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.

Paper STGNN final architecture = MSTCN + GAT (static Spearman topology), without Transformer. DynaTopo new architecture = MSTCN + GAT (static Spearman topology + condition-driven dynamic graph), see below. See Model Architecture and Version Notes below.

Project Structure

RUL_Prediction/
├── data/                # Data storage (raw / processed)
├── core_models/         # Model building blocks
│   ├── stgnn_static.py       # Static-graph STGNN (original paper architecture)
│   ├── stgnn_dynatopo.py     # 🆕 Dual-graph STGNN (static + condition-driven dynamic)
│   ├── topo_generator/       # 🆕 Dynamic graph generator subpackage (similarity/attention)
│   └── topo_fusion/          # 🆕 Graph fusion strategy subpackage (feature/topology)
├── utils/               # Utilities (data processing, loss functions, evaluation metrics)
├── configs/             # Global config + dynatopo experiment config
├── scripts/             # Training & evaluation scripts
├── notebooks/           # Visualization & charting
├── extracted_pdf/       # Extracted content from the research paper PDF
├── saved_models/        # Trained model weights
└── logs/                # Run logs
    └── dynatopo/        # 🆕 DynaTopo experiment 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
static (paper)	MSTCN + GAT (Spearman fixed graph)	_static suffix	Final architecture in the paper, Transformer disabled
dynatopo (new)	MSTCN + GAT (static + condition-driven dynamic)	dynatopo_ prefix	🆕 Switchable A×B combinations
v1 (deprecated)	MSTCN + GAT + Transformer	—	Deleted

Static Scripts (paper reproduction)

• scripts/train_basic_static.py — Static-graph single-condition training
• scripts/ablation_static.py — Static-graph ablation study
• scripts/evaluate_1_static.py — Single-condition evaluation
• scripts/evaluate_2_static.py — Cross-condition evaluation
• scripts/train_transfer.py — Unified transfer script (supports none/global_mmd/lmmd_uda/lmmd_semi)

DynaTopo Scripts (new experiments)

• scripts/train_basic_dynatopo.py --preset A1B1 — Dual-graph training (config-driven)
• scripts/ablation_dynatopo.py — Dual-graph ablation study (4 A×B + ablation controls)
• scripts/evaluate_1_dynatopo.py — Dual-graph single-condition evaluation
• scripts/evaluate_2_dynatopo.py — Dual-graph cross-condition evaluation

Dynamic Graph Generation Strategies (A)

Strategy	ID	Description
A1 Similarity	similarity	Cosine similarity + condition modulation → Top-K sparsification
A2 Attention	attention	Multi-head attention + condition bias → Top-K sparsification

Graph Fusion Strategies (B)

Strategy	ID	Description
B1 Feature Fusion	feature	Static and dynamic graphs each through independent GATs, feature-level concatenation
B2 Topology Fusion	topology	Merge and deduplicate static & dynamic edges, unified single GAT

Experiment Presets

# 4 A×B combinations
python scripts/train_basic_dynatopo.py --preset A1B1  # Similarity × Feature Fusion
python scripts/train_basic_dynatopo.py --preset A1B2  # Similarity × Topology Fusion
python scripts/train_basic_dynatopo.py --preset A2B1  # Attention × Feature Fusion
python scripts/train_basic_dynatopo.py --preset A2B2  # Attention × Topology Fusion

# Ablation controls
python scripts/train_basic_dynatopo.py --preset static_only   # Static only (= original STGNN)
python scripts/train_basic_dynatopo.py --preset dynamic_only  # Dynamic only

# List all presets
python scripts/train_basic_dynatopo.py --list-presets