CUSTOMER DESCRIPTION:

The customer is a globally renowned chemical engineering company that produces fuel cell components in one of their several plants. A crucial part of a fuel cell is the membrane electrode assembly (MEA), which helps produce the electrochemical reaction needed to separate electrons and generate current. Our customer is the only MEA supplier with in-house manufacturing, so the entire process is done in one plant. This means they must track parts throughout the process and make sure that they associate all the right measurements with a specific part in the database.

CHALLENGE:

The customer must track parts throughout the process and make sure that they associate all the right measurements with a specific part in the database. Tracking also aids in error detection by always knowing which part became faulty, and at which step in the process it happened. Therefore, they needed a system that could read the part number after each step of assembly and assign the data from various measurements into the database correctly. Their goal was to detect part numbers with near-100% accuracy and ensure traceability throughout the entire manufacturing process.

SOLUTION:

The vision system had to correctly recognize the digits (0-9) that made up the part ID number using optical character recognition (OCR). The number is made up of a year number and a sequential number of the part, separated by a control sign “–“. After each operation, such as applying a layer of various materials, the part was weighed multiple times and the measurements were associated with the part ID in the database. The correct part ID was determined using OCR. A single database entry included the location (dosing station) of the part, dosing duration, dose quantity, and part status, which was either “OK” or “NOK” with a rejection reason (e.g. measurement outside of acceptable range).

The solution consisted of a Cognex camera and a lighting setup, which had to be set up very precisely to ensure the optimal environment for character recognition as the number is imprinted on a grey, distorted surface, and is hard to read even by a human eye. The system used a Cognex machine learning tool which has been trained on numerous pattern samples to determine the boundary values for each character. Since the part gets coated in various materials during the process, the number is mechanically stamped rather than printed. The membrane is also very fibrous with fibres of varying colour, so OCR based on colour alone could not be done. The optimal type, colour, and angle of lighting were necessary to help the system to recognise similar digits, such as 1 and 7, 3 and 9, or 8 and 9.

We needed to introduce double checking of OCR, as sometimes surface defects appear right on the character area and disable any possibility for correct recognition. In such cases, the system detects poor marking based on matching score to trained samples and with a comparison to correct serial numbers that should be imprinted. It then saves all the data, including a timestamp of the reading attempt, and rejects the part.

IMPACT:

• Over 98% OCR reliability
• Improved cycle time (measurements, reading, dosage) from 36s to 28s

“With Inea’s optical recognition system, we were able to speed up our cycle time significantly and gain important insight into every part’s journey through manufacturing.”

 VP, Manufacturing