中文版本

Aero-Engine Remaining Useful Life (RUL) Prediction

This repository contains the complete code implementation for the research project "Research on a Graph Neural Network-Based Method for Remaining Useful Life Prediction of Aero-Engines", using the NASA C-MAPSS dataset. It is also publicly available on GitHub for study and reference.

Final STGNN architecture in the paper = MSTCN + GAT (v2), without Transformer. See Model Architecture and Version Notes below.

Project Structure

RUL_Prediction/
├── data/                # Data storage (raw / processed)
├── core_models/         # Model building blocks (MSTCN, GAT, Transformer, STGNN assembly)
├── utils/               # Utilities (data processing, loss functions, evaluation metrics)
├── configs/             # Global configuration
├── scripts/             # Training & evaluation scripts
├── notebooks/           # Visualization & charting
├── extracted_pdf/       # Extracted content from the research paper PDF
├── saved_models/        # Trained model weights
└── logs/                # Run logs

Environment Setup

# Create virtual environment
conda create -n rul_env python=3.10 -y
conda activate rul_env

# Install dependencies
pip install -r requirements.txt

Dataset

NASA C-MAPSS dataset (FD001 ~ FD004), located under data/raw/.

Model Architecture

The final STGNN (Spatio-Temporal Graph Neural Network) adopted in the paper consists of the following modules:

• MSTCN (Multi-Scale Temporal Convolutional Network): Three stacked Conv1d layers (kernel 3→5→7, channels 32→64→128) with parameter sharing, extracting multi-scale temporal degradation features across 14 sensors
• GAT (Graph Attention Network): Two-layer multi-head attention (4-head → 1-head), modeling spatial coupling relationships over a sensor topology graph built via Spearman rank correlation
• Global Temporal Context: Time-dimension mean pooling + two Linear layers, providing supplementary global degradation trend information
• Operating Condition Encoding: Conv1d encoding of 3 operating parameters
• Feature Fusion: Concatenation of operating condition encoding, GAT spatial features, and global temporal context → FC → RUL
• Loss Function: MSE + NASA asymmetric scoring joint loss
• Transfer Learning: LMMD (Local Maximum Mean Discrepancy) cross-condition adaptation based on RUL degradation stage subdomain partitioning

Note: The Transformer module (core_models/transformer.py) and its ablation switch use_transformer are retained in the codebase for architecture exploration and comparative experiments. Ablation studies showed that the Transformer branch contributes limitedly to final performance, so the paper's final STGNN model does not use Transformer.

Version Notes

Version	Architecture	Identifier	Description
v2 (main)	MSTCN + GAT	_v2 suffix	Final architecture used in the paper, Transformer branch disabled
v1 (exploratory)	MSTCN + GAT + Transformer	no suffix	Full three-component architecture, kept for comparison

• scripts/train_basic_v2.py, scripts/evaluate_2_v2.py — Single-condition v2 training & evaluation
• scripts/ablation_study_v2.py — v2 ablation study (MSTCN+GAT / MSTCN-only / GAT-only / all-off)
• scripts/train_transfer_v2_*.py — v2 cross-condition transfer experiments (unsupervised UDA / semi-supervised / global MMD comparison)
• Scripts without _v2 suffix are v1 versions that include the Transformer branch