Automated Quality Control for Manufacturing Using Computer Vision

  • #Cloud
  • #Computer Vision
  • #Custom software development
  • #Machine Learning

About the Client

One of the world-leading manufacturers of compressors for refrigeration systems.

Business Challenge

Our client produces thousands of high-quality level compressors daily. To reduce the number of defective items, the company implemented quality control throughout the whole product lifecycle using Industry 4.0 technologies.

The production line had cameras mounted above assembly lines with the embedded computer vision tool that was quite outdated and couldn’t correctly identify compressor tubes that weren’t equipped with caps. Part of why this happened was differences in compressor arrangement on the assembly line were not obvious to the tool. As a result, some compressors were seriously damaged during the painting stage.

Therefore, the client needed to improve the manufacturing process and cut costs by reducing manufacturing defects.

Solution Overview

The Quantum team has come up with a solution to detect a compressor and identify whether the caps were mounted on the tubes regardless of the compressor arrangement and lighting conditions.

Using a video stream from the cameras formerly mounted above the conveyor belts, we could process the incoming video feed in real-time, detect the absent cap on any compressor tube, and warn the assembly line control system. That enables the personnel to conduct additional checks and ensure all the tubes are equipped with caps.

Project Description

The solution operates a FullHD video camera attached to NVIDIA Jetson TX2. This method of Edge Computing allows the processing of the data closer to the original location, thus optimizing infrastructure capabilities and saving costs. If the system detects a missing cap, it immediately stops the conveyor belt and alerts the staff.

Our team reached 99.99% accuracy in absent caps detection, which is ten times more than before.

Let's discuss your idea!

Technological Details

The video stream was processed using OpenCV tools. The input was sent to a trained YOLO object detection model chosen because of a higher accuracy achieved during the model training process. The model’s output was processed using historical data to reduce the False Positives rate and achieve the target accuracy goal. If the object that didn’t meet the set requirements (each compressor has to have a certain amount of caps attached) was detected on the conveyor belt – the system sends the alert.

Python
Python
Darkflow YOLO
Darkflow YOLO
OpenCV
OpenCV

Connect with our experts