rotaguard-vibration-classifier

RotaGuard Vibration Classifier

LightGBM multiclass classifier that maps 19 time/frequency/envelope-spectrum features (extracted from a 4096-sample, 12 kHz bearing-vibration segment) to one of 12 labels: Normal, or {IR,OR,B}_{7,14,21,28} (fault location × severity in mil, i.e. thousandths of an inch defect diameter).

Data honesty

Trained on the public CWRU Bearing Data Center 12 kHz drive-end dataset — a real motor-bearing test rig, not an in-situ oil & gas pump/compressor/turbine. The bearing-fault physics (inner/outer race, ball defects) transfers directly to industrial rotating equipment; the specific test machine does not. See the dataset card for the full evaluation-protocol caveats (matched-condition split, not cross-load).

Held-out test results (2,750 segments, 12 classes)

Full confusion matrix and per-class accuracy are in eval_results.json. A 100% held-out score is plausible (and consistent with the ~99-100% range other public studies report) precisely because train/test is a matched-condition split (same 4 load conditions in both) with a rich hand-engineered feature set — not evidence of leakage. Cross-load generalization (train on some loads, test on unseen loads) is a separately-studied, harder benchmark variant and is not what is certified here; recalibrate/re-validate before any cross-load or field deployment.

Files

• vibration_classifier.joblib — LightGBM LGBMClassifier (150 trees, depth 6)
• eval_results.json — held-out protocol, confusion matrix, published-baseline citations
• manifest.json — feature column order, label set

How to score a segment

import joblib, numpy as np
from scipy.signal import hilbert
from scipy.stats import kurtosis, skew

clf = joblib.load("vibration_classifier.joblib")
# x: raw 4096-sample, 12 kHz vibration segment (1D float array)
# Recompute the 19 features with rotaguard.features_vibration.extract_features(x[None, :])
# (see the Space's src/rotaguard/features_vibration.py for the exact vectorized formulas)
proba = clf.predict_proba(features)[0]
label = clf.classes_[proba.argmax()]

Related

• Dataset: alirezaaminzadeh/rotaguard-vibration-rul-features
• RUL model: alirezaaminzadeh/rotaguard-rul-lstm
• Space: alirezaaminzadeh/rotaguard-predictive-maintenance
• Product: aria-ai.ir

MIT · Aria AI Engineering Team