WF-14 · Process Sentinel

Process Sentinel

Detects anomalies and faults on a generic dynamic chemical process with isolated truth, a frozen 28-day synthetic benchmark and a drift-aware PCA + process-residual ensemble.

Experimental · simulated plantEvery 6 h: 00:50/06:50/12:50/18:50 UTCIndustry · process monitoring · anomaly detection
Demonstration card: workflow summary, not evidence from a real plant.
1 · In one sentence

Detects anomalies and faults on a generic dynamic chemical process with isolated truth, a frozen 28-day synthetic benchmark and a drift-aware PCA + process-residual ensemble.

2 · Why it exists

Study how a monitoring system can recognise anomalies and faults while jointly measuring sensitivity, false alarms and detection time on a controlled benchmark.

3 · What it produces

State, anomaly score, contributors, detected/missed events, recall, false alarms, delay and benchmark readiness.

The images come from demonstration runs included in the attached repository and are shown without dates to illustrate workflow structure, method and evidence.

Workflow overview and primary output.
Calculation method and pipeline.
Evidence, metrics or checks from the demonstration run.
History, sensitivities or validation context.

4 · How it works

From data to output, with explicit controls.

1CSTR simulator and isolated fault truth
2Multivariate features and missing-data handling
3High-specificity PCA T²/Q
4Physical residuals + slow CUSUM
5Voting ensemble and fault hypothesis
6Frozen benchmark and event metrics

Transferable pattern: Multivariate monitoring + frozen synthetic evaluation era.

5 · Data, AI/ML/RPA and method

Data and features

generic CSTR multivariate sensors, noise, missing data, step/drift/intermittent faults

Calculation / inference

drift-aware ensemble: high-specificity PCA T²/Q + steam, cooling and reaction residuals + slow steam-residual CUSUM; Isolation Forest/autoencoder as comparators; fault classifier

Validation

parameters frozen on a separate development simulation; same frozen 28-day synthetic benchmark; event recall, missed rate, false-alarm episodes/24h, P50/P90 delay, v1.4.16 ensemble comparison and anti-leakage

Stack

Python, NumPy, pandas, scikit-learn, SQLite, Matplotlib

6 · Technical dossier

Operating conditions and declared limitations.

Compute cadenceEvery 6 h: 00:50/06:50/12:50/18:50 UTC
Planned publicationMaterial events only; Mon/Fri reserved windows with cooldown
Possible applicationscondition monitoring, early warning and assisted diagnosis
Research usedetection-delay/false-alarm trade-off and model gating
Required conditionperformance applies only to the stated benchmark/simulator
Failure modeperformance applies only to the simulator; test-bed-specific process residuals and an uncalibrated classifier require validation before real-plant transfer
Interpretation boundary

The benchmark and process are synthetic. Performance does not automatically transfer to a real plant, and process residuals must be redesigned for the target process.

7 · Verification and further reading

Read the workflow in the wider laboratory context.