uored_vafcls_42khz_60rec.npz ============================ A repackaged copy of the UORED-VAFCLS bearing vibration dataset, mirrored by BrainChip so that the accompanying example can fetch the data as a single file. Original dataset ---------------- Sehri, Maryam; Dumond, Patrick (2023), "University of Ottawa Rolling-element Dataset - Vibration and Acoustic Faults under Constant Load and Speed conditions (UORED-VAFCLS)", Mendeley Data, V5, doi: 10.17632/y2px5tg92h.5 Source: https://data.mendeley.com/datasets/y2px5tg92h/5 The dataset comprises 60 accelerometer recordings, each 10 s at 42 kHz, from 20 physical bearings. Every bearing contributes three recordings: one healthy, plus two severities of a single fault mode. Bearings 1-5 carry inner-race faults, 6-10 outer race, 11-15 ball, 16-20 cage. License ------- Creative Commons Attribution 4.0 International (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/ You are free to share and adapt this material for any purpose, including commercially, provided you give appropriate credit to the original authors (citation above), provide a link to the license, and indicate if changes were made. This repackaged copy is redistributed under the same license. Changes made ------------ As the license requires us to state, the following changes were made relative to the original Mendeley distribution. No sample values were altered, resampled, filtered or rescaled. 1. Format. The original is distributed as one CSV file per recording, under five subfolders (1_Healthy, 2_Inner_Race_Faults, 3_Outer_Race_Faults, 4_Ball_Faults, 5_Cage_Faults). Those files have been repacked into this single uncompressed NumPy .npz archive. 2. Column selection. Only the "Accelerometer" and "Speed" columns of each CSV are carried over. The accelerometer samples are stored verbatim as float32; from "Speed", only the first row's value is retained, as a per-recording nominal RPM. 3. Added metadata. Per-recording identifiers and multi-label fault targets, derived from the original filenames and folder structure, are stored alongside the signals (see Contents below). Contents -------- Load with numpy.load(). Arrays, all with 60 rows in a fixed row order (folder order above, then sorted filename within each folder): signals float32 (60, 420000) raw accelerometer samples, 42 kHz labels float32 (60, 4) multi-label targets, one-hot per column order in label_columns; healthy recordings are all-zero label_columns str (4,) ('inner', 'outer', 'ball', 'cage') row_idx int32 (60,) 0..59 bearing_id int32 (60,) physical bearing, 1..20 fault_type str (60,) 'Healthy', 'Inner', 'Outer', 'Ball', 'Cage' severity int32 (60,) 0 for healthy, else fault severity waveform_id str (60,) '__' rpm int32 (60,) nominal speed; for analysis only, it is never a model input fs int32 scalar 42000 SHA-256: 4ba82c0c36a38fb2afc1428c46a84264093da8c6084d4e471dbf00c9679c9c98 Size: 101 MB Notes on use ------------ Fault mode is a property of the physical bearing rather than of the signal, so any train/test split that lets windows from one bearing fall on both sides of the boundary leaks the label and inflates the reported score. Split by bearing_id, not by recording or by time segment. The example that uses this file, including the script that rebuilds this archive from the original CSVs (uored_vafcls_data.py --prepare-raw), is at: https://github.com/Brainchip-Inc/brainchip_devhub/tree/main/akida1/model_zoo/uored_vafcls