xijiazhao/welding2023

NSTDA 2023 Motor-Rig Vibration Dataset

2,161 vibration recordings of 10 s from the NSTDA motor test rig (2023 collection, 16 sessions, 27 Jul – 15 Sep 2023). An 11 kW induction motor drives a gearbox (3.15 : 1) and an eddy-current brake that sets the motor torque. Healthy machine, unbalance, parallel and angular misalignment at two severities each, and compound faults, recorded on a crossed grid of shaft speed (600 – 1460 rpm) × torque (20 – 37 N·m). Companion to xijiazhao/Welding202 (the 2022 collection of the same rig: four speeds, no torque control, 1 s files).

What is new compared with 2022: a controlled load axis; 10 s files (1X resolved at 0.1 Hz); two severities per fault type; a second block of sessions (28–29 Aug) with unseen speeds, torques and severities that works as a shifted test set; three more measurement points (V1 / V3 / V5).

Contents

path	content	size
raw/<YYYYMMDD>/<condition folder>/<sensor group>/<file>.csv	the original CSV files, untouched (folder and file names as recorded)	50.6 GB
raw/.../IEPE_S/	point V2 (motor drive-end bearing), channels V2V, V2H, V2A, 51 200 Hz; 1,011 files	33.6 GB
raw/20230817/Unbalance (2P)/additional/	V2, take 1 of the 08-17 U2P recordings; 69 files	2.3 GB
raw/.../CurrentS/	points V1, V3, V5 (motor non-drive end, gearbox input, output-shaft bearing), 9 channels, 10 000 Hz; 1,081 files	14.7 GB
cache2023/x2560.npy, cache2023/meta.parquet	processed V2: every usable V2 recording decimated to 2560 Hz, 10 s, float32 (1077 × 25600 × 3) + one metadata row per recording	331 MB
metadata/recordings.csv	one row per raw file: repo path, labels, measured speed, duration, flags, row in the cache	0.4 MB
metadata/files_full.csv, meta_files.csv, signal_summary.csv	the index tables the above was built from (per-channel mean / std, 1X / 2X / 3X amplitudes, measured shaft frequency)	3.8 MB
docs/DATASET.md, docs/figures/	full description: rig, sessions, file format, data-quality checks, measured properties, confounds	–
scripts/	scan_meta.py, scan_signal.py (index tables), make_figures.py, build_recordings.py, build_cache2023.py (rebuilds cache2023/ bit-identically from raw/)	–

Quick start

# processed V2 only (331 MB); this is what the Interpretable-motor r11 code reads at $NBAK/nstda2023/cache2023/
hf download xijiazhao/welding2023 --repo-type dataset --include "cache2023/*" --local-dir "$NBAK/nstda2023"
# metadata and docs only
hf download xijiazhao/welding2023 --repo-type dataset --include "metadata/*" "docs/*" --local-dir welding2023
# raw V2 only (36.0 GB)
hf download xijiazhao/welding2023 --repo-type dataset --include "raw/*/*/IEPE_S/*" "raw/*/*/additional/*" --local-dir welding2023
# everything (51 GB)
hf download xijiazhao/welding2023 --repo-type dataset --local-dir welding2023

import numpy as np, pandas as pd
X = np.load("cache2023/x2560.npy", mmap_mode="r")   # (1077, 25600, 3) float32: V2V, V2H, V2A at 2560 Hz
meta = pd.read_parquet("cache2023/meta.parquet")    # row i describes X[i]; X[i, :meta.n_valid[i]] is signal, the rest is zero padding

r = pd.read_csv("metadata/recordings.csv", keep_default_na=False)
row = r[(r.sensor_group == "IEPE_S") & (r.exclude_reason == "")].iloc[0]
x = pd.read_csv(row.path, skiprows=5 if row.has_header else 0, header=None).iloc[:, 1:].dropna().to_numpy()  # (n_samples, 3)

Conditions

V2 recordings after the exclusions below (take 1 = a second 10 s capture of the same round, 11–12 s after take 0):

group	family	condition folder	file code	V2 recordings, take 0 / take 1	days (2023)	block
H	H	Normal	Normal	205 / 6	07-27, 08-03, 08-04, 08-07, 08-08, 08-09, 08-10, 08-11, 08-15, 08-16, 08-17, 08-21, 08-28	A/B
U1P	U	Unbalance (1P)	Unbalance(1P)	115 / 6	08-04, 08-07, 08-08, 08-21, 08-29	A/B
U2P	U	Unbalance (2P)	Unbalance(2P)	69 / 69	08-17	A
P25	P	Parallel Misalignment (25mm)	ParallelAlign(25)	69 / 0	08-11	A
P36	P	Parallel Misalignment (36mm)	ParallelAlign(36)	69 / 0	08-15	A
P32	P	Parallel Misalignment (32mm)	ParallelAlign(32)	30 / 0	08-29	B
A15	A	Angular Misalignment (15mm)	AngularAlign(15)	69 / 0	08-10, 08-11	A
A33	A	Angular Misalignment (33mm)	AngularAlign(33)	69 / 0	08-08, 08-09	A
A36	A	Angular Misalignment (36mm)	AngularAlign(36)	30 / 0	08-28	B
UP31	UP	Unbalance (1P) + Parallel Misalignment (31mm)	Unbalance(1P)_ParallelAlign(31)	69 / 0	08-21	A
UP32	UP	Unbalance (1P) + Parallel Misalignment (32mm)	Unbalance(1P)_ParallelAlign(32)	30 / 0	08-29	B
UA29	UA	Unbalance (1P) + Angular Misalignment (29mm)	Unbalance(1P)_AngularAlign(29)	69 / 0	08-18	A
UA36	UA	Unbalance (1P) + Angular Misalignment (36mm)	Unbalance(1P)_AngularAlign(36)	30 / 0	08-28	B
UPA	UPA	Combo(27_23)	Combo(27_23)	69 / 0	09-15	A

• "1P / 2P" = one / two unbalance pieces. "mm" is the misalignment fixture offset (not documented further). Combo(27_23) is presumably angular 27 + parallel 23 + unbalance (unverified).
• Block A (03 Aug – 21 Aug, 15 Sep): speeds 600, 750, 900, 1050, 1200, 1460 rpm (no 1460 at 20 N·m), torques 20, 25, 30, 35 N·m, 3 rounds per cell (rounds 0 and 1 minutes apart, round 2 usually 1–3 h later). A full fault grid is 23 cells × 3 rounds = 69 files.
• Block B (28–29 Aug): speeds 600, 800, 1100, 1200, 1300 rpm, torques 27, 30, 37 N·m, 2 rounds; P32, A36, U+P32, U+A36, U1P and a full Normal grid (plus a few 30 N·m Normal files at 750 / 900 / 1050 and 30 / 35 N·m files at 1460).
• 07-27 is a pilot (Normal, 600 / 750 rpm, different file names).
• Speed is closed-loop: the measured shaft frequency matches the label to < 0.3 % at every torque ("1460" runs at 1458 rpm; "800 / 1100 / 1300" at 798 / 1098 / 1302). Torque therefore cannot be read from the speed.

Raw file format

<Cond>_<Torque>_<Round>_<RPM>_<Take>.csv           sessions from 08-03 on
Normal_20230727_<Torque>_r<Round>_<RPM>_0.csv      07-27, IEPE_S (rounds r1-r3 = 0-2)
Normal_20230727_r<Round>_<Torque>_<RPM>_0.csv      07-27, CurrentS (torque and round swapped)

Header of 5 lines, then data; column 0 is a timestamp, the rest are the channels; the last line is empty commas:

waveform,V2V,V2H,V2A
t0,8/8/2023  09:49:03.907446,...          # start time, M/D/YYYY (two spaces) HH:MM:SS.ffffff
delta t,1.953125E-5,...                    # 1 / fs
<empty>
time,V2V,V2H,V2A
8/8/2023  09:49:03.907446,-6.39E-1,3.73E-1,-3.42E-1

9 files have no header (data from line 1; has_header = False in recordings.csv). Units are not documented for any channel. V2 and V1 / V3 / V5 are captured simultaneously (t0 differs by 0.01–0.22 s). raw/20230817/Unbalance (2P)/IEPE_S/Unbalance(2P)_(Torque)_(Round)_.txt is an empty file kept from the original copy.

metadata/recordings.csv

column	meaning
path	file path inside this dataset
date, session	recording day
sensor_group, point, channels, fs_hz	IEPE_S / additional = V2 at 51 200 Hz; CurrentS = V1 + V3 + V5 at 10 000 Hz
condition, condition_folder	condition code from the file name and the folder it sits in
family, group	H / U / P / A / UP / UA / UPA (KNOCK = impact test) and the group code of the table above
torque_nm, rpm_label, rpm_measured	brake set-point, labelled speed, 60 × shaft frequency measured from the spectrum
round, take, block	repetition 0–2, capture 0–1, block A / B
n_samples, duration_s, has_header, t0	signal length and start time
exclude_reason	rpm_0 (motor stopped or ramping), knock_test (impact test inside 08-04 Normal), torque_label_conflict (08-17 Normal 600 rpm: the V2 files say 35 N·m, the simultaneous V1 / V3 / V5 files say 30 N·m); empty = usable
short	shorter than 9.9 s (21 files, 2–9 s)
cache_row	row of cache2023/x2560.npy holding this recording (−1 if not in the cache)

cache2023/

Every V2 recording (IEPE_S + additional) with rpm > 0 and no knock test (1,077; the 4 torque_label_conflict files are included and flagged mislabel = True): per-channel mean removed, decimated 51.2 kHz → 2560 Hz with zero-phase FIR filters in two stages (×4, then ×5; this removes the 8.1 kHz carrier that holds 62–94 % of the raw V2 power), the first 10 s kept and zero-padded. meta.parquet columns: file, folder, day (MMDD), family, group, torque, rpm, rpm_meas, round, take, block, n_valid, t0, mislabel. No amplitude normalisation (amplitude carries load and fault).

Things that decide what an experiment can claim

• Fault ↔ day. Every fault configuration except U1P is recorded on one day; a classifier can learn the day. Normal is recorded on most days and is the per-day anchor; the healthy machine drifts between days (1X horizontal at 30 N·m, 1200 rpm: 0.60 to 1.10).
• Torque ↔ day for A15 and A33 (25 / 30 N·m and 20 / 35 N·m on different days).
• Healthy coverage is thin away from 30 N·m (20 N·m: 2 files per cell, each speed from another day).
• Torque has a real, fault-dependent signature (torque explains 10–90 % of the variance of log 2X / 3X within a session and speed).
• Block B is a natural shifted test set (unseen speeds, torques and severities, recorded later).
• Split by file and round, never by window: windows of one file are not independent.

Details and figures: docs/DATASET.md.

Source

Recorded on the motor test rig of the NSTDA Center for Cyber-Physical Systems (NECTEC, Thailand); official page https://www.nstda.or.th/cps/motor-data/ (contact smc-business@nectec.or.th). The official release lists 728 recordings; this copy is the raw superset of the sessions (it adds block B and the extra Normal / Unbalance files). Follow NSTDA's terms when using or redistributing the data.