面向 LLM 驱动工程分析的确定性验证门控与卫星振动试验校准响应预测。 Code for the paper "Language Models Meet Vibration Testing for LEO Mega-Constellation Satellites": deterministic validation gates for LLM-driven engineering analysis and calibrated response prediction for sine vibration testing.
| Files | Last commit | Last update |
|---|---|---|
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 3 days ago |
llm-vibration-testing
Code for the paper Language Models Meet Vibration Testing for LEO Constellation Satellites with Calibrated Response Prediction from Few-show Sweeps.
The paper studies deterministic validation gates for LLM-driven engineering analysis. Two LLMs read Chinese vibration test records and produce a typed 76-field configuration; six deterministic gates check every field before it is consumed by the numerical stage that computes the notching decisions of a sine vibration acceptance test. The repository also contains the training code for the calibrated response-prediction models (WarpNet / FiLM-CNN with conformal calibration) evaluated in the paper.
Repository layout
| Path | Content |
|---|---|
load_lms.py |
Reader for the LMS / Siemens Simcenter Testlab measurement exports |
step1_linearity.py … step6_common.py |
Linearity check, dataset building, model definitions and training (WarpNet, FiLM-CNN, baselines), modal-parametric baseline, unified evaluation |
step7_llm/ |
LLM extraction pipeline: schema, prompts, extraction, scoring, evidence mode, perturbation inputs, second annotation, cross checks |
pipeline/ |
Validation gates and the LLM-driven notching pipeline: configuration ingestion, gate checks, driving evaluation, repair analysis, retraining check |
figs/, figs/v7/ |
Scripts for the figures of the paper |
cad/ |
Generic CAD model used for the illustration in Figure 1 (not the tested hardware) |
Requirements
Python 3.12 or later. Install the dependencies with
pip install -r requirements.txt
A CUDA-capable GPU is recommended for the training scripts but not required.
The CAD scripts under cad/ additionally need cadquery, OCP and pyvista; they are optional and only used to rebuild the illustrative model of Figure 1.
LLM API access
The extraction pipeline calls the DeepSeek and Qwen endpoints through the OpenAI-compatible API. Keys are read from the environment, never from the code:
DEEPSEEK_API_KEY # DeepSeek API key
DASHSCOPE_API_KEY # Alibaba Cloud Model Studio key for Qwen
Run python step7_llm/qwen_probe.py once to select a DashScope endpoint; it writes a local qwen_endpoint.json.
Data availability
The vibration test records and the measurement exports are not part of this repository. Access to the data requires the consent of the data owner; requests may be sent to the corresponding author. The scripts look for the records and the LMS exports under a data directory set by the environment variable VIB_DATA_ROOT (default ./data) and fail with a clear message when the files are absent.
License
MIT, see LICENSE.
Citation
If you use this code, please cite the paper:
Language Models Meet Vibration Testing for LEO Constellation Satellites
with Calibrated Response Prediction from Few-show Sweeps. Wu Jun, Deming Li,
Zhenming Zheng, Honghao Yue, Fei Yang, Hao Sun. arXiv preprint, 2026.
Introduction
面向 LLM 驱动工程分析的确定性验证门控与卫星振动试验校准响应预测。 Code for the paper "Language Models Meet Vibration Testing for LEO Mega-Constellation Satellites": deterministic validation gates for LLM-driven engineering analysis and calibrated response prediction for sine vibration testing.
Customize your domain