Simulation and Digital Twins (MCD) (formerly CADFEM CIS, CADFEM Si-I-Es)

2022: Implementation of Pilot Part Recognition System

On August 17, 2022, Modeling and Digital Twins JSC (MCD JSC) announced that it had completed a pilot project for KAMAZ PJSC to develop a system for assessing the number of parts based on machine vision. The system will allow you to maintain more accurate accounting of parts, optimize control over their movement in production, and the effectiveness of production planning. In addition, this will make it possible to reduce the impact of the human factor.

Optimizing the accuracy of part metering in the manufacture of complex products can lead to significant cost savings. However, the nomenclature of parts in production includes more than ten thousand articles. Quick and unmistakable identification of each part, counting and entering data into the logs is a challenge. For example, one finished car contains thousands of unique parts, each of which goes its own way at the enterprise, including production at several alterations, logistics, storage. It is important to follow this path, take into account the number of parts and blanks at each stage in order to ensure stability and rhythm of production, reduce losses, optimize the quality of dispatching and production planning. noted experts of PJSC "KAMAZ"

Machine vision technology allows this process to be automated. The mathematical model underlying it reliably recognizes and counts details, even if a person appears in the frame, lighting or background changes. In this case, you can use ordinary webcams that are not equipped with artificial intelligence, since the main analytics is conducted on standard. computer server Accuracy recognitions provided in the range of 95-99%. The specialists of Modeling and Digital Twins, using their experience in the field of machine vision, made such a solution for KAMAZ PJSC.

The pilot project covers one large production site where parts of trucks are painted. The system recognizes parts moving on a suspended conveyor, counts their number for further accounting and reconciliation. This area is characterized by many unfavorable factors that interfere with the operation of machine vision systems - foreign objects fall into the field of view of video cameras, the conveyor with parts moves unevenly, there is no special lighting. In this case, it was necessary to take into account the following requirements:

• The system should provide stable and fast work with the entire range of parts - identification, counting, comparison.
• The system should not require excessive computing power for its operation.
• The system shall not make errors during counting (not more than during manual counting).
• It should be easy enough to teach the system to identify a part in production and not require large labor costs.

The latter requirement is especially important, since the introduction of details into the system is not just the entry of digital articles, but the introduction of many images and the subsequent training of a mathematical model.

As part of the pilot project, two main scenarios were completed - automatic part recognition on the conveyor and part recognition using a mobile device. In the first case, the camera was fixed in a certain area of ​ ​ the conveyor, continuously capturing the flow of parts and automatically entering its readings into the ERP system. As a result, accurate recognition of parts and counting of their quantity took place, and then the system checked the data with the quantity in the accounting system.

In the second scenario, an employee opened a specially prepared web page on the smartphone of the mobile part of the part recognition system. When the technician had to determine the part during sorting, it was enough to point his camera at the part, take a picture and get its article on the screen.

Production conveyors, such as painting shops , can have many different types of parts at the same time. Their sorting and counting is difficult, as they are often similar to each other. The machine vision system automatically determines the part article number and helps personnel sort them, determine the number of each view in key areas of the assembly. told Ilnar Satdarov, chief technologist of the press and frame plant of PJSC "KAMAZ"

As of August 2022, as part of the training and testing of models of stationary and mobile recognition systems, a number of tasks have been completed:

• Photos of details for training models of stationary and mobile systems have been prepared.
• Part samples for stationary system have been marked up.
• The model was trained on marked samples of parts.
• The model was tested on samples of parts that were not involved in the model training process.

Although the pilot implementation of the system has not yet been completed and ahead - the scaling stage, we can already say that the system confidently copes with the task of counting parts and helps technical personnel on the conveyor quickly determine the part number during shipment to the warehouse or to other areas of the assembly. We are sure that it is suitable for work in factory conditions. told Andrey Krylov, Director of the Center for Digital Technologies of the company "Modeling and Digital Twins"

In order to scale the system, experts have to solve one of the key tasks - to quickly introduce images of all parts produced by the enterprise into the model. Tens of thousands of parts cannot be manually entered by piece - this process will require significant time. In addition, this minimizes the effects of the system, despite the fact that, according to experts, the target payback period is no more than 1.5 years. To solve this problem, a software "conveyor" will be created to automate the introduction of parts - it will allow the part image to go from a 3D model developed by the designer to a machine vision model with minimal human participation.