GeorgeWJJ/steam-generator-multimodal-prognostics

Steam Generator Multimodal Prognostics

中文说明 · Project code · Original benchmark

An attributed mirror of the synthetic steam-generator benchmark introduced by Zhe Yang, Piero Baraldi, and Enrico Zio. Images, maintenance records, and four numerical variables are aligned to the target degradation at the next scheduled inspection. These are simulated observations, not measurements from an operating nuclear facility.

The uploader maintains the packaging, integrity manifests, and source-aware evaluation metadata. The uploader is not the original dataset author. The original source is the LASAR public benchmark page and its Google Drive archive.

Contents

Asset	Purpose
steam-generator-multimodal-prognostics.zip	All 7,977 original local files: 50 equipment folders, TIFF images, text/numerical sequences, labels, and the original Readme.docx. Extraction preserves the folder Multimodal data for prognostics/.
data/observations.parquet	7,626 aligned observations with embedded original TIFF bytes, numerical fields, text, targets, and audit identifiers.
raw_checksums.json	Per-file SHA-256, relative path, and byte size for verifying the unmodified raw files.
paper_splits.json / clean_sources_splits.json	Two fixed five-fold evaluation schedules used by the accompanying project.
*_source_overlap.json	Actual source intersections between training, validation, and test partitions.
*_data_audit.json	Image, target, alignment, and duplicate-history audit.

The Parquet train split is a storage split containing the entire dataset. It is not a prescribed experiment training partition. Use the supplied split schedules to construct train/validation/test subsets.

Schema

Field	Meaning
row_id	Zero-based aligned observation identifier.
trajectory	Original equipment-folder identifier, 1–50.
source_group	Input-history group identifier, formed from time/text/numerical histories without using targets.
sample	One-based position in the original trajectory.
time	Collection time, in arbitrary time units (atu).
image	Original 194×325 binary TIFF; tensor convention is H=325, W=194.
text	Inspection/maintenance record; the previous record is repeated when no new record exists.
chemical_cleanings	Number of preceding chemical cleanings.
mechanical_cleanings_since_chemical	Mechanical cleanings since the previous chemical cleaning.
time_since_repair	Time since the previous repair, in atu.
time_to_next_inspection	Time until the pre-planned next inspection, in atu.
next_inspection_degradation	Regression target, already aligned to the next inspection; do not shift it again.
blank_image	Whether the image matches the released no-signal template, which still contains axes.

Only image, text, and the four numerical fields are inputs in the project. Identifiers, absolute collection time, source labels, and target values must not become extra predictor inputs. A planned inspection interval is not remaining useful life (RUL).

Dataset statistics and source overlap

• 7,626 aligned rows, 50 original folders, 30 distinct time/text/numerical input histories.
• 6,092 signal images and 1,534 no-signal template images (20.12%).
• 794 distinct complete text strings; at most 27 tokens under the project's documented tokenizer.
• Target range: 12.75–104.83524.
• Folders 1–20 pair with 31–50. All 20 pairs share time, text, numerical, and target sequences. Images match in 17 pairs; 8/38, 9/39, and 16/46 each differ in one image.

Directory-level splitting can therefore place copies of the same input history in different partitions. clean_sources keeps those groups together across training, validation, and test. All raw duplicates are retained; duplicated histories still receive extra weight in sample-wise training and evaluation. This dataset contains 30 distinct input-history groups, not 50 independent histories.

Load

from datasets import load_dataset

data = load_dataset("GeorgeWJJ/steam-generator-multimodal-prognostics", split="train")
print(data[0]["text"])
print(data[0]["next_inspection_degradation"])

For the accompanying MB-DNN implementation, download the ZIP using scripts/download_dataset.py in the project. It validates the archive hash and every extracted raw file before training preparation.

Provenance and usage terms

The raw observations and original documentation are attributed to the benchmark authors. No explicit dataset redistribution license was found in the downloaded Readme.docx or the linked benchmark page during packaging on 2026-10-14. upstream-unspecified records that uncertainty; it is not a new license grant, a claim that the data are public domain, or an MIT license for the dataset. Consult the original authors/source for permissions required by your intended use. See LICENSING.md.

The derived Parquet representation retains original image bytes and row values. It adds schema names and audit fields; it does not create new measurements, remove duplicate histories, shift labels, or rescale targets. The supplied splits are project evaluation metadata, not author-provided original fold assignments.

Citation

@article{yang2021multibranch,
  author = {Yang, Zhe and Baraldi, Piero and Zio, Enrico},
  title = {A multi-branch deep neural network model for failure prognostics based on multimodal data},
  journal = {Journal of Manufacturing Systems},
  volume = {59},
  pages = {42--50},
  year = {2021},
  doi = {10.1016/j.jmsy.2021.01.007}
}

中文说明

这是 Yang、Baraldi、Zio 发布的蒸汽发生器合成多模态数据集的署名镜像，包含图像、检修文本、四个数值变量及下一次计划检查时的退化程度。上传者负责整理、完整性校验和评估分组，不是原始数据作者。

原始 ZIP 保留全部 7,977 个文件；Parquet 提供 7,626 条对齐样本和原始 TIFF 字节，方便浏览与加载。train 是存储分区，包含全量数据，实际实验划分必须使用分组索引。目标已经表示下一次检查退化，不再平移标签，也不转换为 RUL。

50 个目录对应 30 种不同输入历史。1–20 与 31–50 成对重复；17 对图像也完全相同，另外3对各有一张不同图像。按目录划分存在同源数据跨集合问题；来源隔离配置将复制历史作为同组，仍保留所有发布样本。设备编号、来源编号、绝对时间和目标不应作为额外模型输入。

原下载说明和发布页没有明确提供数据集再分发许可证。本页保留来源与引用，不替原作者授予新许可；具体用途所需权限应向原始来源确认。