Predictive Maintenance
Predictive maintenance you can explain. No black box.
Deterministic anomaly detection plus a formula-driven health score, so your team trusts the alerts and your auditors get the math. Every score traces back to a specific event and a published deduction.
WHAT YOU GET
Static baseline z-score
Each metric is compared to its rolling mean plus standard deviation. Readings that drift outside the expected envelope are flagged with a z-score, even when the absolute value is still in machine tolerance.
EWMA for gradual drift
The exponentially weighted moving average adapts to slow drift. It catches the temperature climb on a hydraulic system or the creeping cycle-time before a static threshold would fire.
Rate-of-change detector
Flags abnormal first-difference velocity. Sudden jumps and steep slopes register even when the absolute value is still in range.
TECHNICAL DEPTH
How the health score is computed
A documented formula and a published deduction schedule. Every score traces back to specific events. Your team can reproduce the math on paper.
health_score = (1 − w) × alert_score + w × anomaly_score
Where w is the configurable anomaly weight per tenant. alert_score reflects the open alerts on the machine; anomaly_score reflects the detector activity.
Per-event deductions
| Event | Deduction | Cap |
|---|---|---|
| Critical alert, past 7 days | −15 pts | 3 |
| Critical alert, days 8 to 30 | −6 pts | 3 |
| Warning alert, past 7 days | −5 pts | 4 |
| Warning alert, days 8 to 30 | −2 pts | 4 |
| Unacknowledged active alert | −4 pts | 5 |
WHAT IT ALSO DOES
Composite alert rules
Multi-metric conditions with AND or OR logic and duration windows. Example: vibration > 5 mm/s AND temperature > 78 °C for 5 minutes triggers a critical alert.
Alert to work-order chain
An anomaly fires an alert; the alert opens a CMMS work order with parts pulled from inventory; the operator signs the completion. One audit trail end to end.
Per-machine baselines
MetricBaseline rebuilds nightly per machine, so a CNC at Plant A does not share thresholds with a CNC at Plant B even when the model number is identical.
Severity model
Info, warning, and critical levels. Operators see context on every alert and can acknowledge, snooze, or comment without leaving the dashboard.
No training period
Detectors run from the first reading. Baselines refine over time, but we do not need 90 days of data before the platform sends a useful alert.
Tamper-evident trail
Every score change, alert, acknowledgement, and signoff is HMAC-SHA256 chained in the audit log. Your auditor sees a reproducible record.
FREQUENTLY ASKED
Is this AI?
No, by design. AI models drift, need retraining, and cannot explain individual decisions to an auditor. The Haltless score is a transparent formula: fast to tune, easy to defend, immune to silent accuracy decay.
How long before alerts become useful?
Detectors run from the first reading. The static baseline z-score and rate-of-change settle within a few hours of history per metric; the EWMA adapts continuously. There is no 90-day learning window.
Can we tune the detector parameters?
Yes. The z-score window, EWMA smoothing factor, and rate-of-change sensitivity are per-tenant settings, and the deduction schedule above is the documented table that drives the health score.
