Jumeau numérique de suivi de barrage
Dam owner, one large embankment. Survey, geotechnical and telemetry records fused into a living twin with alerting and an inspection-ready evidence trail.
La situation.
Un propriétaire de barrage fusionnant dossiers de levés, données géotechniques et télémétrie pour un grand ouvrage en remblai.
PérimètreFusion des campagnes de levés, de l’instrumentation piézométrique et de suintement, et de la télémétrie structurelle en un modèle vivant unique de l’ouvrage, avec alertes à seuils et piste de preuves prête pour l’inspection.
Une vue opérationnelle unique remplaçant 5 feuilles de calcul, avec une piste de preuves prête pour l’inspection derrière chaque chiffre.
Three source families converge on the twin. Repeat laser surveys register geometry in 3D tiles and act as the spatial backbone. Piezometric and seepage instrumentation plus structural telemetry land in TimescaleDB as dated series with QC history. Datum lineage for each instrumentation generation is reconstructed before fusion, so a threshold evaluated anywhere in the record means one thing. Alerting evaluates thresholds per instrument and writes exceedances into the same evidence schema the series live in. As-of-date views combine the latest survey epoch up to any date with the instrumentation active then. The operations view renders structure and series together, and every number on it links back to its source record.
Envisagé : Normalising all three instrumentation generations to a single datum at load
Choisi : Reconstructing datum lineage per generation before fusing
Parce que : The generations used different datums, and one mapping applied across them would have made identical thresholds mean different things.
Envisagé : A latest-state model that always shows current conditions
Choisi : As-of-date views over the full record
Parce que : Dam safety questions are historical, and the correct answer to what did we know is the state known on that date, not the corrected present.
Envisagé : A bespoke 3D inspection platform
Choisi : 3D tiles
Parce que : Survey geometry had to render in ordinary browser sessions for inspection reviews, on hardware the dam owner already had.
Envisagé : A separate document store for inspection evidence
Choisi : Evidence bound to the same records that produce the numbers
Parce que : A second store means a second reconciliation, and the trail had to be inspection-ready without one.
TimescaleDB earns its place through dated series with QC history and interval scans across irregular epochs. 3D tiles carry repeat laser scans, the geometric backbone, at a cost a browser can pay. Alerting sits inside the same store as the series, so an exceedance and its context are one query apart.
The verified figures: five spreadsheets consolidated into a single operations view, and three instrumentation generations reconciled to one threshold meaning. That is the whole numeric record. No trend rates, exceedance counts or latency figures were certified on delivery, and none are stated here. The verified acceptance outcome is qualitative: an inspection-ready evidence trail stands behind every number on the operations view, the criterion being that any number resolves to its source record, its QC history and its datum lineage on request.
Discovery found the structure split across generations
A dam owner held survey campaigns, geotechnical instrumentation and structural telemetry for a large embankment. Nothing spoke to anything else. Three generations of instruments had been installed over the structure's life, each on its own datum, so a piezometric reading from one generation and a threshold set against another did not share a meaning. Survey epochs arrived with each mobilisation, on their own cadence. Five hand-maintained spreadsheets carried the operational picture, light on the evidence behind their numbers. And when an inspection asked a historical question, the spreadsheets showed today's understanding, against the knowledge of that date. The brief was one living model of the structure, with alerting and an evidence trail.
The hard problem was datum lineage, then time
Two problems ordered the work. Datums first. Fusing before resolving lineage would have produced a record where thresholds meant three things depending on instrument age, and a threshold that means three things is worse than none. Datum lineage was reconstructed per instrumentation generation from installation records and tie surveys, and only then did fusion run, so every series carries its lineage and thresholds evaluate against one meaning. Time second. Survey epochs arrived at irregular intervals, so a question about a decade ago could not be answered with today’s geometry. As-of-date views settled it: any historical question returns the state known at that date, latest epoch up to then, instruments active then, lineage intact. The corrected present stays available, but the historical answer is the one an inspection trusts.
The build fused geometry, series and evidence into one model
Repeat laser scans anchor the model. They come from Epcotec's geomatics practice, where scan differencing is routine, and they give the twin geometry that survives review: registered in 3D tiles, browser-rendered, comparable epoch to epoch. Piezometric, seepage and structural telemetry land as dated series in TimescaleDB with QC history attached, lineage-resolved before storage. Threshold alerting writes exceedances into the evidence schema, so an exceedance carries its context without a manual step. The operations view draws structure and series from the twin onto one screen.
Verification asked the twin the questions an inspector asks
We verified against inspection questions, not a demo script. A threshold reading had to resolve, on request, to its source series, its QC history and the datum lineage of its generation. A historical question had to return the as-of-date state, and the same question answered on two different days had to return two different states only when knowledge had actually changed between them. Survey epochs had to render against the instrumentation active at each date. Alerting had to flag departures and land them in the evidence schema. The engagement closed when the single operations view stood in place of the spreadsheets, with the evidence trail behind every number.
Handover, and what changed for the owner
Handover covered the twin, the alerting configuration and the as-of-date interface, with documentation, a walkthrough and 30 days of post-launch support per the studio's standard. What changed for the owner was the shape of an answer. A question about the embankment used to mean hand-assembly across records, hoping the generations compared fairly. Now the same question runs against the twin and carries its own evidence: source record, QC history, datum lineage, survey epoch. And because repeat laser scans anchor the geometry, the next survey campaign extends that record with every new epoch. The structure finally has a model that ages with it.
Un problème similaire ?
Périmètre fixe, ingénieurs seniors, staging en ligne dès la première semaine.
Réserver une revue de périmètre →