A reproducible demonstration of detection, prioritization, and repair verification across 120 assets and six inspection campaigns.
Prepared by: Hadi Farahani, PhD
Demonstration project — synthetic operational data. This report demonstrates KANON's advanced applied industrial analytics methodology, validation architecture, and decision-support capabilities. It does not represent measured performance from an actual industrial facility.
Executive Intelligence Brief
v1.0 | SYNTHETIC
208.98
Total Modelled Mass (kg CH4/h)
4.59%
Invisible Mass (Modelled %)
6/6
Planted Events (54-100% CI)
85.7%
May Repair Effect (Efficiency)
Prioritize the six highest-ranked assets (ME-081, ME-082, ME-041, ME-002, ME-001, ME-042) for immediate verification.
What this analysis does NOT establish: field-validated detection accuracy, verified abatement, or regulatory compliance.
KANON makes industrial measurement data reliable enough to support important decisions.
Q1 · Where are we leaking — and should we worry?
If we only have limited inspection capacity, where should we look first — and how confident can we be that we're right?
Decision Dashboard
Targeted weekly inspection of 6 priority assets yields the highest ROI.
208.98
Total Modelled Mass (kg/h)
4.59%
Invisible Mass (Modelled)
6/6
Planted Events Detected
336h
Median Detection Latency
Contents
⚠️
THE PROBLEM
Methane leaks are invisible, odorless, and financially material.
Operators must prioritize limited inspections despite imperfect detection.
A single missed high-rate source can dominate the annual emissions inventory.
Unmonitored leaks = unbudgeted liabilities.
⚙️
THE SOLUTION
A reproducible DuckDB analytical pipeline over six monthly inspection campaigns.
Priority ranking with quantified uncertainty bounds and confidence tiers.
Validation at event level, duration level, and out-of-sample.
Targeted inspection based on data, not guesswork.
✅
THE RESULT
All 6 planted leaks detected — recall 100%.
Only 4.59% of methane mass is invisible to the survey.
May repair reduced detected rate by 86% through efficiency, not activity.
Defensible, reproducible, and audit-ready.
84.6%
The Discovery: Detection Count is not Detection Mass.
84.6% of inspection events resulted in a nondetection. However, because those events were overwhelmingly minor baseline leaks, they accounted for only 4.59% of the total modelled methane mass. The survey is highly effective at catching the mass-dominant leaks, even if it "misses" the small ones.
Key Definitions & Abbreviations LOD (Limit of Detection): The minimum source rate the sensor can reliably see. IQR / MAD: Interquartile Range and Median Absolute Deviation (robust statistical bounds). GUM: Guide to the Expression of Uncertainty in Measurement (ISO standard). DuckDB & Polars: In-process analytical database and high-speed DataFrame library used to process the 21,720 latent-rate records in milliseconds.
2 · Data Architecture
01
Industrial Data Sources
SCADA · Sensors · Throughput · Operational records
Q2 · How does the method separate size from efficiency?
Q: What data do we analyze?
A: 120 assets (40 valves, 40 flanges, 40 compressors), 720 inspections across six monthly campaigns, 21,720 daily latent-rate records, and 6 planted leak events with repair on May 1.
Q: What is the detection method?
A: Sensor baseline with IQR/MAD anomaly detection. Detection probability follows a logistic curve: p = 1 − exp(−ln(2) · rate / scale), where scale grows with wind. Quantification error is log-normal with CV 0.25.
Q: What tools did we use?
A: The pipeline was built for speed and auditability. We used DuckDB as the analytical engine to join the truth and analyst schemas without loading everything into memory, and Polars for high-speed DataFrame manipulation. The entire 21,720-row dataset processes in under 2 seconds.
Q: What is the decomposition?
A: Total emissions = Activity × Emission Factor × (1 − Control Efficiency). Period-over-period change is split into an activity effect (which assets were inspected) and an efficiency effect (how much they leaked).
Figure 1 — Detection rate by asset type and wind condition. Flanges are systematically under-detected (4.5–5.0%) because their true rates are smallest.
4 · Validation
Q3 · Does the algorithm find real events — or just noise?
Q: Did the method find the planted leaks?
A: Yes — recall 100% (6/6). However, with only 6 events, the 95% Clopper-Pearson lower bound is 54%. We have high confidence, but not absolute certainty.
Q: Why is precision 19.4%?
A: Precision against planted-event truth is 19.4%. However, the 25 apparent false positives are correctly detected baseline leaks. Precision against physical reality is 100%. We report both: the former tests the algorithm's alignment with the event catalog; the latter tests its real-world detection capability.
Q: How fast do we detect a leak?
A: 336 hours (14 days). This is a cadence artifact, not a method limitation. Because all events start on March 1 and the first inspection is March 15, the latency is fixed. A weekly cadence would reduce the expected detection wait to 3.5 days.
Figure 2 — Validation summary.
Event
Asset
Start
First detected
Latency
LEAK-01
ME-001
2025-03-01 12:00:00
2025-03-15 12:00:00
336 h
LEAK-02
ME-002
2025-03-01 12:00:00
2025-03-15 12:00:00
336 h
LEAK-03
ME-041
2025-03-01 12:00:00
2025-03-15 12:00:00
336 h
LEAK-04
ME-042
2025-03-01 12:00:00
2025-03-15 12:00:00
336 h
LEAK-05
ME-081
2025-03-01 12:00:00
2025-03-15 12:00:00
336 h
LEAK-06
ME-082
2025-03-01 12:00:00
2025-03-15 12:00:00
336 h
5 · The Business Case
Q4 · What does a missed leak actually cost?
THE REALITY CHECK (MODELLED)
A missed high-rate leak (30 kg CH4/h) equals ~0.72 tonnes/day. At $2/MMBtu (approx. $75/day in direct gas value), the immediate loss is modest. However, under a $50/t CO2e carbon price (GWP100 = 28), the penalty is $1,000+ per day. Adding regulatory fines and safety risks brings the blended cost to ~$2,500/day.
The ROI calculation (The "Targeted" Strategy)
Broad weekly inspection (all 120 assets): Costs $144,000 per extra campaign, avoids $26,250 per event. Requires 5.5 prevented events to break even. Not recommended. Targeted weekly inspection (Top 6 assets): Costs ~$93,600 per extra campaign, avoids ~$157,500 across 6 potential events. Best-case ROI: 1.7×. Requires at least 3 of the 6 assets to actually leak to justify the cost.
Figure 3b shows the threshold. At a $2,500/day blended loss, an asset must have a 59% chance of leaking to justify weekly inspection. If the loss is $10,000/day, the threshold drops to 15%.
Figure 3b — Break-even probability. Weekly inspection pays back only if the asset has a high enough chance of leaking.
Does 17× cadence mean 17× detection?
No. Figure 3c shows that detection probability saturates. At ~15% per-visit detection, monthly catches 48%, weekly catches 94%. Faster cadence buys shorter latency, not infinite detection.
Figure 3c — Detection calibration. Probability saturates as cadence increases.
6 · Limitations
Q5 · What could make this conclusion wrong?
Blind Spots Register — six substantive, quantified limitations.
L1 — Sub-threshold leaks are invisible. Only 4.59% of methane mass is invisible, but 84.6% of inspection events are nondetections. Count and mass tell different stories.
L2 — Flanges are systematically under-detected. 4.5–5.0% detection rate vs. 27–28% for compressors. A field pilot would need to recalibrate the flange detection curve.
L3 — Valves carry the highest invisible fraction. 8.6–13.1% of valve mass is missed, vs. 0.7–3.3% for flanges. Blind spots concentrate in a specific asset class.
L4 — No false positives are modeled. The V1 model has false_positive_probability = 0.0. The reported 173.6 false alarms per 1,000 hours uses calendar-days, not asset-hours; the corrected metric is ~0.048 per 1,000 asset-hours.
L5 — Even 97.8%-probability sources can be missed. ME-001 on May 15 had detection probability 0.9783 and was not detected (uniform draw 0.9851). A single missed inspection is not proof of absence.
L6 — Nondetection is not zero. The blueprint states this explicitly; the data confirms it. A missed inspection is not a zero-emission observation.
Figure 4 — Top 10 invisible methane sources. These are real emissions the survey did not see.
7 · Reproducibility
Q6 · Can a client re-run this analysis and verify the numbers?
Q: Can a client re-run the analysis?
A: Yes. The generator is seeded (seed 1003), the DuckDB database is built from six CSVs with explicit column mapping, and the six-step analytical chain is documented in reproducibility_chain.csv.
Q: Can a client verify the numbers?
A: Yes. Every source file and every output CSV has a SHA-256 hash in reproducibility_manifest.csv. The master hash of all outputs appears below.
MASTER HASH (SHA-256 of all analytical outputs): a6b91c19ec2e8413e3b0f592cf18c1cc1388bb5e1ca953ea013cc277b88316cf
Source Data
9
Files fingerprinted
Analytical Scripts
6
Steps documented
Output Artifacts
17
CSVs generated
Audit Status
100%
Verified
Full Manifest: The complete SHA-256 hashes for every file, along with the one-command execution chain, are available in the downloadable evidence package (reproducibility_manifest.csv).
8 · Statement of Integrity
Statement of Method Integrity. I certify that the analytical methods, synthetic generation, and metrics presented in this report are accurate to the best of my knowledge. The limitations and uncertainties of this synthetic demonstration have been explicitly disclosed in Section 6. The pipeline, data, and outputs are fully auditable via the SHA-256 hashes in Section 7.
Conflict of Interest. This project was conducted independently as a methodological demonstration. No funding or analysis was influenced by equipment manufacturers or operators. This report does not claim regulatory compliance or field-validated performance.
9 · Decision Implications
Q7 · What should we do Monday morning?
Q: What should we do Monday morning?
A: Inspect the top 6 assets first: ME-081, ME-082, ME-041, ME-002, ME-001, ME-042. They carry 6–20× the priority of the next tier, and the ranking is robust to distance-weighting assumptions.
Q: How often should we inspect?
A: Weekly campaigns give 17 chances per event; monthly give 4; quarterly give 1. The detection risk changes by 17× with cadence. Cadence is a budget decision, not a technical one.
Q: Did the May repair work?
A: Yes. 86% of the May decline is efficiency, not activity. The same assets were inspected; they leaked less.
Q: How much methane are we still missing?
A: 4.59% of total mass. That is the honest number.
Q: What would change our answer?
A: A 20% wind-model error changes mean detection probability by ±3%. A 20% instrument-sensitivity error changes it by ±8–9%. Both are bounded; the recommendations are robust.
Figure 5 — Top 10 assets by median priority, with 95% uncertainty bounds.Figure 6 — Activity vs. efficiency decomposition. May repair is 85.7% efficiency.
Figure 7 — Detection latency, all events at 336 hours (14 days).
10 · References
IEA (2025) - Global Methane Tracker 2025. International Energy Agency, Paris.
European Commission (2024) - Regulation (EU) 2024/1787 on the reduction of methane emissions in the energy sector.
OGMP 2.0 (2025) - Oil and Gas Methane Partnership 2.0 Reporting Framework. UNEP.
US EPA (2024) - Standards of Performance for New, Reconstructed, and Modified Sources and Emissions Guidelines for Existing Sources: Oil and Natural Gas Sector Climate Review.
JCGM 100:2008 - Evaluation of measurement data: Guide to the expression of uncertainty in measurement (GUM).
ISO 8000-61:2016 - Data quality - Process reference model.
Founded and led by Hadi Farahani, PhD. Combining measurement science, analytical chemistry, and advanced data analytics to solve complex industrial data challenges.