VSD, a.s. is an energy company, the main activity of which is electricity distribution to the end consumer through its own distribution system of almost 22,600 kilometers and 600,000 off-take points.
Summary
VSD aimed to develop an AI solution to extract and record data from fusebox photographs with high accuracy. We implemented modern classification and object detection models, achieving over 90% accuracy and 80% precision. The solution, utilizing InceptionV3 and YOLOv9 models, significantly exceeded expectations, analyzing 10,000 fuseboxes and delivering a 95.7% mAP success rate.
Challenge
Motivation to change
The goal of the project was the development of AI solution responsible for extraction of useful information from fusebox photographs and automatically recording it to an IT system.
Fusebox data are regularly being collected by employees with the purpose of documenting a box category, fuse status (ON/OFF), fuse amperic values and other data.
Solution
Change delivery
The core challenge is a creation of a solution capable of performing the reading with above 90% accuracy, recall and 80% precision.
Modern classification and object detection InceptionV3 and YoloV9 models have been transfer-learnt for this purpose scoring 92%-96% on all recorded metrics, exceeding the initial expectations by approx. 6% and delivered as part of a reusable Python service.
Tools & means
- Python
- InceptionV3
- YOLOv9
- PyTorch
- FastAPI
- Label Studio
Outcomes
Change outcomes
We have developed AI solution for extracting and recording fusebox data with over 90% accuracy and 80% precision.