Raw vibration and motor-current recordings of an induction motor running under 8 health conditions, 2 drive frequencies and 2 load settings. The same 32 operating points were recorded in two separate sessions (trial=1, trial=2).

Layout

trial=1/condition={condition}-frequency={frequency}-load={load}.npy   (32 files)
trial=2/condition={condition}-frequency={frequency}-load={load}.npy   (32 files)

Field	Values
`condition`	`normal`, `faulted_bearings`, `phase_loss`, `rotor_bowed`, `rotor_broken`, `rotor_misalignment`, `rotor_unbalance`, `voltage_unbalance`
`frequency`	`33.3`, `50.0` (motor drive frequency, Hz)
`load`	`0`, `3` (load setting)

Each file is a NumPy array of shape (3, 1440000), dtype float64:

Channel	Signal
0	accelerometer, y-axis
1	accelerometer, x-axis
2	motor current

The channel assignment is taken from the mfs-datamodule loader documentation. Physical units are not recorded in the files.

Sampling rate: 12.8 kHz (1,440,000 samples = 112.5 s per recording). The rate is not stored in the files. It is inferred from the data: at 12.8 kHz the fundamental of the current channel falls at exactly 33.31 Hz and 50.00 Hz in the respective files.

What a recording contains

Each 112.5 s recording covers the whole run, not only steady-state operation: motor start-up (roughly the first 10–20 s), steady running, spin-down, and then a stretch with the motor off (current near zero) until the end of the file. The start and stop times differ from file to file and between trials.

The mfs-datamodule loader keeps samples 200000:920000 (15.6–71.9 s) to drop the start-up and spin-down portions. That crop is almost entirely steady-state in both trials: on average 0.5% of the crop in trial 1 and 1.3% in trial 2 is start-up or spin-down.

Trial 1 vs trial 2

The two trials are independent recordings of the same 32 operating points, not copies of each other. No files are byte-identical, the vibration channels are uncorrelated between trials (|r| ≤ 0.11 in all 32 pairs), and a time-shift search finds no alignment. The source documentation does not describe how the sessions differed. Observed differences:

Run timing. Steady running starts at 9–23 s in trial 1 and 11–20 s in trial 2, and ends at 76–104 s in trial 1 and 69–90 s in trial 2.

Vibration amplitude of several mechanical faults. The table below gives steady-state RMS on channel 0, trial 2 ÷ trial 1:

Condition	Trial 2 / trial 1
`faulted_bearings`	2.1–3.8×
`rotor_misalignment`	1.3–3.6×
`rotor_bowed`	2.0–2.6×
`rotor_broken`	0.54–0.72×
`phase_loss`	0.57–1.36×
`normal`, `rotor_unbalance`, `voltage_unbalance`	≈ 1.0–1.2×

Unchanged: the drive frequency. The shaft speed agrees within 0.2 Hz for 29 of 32 operating points; the three exceptions (normal, phase_loss, voltage_unbalance at 33.3 Hz, load 3) differ by 0.8–1.3 Hz.

These differences cause a cross-session shift. A logistic-regression classifier on log-PSD features of 1024-sample steady-state windows reaches 99.5% / 99.8% accuracy within trial 1 / trial 2 (temporal 70/30 split). Trained on one trial and tested on the other, it drops to 77% (trial 1 → trial 2) and 83% (trial 2 → trial 1). Most of the errors fall on rotor_broken and rotor_misalignment.

Usage

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xijiazhao/welding2",
    repo_type="dataset",
    filename="trial=1/condition=normal-frequency=50.0-load=0.npy",
)
x = np.load(path)            # (3, 1440000): accel-y, accel-x, current
steady = x[:, 200000:920000]  # drop start-up / spin-down

To fetch everything: hf download xijiazhao/welding2 --repo-type dataset --local-dir mfs_motor_fault_data_set.

With mfs-datamodule, point data_dir at a single trial folder (for example .../trial=1). The loader searches for .npy files recursively and takes labels only from file names. Given the dataset root, it will silently merge both trials with no trial label.

Downloads last month: 45

Total file size:

2.21 GB