"An electric motor doesn't fail suddenly. It sends signals for weeks — vibration that creeps upward, bearings that run progressively hotter, current that increases as mechanical resistance builds. These patterns exist in sensor data long before the motor trips a protection relay."
Unplanned motor failure is one of the most expensive events in industrial operations. Not because motors are cheap to replace — but because of what stops with them: production lines, conveyor systems, HVAC, pumps. The downtime cost dwarfs the repair cost, every time.
The shift from reactive to predictive maintenance doesn't require exotic technology. It requires reading the signals that are already there. This project applies KNN to classify motor operating states as normal or incipient failure, using six condition-monitoring signals that any SCADA or PLC system already captures.
This project is free and fully open. Complete dataset, full notebook, and working simulator included — because predictive maintenance is too important to sit behind a paywall.
- 1,847 motor operating records from a condition monitoring system
- Target:
incipient_failure(binary) — 0 = normal operation, 1 = incipient failure state - Class balance: 29.0% incipient failure — reflecting typical monitoring coverage where failure states are captured during scheduled condition assessments, not continuous fault logging
- Source: Simulated sensor data modeling two distinct motor operating regimes
| Signal | Normal Range | Failure Range | Physical Meaning |
|---|---|---|---|
vib_rms_mms |
~2.2 mm/s | ~4.3 mm/s | Overall vibration energy |
vib_peak_to_peak_mms |
~5.0 mm/s | ~9.0 mm/s | Impact events, looseness |
bearing_temp_c |
~60°C | ~82°C | Friction buildup |
motor_current_a |
~28A | ~35A | Mechanical resistance |
dominant_freq_hz |
~29.5 Hz | ~24.5 Hz | Misalignment signature |
load_pct |
~70% | ~85% | Operating load context |
The class separation in this dataset is physically grounded — the feature distributions between normal and failure states reflect real ISO 10816 alarm thresholds. This is not arbitrary labeling.
Algorithm: K-Nearest Neighbors — sklearn.neighbors.KNeighborsClassifier
KNN is the right choice here because of a simple physical argument: if two motors have similar vibration, temperature, and current profiles, they're likely in the same operating state. Distance in sensor space maps directly to operational similarity. No linear assumption, no parameterized boundary — just proximity in a physically meaningful feature space.
Why KNN specifically requires StandardScaler: A feature in Amperes (range 25–40) would dominate distance calculations over one in mm/s (range 1–6) without normalization. Scaling isn't optional — it's what makes the distance metric physically meaningful.
Preprocessing: StandardScaler on all features, wrapped in a Pipeline.
Tuning: GridSearchCV over k × weights × distance metric — best: k=7, weights='distance', p=1 (Manhattan).
| Metric | Value |
|---|---|
| Accuracy | 88.3% |
| ROC-AUC | 0.940 |
| Recall (Failure) | 71.4% |
| F1 (Failure) | 0.780 |
| Precision (Failure) | 0.858 |
Confusion matrix (555 test records):
| Pred: Normal | Pred: Failure | |
|---|---|---|
| Actual: Normal | 375 (TN) | 19 (FP) |
| Actual: Failure | 46 (FN) | 115 (TP) |
46 failures missed across 555 test records. Post-analysis of these false negatives reveals a consistent pattern: motors running at 85–90% load with moderate vibration elevation (3.0–3.5 mm/s) — a zone where thermal signals become ambiguous because high-load motors run naturally warmer, regardless of mechanical condition. In production, this zone would be flagged for increased monitoring frequency rather than binary classification.
| Feature | Accuracy Drop | Role |
|---|---|---|
vib_rms_mms |
Δ −1.3% | Primary alarm signal |
vib_peak_to_peak_mms |
Δ −0.7% | Impact event detector |
load_pct |
Δ −0.2% | Operating context |
motor_current_a |
Δ −0.1% | Electrical context |
dominant_freq_hz |
Δ −0.0% | Misalignment type |
bearing_temp_c |
Δ +0.1% | Thermal — ambiguous under high load |
Vibration RMS remains the leading signal — elevation from ~2.2 to ~4.3 mm/s reflects a shift beyond ISO 10816 alarm thresholds. Bearing temperature, however, shows near-zero permutation importance: at load levels above 85%, motors run consistently warmer regardless of mechanical state, reducing its discriminative value in this dataset.
Motor_Failures_Prediction/
├── 03_KNN_Motor_Failure_Prediction.ipynb # Notebook (no outputs)
├── motor_failure_prediction_data.csv # Complete dataset (1,847 records)
├── app.py # Motor condition simulator
├── README.md
└── requirements.txt
✅ This project is completely free. Full dataset and simulator included. If this helped you, check out the rest of the portfolio at lozanolsa.gumroad.com.
Option 1 — Google Colab: Click the badge above.
Option 2 — Local:
pip install -r requirements.txt
jupyter notebook 03_KNN_Motor_Failure_Prediction.ipynbOption 3 — Run the simulator:
python app.py- Distance = physics — in condition monitoring, sensor similarity is operational similarity. KNN formalizes an intuition every maintenance engineer already has.
- Permutation importance works on any model — KNN has no coefficients, but shuffling features and measuring accuracy drop gives model-agnostic interpretability that stakeholders can understand.
- StandardScaler is not optional for KNN — without it, the distance metric is dominated by whichever feature has the largest absolute range. Scale before you measure.
- GridSearchCV on distance metric matters — Manhattan distance (p=1) outperformed Euclidean (p=2) here, consistent with sensor data that often contains outliers where L1 is more robust.
- The 46 missed failures tell the real story — motors running near threshold (85–90% load, moderate vibration) occupy a zone where the sensor profile is genuinely ambiguous. A model that claims to catch 100% of failures on industrial sensor data is not a better model — it's a cleaner dataset. In practice, these borderline cases would escalate to increased sampling frequency, not binary classification.
Luis Lozano | Operational Excellence Manager · Master Black Belt · Machine Learning
GitHub: LozanoLsa · Gumroad: lozanolsa.gumroad.com
Turning Operations into Predictive Systems — Clone it. Fork it. Improve it.