Condition-Aware Multiscale Hourglass-Transformer for Remaining Useful Life Prediction of Turbofan Engines

Technical Summary: Condition-Aware Multiscale Hourglass-Transformer for RUL Prediction

Problem Statement
Accurate Remaining Useful Life (RUL) prediction is critical for Prognostics and Health Management (PHM) in safety-critical systems like aircraft turbofan engines. While deep learning models have achieved strong performance on the NASA C-MAPSS dataset, they face significant challenges in multi-condition scenarios (specifically subsets FD002 and FD004). In these environments, sensor distributions vary across different operating regimes. Existing approaches often apply a single global normalization scheme across all conditions, which can conflate variations caused by operating settings with those caused by actual engine degradation. This confusion weakens the underlying degradation signal and reduces prediction reliability. Furthermore, the "black box" nature of deep learning models limits their adoption in safety-critical applications where understanding the decision-making process is essential.

Methodology
The paper proposes the Condition-Aware Multiscale Hourglass-Transformer (CA-MHT), a framework designed to address operating condition heterogeneity and improve model transparency. The methodology consists of four key stages:

1. Operating-Regime Identification: The framework utilizes K-Means clustering on the three operating setting vectors to group training samples into distinct operating regimes. For the multi-condition subsets FD002 and FD004, 7 and 8 clusters are identified, respectively, while single-condition subsets (FD001, FD003) use a single cluster.
2. Condition-Aware Sensor Normalization: Unlike global normalization, sensor measurements are normalized separately within each identified operating regime. Minimum and maximum values for normalization are calculated exclusively from the training data assigned to a specific cluster. This ensures that sensor values are scaled relative to their specific operating context, preserving the degradation signal. Operating settings themselves are normalized globally and retained as model inputs.
3. Multiscale Temporal Modeling: The normalized data is processed by a Multiscale Hourglass-Transformer (MHT) backbone. This architecture employs convolutional embedding, hourglass-based downsampling to create multi-resolution representations, and pyramid self-attention. This structure allows the model to capture both short-term fluctuations and long-term degradation trends.
4. Explainability Analysis: To enhance transparency, the trained models are analyzed using SHAP (SHapley Additive exPlanations) and Integrated Gradients. These post-hoc attribution methods identify influential input features, examine temporal behavior, and assess the stability of feature rankings across repeated runs.

Key Contributions
The paper outlines four primary contributions:

• Preprocessing Framework: A novel condition-aware preprocessing strategy that identifies operating regimes and applies regime-specific sensor normalization to prevent the mixing of condition-dependent and degradation-dependent variations.
• Model Integration: The successful integration of this preprocessing strategy with the Multiscale Hourglass-Transformer architecture for turbofan engine RUL prediction.
• Comprehensive Evaluation: Extensive evaluation on all four C-MAPSS subsets, with a specific focus on demonstrating improvements in the challenging multi-condition FD002 and FD004 subsets.
• Explainability Analysis: A dual-method analysis (SHAP and Integrated Gradients) that identifies influential sensors, investigates regime-specific behaviors, and compares the agreement between different explanation techniques.

Results
The proposed CA-MHT was evaluated using RMSE and the NASA scoring function (where lower values indicate better performance) on the C-MAPSS dataset.

• Performance Gains: The method demonstrated significant improvements on multi-condition subsets. On FD002, RMSE decreased from 13.70 (original MHT) to 13.41, and the NASA Score improved from 746.70 to 661.71. On FD004, the improvements were more pronounced, with RMSE dropping from 17.73 to 13.66 and the NASA Score falling from 1572.00 to 782.54.
• Single-Condition Subsets: On FD001 and FD003, where operating conditions are uniform, CA-MHT did not outperform the original MHT, confirming that the benefits of condition-aware normalization are specific to heterogeneous operating environments.
• Comparative Performance: Compared to representative literature methods (e.g., ATCN, DA-LSTM, MSIDSN), CA-MHT achieved the lowest NASA Score on FD002 and the lowest RMSE and NASA Score on FD004.
• Ablation Studies: Experiments confirmed that the performance gains are driven by the combination of condition-aware normalization and the multiscale Transformer architecture. Using condition-aware normalization with a TCN baseline also improved performance, but the full CA-MHT yielded the best results.
• Explainability Findings:
  • Feature Importance: Sensor s11 was identified as the most influential input (accounting for ~20% of global importance), followed by s4, s9, operating setting 2, and s15.
  • Regime Consistency: s11 ranked first across all eight operating regimes in FD004, suggesting it is a robust degradation indicator, while the ranking of other features varied by regime.
  • Temporal Reliance: The model relied heavily (98.07%) on the original temporal sensor sequence rather than the appended mean and slope statistics.
  • Stability and Agreement: SHAP results were highly stable across repeated runs (correlation > 0.99). SHAP and Integrated Gradients largely agreed on the set of top-10 influential features (9 out of 10), though their specific rankings differed.

Significance and Claims
The paper claims that condition-aware preprocessing is a vital step for improving predictive performance in multi-condition scenarios, as it prevents operating variations from obscuring degradation signals. Additionally, the study demonstrates that explainability methods (SHAP and Integrated Gradients) provide valuable insights into model behavior, confirming that the model relies on temporal sensor dynamics and that feature importance rankings are stable.

The author maintains a modest stance regarding the interpretation of these results. They explicitly state that attribution values describe the behavior of the trained model and should not be interpreted as direct evidence of physical causality. The work concludes that while the proposed framework improves accuracy and transparency, future work is needed to explore alternative regime identification methods, uncertainty-aware estimation, and the relationship between model explanations and known physical degradation mechanisms.