Computer vision monitoring solution

About the client

A manufacturing company that manages complex and highly monitored engineering systems worldwide.

How it all started

The collaboration began in 2024 when the client approached XPG Factor and their partner IOX to improve and automate the monitoring of heat transfer systems. 

The client’s trace heating systems were monitored manually by duty staff who checked multiple control rooms every 8 hours. Our mission was to automate this process by using cameras and computer vision algorithms. The developed solution had to determine the on/off status of the red LEDs, interpret the numerical values on the displays, detect anomalies and generate automatic alerts, ensuring that the data was updated at least once an hour.

Challenges we faced

One of the key technical problems was the placement of cameras near the ceiling, while the displays were positioned vertically. As a result, the image was heavily distorted and needed advanced perspective adjustment. In addition, unstable lighting conditions such as bright lights, harsh shadows and glare made it difficult to accurately recognize both the red LED indicators and the numerical values on the displays.

How we solved it

The solution developed by XPG factor together with IOX was based on modern computer vision and neural network technologies. Our team used the YOLOv5 model, a fast and efficient real-time object detection algorithm, to detect LED indicators and determine whether they are on or off.

For reading numbers from LED displays, we integrated an OCR module based on Tesseract. In addition, the algorithm was trained on more than 1000 images. All of this helped the model adapt to camera angles and lighting conditions, providing reliable and accurate results.

Core features

• Display of equipment status
  Real-time monitoring of ten control rooms.
• History of changes
  Automatic recording of all changes in indicator status, available for further review.
• ID-based event tracking
  Each LED indicator has a distinct identifier (e.g., “LED 12”) to help locate problems precisely.
• Regular data updates
  Image processing and status updates are performed once per hour.
• Automatic alerts
  When a fault is detected, the system generates an alert and sends it to the responsible staff.

Testing

Our team tested the system using real images collected on-site via Raspberry Pi devices. The dataset was split into training, validation and test sets.

To make sure the system works well in real conditions, we also tested it with poor lighting and different kinds of displays.

As a result, the solution demonstrated high accuracy and confirmed that it was fully ready for live operation.