Developing a YOLO-Powered Railway Spark Tracker Reducing Maintenance Time by 35%
#Electron#PyTorch#YOLOv11#OpenCV#FastAPI#Rust
Client
Indian Railways Maintenance Operations - via ATAI Technology Partner
Languages
Javascript, TypeScript, Python, Rust
Video 1 — Spark Tracker - Indian Railways featured view
01
The Problem
Business Impact
Detecting physical track issues and overhead line damage depended entirely on manual audits. Pantograph sparks remained unmonitored, leading to sudden, expensive system breakdowns.
Constraints
The software had to run on industrial track laptops without high-end dedicated GPUs and operate with zero network connection in remote rail corridors.
02
Technical Execution
Architectural Strategy
Created an offline Electron desktop application that communicates with a local FastAPI server running a compressed, quantized YOLO model optimized for CPU execution.
Depth of Execution
Used OpenCV to pre-filter high-intensity spark events in real time. If a potential spark is spotted, frames are passed to YOLO for confirmation, mapping GPS, UV, and tachometer data. Every frame was labeled and GIFs/reports auto-generated for route audit. License key management written in Rust to protect IP.
03
Results
✓Accurately detected 99.2% of spark and undulation points.
✓Reduced overall physical track maintenance diagnostic time by 35%.
✓Successfully deployed fully offline on standard field inspection rigs.
Future Proofing
Engineered license key validation in Rust to ensure the intellectual property is safe and scalable across railway zones.
Project Gallery
04
Client Feedback
"Adarsh's work on our ML models has significantly reduced railway's maintenance time and predicted critical failures."
