README.md

---
license: cc-by-4.0
task_categories:
- image-classification
pretty_name: UORED-VAFCLS Bearing — Perception Representations (signal→VLM)
tags:
- bearing-fault-diagnosis
- vibration
- acoustic
- signal-to-image
- uored-vafcls
---
# UORED-VAFCLS — perception representations (visual grounding)

Rolling-element bearing records from the University of Ottawa constant-load, constant-speed rig,
rendered as **perception** images — one HF **config** per representation. Part of the
AI4Manufacturing FORGE corpus (Category **C**, task **T-C1**).

**480 records per config**, from 20 physical bearings; labels
{'ball': 80, 'cage': 80, 'normal': 160, 'inner_race': 80, 'outer_race': 80}; channels {'acc': 240, 'mic': 240}.

## Configs

```python
load_dataset("AI4Manufacturing/UORED-perception", "spectrogram")
```

| config | records |
|---|---|
| `spectrogram` | 480 |
| `scalogram` | 480 |
| `waveform` | 480 |
| `reshaped` | 480 |

## ⚠️ There is no train/test split — you must make one, and it must be BY BEARING

This dataset is shipped **unsplit, deliberately**. Every bearing is used for exactly **one** fault
type:

| bearings | class |
|---|---|
| 1–5 | `inner_race` |
| 6–10 | `outer_race` |
| 11–15 | `ball` |
| 16–20 | `cage` |

Each bearing contributes one healthy record (`H_<n>_0`) and two fault records, and each record is cut
into 4 windows × 2 sensor channels × 4 representations. **A random split therefore puts windows from
the same physical bearing — often the same 2.5 s window, merely rendered differently or read on the
other sensor — on both sides, and the reported accuracy will be an artefact.**

> **Note on the `train` split name.** HuggingFace requires every split to be named, so the single
> unsplit set is served as `train`. **It is the whole dataset, not a training portion — there is no
> matching `test`.** `load_dataset("AI4Manufacturing/UORED-perception", "spectrogram")["train"]`
> returns all 480 records, and it is on you to divide them.

Group by `metadata.bearing_id` and hold out whole bearings, stratified across the four classes. Use
`metadata.record`, `metadata.channel` and `metadata.window_idx` if you need finer grouping. We do not
publish our own holdout: any particular choice would read as the only defensible one, and the honest
constraint is the grouping rule, not one instance of it.

## Why perception-only

`reasoning` is empty on every row and there is no `-annotated` sibling planned. Three of the four
fault classes on this rig **cannot be attributed** from the signal, and the cause is the bearing's own
geometry rather than the recording quality:

- 6203 has **8 balls**, so the outer- and inner-race orders sit symmetrically about 4; and
  ball-diameter ÷ pitch-diameter = 0.2375 is within 5% of ¼. Together these put **BPFO 3.0498×
  (1.63% from 3×)**, **2×BSF 3.9722× (0.70% from 4×)** and **BPFI 4.9502× (1.01% from 5×)**. A line
  matcher has to allow 1–2% for real bearing-to-bearing geometric scatter, so its window contains both
  the bearing line and the shaft harmonic — and shaft harmonics are present on healthy machines.
- Measured without presupposing the answer: the brightest harmonic comb near the inner-race order sits
  at **4.998–5.036 (median 5.004)** on 10/10 records — 0.1% from the integer, 0.9% from BPFI, with a
  spread of 0.8% while bearing-to-bearing scatter is 1–2%. Outer race lands at 3.078 and **is**
  attributable (7/10). Ball lands at 4.012. The cage line has no integer neighbour at all and its
  window is simply empty.
- Ball has a second, documented cause: the dataset paper states *"For ball fault data, no load was
  applied"* — no load zone, the balls slip, and the impacts are not repeatable.

**The defects themselves are real and detectable** — a paired-baseline band-energy test separates
30/40 fault records at zero false alarms over 20 healthy records. What this dataset cannot support is
the *attribution* step, which is exactly what a faithful chain-of-thought would have to perform. So
the labels ship as **implanted gold** (where the rig operator installed the defect), the computed
evidence rides along as **non-gating** metadata (`evidence_is_gate: false` on every row — nothing was
dropped, relabelled or reordered by it), and **no envelope-spectrum representation is rendered**: an
image with the bearing lines drawn on top of the shaft harmonics would invite exactly the
confabulation this decision refuses.

## ⚠️ Two disclosures that affect how you should train and report

**1. Bearings 16–19 are a different recording batch.** They run at **1157–1172 rpm** while every other
bearing runs at **1769–1784 rpm**, and their records carry roughly **30× the broadband RMS**. The
source paper documents a single constant nominal speed of 1750 rpm and does not mention this. Since
those are four of the five `cage` bearings, **amplitude or texture alone can identify the `cage` class
without any bearing physics.** Every row carries `metadata.recording_batch`; audit against it, and
treat a high `cage` score with suspicion. This is recorded, not corrected — the cause is unknown.

**2. The healthy records are "not yet broken", not "known good".** The paper states they are taken
from the first files of the same run-to-failure sequence, on bearings whose seals had been removed and
which had been degreased to accelerate deterioration. They are a usable reference for the same
bearing's own earlier state; they are **not** a clean negative control.

## Other caveats

- **Both sensor channels are shipped** (`metadata.channel` ∈ {`acc`, `mic`}): a PCB 623C01
  accelerometer inside the housing and a PCB 130F20 microphone 2 cm away. Rows sharing a `record` and
  `window_idx` are the *same physical event* on two sensors — keep them together when splitting. The
  microphone is the stronger detector on this rig (30/40 vs 15/40 at the zero-false-alarm point), and
  the corpus has no other acoustic dataset.
- **The differential-temperature channel is not rendered** and must not be used as an input feature: a
  single 6.8 °C threshold separates healthy from faulty 59/60, while the four fault classes' medians
  all fall within 21.9–26.1 °C. It encodes how long the rig ran, not what broke. It is summarised per
  row as `d_temp_c` with `d_temp_leaks_label: true`.
- **`state` (the `_0/_1/_2` file suffix) is not a severity label.** The three states are taken by file
  index within each run, not by measured damage size; only 7 of 20 pairs move in the expected
  direction. Do not train on it.
- **Renderer parameters are not the library defaults.** Every default was calibrated on CWRU at
  12 kHz; at 42 kHz the same code means something else. Decimating to 12 kHz to restore those numbers
  would have discarded 6–20 kHz, which is where this rig's bearing impacts live, so the full rate is
  kept and each renderer is given the resolution or slice it needs. The full conversion is in
  `provenance.json` under `raw_form_provenance.render_calibration`.

## Provenance & reproducibility

Generated deterministically by `forge_agent/examples/uored_vafcls/convert.py`
(`4aab7b24dd`) → `forge_model/UORED/convert_uored.py`
(`2e4f6ba156`). Shaft speed is **measured per record** from the raw
spectrum — the stamped rpm column in the source files points at empty spectrum, including on healthy
records — using a 1–6 harmonic ladder that skips terms falling on mains harmonics. See
`provenance.json` for the full record.

## Source & license

Sehri, M. & Dumond, P., *University of Ottawa constant load and speed rolling-element bearing
vibration and acoustic fault signature datasets*, **Data in Brief 49 (2023) 109327**,
doi:[10.1016/j.dib.2023.109327](https://doi.org/10.1016/j.dib.2023.109327). Data:
doi:[10.17632/y2px5tg92h.5](https://doi.org/10.17632/y2px5tg92h.5), **CC BY 4.0**.