xijiazhao/Welding202

NSTDA 2022 Rotating-Machinery Vibration Dataset

7,650 CSV recordings from the NSTDA rotating-machinery test rig (2022 collection). Single faults, compound faults, and healthy baseline, recorded at four shaft speeds. Unzip 2022.zip to get a single flat folder 2022/ of CSV files.

Filename convention

UNB_ANG_PAR_LOAD_RPM_IDX.csv        e.g. 4301_16_00_00_1200_041.csv

Field	Meaning	Units / notes
UNB	Unbalance severity	grams x 10 (4301 = 430.1 g); 0000 = absent
ANG	Angular misalignment	mm/100mm (NOT degrees); 00 = absent
PAR	Parallel misalignment	mm; 00 = absent
LOAD	Load code	00 throughout this set
RPM	Shaft speed	one of 0600 / 0900 / 1200 / 1480
IDX	Recording index	repeat number within the condition

• Healthy baseline: 0000_00_00_...
• Single fault: exactly one of UNB/ANG/PAR nonzero.
• Compound fault: two or three nonzero (e.g. 1961_21_22_... is a triple: Unb 196.1 g + Ang 21 + Par 22).

CSV format

Each file is ~1 second of vibration at 51.2 kHz (delta t = 1.953125e-5 s, ~51,180 samples). Layout:

• Rows 1-3: metadata (waveform column names, t0 timestamp, delta t).
• Row 4: blank. Row 5: repeated column header (time,V2V,V2H,V2A).
• Row 6 onward: data. Column 0 = timestamp; columns 1-3 = the signals.

Channels (accelerometer at sensor position 2):

• V2V = vertical, V2H = horizontal, V2A = axial.

Load with pandas by skipping the 5 header rows and dropping the time column:

import pandas as pd
df = pd.read_csv("2022/0000_00_00_00_0600_001.csv",
                 skiprows=5, header=None, usecols=[1, 2, 3])
x = df.to_numpy()   # shape (~51180, 3): V2V, V2H, V2A

Gotchas

• ANG is misalignment in mm per 100 mm, not an angle in degrees.
• UNB in the filename is 10x the gram value.
• If you see a .nstda_cache/ folder after processing, it is derived cache, not source data - safe to delete, regenerated on demand.