Mass balance catches the simulated leaks.
Broader calibration detected 36/36 injected events. The check uses the same known volumes and gas properties as the simulator. Its 0.245 false alarms/hour still exceeds the 0.1 calibration target.
FROM A CNN–LSTM BASELINE TO TESTABLE CHECKS
Test the pipeline.
Compare the models.
Change the operating conditions, follow the sensor readings, and see where each detector helps — and where it fails.
RESEARCH PROTOTYPE
These models were trained on a three-volume gas simulator. Field accuracy and leak localization have not been established.
01 / SET UP
Your experiment
OR / BRING SENSOR READINGS
Use the eight-feature SI schema. Uploads are scored against the simulation reference and kept in memory for at most one hour; export the result to keep it.
32–1,200 rows · 5 s cadence · no label columns used02 / OBSERVE
Known simulated layout
Measured pressure & flow
Loading an experiment generated by the actual inference engine…
00:0060:00
One hour of simulated operation · readings every 5 sReplay, not a live field feed
Pressure at each volume
ABSOLUTE / MPa A B CDashed line: injected leak onset
03 / COMPARE
Five checks, the same readings
Alerts require three consecutive scores above threshold.
The D9 context refiner needs 60 readings; the original four need 32.
Evidence over time
Score divided by the frozen threshold. A ratio above 1 is an exceedance, not a probability. Ratios below 0.01 sit at the plot floor.
04 / INSPECT
Keep the experiment
Export every reading, score and alert, including healthy periods and missing sensors. Ground truth is marked separately.
PAPERS → ALGORITHMS → MEASURED RESULTS
Test the methods from the papers.
Dr. Dankers' network papers ask whether measurements can reveal the particular physical connection we care about. His acoustic work asks how to separate sources that arrive mixed together. These experiments implement finite versions of those methods; the gas detector also includes an adapted D9 context-refinement framework.
The network and five-channel acoustic layouts below are controlled experiments. Their results do not establish field pipeline localization.
D2 / INPUT SELECTION + TWO-STAGE IDENTIFICATION
Estimate a connection despite shared noise.
Imagine two sensors both disturbed by the same vibration. Ordinary regression can mistake that shared disturbance for a strong connection. Two-stage identification first extracts the response to independent external test signals, then estimates the connection from those projected measurements.
D4 + D6 / TARGET IDENTIFIABILITY + DATA INFORMATIVITY
The right test signal in the right place.
You may be able to identify one connection without identifying every connection. But adding more data at the wrong input does not create the missing information. This reproduces the D6 input network and D4 target-rank example.
D8 / DELAYED SPATIAL FIR SOURCE ESTIMATION
Separate delayed echoes across channels.
A sound reaches several sensors with delays and reflections. A local FIR filter learns those delayed relationships from healthy measurements. The frozen filter estimates sources in subsequent measurements. A separate healthy segment checks remaining cross-correlation; passing that screen does not prove independence.
Analyze your synchronized acoustic channels
CSV headers must be m1,m2,… with 2–8 simultaneous channels and 768–4,096 rows. All rows must be finite and evenly sampled. Mark an initial segment known to be healthy: its first half fits the filter, its second half verifies it, and only later samples are monitored. Row index is used; no physical distance or sampling rate is inferred.
EXECUTED BENCHMARKS / INCLUDING FAILURE CASES
What the tests actually show
D9 / CONTEXT REFINEMENT IN THE GAS DETECTOR
Different normal behavior, different refiners.
The new fifth detector summarizes the previous 30 command readings, routes them to the nearest learned context, and refines a base forecast. DBSCAN centers, balanced assignment and augmented forecasts follow D9; ridge forecasting replaces its CatBoost models. Every context uses past readings.
RECORDED GAS SOUNDS / CONTROLLED MIXING TRANSFER
684 acquired GPLA-12 waveforms.
FROZEN RESULTS / INDEPENDENT SIMULATED RUNS
Calibration changes the story.
After the initial detectors produced too many alerts under command changes, healthy calibration coverage was expanded. The weights stayed frozen. The new thresholds were evaluated on 66 fresh runs, including 36 injected leaks.
36injected leak events
28.58 hscored healthy test exposure
51.675 hhealthy calibration exposure
Same fresh test runs · neural seed 11
| Method | New leak alarms | Event recall · 95% CI | False alarms / healthy hour | Median detected delay | Sensor drift cases alerted |
|---|
Sensor drift needs its own check.
Mass balance alerted on 0/6 pressure-drift cases; the dynamic and hybrid checks alerted on 6/6. A single detector should not replace sensor-quality monitoring.
A stricter threshold can hide leaks.
The CNN–LSTM's low false-alarm rate came with 0/36 detected events after broader calibration. The hybrid did not improve event recall over the dynamic predictor in this follow-up.
Only a new confirmed alarm at or after the injected onset earns detection credit. A pre-existing false alert receives no credit. Delays describe detected events only; missed events remain in recall. Confidence intervals are conditional on this simulator and event set.
SEPARATE DATA MODALITY
A small real gas-acoustic pilot.
150 unique one-second recordings were acquired. The provisional prefix split gives 30 test clips: 20 leak clips and 10 healthy clips. This is acoustic data, not pressure/flow SCADA. Acquisition sessions are unknown, and the recordings are not redistributed here.
Clip-level results do not establish field event delays, alarms per hour, independent acquisition sessions, or blind source separation. The single-channel pilot cannot identify leak location.
Review the complete record
INSPIRED BY THE WORK OF DR. ARNE DANKERS & COLLABORATORS
Algorithms with explicit assumptions.
Model the connected process.
The paper lab implements D2 graph-preserving input selection and external-reference two-stage identification. Shared-disturbance experiments compare bias against ordinary regression. The gas ARX remains a separate prediction baseline.
D1 · D2 · D5D2 is specialized to a declared one-tap network with independent external references. Realized feedback commands are not automatically instruments.Separate detection from attribution.
The paper lab reproduces D4 target rank/path conditions and D6 disconnecting-set excitation placement. Missing graph paths or weak external excitation withhold network attribution. Gas leak location remains unestablished.
D3 · D4 · D6Generic graph results assume known topology and independently parameterized modules. Finite frequency/SVD checks are separately labelled.Test context and the deployment boundary.
The fifth gas detector implements adapted D9 DBSCAN contexts, balanced assignment, augmented forecasts and L1 routing using only past commands. Missing sensors and gaps restart warm-up; tests check prefix invariance.
D9 · D10Ridge replaces CatBoost and numeric commands replace text/outage context. Matched-data clustering superiority is not established. D10 VMD is not implemented.Earn the acoustic extension.
D8 delayed spatial FIR filtering runs on synchronized simulated channels and uploaded multichannel CSVs. All 684 acquired GPLA waveforms are tested under controlled mixing with recorded machinery interference.
D7 · D8D8 finite specialization is implemented. Original field validation and D7 full sparse optimization are not reproduced; correlation cannot certify independence.DATA ACCESS / WHAT WE ACTUALLY HAVE
A qualified data plan
| Source | Fit for PipeGuard | Access & limits |
|---|---|---|
| Three-volume gas simulatorUsed in this online lab | Multichannel pressure, flow, temperature, commands; repeatable leaks and sensor faults. | Available and reproducible. Isothermal, quasi-steady surrogate; no wave, thermal, multiphase or field validation. |
| NGPOD / MKTCN ↗Candidate gas SCADA data | 17.95 km pipeline; four stations; 20 s data. Much closer to the actual pressure/flow task. | Raw arrays were absent from the linked repository when audited. Not acquired. Original records and incident/maintenance metadata are needed. |
| Gas acoustic recordings ↗150 clips acquired | Actual gas-leak sounds for a separate acoustic pilot; 96 kHz mono audio. | Small provisional subset; session independence and redistribution rights unclear. Audio is not bundled. |
| GPLA-12 gas acoustics ↗684 waveform rows acquired | Recorded source waveforms for controlled acoustic mixing transfer; separate from gas SCADA. | v1 CSV release hashed; v3 is a placeholder. No healthy class, verified session independence or released synchronization. Used for waveform recovery, not field leak accuracy. |
| OLGA multiphase study ↗Access request required | Useful for testing liquid/gas multiphase behavior beyond this surrogate. | Not acquired; simulated data with provider/licensing constraints. Not an open field validation set. |
READING RECORD
Ten papers behind the research
FROZEN ARTIFACTS
Weights you can inspect. Experiments you can repeat.
Actual trained checkpoints are exported to ONNX for CPU inference. Checksums, seed, thresholds and export parity checks are available through the API; training stays outside web requests.