Forgis/Mechanical-Components

---
license: mit
configs:
- config_name: source_metadata
data_files:
- split: train
path: source_metadata/*.parquet
- config_name: bearings
data_files:
- split: train
path: bearings/*.parquet
- config_name: gearboxes
data_files:
- split: train
path: gearboxes/*.parquet
- config_name: v2_train
data_files:
- split: train
path: v2_train/*.parquet
---

# Mechanical Components Vibration Dataset

Comprehensive multi-source mechanical vibration dataset for training cross-component fault diagnosis and prognostics models. Designed for the Mechanical-JEPA project.

**Total: ~12,000+ samples | 9.5 GB | 16 sources | 5 component types**

## Quick Start

```python
from datasets import load_dataset

bearings = load_dataset("Forgis/Mechanical-Components", "bearings", split="train")
gearboxes = load_dataset("Forgis/Mechanical-Components", "gearboxes", split="train")
sources = load_dataset("Forgis/Mechanical-Components", "source_metadata", split="train")
```

## Two-Level Schema

**source_metadata** (16 entries): One row per source dataset with constant properties.
**bearings/gearboxes** configs: Per-sample data linked via `source_id` foreign key.

## Dataset Sources

### Bearings (~10,000 samples from 10 sources)

| Source | Samples | Component | Sensors | Unique Value |
|--------|---------|-----------|---------|--------------|
| [CWRU](https://engineering.case.edu/bearingdatacenter) | 40 | Ball bearing | Vibration | Standard benchmark |
| [MFPT](https://www.mfpt.org/fault-data-sets/) | 20 | Ball bearing | Vibration | Variable load |
| [FEMTO](https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository) | 3,569 | Ball bearing | Vibration, temperature | Run-to-failure (RUL) |
| [Mendeley](https://data.mendeley.com/datasets/vxkj334rzv/7) | 280 | Ball bearing | Vibration | **Speed transitions** (action-conditioning) |
| [XJTU-SY](https://github.com/WangBiaoXJTU/xjtu-sy-bearing-datasets) | 1,370 | Ball bearing | Vibration | Run-to-failure (RUL) |
| [IMS/NASA](https://data.nasa.gov/dataset/ims-bearings) | 1,256 | Ball bearing | Vibration | Run-to-failure (RUL) |
| [Paderborn](https://mb.uni-paderborn.de/kat/forschung/bearing-datacenter/) | 384 | Ball bearing | Vibration, **current** | Real + artificial faults |
| [MAFAULDA](https://www02.smt.ufrj.br/~offshore/mfs/page_01.html) | 800 | **Shaft+bearing** | Vibration, **acoustic**, tachometer | **Imbalance, misalignment** (shaft faults!) |
| [Ottawa](https://data.mendeley.com/datasets/y2px5tg92h/1) | 180 | Ball bearing | Vibration, **acoustic** | **Cage faults**, 3 health stages |
| [SCA Pulp Mill](https://data.mendeley.com/datasets/tdn96mkkpt/2) | 2,663 | Industrial bearing | Vibration | **Real industrial data** |
| [VBL-VA001](https://zenodo.org/records/7006575) | 800 | Shaft+bearing | Vibration (triaxial) | Misalignment, unbalance |
| [SEU](https://github.com/cathysiyu/Mechanical-datasets) | 140 | Drivetrain bearing | 8-ch (motor+gearbox) | Cross-component rig |

### Gearboxes (~1,225 samples from 4 sources)

| Source | Samples | Component | Sensors | Unique Value |
|--------|---------|-----------|---------|--------------|
| [OEDI](https://data.openei.org/submissions/623) | 20 | Spur gear | Vibration (4-ch) | Healthy vs gear crack |
| [PHM 2009](https://phmsociety.org/public-data-sets/) | 109 | Spur gear | Vibration, tachometer | Challenge data |
| [MCC5-THU](https://github.com/liuzy0708/MCC5-THU-Gearbox-Benchmark-Datasets) | 956 | Spur gear | Vibration | **Speed/load transitions** |
| [SEU](https://github.com/cathysiyu/Mechanical-datasets) | 140 | Planetary+parallel | 8-ch (motor+gearbox) | Cross-component rig |

## Component Types Covered

| Component | Sources | Fault Types |
|-----------|---------|-------------|
| **Bearings** | CWRU, MFPT, FEMTO, Mendeley, XJTU-SY, IMS, Paderborn, Ottawa, SCA | inner_race, outer_race, ball, cage, compound, degrading |
| **Gears** | OEDI, PHM2009, MCC5-THU, SEU | gear_crack, gear_wear, missing_tooth, tooth_break |
| **Shafts** | MAFAULDA, VBL-VA001 | imbalance, misalignment_horizontal, misalignment_vertical |
| **Drivetrains** | SEU | Combined motor+gearbox+bearing from single rig |
| **Industrial** | SCA Pulp Mill | Naturally occurring faults in real machinery |

## Sensor Modalities

| Modality | Sources | Channels |
|----------|---------|----------|
| **Vibration** (accelerometer) | All 16 | 1-8 channels per sample |
| **Motor current** | Paderborn, Mendeley (partial) | 2-3 phase current |
| **Acoustic** (microphone) | MAFAULDA, Ottawa | 1 channel |
| **Tachometer** | MAFAULDA, PHM2009, OEDI | 1 channel |
| **Temperature** | FEMTO | Scalar in slow_signals |
| **Torque** | SEU, MCC5-THU | 1 channel |

## Key Features

- **412 transition samples** for action-conditioning (Mendeley speed ramps + MCC5-THU speed/load transitions)
- **Episode/RUL fields** for prognostics (FEMTO, XJTU-SY, IMS, SCA)
- **Real industrial data** from SCA pulp mill (not lab)
- **Shaft faults** (imbalance, misalignment) from MAFAULDA and VBL-VA001
- **Acoustic data** from MAFAULDA and Ottawa (microphone alongside accelerometer)
- **Cross-component** drivetrain data from SEU (motor+gearbox+bearing on single rig)

## Per-Sample Schema

```python
{
"source_id": "cwru", # FK to source_metadata
"sample_id": "cwru_105",
"signal": [[0.1, 0.2, ...]], # (n_channels, signal_length)
"n_channels": 2,
"channel_names": ["DE_accel", "FE_accel"],
"channel_modalities": ["vibration", "vibration"],
"health_state": "faulty", # healthy | faulty | degrading
"fault_type": "inner_race",
"fault_severity": None,
"rpm": 1750,
"load": 2.0,
"load_unit": "hp",
"episode_id": None, # For run-to-failure
"episode_position": None, # 0.0 to 1.0
"rul_percent": None, # Remaining useful life
"is_transition": False, # Speed/load change
"transition_type": None, # ramp_speed | ramp_load
}
```

## v2 Training-Ready Config (Planned)

A standardized config for direct model training:
- Fixed sampling rate: 12,800 Hz
- Fixed window: 16,384 samples (1.28 seconds)
- Vibration-only (single modality)
- Per-sample instance normalization
- Source-disjoint train/val/test splits

## v2 Training-Ready Config (LIVE)

Standardized for direct model training. All sources resampled to common format.

```python
# Load training-ready data (all splits in one, filter by 'split' column)
v2 = load_dataset("Forgis/Mechanical-Components", "v2_train", split="train")
v2_train = v2.filter(lambda x: x["split"] == "train") # 20,143 samples
v2_val = v2.filter(lambda x: x["split"] == "val") # 1,332 samples
v2_test = v2.filter(lambda x: x["split"] == "test") # 6,363 samples
```

| Parameter | Value |
|-----------|-------|
| Sampling rate | 12,800 Hz |
| Window length | 16,384 samples (1.28 seconds) |
| Channels | 1 (primary vibration) |
| Normalization | Per-sample instance norm |
| Splits | Source-disjoint (train/val/test) |

**Train** (12 sources): CWRU, MFPT, FEMTO, XJTU-SY, IMS, OEDI, PHM2009, MCC5-THU, SEU, MAFAULDA, VBL, SCA-train
**Val** (2 sources): Paderborn, Ottawa
**Test** (2 sources): Mendeley, SCA-test

## Citations

Please cite the original datasets. See source_metadata config for full citations per source.

许可证：MIT协议
配置项：
- 配置名称：source_metadata
数据文件：
- 拆分方式：训练集
路径：source_metadata/*.parquet
- 配置名称：bearings
数据文件：
- 拆分方式：训练集
路径：bearings/*.parquet
- 配置名称：gearboxes
数据文件：
- 拆分方式：训练集
路径：gearboxes/*.parquet
- 配置名称：v2_train
数据文件：
- 拆分方式：训练集
路径：v2_train/*.parquet

# 机械组件振动数据集
## 数据集简介
本数据集为多源综合机械振动数据集，用于训练跨组件故障诊断与预后模型，专为Mechanical-JEPA项目打造。
**总规模：约12000+样本 | 9.5 GB | 16个数据源 | 5类组件**

## 快速入门
python
from datasets import load_dataset

bearings = load_dataset("Forgis/Mechanical-Components", "bearings", split="train")
gearboxes = load_dataset("Forgis/Mechanical-Components", "gearboxes", split="train")
sources = load_dataset("Forgis/Mechanical-Components", "source_metadata", split="train")

## 二级架构
**source_metadata**（共16条条目）：每个数据源对应一行，包含固定属性。
**bearings/gearboxes** 配置：通过`source_id`外键关联的单样本数据。

## 数据集数据源
### 轴承数据集（来自10个数据源，约10000样本）
| 数据源 | 样本量 | 组件类型 | 传感器类型 | 特色参数 |
|--------|---------|-----------|---------|--------------|
| [CWRU](https://engineering.case.edu/bearingdatacenter) | 40 | 滚动轴承（Ball bearing） | 振动 | 标准基准数据集 |
| [MFPT](https://www.mfpt.org/fault-data-sets/) | 20 | 滚动轴承 | 振动 | 变负载工况 |
| [FEMTO](https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository) | 3,569 | 滚动轴承 | 振动、温度 | 全寿命退化数据（剩余使用寿命（Remaining Useful Life, RUL）） |
| [Mendeley](https://data.mendeley.com/datasets/vxkj334rzv/7) | 280 | 滚动轴承 | 振动 | **转速过渡工况**（动作调节型） |
| [XJTU-SY](https://github.com/WangBiaoXJTU/xjtu-sy-bearing-datasets) | 1,370 | 滚动轴承 | 振动 | 全寿命退化数据（RUL） |
| [IMS/NASA](https://data.nasa.gov/dataset/ims-bearings) | 1,256 | 滚动轴承 | 振动 | 全寿命退化数据（RUL） |
| [Paderborn](https://mb.uni-paderborn.de/kat/forschung/bearing-datacenter/) | 384 | 滚动轴承 | 振动、电机电流 | 真实+人工故障样本 |
| [MAFAULDA](https://www02.smt.ufrj.br/~offshore/mfs/page_01.html) | 800 | 轴+轴承 | 振动、声学（麦克风（microphone））、转速计（tachometer） | 不平衡、不对中（轴类故障） |
| [Ottawa](https://data.mendeley.com/datasets/y2px5tg92h/1) | 180 | 滚动轴承 | 振动、声学 | 保持架故障、3种健康状态 |
| [SCA Pulp Mill](https://data.mendeley.com/datasets/tdn96mkkpt/2) | 2,663 | 工业轴承 | 振动 | 真实工业现场数据 |
| [VBL-VA001](https://zenodo.org/records/7006575) | 800 | 轴+轴承 | 三轴振动 | 不对中、不平衡 |
| [SEU](https://github.com/cathysiyu/Mechanical-datasets) | 140 | 传动系统轴承 | 8通道（电机+齿轮箱） | 跨组件试验台 |

### 齿轮箱数据集（来自4个数据源，约1225样本）
| 数据源 | 样本量 | 组件类型 | 传感器类型 | 特色参数 |
|--------|---------|-----------|---------|--------------|
| [OEDI](https://data.openei.org/submissions/623) | 20 | 直齿轮 | 4通道振动 | 健康vs齿轮裂纹 |
| [PHM 2009](https://phmsociety.org/public-data-sets/) | 109 | 直齿轮 | 振动、转速计 | 竞赛数据集 |
| [MCC5-THU](https://github.com/liuzy0708/MCC5-THU-Gearbox-Benchmark-Datasets) | 956 | 直齿轮 | 振动 | 转速/负载过渡工况 |
| [SEU](https://github.com/cathysiyu/Mechanical-datasets) | 140 | 行星+平行轴齿轮箱 | 8通道（电机+齿轮箱） | 跨组件试验台 |

## 覆盖的组件类型
| 组件类型 | 数据源 | 故障类型 |
|-----------|---------|-------------|
| **轴承** | CWRU、MFPT、FEMTO、Mendeley、XJTU-SY、IMS、Paderborn、Ottawa、SCA | 内圈故障（inner_race）、外圈故障（outer_race）、滚珠故障（ball）、保持架故障（cage）、复合故障（compound）、退化故障（degrading） |
| **齿轮** | OEDI、PHM2009、MCC5-THU、SEU | 齿轮裂纹（gear_crack）、齿轮磨损（gear_wear）、断齿（missing_tooth）、齿断裂（tooth_break） |
| **轴类** | MAFAULDA、VBL-VA001 | 不平衡（imbalance）、水平不对中（misalignment_horizontal）、垂直不对中（misalignment_vertical） |
| **传动系统** | SEU | 单试验台集成电机+齿轮箱+轴承的复合故障 |
| **工业组件** | SCA纸浆厂 | 真实工业设备的自然故障 |

## 传感器模态
| 模态类型 | 数据源 | 通道数 |
|----------|---------|----------|
| **振动（加速度计（accelerometer））** | 全部16个数据源 | 每样本1-8通道 |
| **电机电流** | Paderborn、Mendeley（部分数据源） | 2-3相电流 |
| **声学（麦克风）** | MAFAULDA、Ottawa | 1通道 |
| **转速计** | MAFAULDA、PHM2009、OEDI | 1通道 |
| **温度** | FEMTO | 慢信号中的标量数据 |
| **扭矩** | SEU、MCC5-THU | 1通道 |

## 核心特性
- **412个工况过渡样本**：涵盖Mendeley的转速斜坡数据 + MCC5-THU的转速/负载过渡样本
- **预后任务专用字段**：包含Episode/RUL字段，覆盖FEMTO、XJTU-SY、IMS、SCA数据源
- **真实工业现场数据**：来自SCA纸浆厂，非实验室采集的工业设备数据
- **轴类故障数据集**：来自MAFAULDA与VBL-VA001，包含不平衡、不对中故障样本
- **同步声学数据**：来自MAFAULDA与Ottawa，与加速度计信号同步采集
- **跨组件传动系统数据**：来自SEU试验台，集成电机、齿轮箱与轴承组件

## 单样本数据架构
python
{
"source_id": "cwru", # 指向source_metadata的外键
"sample_id": "cwru_105",
"signal": [[0.1, 0.2, ...]], # 形状为(n_channels, signal_length)的信号数组
"n_channels": 2,
"channel_names": ["DE_accel", "FE_accel"],
"channel_modalities": ["vibration", "vibration"],
"health_state": "faulty", # 可选值：healthy | faulty | degrading
"fault_type": "inner_race",
"fault_severity": None,
"rpm": 1750,
"load": 2.0,
"load_unit": "hp",
"episode_id": None, # 全寿命退化任务专用ID
"episode_position": None, # 退化进度，范围0.0到1.0
"rul_percent": None, # 剩余使用寿命百分比
"is_transition": False, # 是否为工况过渡样本
"transition_type": None, # 过渡类型：ramp_speed | ramp_load
}

## v2 训练就绪配置（规划版）
专为直接模型训练设计的标准化配置：
- 固定采样率：12800 Hz
- 固定窗口长度：16384样本（对应1.28秒）
- 仅振动单模态数据
- 逐样本实例归一化
- 按数据源划分的训练/验证/测试集

## v2 训练就绪配置（正式上线版）
已标准化以支持直接模型训练，所有数据源均重采样为统一格式。
python
# 加载训练就绪数据（包含全部拆分，可通过'split'字段筛选）
v2 = load_dataset("Forgis/Mechanical-Components", "v2_train", split="train")
v2_train = v2.filter(lambda x: x["split"] == "train") # 20143个样本
v2_val = v2.filter(lambda x: x["split"] == "val") # 1332个样本
v2_test = v2.filter(lambda x: x["split"] == "test") # 6363个样本

| 参数项 | 参数值 |
|-----------|-------|
| 采样率 | 12800 Hz |
| 窗口长度 | 16384样本（1.28秒） |
| 通道数 | 1（主振动通道） |
| 归一化方式 | 逐样本实例归一化 |
| 拆分规则 | 按数据源划分（训练/验证/测试集互不重叠） |

**训练集（12个数据源）**：CWRU、MFPT、FEMTO、XJTU-SY、IMS、OEDI、PHM2009、MCC5-THU、SEU、MAFAULDA、VBL、SCA-train
**验证集（2个数据源）**：Paderborn、Ottawa
**测试集（2个数据源）**：Mendeley、SCA-test

## 引用说明
请引用各原始数据集的官方文献。详见source_metadata配置中各数据源的完整引用信息。