Infrastructure Asset Risk and Investment Decision Support System
A solo capstone project that scores and ranks 6,554 Victorian road bridges by risk, to support maintenance and investment prioritisation decisions — the kind of analysis a road authority or infrastructure consultancy would use to decide where limited maintenance budget goes first.
Live demo: https://infrapulse.streamlit.app/ 📖 Read the full INFRAPULSE Report
KPI summary and top-20 ranked table
Risk map, colored by tier
Victoria's own 2011 Auditor-General review found that maintenance funding for road bridges was running at roughly half of annual depreciation, even as over half the state-managed bridge network was already inside the 30–60-year window where major repairs are typically needed. Prioritisation decisions have to be made with imperfect information, on a network of thousands of structures, by a small number of engineers. InfraPulse is a working model of the kind of tool that supports that decision — score every bridge consistently, surface the highest-risk ones, and let filtering narrow the view by region or road class.
- Risk scoring for 6,554 bridges — a transparent, rule-based multi-criteria formula (not a black-box model), so every score can be explained.
- Interactive risk map — colored by risk tier, colorblind-safe palette with text labels, not color alone.
- Ranked table — top-risk bridges surfaced first, filterable.
- Live filters — region, road/state class, risk tier, inspection priority, and flood history.
- Inspection Priority — maps risk score to standard 3-tier inspection cadences.
- Interactive Sandbox — dynamically adjust likelihood weights (Age, Traffic, Climate) to see real-time what-if impacts on the risk scores.
- PDF Export — generate and download static reports for the current filtered region view or individual bridge profiles.
- In-app methodology page — documents data sources, the scoring formula, and — deliberately — the data-quality failures that shaped the design.
infrapulse/
├── data/
│ └── processed/
│ └── bridges_risk_scored.csv # pipeline output; the app's only data source
├── pipeline/ # batch scoring pipeline, run independently of the app
│ ├── run_pipeline.py
│ ├── features_weather.py # Open-Meteo historical weather, grid-deduplicated
│ ├── features_traffic.py # TIRTL sensor data, spatial-joined to nearest bridge
│ ├── features_condition.py # condition-data sourcing attempts (see Limitations)
│ └── risk_scoring.py # the likelihood/consequence formula
├── assets/screenshots/ # README images
├── pages/
│ ├── home.py # overview + headline stats
│ ├── dashboard.py # filters, KPIs, map, ranked table
│ └── methodology.py # full data-source and scoring writeup
├── docs/
│ ├── infrapulse_vic_briefing.md # research briefing backing the methodology
│ └── build-prompts/ # Antigravity build-context prompts, archived
├── streamlit_app.py # entrypoint, defines navigation
├── data_loader.py # single source of truth for loading + risk tiers
├── .streamlit/config.toml # theme
└── requirements.txt
The architecture is deliberately minimal: a batch pipeline writes a flat CSV, and a read-only Streamlit dashboard consumes it. No cloud dependency, no authentication, and no live API calls from the app itself — every number on screen traces back to a pre-computed file, not a runtime request.
- Bridge registry — Transport Victoria open data portal, 6,554 bridges after cleaning, including a fix for 61 construction-year values that had been misclassified as unknown due to inconsistent date formats.
- Weather — Open-Meteo Historical Weather API, grid-deduplicated at 0.25° resolution.
- Traffic — TIRTL 15-minute sensor data from 406 Melbourne freeway sites, Austroads vehicle classes 3–12 treated as heavy vehicles, spatially joined to the nearest bridge.
- Condition data — attempted via three independent sources; all three proved unusable (see Known limitations).
A transparent, rule-based multi-criteria score (0–100, min-max rescaled) — explicitly not a trained predictive model:
Likelihood = age (45%) + traffic (30%) + climate (25%)
reweighted proportionally when a component is missing
The architecture is deliberately minimal: a batch pipeline writes a flat CSV,
and a read-only Streamlit dashboard consumes it. No cloud dependency, no
authentication, and no live API calls from the app itself — every number on
screen traces back to a pre-computed file, not a runtime request.
## Data sources
- **Bridge registry** — Transport Victoria open data portal, 6,554 bridges after
cleaning, including a fix for 61 construction-year values that had been
misclassified as unknown due to inconsistent date formats.
- **Weather** — Open-Meteo Historical Weather API, grid-deduplicated at 0.25°
resolution.
- **Traffic** — TIRTL 15-minute sensor data from 406 Melbourne freeway sites,
Austroads vehicle classes 3–12 treated as heavy vehicles, spatially joined to
the nearest bridge.
- **Condition data** — attempted via three independent sources; all three
proved unusable (see *Known limitations*).
## Risk scoring methodology
A transparent, rule-based multi-criteria score (0–100, min-max rescaled) —
explicitly **not** a trained predictive model:
Likelihood = age (45%) + traffic (30%) + climate (25%) reweighted proportionally when a component is missing
Consequence = multiplier from road/state class (CD_STATE_CLASS) HF = 1.5 MR = 1.2 TR = 1.1 RA / FR / unknown = 1.0
Risk score = Likelihood × Consequence, rescaled to 0–100
The weighting is grounded in the sector's own findings, not chosen arbitrarily
— the 45% age weight mirrors the same 30–60-year repair-window statistic
Victoria's Auditor-General uses as a headline risk indicator, and the climate
feature specifically targets extreme-rainfall days rather than total rainfall,
because average rainfall has been *declining* in Victoria even as extreme
rainfall intensity rises. Full reasoning, with citations, is on the in-app
**Methodology** page and in `docs/infrapulse_vic_briefing.md`.
## Known limitations
The original design planned to use physical condition ratings as a model
input. That plan was abandoned after three separate sourcing attempts failed:
the legacy VicRoads "Bridge Structures" dataset was confirmed dead, the Vicmap
Transport ArcGIS FeatureServer's condition field returned near-zero variance,
and an AURIN portal download intended for Victoria returned Western
Australia's data instead. Victoria's own 2011 Auditor-General audit found
condition-rating data inconsistent and unreliable sector-wide, independent of
this project — which is the basis for treating the pivot to rule-based scoring
as a methodologically sound choice, not a fallback. The `physical_condition`
field is retained in the dataset for reference only and does not feed the
score.
## Future Expansion: AI-Driven Data Centre Intelligence
Expand InfraPulse beyond public infrastructure to support data centre asset management and strategic planning. The platform would monitor key operational metrics such as Power Usage Effectiveness (PUE), energy consumption, cooling efficiency, carbon emissions, and equipment health through real-time analytics and predictive AI.
By integrating electricity pricing, renewable energy availability, climate conditions, fibre connectivity, land availability, and demand forecasts, InfraPulse could also recommend optimal locations for future data centre developments, enabling organisations to improve operational efficiency while supporting sustainable digital infrastructure.
Consequence = multiplier from road/state class (CD_STATE_CLASS)
HF = 1.5 MR = 1.2 TR = 1.1 RA / FR / unknown = 1.0
Risk score = Likelihood × Consequence, rescaled to 0–100
The weighting is grounded in the sector's own findings, not chosen arbitrarily
— the 45% age weight mirrors the same 30–60-year repair-window statistic
Victoria's Auditor-General uses as a headline risk indicator, and the climate
feature specifically targets extreme-rainfall days rather than total rainfall,
because average rainfall has been declining in Victoria even as extreme
rainfall intensity rises. Full reasoning, with citations, is on the in-app
Methodology page and in docs/infrapulse_vic_briefing.md.
The original design planned to use physical condition ratings as a model
input. That plan was abandoned after three separate sourcing attempts failed:
the legacy VicRoads "Bridge Structures" dataset was confirmed dead, the Vicmap
Transport ArcGIS FeatureServer's condition field returned near-zero variance,
and an AURIN portal download intended for Victoria returned Western
Australia's data instead. Victoria's own 2011 Auditor-General audit found
condition-rating data inconsistent and unreliable sector-wide, independent of
this project — which is the basis for treating the pivot to rule-based scoring
as a methodologically sound choice, not a fallback. The physical_condition
field is retained in the dataset for reference only and does not feed the
score.
Python, Pandas, Streamlit, Folium, Scikit-learn (pipeline utilities). Built
with assistance from Claude and Antigravity; all data-quality findings and
scoring decisions verified against primary/secondary sources listed in
docs/infrapulse_vic_briefing.md.
git clone https://github.com/AJ-ing/InfraPulse
cd infrapulse
pip install -r requirements.txt
streamlit run streamlit_app.pyDeployed on Streamlit Community Cloud, connected
directly to this GitHub repository. Any push to main redeploys automatically.
Note: data/processed/bridges_risk_scored.csv must be committed to the repo —
the app has no live database, so this file is its only data source.
This project publishes coordinates and risk rankings for real public infrastructure. Before treating any deployed instance as a public-facing production tool, confirm Transport Victoria's open data licence terms permit this kind of republication.
Victorian Auditor-General's Office (2011), Management of Road Bridges, and
related sources cited in full in docs/infrapulse_vic_briefing.md.
Aayush Jain