Everything you need to go from raw sensor readings to a 24-hour failure prediction — no ML knowledge required.
YieldGuard is a predictive maintenance system — think of it as a doctor's check-up for your machines. Instead of waiting for a machine to break down and halt production, YieldGuard watches the machine's vital signs (vibration, temperature, current, pressure, RPM, and acoustic level) and learns to recognize the early patterns that appear before a failure.
A trained AI model analyzes these readings and returns a single number: the probability that this machine will fail in the next 24 hours. If that probability is high, your maintenance team can schedule a repair during planned downtime — not at 2am during an emergency.
Explore 5 pre-loaded machines with real degradation scenarios. No data needed. Play the timeline and watch the risk gauge respond.
Open demoUpload a CSV from your historian or SCADA, or enter current readings with the quick-try sliders. Get an instant prediction.
Go to dashboardAfter running an analysis, you'll see a result panel with three key parts:
The gauge shows the model's predicted probability of failure in the next 24 hours. The color tells you the severity: teal = healthy, amber = caution, red = danger.
Readings within normal range. Continue routine monitoring.
Early warning signals detected. Not urgent but schedule an inspection.
Multiple sensors show abnormal patterns. Arrange maintenance soon.
Imminent failure pattern. Stop machine if safe and dispatch a crew.
Below the gauge you'll see the top contributors to the prediction — which sensors are behaving abnormally and in which direction. Each factor shows the sensor metric name, whether it's trending up or down, and how strongly it influenced the risk score.
For example: "Vibration variability trending up" means the rolling standard deviation of vibration has increased sharply relative to the machine's normal baseline — a classic bearing wear signature.
YieldGuard expects a time-ordered CSV with one row per reading. Readings should be at fixed intervals — 10 minutes is recommended (matching PLC historian defaults), but any fixed interval works.
| Column | Required | Example | Description |
|---|---|---|---|
| timestamp | — | 2024-01-01T08:00 | ISO 8601 datetime (optional). If omitted, YieldGuard treats rows as consecutive 10-minute readings. |
| vibration_mm_s | ✓ | 2.45 | Overall vibration amplitude in mm/s (RMS). Typically from an accelerometer on bearing housings. Normal: 1–5 mm/s. |
| temperature_c | ✓ | 64.8 | Operating temperature in °C measured at bearing or motor casing. Normal: 50–80 °C for most drives. |
| pressure_bar | ✓ | 8.1 | Hydraulic or process pressure in bar. Normal varies by application — stable values matter more than absolute numbers. |
| current_a | ✓ | 12.1 | Motor phase current in amps (RMS, any one phase). Rising current at constant load signals mechanical friction or winding degradation. |
| rpm | ✓ | 1476 | Shaft rotation speed in revolutions per minute. Unexpected slowdown relative to setpoint is an early wear indicator. |
| acoustic_db | ✓ | 71.8 | Acoustic emission or airborne sound level in decibels (dB SPL). Elevated levels may indicate bearing defects or cavitation. |
The model needs all 6 columns. If your historian doesn't log one (e.g. acoustic), fill it with typical values for your machine class — the model will still capture trends in the channels you do have, though accuracy for acoustic-related failures may reduce.
At minimum 144 rows (24 hours at 10-minute intervals). More is better — 288 rows (48h) gives the rolling-window features their full context. The upload panel tells you if you have enough.
Yes. Upload as-is. The model's rolling windows are sample-count based, so a 5-minute interval means your 144-reading window covers 12 hours instead of 24. Results are still valid — just scale your interpretation of the horizon accordingly.
It means the model estimates a 73% probability of a failure event occurring within the next 24 hours, given the pattern in your uploaded data. It is not a certainty — but it means the sensor pattern closely resembles machines that failed within 24 hours in training.
No. The model runs entirely in your browser (WebAssembly-like tree scoring in TypeScript). Your CSV never touches a server. You can verify this in your browser's Network tab — no outbound requests are made when you hit 'Run Analysis'.
The model was trained on synthetic data spanning hydraulic presses, CNC spindles, pumps, and conveyor drives. For best accuracy on your specific machine, the trained Python pipeline can be fine-tuned on your historical failure data — see the About page.
The demo runs the real model on sample machine data — no signup, no uploads needed.