🚀 GEBiLSTM-Attn: Novel Architecture for RUL Prediction

📌 Overview

This repository presents GEBiLSTM-Attn, a novel deep learning architecture designed for Remaining Useful Life (RUL) prediction using the NASA CMAPSS dataset.

The model combines the strengths of:

• Bidirectional LSTM (BiLSTM) → Captures forward & backward temporal dependencies
• Attention Mechanism → Focuses on the most important time steps
• Gated Enhancements (Proposed) → Improves feature selection and information flow

👉 Traditional sequence models often fail to prioritize critical temporal features.
👉 Our architecture addresses this using attention-driven learning and enhanced gating mechanisms.

BiLSTM processes sequences in both directions, improving context understanding, while attention assigns higher importance to relevant inputs, boosting performance.

---

🧠 Proposed Architecture: GEBiLSTM-Attn

🔹 Key Idea

We introduce a Gated Enhanced BiLSTM with Attention (GEBiLSTM-Attn) that improves:

• Feature extraction
• Temporal dependency modeling
• Interpretability

---

🔹 Architecture Flow

Input Time-Series Data
↓
3 Parallel Branches (LSTM, GRU, CNN)
↓
Data Preprocessing & Normalization
↓
Gated Feature Enhancement Layer (Proposed)
↓
BiLSTM Layer (Bidirectional Temporal Learning)
↓
Attention Layer (Feature Importance Weighting)
↓
Fully Connected Layers
↓
RUL Prediction (Regression Output)

---

⚙️ Why This Architecture?

🚫 Problems with Existing Models

• Standard LSTM → Struggles with long-term dependencies
• BiLSTM → Treats all time steps equally
• Lack of focus on critical degradation signals

✅ Our Improvements

• Gated Enhancement Layer → Filters noisy sensor data
• Attention Mechanism → Prioritizes important degradation patterns
• BiLSTM → Captures bidirectional dependencies

👉 Hybrid architectures combining BiLSTM and attention significantly improve sequence modeling and prediction accuracy.

---

📊 Dataset

• Dataset: NASA CMAPSS (Commercial Modular Aero-Propulsion System Simulation)
• Type: Multivariate time-series
• Task: Predict Remaining Useful Life (RUL) of engines

---

🧪 Methodology

🔹 Steps

1. Data preprocessing & normalization
2. Sequence generation
3. Model training
4. Evaluation using regression metrics

🔹 Model Training

• Loss Functions: MAE, RMSE
• Optimizer: Adam
• Framework: TensorFlow

---

📈 Results

✔ Improved prediction accuracy compared to baseline models
✔ Better handling of long-term dependencies
✔ Enhanced interpretability via attention

🔹 Performance Metrics

Dataset	MAE (cycles)	RMSE (cycles)	R² Score
FD001	10.73	14.63	0.8726
FD002	13.82	21.03	0.8346
FD003	9.80	13.51	0.8893
FD004	14.65	21.02	0.8355

---

🔬 Novel Contributions

⭐ Proposed GEBiLSTM-Attn architecture
⭐ Introduced Gated Feature Enhancement Layer
⭐ Improved RUL prediction performance on CMAPSS dataset
⭐ Built end-to-end pipeline from preprocessing to evaluation

---

👩‍💻 Author Contribution

• Designed and proposed GEBiLSTM-Attn architecture
• Built complete deep learning pipeline from scratch
• Implemented preprocessing, training, and evaluation
• Conducted experiments and performance analysis

---

🔗 Future Work

• Extend architecture with Transformer-based hybrid models
• Build real-time predictive maintenance systems
• Deploy as scalable API for industrial applications