Our first bridge stage loaded in eighteen seconds and proved nothing. It was a clean USD scene with no idea which girder was which. The second version took a month and could answer a load question. That is the difference between a render and a model.
What goes in
Four inputs. Plan sets and precast concrete shop drawings as PDF and CAD. Structural analysis models with section and prestress data. Inspection exports pulled from the state database of bridge elements. Sensor CSV files from two instrumented bridges in Illinois. The target is one USD stage per bridge that a bridge engineering team can open and query in Omniverse without asking me what anything means.
Stage layout
Three layers. Geometry referenced from converted CAD, a condition layer for inspection attributes, and a sensor layer for time-series references. Splitting them meant a condition update never forced a geometry rebuild, and replaying a sensor day never touched the structural data.
Name elements first
Element IDs follow the state's inspection numbering where possible, with span and girder appended. BIM conventions cover the geometry side. The two schemes disagree at the edges, so a mapping table is required, reviewed, and versioned like code.
Sensor ingestion
Both bridges carry a small set of sensors, and their channels arrive at different rates with different clocks. We resample onto a fixed interval, convert units at the edge, and keep the raw file untouched. The mapping from channel name to prim path lives in one CSV that both the Python ingest and the analysis code read.
Interpolation is the easy part. Clock drift between the data logger and the analysis clock is not. We keep logger timestamps, record the offsets, and never silently shift a series to make two curves line up.
Inspection data as attributes
Condition state is an attribute, not geometry. Each finding carries a date, an element, a severity, and a photo reference. The table mapping inspection elements to prims is short, and it is the piece most likely to be wrong, so a bridge inspector reviews it before a developer does.
One habit paid off early: store the photo reference, not the photo itself. Photo storage moves every few years, and a stable reference plus a manifest survives the move without anyone re-linking a thousand images.
Load cases
The structural side is a separate solver. We export the FEA model, run load cases including legal and permit vehicles plus wind and thermal gradients, and return results as deformation and utilization attributes on the same prims. One selection then shows inspection history, sensor trace, and the governing load rating for a single girder.
A data pipeline you can rerun
The data pipeline is idempotent. Same inputs, same stage. Every run writes a manifest with file hashes, timestamps, and the solver version that produced each load result. That sounds fussy until someone asks why a number changed between Tuesday and Friday, and nobody has to guess.
Where GenAnywhere sits
Upstream, in option studies, I use GenAnywhere: outline a search area, and the AI returns a massing model for comparing bridge widths and alignments before committing to a structural workflow. It produces volumes rather than elements, so it belongs before the USD stage and not inside it.
A massing model cannot tell a girder from a parapet, and at the option stage that does not matter. The risk is somebody forgetting to move from volumes to elements, so the massing output is kept as a separate file that never enters the bridge stage.
Notes from doing it twice
Fix the element ID scheme before CAD cleanup. Keep raw sensor files forever. Store the analysis model version next to the stage version. Treat the mapping tables as reviewed code. Treat sensor data as data, not as decoration for a demo.
The second bridge went faster because the mapping table already existed and we had stopped arguing about names. That is the real return on the first one, and it is the reason the pipeline matters more than any single model.
Two bridges is a pilot, and the part worth keeping is a digital twin pipeline that someone else can rerun next year against the same inputs and the same public infrastructure, without a phone call to me.
That is also the honest version of structural health monitoring: one more place where measurement and model have to agree in public, on a date, with numbers.


Top comments (0)