mudestreda / README.md
Kriscamp's picture
Upload README.md with huggingface_hub
8e74e18 verified
|
Raw
History Blame Contribute Delete
6.74 kB
---
dataset_info:
features:
- name: query
dtype: string
- name: image
dtype: image
- name: images
list: image
- name: annot
dtype: string
- name: reasoning
dtype: string
- name: cate
dtype: string
- name: task
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 307973991
num_examples: 404
- name: val
num_bytes: 29696351
num_examples: 52
- name: test
num_bytes: 35860720
num_examples: 56
download_size: 373200637
dataset_size: 373531062
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: val
path: data/val-*
- split: test
path: data/test-*
tags:
- smart-manufacturing
- cnc
- tool-wear
- industrial
pretty_name: Mudestreda
extra_gated_prompt: Released for research use; access granted manually. State your
name, affiliation, and intended use.
license: gpl-3.0
extra_gated_fields:
Name: text
Affiliation: text
Intended use: text
---
# Mudestreda (CNC 铣削刀具磨损 · 多模态)
Mudestreda (Multimodal Device State Recognition) 转成团队统一 7 字段 SFT schema。
一个 sample =「某把刀 T + 某刀次 R + 某刀刃 B」的一次观测(id 形如 T1R2B3):
1 张刀刃显微图 + 该刀次的 3 张力谱图(X/Y/Z 轴 STFT 热图), 目标是三选一的磨损状态
**sharp / used / dulled**(显微镜实测磨损量的 µm 硬阈值分档; 本集只发离散三类, 不发连续值)。
> **task 码依据(照团队官方文档, 非臆断)**: `sft_task_merge_plan_v3` 定义
> **`T-C2` = Remaining-Useful-Life & Health Estimation (cate C)**, 原文
> *"From a degradation signal history, predict remaining useful life and assess component/engine
> health state… Regression / assessment over **run-to-failure trajectories** — a different target
> from the categorical diagnosis of T-C1"*, 输出为 `RUL / health state`。
> 本集正是刀具**跑到报废(run-to-failure)**的退化轨迹, 磨损档位即 **health state**,
> 与文档中同归 T-C2 的 NASA-IMS、XJTU-SY(轴承跑到报废)同构。
> (但本集输出形态是一次性三分类, 也像 T-C1 的类别式诊断; T-C1/T-C2 归类待学长拍板。)
## 标签: annot 用的是 tool_label(先看这段再用)
原始 CSV 三列 `id,image_label,tool_label`, 后两列的名字都很误导:
- **`tool_label` 不是"第几把刀", 是磨损类别**, 取值 {1,2,3} 且 1-indexed:
**1=sharp / 2=used / 3=dulled**。官方训练代码用的就是它
(`labels_train = df.tool_label.values - 1`), 粒度是**整把刀/逐刀次** ——
128/128 个 ride 内 4 个刀刃的 tool_label 完全一致。
- `image_label` 才是逐刀刃的类别字符串, 粒度更细。两列在本次构建的
**71/512** 个样本上不一致(核查员实测 71/512 ≈ 13.9%,
绝大多数是相邻类错位) —— 不是笔误, 是两个粒度的标注。
本库 **`annot` = tool_label 映射出的类名**(与论文/官方代码的监督信号对齐), 所以一条样本预测的是
**这一刀次整把刀的状态**, 不是这一个刀刃自己的状态。`image_label` 原样保留在 `metadata`,
要做逐刃诊断的下游自行改用(代价: 与论文数字不可比, 且谱图重复组里会出现标签自相矛盾的样本, 见下)。
## split: 默认 tool_distribution; random_distribution 有坑, 要显式选
本 config 用官方 `tool_distribution`(**train=T1–T8 / val=T9 / test=T10**):
train 404 / val 52 / test 56, 共 512 个样本;
ride 跨 split 数 = 0(运行时校验, 应为 0)。
官方另一套 `random_distribution` 默认****推, 因为(核查员逐位核过):
- 它的 CSV 共 512 行但只有 508 个唯一 id: T6R15B1–B4 各出现两次(都在 train, 不去重会被训练两遍);
同时漏掉了 tool split 里存在的 T10R13B3/B4、T10R14B3/B4 —— 两套划分覆盖的样本集合并不相同,
所以"512 samples"这句话只在 tool_distribution 下成立。
- **强相关跨集**: 63/128 个 ride 横跨不同 split, test 48/48 的样本在 train 里都有
同一把刀同一刀次的相邻刀刃"兄弟"。拿它做泛化评测数字会虚高(上游 README 那组接近满分的
结果就是在这个 split 上报的; 同组后继 Nonastreda 用 leave-one-tool-out 自报只有约 75%)。
- 确实要对比: 用 `--distribution random` 重跑本转换器, 会自动去重并推成独立的
`random_distribution` config, `metadata.distribution` 如实标注。
## 谱图冗余(做统计前必读)
128/128 个 ride 的 B4 谱图与 B3 **字节完全相同**(x/y/z 分组一致, 核查员 CRC32 核过;
为什么相同, 上游文档和代码都没交代)。本次运行时校验: B3+B4 齐全的 128 个 ride 中
**128** 个 B4 与 B3 三张谱图内容相同; 512 个谱图三元组只有
**378** 种不同内容(刀图倒是全唯一)。
后果: **谱图侧真实独立观测是 378 个, 不是 512**, 按独立样本做统计会高估有效数据量。
重复组内标签冲突(运行时实算): image_label 冲突 **24** 组 / tool_label 冲突 **0**
(核查员实测 24/132 组 vs 0/132 组 —— 这也是 annot 选 tool_label 的理由之一)。
## 7-field schema
`query`(固定 prompt) · `image`(刀刃显微图) · `images`(全部 4 模态: tool + spec_x/y/z) ·
`annot`(sharp/used/dulled, 来自 **tool_label**) · `reasoning`(null) · `cate`/`task` ·
`metadata`(id, tool_id, ride_id, **tool_label**, **image_label**, **label_source**, **label_note**,
distribution, split, 模态相对路径)。
## Load
```python
from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/mudestreda") # 默认 config = tool_distribution
# ds = load_dataset("AI4Manufacturing/mudestreda", "random_distribution") # 仅当另行推送过该 config
```
_Gated — request access; granted manually._
## Provenance & license
- 数据源: Zenodo 8238653(original, 非增广版), https://zenodo.org/records/8238653。
- **许可证有冲突**: Zenodo 数据记录标 **GPL-3.0-or-later**, 但 GitHub 代码仓(hubtru/Minape)的
LICENSE 是 **MIT**(版权行写的还是另一篇论文的标题, 疑似复制粘贴残留), 作者未澄清。
本库按更严的一侧标 GPL-3.0。我们是研究用途、非商用, 不构成阻碍; **商用前必须先厘清**。
- 论文: *Multimodal Isotropic Neural Architecture with Patch Embedding* (ICONIP 2023);
代码仓 github.com/hubtru/Minape。研究用途转载, 请引用上游。
- 10 把刀(T1–T10) × 128 个唯一 (Tool,Ride) × 每 ride 4 个刀刃, 3 类;
tool_distribution 下 512 个 id 全唯一。