CHALLENGE
Quality control is one of the most important production steps that companies and industries face, relying on the expert eye of their workers and employees.
Timing and possible human distractions have a considerable impact on the standards to be met and on productivity.
TASK
An international company in the production and manufacture of safe and functional food packaging for the large-scale retail trade and the food industry, asked us to create a customised quality control system that could be adapted to all their production areas by means of insertion and integration into their production lines.
A secondary objective was to be able to monitor the data analysed by the system.
SOLUTION
We proposed a customised quality control system based on artificial intelligence guided vision. The objective is to create a high-performance machine vision platform customised to the customer’s needs.
The proposed technology is based on the use of neural networks that through the vision tools (cameras, optics and illuminators) perform image processing in order to achieve classification and recognition of the rejected product without the need for personnel.
The platform uses our K-VISION software.
The use of automated inspection systems makes it possible to:
1.
Not slow down production, but speed it up and improve it by eliminating human errors;
2.
Improve the reliability and accuracy of measurements even for complex parts where volumes are high;
3.
Obtain an important database of information to make better decisions for the company; Collect real-time dynamic analysis of production quality as well as comprehensive compliance testing.
4.
Obtain a system that increases its performance over time in order to have continuous algorithm improvement, thanks to the use of neural networks.
5.
Not having to modify your production lines, optimising them, because it is a fully customisable solution that can be integrated.
In order to achieve the goal of classifying suitable products from unsuitable ones with a high degree of confidence, it was decided to proceed in steps of increasing complexity.
Dataset
The dataset was divided by product colour and the unsuitabilities were divided into two classes (shape problem, contaminant presence).
Then two different applications were developed to classify:
- Product with a white background (two classes, between compliant and contaminant presence);
- Product with a black background (two classes, between compliant and presence of contaminant).
Parameters
Accuracy: a measure of how well the model is able to classify images correctly overall;
Precision: the proportion of instances correctly classified as belonging to a given class compared to the total number of instances classified as that class;
Recall: the proportion of instances of a given class that have been correctly identified by the model;
F1 Score: the harmonic mean between accuracy and recall;
Confidence: the confidence with which the algorithm predicts the given result, the higher it is, the more certain the result is (according to the algorithm).
The results of this first phase determined that the developed algorithm:
is very robust to different variations, shape, light, etc. This means that the customer does not have to modify the production line because, being immune to environmental conditions, the vision tool integrates itself into the existing machine;
it provides many details (not only compliant or non-compliant) that make it easy to decide whether to rely or not. Unlike an operator-operated quality control, the vision system shows details that are difficult to detect by a person in charge and allows to historise the problems the plant is suffering from;
it adapts quickly to the set recipe with high flexibility. Given the initial dataset, it does not require a format change;
works in real time, so speed of execution and no slowdown in production capacity.
Go into action
Tell us your idea!
Send an email, call the office or if you prefer send us a Whatsapp to speak to one of our consultants.
Mail: info@kablator.com – Tel: +39 331 3466711
