, especially during early and middle degradation stages.
• The model demonstrates robustness under varying operating conditions, as shown by consistent performance across different engine types in the C-MAPSS dataset.
• Sensitivity analysis confirms that moving window size significantly influences the model’s ability to capture degradation dynamics.