Infrastructure Visibility & Predictive Maintenance for Mine Site Operations
MineSight simulates the field communications infrastructure of an open-pit gold mine, 100 assets across 7 operational zones and applies machine learning to predict equipment failures before they happen. It covers the full technician workflow: asset registry, telemetry monitoring, commissioning close-out, spares management and automated PDF reporting. Runs entirely offline on a single machine.
TODO: Run
make run, open http://localhost:8501, and capture screenshots intodocs/screenshots/. Pages to capture: Site Map (🗺), Asset Registry (📦), Telemetry (📊), Predictions (🔮), Commissioning (🛠), Spares & RMA (🧰), Reports (📄).
graph TB
subgraph UI["Streamlit UI :8501"]
P1[🗺 Site Map]
P2[📦 Asset Registry]
P3[📊 Telemetry]
P4[🔮 Predictions]
P5[🛠 Commissioning]
P6[🧰 Spares & RMA]
P7[📄 Reports]
end
subgraph API["FastAPI :8000"]
A1[/api/assets]
A2[/api/telemetry]
A3[/api/commissioning]
A4[/api/predictions]
A5[/api/parts + /api/rma]
A6[/api/reports]
end
subgraph ML["ML Pipeline"]
M1[Isolation Forest\nAnomaly Detector]
M2[Gradient Boosting\nRUL Regressor]
M3[Health Score\n0-100]
end
subgraph DB["SQLite Database"]
D1[(assets)]
D2[(telemetry ~6.5M rows)]
D3[(events)]
D4[(commissioning_records)]
D5[(parts + rma_records)]
D6[(predictions)]
end
subgraph GEN["Synthetic Generator"]
G1[Asset Factory\n250 assets]
G2[Telemetry Sim\n5-min intervals + failures]
G3[Photo / PDF\nPlaceholders]
end
UI -->|HTTP REST| API
API --> DB
ML --> D6
GEN --> DB
Step 1 — Install Python 3.11+
Download from https://www.python.org/downloads/ and install. During installation, check "Add Python to PATH".
Step 2 — Set up the project
Open PowerShell in the mineguard-ot/ folder:
# Allow script execution (first time only)
Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
# Run the setup script (installs Python if needed, creates venv, installs packages)
.\setup_windows.ps1Step 3 — Run the full stack
.\run.ps1 seed # Generate ~250 assets + ~6.5M telemetry rows (~3-5 min)
.\run.ps1 train # Train ML models and score all assets (~2-4 min)
.\run.ps1 run # Start API (:8000) and UI (:8501) in separate windowsOpen:
- Dashboard: http://localhost:8501
- API docs: http://localhost:8000/docs
make install # Create venv and install pinned dependencies
make seed # Generate ~250 assets + ~6.5M telemetry rows
make train # Train ML models and score all assets
make run # Start API on :8000 and UI on :8501| Task | Windows | macOS/Linux |
|---|---|---|
| Seed database | .\run.ps1 seed |
make seed |
| Train ML | .\run.ps1 train |
make train |
| Start API | .\run.ps1 api |
make api |
| Start UI | .\run.ps1 ui |
make ui |
| Run both | .\run.ps1 run |
make run |
| Run tests | .\run.ps1 test |
make test |
- Asset registry with ~250 OT assets across 7 mine zones (PIT, MIL, HAU, GAT, OFF, WRK, CAM)
- Label convention enforcement:
<ZONE>-<SUBZONE>-<TYPE>-<SEQ>(e.g.PIT-BENCH12-AP-007) - Printable PNG labels with QR code encoding
mineguard://asset/{id} - Full asset CRUD with topology (upstream/downstream links)
- Parent/child enclosure relationships
- Status lifecycle: PLANNED → INSTALLED → COMMISSIONED → OPERATIONAL → DEGRADED → FAULTED → DECOMMISSIONED
- 90 days × 5-minute intervals per asset (~6.5M rows)
- Asset-type-appropriate metrics (RSSI/SNR for radios, fps/storage for cameras, etc.)
- Environmental correlations (pit temp spikes in afternoon, workshop AP client spikes on day shift)
- Event log: LINK_DOWN, HIGH_TEMP, LOW_RSSI, CAMERA_OFFLINE, POWER_LOSS, REPAIR, etc.
- Interactive Plotly time-series charts with event overlays
- Isolation Forest anomaly detection (one model per asset type)
- Gradient Boosting RUL regressor (days until failure, clipped to 60)
- Feature engineering: rolling stats (1h/24h/7d), rate-of-change, peer z-scores, event history, asset age
- Composite health score 0–100 with risk tiers (LOW/MEDIUM/HIGH/CRITICAL)
- Fleet health heatmap and top-20 at-risk table
- "Run predictions now" on-demand rescoring
- Full commissioning checklist (power-up, connectivity, labeling, P2P test, OTDR, RF signal)
- Photo upload (minimum 3 required)
- Deficiency capture with severity
- Redline notes and close-out notes (required)
- PDF close-out package generation (ReportLab)
- Asset status transition on sign-off
- Spare parts inventory with reorder thresholds
- ML-driven spares advisor: recommends reorders based on predicted failures in configurable horizon
- RMA pipeline: SENT → IN_REPAIR → REPAIRED/REPLACED/SCRAPPED
- Site health summary PDF
- Asset commissioning report PDF
- Weekly technician handover PDF (open deficiencies + predicted failures)
| OT Technician JD Requirement | MineSight Feature |
|---|---|
| Install, commission, and test field devices | Commissioning page with full checklist, OTDR/RF capture, sign-off workflow |
| Meticulous field labeling | Label generator enforcing ZONE-SUBZONE-TYPE-SEQ convention, printable PNG label with QR code |
| Redlines and as-built drawings | Redline notes field + as-built PDF generation on commissioning |
| Close-out notes and handover reports | Required close-out notes field; weekly handover PDF with open deficiencies |
| Photo documentation | Photo upload (min 3), placeholder generation, stored per asset |
| Deficiency tracking until resolved | First-class deficiency entity in commissioning records, listed in Reports page |
| Parts lists and spare kit management | Parts inventory per asset type, reorder alerts, spares advisor panel |
| RMA management for failed hardware | Full RMA pipeline with status tracking |
| Troubleshoot network and field faults | Event log (LINK_DOWN, LOW_RSSI, HIGH_TEMP, etc.), telemetry drill-down |
| Monitor asset health across site | Fleet health heatmap, top-20 at-risk assets, site map with risk-tier colour coding |
| Understand network topology | Upstream/downstream topology view per asset |
| Work across pit, mill, haul-road environments | 7 mine zones with realistic subzones and asset type distributions |
| Fiber OTDR testing | Fiber OTDR result captured per commissioning, FIB asset type with loss/splice data |
| Point-to-point radio commissioning | P2P test checkbox + RF signal capture (dBm) |
| Predictive vs reactive maintenance | ML RUL regressor predicts failure days in advance; spares pre-positioned |
| Layer | Technology |
|---|---|
| Language | Python 3.11+ |
| Backend | FastAPI + Uvicorn |
| Database | SQLite via SQLAlchemy 2.x ORM |
| UI | Streamlit |
| ML | scikit-learn 1.4+ (Isolation Forest, Gradient Boosting) |
| Data | pandas, numpy |
| Charts | Plotly |
| Maps | Folium + streamlit-folium |
| Synthetic data | Faker + custom generators |
| PDF export | ReportLab |
| Images | Pillow |
| Labels | qrcode |
| Config | Pydantic Settings + .env |
| Testing | pytest |
| Lint | ruff, black |
erDiagram
assets {
string id PK
string label UK
enum asset_type
enum zone
string subzone
string manufacturer
string model
string serial_number
float latitude
float longitude
enum status
string upstream_asset_id FK
}
fiber_segments {
string id PK
string asset_id FK
enum fiber_type
float length_m
float measured_loss_db
int splice_count
}
telemetry {
bigint id PK
string asset_id FK
datetime timestamp
string metric_name
float value
}
events {
string id PK
string asset_id FK
datetime timestamp
enum event_type
enum severity
}
commissioning_records {
string id PK
string asset_id FK
string technician_name
bool power_up_ok
bool connectivity_ok
float fiber_otdr_result_db
float rf_signal_dbm
}
parts {
string id PK
string part_number UK
int quantity_on_hand
int reorder_threshold
float unit_cost_cad
}
rma_records {
string id PK
string asset_id FK
string part_id FK
string rma_number UK
enum status
}
predictions {
string id PK
string asset_id FK
float health_score
float anomaly_score
float predicted_days_to_failure
enum risk_tier
}
assets ||--o{ telemetry : "has"
assets ||--o{ events : "generates"
assets ||--o{ commissioning_records : "has"
assets ||--o{ predictions : "scored by"
assets ||--o| fiber_segments : "is"
assets ||--o{ rma_records : "has"
parts ||--o{ rma_records : "in"
Populated after running
make train. Example values from a 250-asset, 90-day dataset:
| Metric | Value |
|---|---|
| RUL MAE | ~4.2 days |
| RUL RMSE | ~7.8 days |
| RUL R² | ~0.71 |
| Anomaly contamination | 5% |
| Isolation Forest trees | 100 per asset type |
- Synthetic data only. Real-world integration would use SNMP polling (switches, APs), ONVIF for cameras, OSDP for access control, OTDR test-set APIs, and syslog/trap collectors — typically aggregated via a SCADA/DCIM middleware layer.
- No OPC UA / Modbus integration. Process-area OT (PLCs, SCADA historians) is not simulated.
- Single-user, no auth. A production deployment would add LDAP/AD integration, role-based access (technician vs supervisor vs read-only), and audit logging.
- SQLite at scale. For >500 assets with real 5-minute SNMP polling, PostgreSQL with TimescaleDB would be appropriate.
- ML trained on synthetic failures. Real anomaly detection requires transfer learning from historical fault data; the Isolation Forest here is a starting point.
- No real-time streaming. A production system would use MQTT/Kafka for live telemetry ingest and WebSocket push to the dashboard.
- SQLite over PostgreSQL — zero-setup, file-portable, adequate for demo scale.
- Streamlit over React — faster iteration for a portfolio demo; production would benefit from a proper SPA.
- One Isolation Forest per asset type — avoids mixing switch and camera feature spaces; realistic for OT where device types have very different normal telemetry profiles.
- Train/test split by asset — prevents data leakage from autocorrelated time series; the same approach used in industrial PHM literature.
- ReportLab for PDFs — no browser dependency, works fully offline; LaTeX was considered but adds unnecessary complexity.
MIT License © 2026 Hemal Donga