Click any name or subtitle to editClick badge to cycle layer

CI Monitoring โ€” Pipeline Architecture

cimonitoring ยท simulation engine + detectors ยท click any component for details
Inputs โˆ’
SIM
Config โ€” Simulation Parameters
Shift hours ยท power levels ยท sampling ยท random seed (deterministic)
parameters
Simulation Engine ยท Modules 1โ€“3 โˆ’
SIM
โ‘  Energy Substrate
Per-second kW = aux + spindle idle + cutting (Gutowski; anchored to Brillinger)
SIM
โ‘ก Carbon Layer
Emissions + rolling CI per piece โ€” the monitoring signal
SIM
โ‘ข Anomaly Model
4 fault archetypes ยท additive excess ยท ground truth
Sensor & Observation Model ยท downsample + meter noise โ†’ observed CI (shared)
Detection & Evaluation ยท Module 4 / 4b โˆ’
DET
โ‘ฃ Deployed Detector
Rolling baseline + threshold + persistence โ€” has the inertia blind spot
DET
โ‘ฃแต‡ Proposed Detector โ˜…
Event-anchored held baseline + residual CUSUM โ€” closes the blind spot
DET
Evaluation
Scores alerts vs ground truth: latency ยท TP/FP ยท attribution
per-fault metrics
Experiment & Reproducibility ยท Module 5 โˆ’
REPRO
โ‘ค Sensitivity Harness
9 sweeps ยท 4,356 runs ยท no tuning
REPRO
Raw Results
data/ ยท 4,356 runs ยท authoritative
REPRO
Figure Reproduction
plot_paper_figs.py ยท from CSV only
REPRO
Colab Quickstart
Browser ยท zero setup
Quality & Distribution
Quality &
Distribution
โˆ’
Test Suite
pytest ยท headline regression test
Continuous Integration
GitHub Actions ยท Py 3.9/3.11/3.12
PyPI Package
ci-monitoring-simulation ยท pip install
Zenodo Archive
Concept DOI ยท versioned
Parameter Provenance
ANCHORED ยท LITERATURE ยท ASSUMPTION
Simulation โ€” substrate, carbon, anomaly (Modules 1โ€“3)
Detection โ€” deployed + proposed detectors, evaluation (Module 4 / 4b)
Reproducibility โ€” harness, data, tests, packaging, archive
Click any component for details ยท Esc to close
Flow types: Pipeline data Reproducibility / packaging Hover a component to highlight its flows
CI Monitoring โ€” Key Results
What the framework establishes ยท simulation study ยท click architecture to return
Characterisation (Paper 1)
FindingDetailSource
Inertia blind spotThe deployed adaptive-baseline detector misses faults whose onset timescale is comparable to its baseline window โ€” the rolling reference tracks the fault and it never crosses threshold.ยง4.7 sigmoid
Detection floor80%-detection threshold โ‰ˆ 47% of baseline load (~1.6 kW); sudden faults are caught in minutes with correct attribution.ยง4.1
Metering cadenceCoarse utility-grade metering (~15-min cadence) is too slow for this style of monitoring โ€” a measurement-resolution limit, not a tuning one.ยง4.2
Proposed detector (Paper 2)
FindingDetailSource
Closes the blind spotEvent-anchored held baseline + residual CUSUM detects a slow ramp the deployed detector misses entirely (warning โ‰ˆ 47 min after onset in the demo seed), with correct attribution and 0 false positives.head-to-head
Robust across seedsEffect established across 200-seed comparisons with confidence intervals, a detector ablation, and multiple severities โ€” not a single lucky run.ยง evaluation
Controlled comparisonBoth detectors consume the identical observed signal (shared sensor model), so the difference is attributable to detection logic alone.by design
Real-data groundingParameters anchored to the open Brillinger et al. (2025) CNC dataset; the detector is exercised on real spindle-power traces with controlled fault injection.validation
Scope: this is a calibrated simulation study. It characterises how the monitoring mechanism behaves and validates the proposed detector on real spindle-power traces with controlled (injected) faults โ€” not a real-world industrial deployment.