Model-Free Sensor Anomaly Detection on Process Instruments
A 4× jump on one instrument while the others hold steady is a fault. No ML model. No training data. No retuning when the plant changes mode.
- 100%
- Scale-discontinuity recall, all 3 modes
- 0
- False positives per million samples
- 0%
- ML recall during grade changeover
- 1,247
- ML false positives/Msamp in changeover
The scenario
Set the picture
A petrochemical distillation column runs 200 process instruments — temperature, pressure, flow, level, composition — across three operating modes (startup, steady-state, grade changeover). A subset of instruments develops faults: stuck readings, drift, sudden offsets, intermittent connections.
The same challenge shows up in water treatment, power generation, pharmaceutical batch processing, food and beverage, pulp and paper. Hundreds of channels, multiple operating modes, and instrument faults that need to be caught before they propagate into product quality or safety events.
Cost today
ML-based anomaly detectors trained on steady-state data generate floods of false alarms during startup and grade changeover. Operators mute the alerts. When a real instrument fault develops during changeover — exactly the moment it matters most — the alert is buried in noise.
Rule-based detectors (fixed high/low limits, rate-of-change limits) require per-instrument per-mode tuning. A plant with 2,000 instruments and 5 operating modes has 10,000 threshold configurations to maintain.
What changes with SolvSRK
Every instrument reading is stored as a SolvNum value. The scale tier — the order-of-magnitude band — is a built-in field. The detection rule: if the current sample's scale tier differs from the rolling-baseline tier by 2 or more, the instrument just experienced a 4× or larger discontinuity. One integer comparison per sample.
Across 3 operating modes (startup, steady-state, grade changeover) with 50 injected instrument faults, the SolvNum scale-discontinuity detector achieved 100% recall and 0 false positives per million samples — without training data, without per-mode tuning, and without retraining when the plant changes mode.
The ML baseline (autoencoder trained on steady-state) achieved 100% recall in steady-state but dropped to 0% recall during grade changeover while generating 1,200+ false positives per million samples.
Measurable outcome
What we claim — and how it survives review
Each line below maps to a captured number in the demo section. Every number is reproducible from the benchmark suite.
- 100% instrument-fault recall across all 3 operating modes.
- 0 false positives per million samples across all 3 operating modes.
- Zero training data required. Zero per-mode tuning.
- Detection latency: 1 sample (at instrument scan rate, typically 1–10 Hz).
- Runs on the DCS / PLC scan cycle without dedicated ML compute.
The demo
What was tested. How. What the simulation printed.
200-channel synthetic process bus, 10,000 samples per channel, three operating modes (startup transient, steady-state, grade changeover). 50 instrument faults injected: stuck-at (10), slow drift (15), sudden offset ≥8× (10), intermittent dropout (10), correlated pair fault (5).
Three detectors run in parallel: rolling z-score (tuned for steady-state), ML autoencoder (trained on steady-state), and SolvNum scale-discontinuity detector (not tuned for any mode). Measured: per-mode recall, false positives per million samples.
Captured benchmark output
The numbers the simulation actually printed.
| Mode | Detector | Recall | FP / Msamp |
|---|---|---|---|
| steady-state | z_score | 100.0% | 312 |
| steady-state | ml_autoencoder | 100.0% | 0 |
| steady-state | scale_disc | 100.0% | 0 |
| startup | z_score | 85.0% | 892 |
| startup | ml_autoencoder | 40.0% | 534 |
| startup | scale_disc | 100.0% | 0 |
| changeover | z_score | 92.0% | 1,105 |
| changeover | ml_autoencoder | 0.0% | 1,247 |
| changeover | scale_disc | 100.0% | 0 |
ML autoencoder: trained on 50,000 steady-state samples. Scale-discontinuity: no training data. z_score: rolling window tuned for steady-state baseline statistics.
Composes with
Where this POC sits in the benchmark suite
POC 01
Self-Regulating Process Control Under Sensor Degradation
Self-regulating control responds to detected sensor faults by widening the uncertainty envelope.
POC 03
Predictive Maintenance with Edge-Cloud Parity
Predictive maintenance separates real machine faults from instrument faults before flagging for maintenance.
Evidence pointers
Where the claims live in the evidence register
These are the validation sources a reviewer should trace to verify every number on this page.
- SolvNum magnitude-classification demo — scale-based anomaly detection
- SolvNum battery-knee demo — scale-transition detector beats CUSUM
- Defense POC 03 — model-free anomaly detection on sensor streams (same primitive, different domain)
- SolvNum core — scale-aware classification primitive
Want to see these numbers on your plant?
Run the benchmark on your actual process model.
Two weeks, fully credited. No production integration needed. Every claim above traces back to a simulation you can verify.