GFD:基于元学习的振动故障少样本诊断项目

用户可利用该项目进行振动故障的少样本诊断研究,支持同域和跨域诊断任务。项目构建任务特定子图,实现无限制训练采样与可配置测试分割,提供JSON配置及多种实验参数,便于基线与数据混合实验对比。【此简介由AI生成】

Branch1Tags0

GNNFD: Fault Diagnosis with Task Subgraph Meta-Learning

GNNFD is a refactored fork of SMILE tailored for vibration-fault few-shot learning on the CWRU dataset. The pipeline now builds task-specific subgraphs on-the-fly and focuses on two evaluation protocols:

  • Same-domain diagnosis (0HP → 0HP)
  • Cross-domain diagnosis (0HP → 1HP)

Key implementation changes

  • Task-level kNN subgraph construction with configurable k
  • Unlimited training sampling (non-overlapping 1024 windows)
  • Test split capped at 5 windows per class (default shot=1, qry=4)
  • JSON-based configs for repeatable baseline vs. mixup experiments

Environment

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Running Experiments

Use the provided configs under configs/ and the --config flag:

python train.py --config configs/cwru_same_domain_baseline.json
python train.py --config configs/cwru_same_domain_mix.json

python train.py --config configs/cwru_cross_domain_baseline.json
python train.py --config configs/cwru_cross_domain_mix.json

Each run appends a JSON summary to the configured output_json file (paths under results/). Command-line flags override config values if provided.

Important Arguments

  • --cwru_train_loads, --cwru_valid_loads, --cwru_test_loads — comma-separated load ids
  • --shot, --qry — per-class support/query counts (ensure shot + qry ≤ test_samples_per_class)
  • --disable_augmentation — baseline (true) vs data-mixing run (false)
  • --task_graph_k — neighbours for task subgraph construction

See train.py for additional options (meta-task count, optimiser settings, etc.).

Introduction

用户可利用该项目进行振动故障的少样本诊断研究,支持同域和跨域诊断任务。项目构建任务特定子图,实现无限制训练采样与可配置测试分割,提供JSON配置及多种实验参数,便于基线与数据混合实验对比。【此简介由AI生成】

Customize your domain