Protostar Labs

Use Cases

Visual quality inspection of bottle packaging

Visual quality inspection of bottle packaging

Production defects are a common problem in the food and beverage industry, but most of the issues that arise can be fixed using modern solutions such as machine learning and computer vision. These issues can include, but are not limited to, problems with bottle caps, liquid levels, foreign objects in the product, etc. Protostar Vision Box combines machine learning with on-board processing and direct integration into existing production lines to allow an immediate increase in production efficiency and worker satisfaction during which we are also reducing production costs and the rate of product returns. Machine learning allows us to track even the smallest imperfections if the environment demands it while keeping the false positive rate under 0.01%. Finally, we have access to state-of-the-art vision systems that allow us to detect foreign bodies in the liquid even if the packaging is already sealed.

Goals

The goal of the project was to develop a system that would detect defects in five different bottle types, all of which can come in three different colors. Because of the large throughput rate, the error rate had to be minimal, and the latency plus inference time had to be under 100 ms. The developed system also had to be integrated into an existing production line that already had a bottle ejection system integrated into it. 

Solutions

The Protostar Vision Box contains two industrial cameras mounted on each side to allow a near360-degree view of the bottle. On top of that, there are six optical sensors mounted at various points that keep track of the bottle as it passes through the Vision Box. Sensors trigger the cameras, after which the PC grabs the images and processes them using a deep-learning model inspired by Yolo architecture. After the processing is done and the algorithm finds the bottle as bad, the number of bottles is sent to the PLC, which turns on the ejection when the bottle reaches a certain point. There is also a photoelectric through-beam sensor mounted that controls if the liquid level is above minimal.

Results

  • Achieved a model accuracy of 99.7% with a false positive rate of under 0.01%. 
  • So far, we have processed over 10 million bottles without any hardware or software errors.

Related Use Cases