Successful Stratis participation at the Hungarian Special Industrial Machine Grand Prix

Our entry work was received with great interest, even though we failed to win the first prize in our category.

Objective of participation and offered solution

The idea for our quality control system using artificial intelligence came from a request from one of our partners. This partner wanted to find a solution to the problem of crack detection in ceramic parts. This issue prompted the development of the equipment presented at the Special Machine Grand Prix. Once we were convinced that the AI-based machine vision approach can reliably meet our partner's needs, we are currently designing and manufacturing the mass production equipment.

Our aim is to replace the monotonous and highly attention-demanding work of several quality inspectors per shift with a single piece of equipment, while increasing efficiency and improving the quality of inspection at the same time. We calculate the payback period for such a solution to be 1-1.5 years.

The main advantages of mechanical sorting over human selection are:

• more reliable,

• faster,

• 24/7 availability

• produces accurate statistics,

• more economical to run.

The prototype presented as a pitch project performs optical quality control using artificial intelligence. The aim is to demonstrate to our customers the possibilities of using artificial intelligence and to perform tests on different machine parts of our partners. If the test results are positive (i.e. if the AI and machine vision show encouraging results), we will design and manufacture a series production machine to meet our customer's specific requirements. The equipment is capable of testing parts of ~10×10mm to 150×150mm in size, in a wide variety of shapes and materials (e.g. plastic, rubber, ceramics, leather, wood, etc.).

A revolutionary procedure in quality control

The special feature of this equipment is the use of artificial intelligence (AI): The compatibility of the products is not based on pre-established and programmed attributes. The equipment is “trained” (based on photos of suitable, matching, and then defective products) to be able to distinguish between rejected scrap and flawless, compatible parts. The larger the training set (the more parts are taught), the more accurate the training. This allows the equipment to be flexibly taught to handle changing customer and market requirements. (Our partner carries out crack tests on 150 different ceramic parts with a single piece of our equipment.) The dedicated machine adapts to the quality criteria of our clients via machine learning.

For this, we "train" it in the quality requirements of the parts based on a sufficiently large sample (~1,000-5,000 pieces). This is done by first running a sufficient quantity of "good" products, followed by "incompatible" products through the equipment in a "learning mode". Increasing the amount of sample data enhances the reliability of subsequent sorting.

After training, in "selection mode', the inspection equipment can distinguish between compliant and non-compliant part quality. In our experience, one product in twenty has a component part that borders on the limit between passed and failed categories. Reasonably these components should be subjected to a posterior test. By expanding the training set, the accuracy of selection can be improved, and the proportion of "to be decided" parts can be reduced.

AI is also able to detect patterns in natural materials (e.g. leather, wood, etc.) that are acceptable or unacceptable for the client. This makes it easy to change the selection requirements. In addition, the equipment becomes more accurate and confident in sorting parts as it operates, likely making it the most reliable and cost-effective quality control system for the plant in a short time.