Forgis/Mechanical-Components

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

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)

Gearboxes (~1,225 samples from 4 sources)

Component Types Covered

Sensor Modalities

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

{
    "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.

# 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

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.