Detection limits of MES-embedded carbon-intensity monitoring

Simulation code, raw sweep data, and figure-generation scripts for the paper:

Yanytska, L. (2026). Detection limits of MES-embedded carbon-intensity monitoring for energy anomalies: a calibrated simulation study in machining-style processes. The International Journal of Advanced Manufacturing Technology (accepted, in press). Manuscript JAMT-D-26-03924.

This repository reproduces every figure in the paper from the 4,356 raw simulation runs released in data/.

Companion method (Paper 2): the paper2_anchored_detector/ folder contains the proposed event-anchored + residual-CUSUM detector that closes the adaptive-baseline inertia blind spot characterised in this study, together with its full evaluation — the operating-point calibration, the detector ablation, generalisation across severity, and validation on real Brillinger spindle-power traces. See that folder's README.

---

Try it first (no install)

The fastest way to see the headline result is the Colab quickstart notebook — it runs entirely in your browser with nothing to install. Open it, choose Runtime ▸ Run all, and you'll reproduce the result in about a minute.

Requirements (for local use)

• Python 3.9 or newer. Check yours with python3 --version. If you don't have Python, install it from python.org (macOS/Windows) or your package manager. pip ships with Python 3.4+; if pip is somehow missing, run python3 -m ensurepip --upgrade.
• Dependencies (numpy, pandas, matplotlib) install automatically with the package below — you don't need to install them by hand.

Install

The simulation framework (Modules 1–5) and the proposed detector are available as an installable package, cimonitoring:

pip install ci-monitoring-simulation
# or, the latest development version straight from GitHub:
pip install git+https://github.com/lesiayanytska78/ci-monitoring-simulation.git
# or, from a local clone:
pip install .

Tip: to keep this isolated from your other Python projects, create a virtual environment first:

python3 -m venv venv && source venv/bin/activate   # Windows: venv\Scripts\activate
pip install ci-monitoring-simulation

import cimonitoring as ci
sub = ci.simulate_work_center(ci.Config(seed=1))
sub = ci.inject_anomalies(sub, ci.AnomalyConfig([
    ci.AnomalySpec(onset_hour=10, duration_minutes=240, magnitude_kw=2.0,
                   onset_profile="ramp", onset_ramp_seconds=3600,
                   affects="spindle", label="slow ramp")]))
obs = ci.run_monitoring_anchored(
    sub, ci.AnchoredMonitorConfig(detector="anchored_cusum"),
    ci.CarbonConfig().static_emission_factor_kg_per_kwh)

The figure-reproduction below reads the released CSVs and does not require installation.

---

Quick reproduction

# 1. Clone and enter the repository
git clone https://github.com/lesiayanytska78/ci-monitoring-simulation.git
cd ci-monitoring-simulation

# 2. Install dependencies (Python 3.9 or newer)
pip install -r requirements.txt

# 3. Regenerate all seven figures from the released CSVs
python plot_paper_figs.py

Expected output: seven PNGs written to figures/. Wall time on a single CPU core: under one minute. plot_paper_figs.py reads only the CSVs in data/ and writes only to figures/ — it does not require the simulation modules.

A Makefile provides shortcuts: make install, make figures, make test.

Tests

A pytest suite in tests/ exercises the full pipeline end to end — the energy substrate, carbon layer, anomaly model, the deployed detector, and the proposed event-anchored + CUSUM detector — and includes a regression test for the paper's headline result (the fixed-reference detector fires on a slow ramp the deployed detector misses) and an integrity check that the released CSVs load and the shared simulation modules do not drift between simulation/ and paper2_anchored_detector/. Run with:

pip install -r requirements.txt pytest
pytest -q

Continuous integration runs the suite on Python 3.9, 3.11, and 3.12 on every push (see the badge above).

Re-running the simulations from scratch (optional)

The five simulation modules live in simulation/. The sweep functions that produced data/ are in simulation/sensitivity.py; the default sweep set (Sweeps 1–6) runs from its __main__ block, and the ramp-time, boundary, and ramp-transition sweeps (Sweeps 7–9) are provided as functions in the same module. Output paths are set at the bottom of the script — adjust them to your environment before running. The released CSVs in data/ are the authoritative results behind every reported number.

---

Repository layout

.
├── README.md                                  this file
├── LICENSE                                    MIT (code)
├── requirements.txt                           numpy, pandas, matplotlib
├── CITATION.cff                               machine-readable citation
│
├── plot_paper_figs.py                         generates Figures 1–7 from data/
│
├── simulation/                                Module 1–5 simulation harness
│   ├── energy_substrate.py                    Module 1 — energy substrate (Gutowski decomposition)
│   ├── carbon_layer.py                        Module 2 — carbon layer (rolling CI per piece)
│   ├── anomaly_model.py                       Module 3 — four-archetype anomaly model
│   ├── monitoring.py                          Module 4 — MES/SCADA sensor + rule-based detector
│   └── sensitivity.py                         Module 5 — sensitivity-sweep harness
│
├── data/                                      raw per-run results (4,356 rows total)
│   ├── sweep1_latency.csv                     severity × duration (300 runs)
│   ├── sweep2_sampling.csv                    severity × meter cadence (240)
│   ├── sweep3_roc.csv                         relative-threshold tightness (240)
│   ├── sweep4_archetypes.csv                  severity × four archetypes (240)
│   ├── sweep5_threshold_types.csv             absolute / relative / statistical (496)
│   ├── sweep6_attribution.csv                 severity × affected channel (140)
│   ├── sweep7_ramp_time.csv                   severity × ramp time × seed (650)
│   ├── sweep8_boundary.csv                    50-seed boundary, 1.0–2.0 kW (250)
│   └── sweep9_ramp_transition_200seed.csv     200-seed fine transition (1,800)
│
├── figures/                                   the seven paper figures (PNG)
│
└── docs/
    └── parameter_provenance.md                full [ANCHORED]/[LITERATURE]/[ASSUMPTION] table

---

Architecture

The system is a layered simulation pipeline: each stage produces a pandas.DataFrame that the next consumes, so any stage can be inspected, swapped, or re-run in isolation. The same observed signal is fed to two interchangeable detectors — the deployed adaptive-baseline detector and the proposed event-anchored detector — which makes the head-to-head comparison a controlled experiment rather than two separate runs.

flowchart TD
    CFG["Config<br/>simulation parameters (seed, shift hours, power levels)"]
    CFG --> M1

    M1["Module 1 — energy_substrate.py<br/>simulate_work_center()<br/>per-second power = aux base + spindle idle + cutting increment<br/>(Gutowski decomposition; anchored to Brillinger 2025)"]
    M1 --> M3
    M1 --> M2

    M2["Module 2 — carbon_layer.py<br/>compute_carbon_layer()<br/>defines the monitoring metric: rolling CI per piece (kg CO2e/piece)"]

    M3["Module 3 — anomaly_model.py<br/>inject_anomalies()<br/>additive excess power, 4 archetypes, with ground truth"]
    M3 --> SENSOR

    SENSOR{"shared sensor model<br/>sample_and_noise()<br/>downsample + meter noise → observed CI per piece"}
    M2 -. metric definition .-> SENSOR

    SENSOR --> M4
    SENSOR --> M4B

    M4["Module 4 — monitoring.py<br/>run_monitoring() — DEPLOYED<br/>rolling baseline + threshold + persistence + attribution"]
    M4B["Module 4b — monitoring_anchored.py<br/>run_monitoring_anchored() — PROPOSED<br/>event-anchored held baseline + residual CUSUM"]

    M4 --> EVAL
    M4B --> EVAL

    EVAL["evaluate()<br/>scores alerts vs ground truth:<br/>detection latency, TP/FP, attribution correctness"]
    EVAL --> M5

    M5["Module 5 — sensitivity.py<br/>sweep harness — 9 parameter sweeps, 4,356 runs"]
    M5 --> DATA["data/*.csv<br/>raw per-run results (authoritative)"]
    DATA --> FIGS["plot_paper_figs.py → figures/<br/>regenerates every paper figure from data/"]

Loading

The key design property is the shared sensor stage: sample_and_noise() lives in Module 4 and is reused unchanged by Module 4b, so both detectors see the identical observed signal and any difference in detection performance is attributable to the detection logic alone.

Layer	File	Responsibility	Key public API
Module 1 — Energy substrate	energy_substrate.py	Per-second power for one work center: auxiliary base + spindle no-load + cutting increment (engaged only while cutting). Structure follows Gutowski et al. (2006); the no-load spindle range is anchored to the Brillinger et al. (2025) CNC dataset.	Config, simulate_work_center()
Module 2 — Carbon layer	carbon_layer.py	Layer emissions and carbon intensity. Defines the central monitoring signal — rolling CI per piece — which rises when energy climbs but output does not.	CarbonConfig, compute_carbon_layer()
Module 3 — Anomaly model	anomaly_model.py	Inject parametrized faults as additive excess power with no extra output, recording exact onset/magnitude/duration as ground truth. Four archetypes.	AnomalyConfig, AnomalySpec, inject_anomalies()
Module 4 — Monitoring (deployed)	monitoring.py	Model the MES/SCADA view (downsampled, noisy), then the deployed detector: rolling baseline + threshold + tiered persistence + attribution. Includes the evaluation scorer.	MonitorConfig, run_monitoring(), sample_and_noise(), evaluate()
Module 4b — Monitoring (proposed)	monitoring_anchored.py	The proposed detector closing the inertia blind spot: event-anchored held baseline + residual CUSUM on the dimensionless residual. Reuses Module 4's sensor model for a controlled comparison.	AnchoredMonitorConfig, run_monitoring_anchored()
Module 5 — Sensitivity harness	sensitivity.py	Run the substrate → anomaly → monitor → evaluate pipeline across parameter sweeps; write raw per-run results to CSV. No tuning to targets.	sweep functions
Reproduction	plot_paper_figs.py	Regenerate every figure purely from the released CSVs (no simulation modules required).	—

▶ Open the interactive diagram — clickable components, flow overlay, and themes (source: architecture.html). A fuller write-up with design rationale and extension points is in ARCHITECTURE.md.

---

How paper claims map to data

Every figure and headline number is reproducible from one CSV in data/.

Paper claim	Section	CSV
Detection rate vs severity, three durations	§4.1, Fig. 1	sweep1_latency.csv + sweep8_boundary.csv
80%-detection threshold ≈ 47% of baseline (1.6 kW)	§4.1	sweep8_boundary.csv
Detection rate at four meter cadences	§4.2, Fig. 2	sweep2_sampling.csv
Threshold-tightness ROC	§4.3, Fig. 3	sweep3_roc.csv
Four-archetype comparison	§4.4, Fig. 4	sweep4_archetypes.csv
Threshold-family ROC (abs / rel / stat)	§4.5, Fig. 5	sweep5_threshold_types.csv
Attribution accuracy by channel	§4.6, Fig. 6	sweep6_attribution.csv
Inertia-trade-off sigmoid (200 seeds)	§4.7, Fig. 7	sweep9_ramp_transition_200seed.csv
Coarse multi-severity inertia behaviour	§4.7	sweep7_ramp_time.csv

---

Parameter provenance

Every numeric parameter in the simulation carries one of three provenance tags, documented in docs/parameter_provenance.md:

• [ANCHORED] — fitted to a real measurement in the Brillinger et al. (2025) open CNC dataset.
• [LITERATURE] — taken from published sources cited in the paper.
• [ASSUMPTION] — engineering estimate, exercised in the sensitivity analysis.

The two [ANCHORED] values — the 0.80–0.99 kW no-load spindle range and the regenerative-braking behaviour — were extracted by the author from the raw Brillinger dataset; the extraction procedure is described in the paper's Methods (§3.2). The Brillinger dataset itself is not redistributed here: it is available from Mendeley Data (DOI 10.17632/gtvvwmz7r7.2, CC BY-NC) and must be downloaded separately to re-derive the anchored values.

---

Citation

If you use this code or data, please cite the paper and the archived repository:

@article{yanytska2026detection,
  title   = {Detection limits of MES-embedded carbon-intensity monitoring for energy anomalies:
             a calibrated simulation study in machining-style processes},
  author  = {Yanytska, Lesia},
  journal = {The International Journal of Advanced Manufacturing Technology},
  year    = {2026},
  note    = {In press; manuscript JAMT-D-26-03924}
}

@software{yanytska2026code,
  author    = {Yanytska, Lesia},
  title     = {Detection limits of MES-embedded carbon-intensity monitoring: simulation code and data},
  year      = {2026},
  publisher = {Zenodo},
  version   = {v2.2.5},
  doi       = {10.5281/zenodo.21268863},
  url       = {https://doi.org/10.5281/zenodo.21268863}
}

This release is permanently archived on Zenodo: DOI 10.5281/zenodo.21268863. A machine-readable CITATION.cff is included at the repository root.

---

Licence

• Code (simulation/, plot_paper_figs.py): MIT — see LICENSE.
• Data (data/): CC BY 4.0 — free to reuse with attribution.
• The upstream Brillinger et al. (2025) dataset, from which the anchored spindle values were derived, is CC BY-NC; commercial reuse of those specific anchored values must comply with the upstream licence.

---

Contact

Lesia Yanytska — lesiayanytska@gmail.com

Issues and pull requests are welcome.