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.
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 cadence | Every 6 h: 00:50/06:50/12:50/18:50 UTC |
|---|---|
| Planned publication | Material events only; Mon/Fri reserved windows with cooldown |
| Possible applications | condition monitoring, early warning and assisted diagnosis |
| Research use | detection-delay/false-alarm trade-off and model gating |
| Required condition | performance applies only to the stated benchmark/simulator |
| Failure mode | performance 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