REACHbyMortoff

Case study · AI Vision – Manufacturing

AI-based quality control on a poultry packaging line

REACH OEE measurement and AI-based video scrap detection on a previously non-digitalised packaging line.

Customer
Varga Szárnyas Kft.
Industry
Poultry processing
AIvisual scrap detection
Real-timeOEE and scrap data
Edgeinternet-independent operation next to the line

In brief

Challenge

The packaging line was not digitalised, the rejection of defective packages relied entirely on human attention, and scrap was recorded on paper.

Solution

REACH OEE measurement and AI-based video quality control on an industrial edge computer, with feedback to operators on a tablet and OEE shown on a TV display.

Result

Consistent, objective scrap detection, shorter reaction time, and OEE that is measurable in real time and auditable.

Starting point – non-digitalised packaging line and manual quality control

When the project started, the production and packaging process at Varga Szárnyas Kft.'s poultry packaging unit was not digitalised at the data level.

The packaging involved two main product types:

  • vacuum-packed poultry meat products
  • tray-packed poultry products in a protective atmosphere

The main characteristics of the operation:

  • There was no real-time OEE measurement or display
  • The detection of defective packages relied entirely on human labour
  • Scrap quantities were recorded on paper
  • The recorded data was processed manually, after the fact

The root of the problem

At machine level the packaging line would have been able to generate a lot of operating data, but this data:

  • was not collected in a structured way
  • was not linked to quality metrics
  • did not support real-time decision-making

Manual scrap detection and paper-based data entry:

  • carried a risk of human error
  • did not provide immediate feedback
  • did not allow accurate, real-time calculation of the Quality component
  • so the OEE metric could only be evaluated after the fact, with limited accuracy

As a consequence:

  • the actual level of efficiency was not transparent
  • interventions took place with a delay
  • scrap trends could only be analysed after the fact

The lack of digitalisation was not only a measurement gap but also a business risk: the accuracy of quality control relied on operator attention, scrap trends only became visible after the fact, and there was no real-time data available for efficiency interventions. This directly affected capacity utilisation, cost levels and the transparency of quality assurance.

Implemented solution – REACH OEE and AI-based quality control

For real-time efficiency measurement and scrap detection, the REACH industrial digitalisation platform was introduced on the packaging line of Varga Szárnyas Kft.

OEE measurement and production line data collection

With the REACH system, we were able to read and process the technical data of the packaging line in a structured way.

During the integration, the following information became available in real time:

  • Machine state (running / stopped)
  • Error code when stopped
  • Tracking of product flow using optical sensors
  • Performance measurement based on cycle time and piece count
  • Calculation of the OEE components (Availability, Performance, Quality)

The previously non-digitalised packaging line thus switched to data-driven operation.

AI-based video quality control

For the automatic detection of scrap, we built an AI-based visual inspection system.

The key to the system is the right:

  • camera positioning
  • stable mounting
  • controlled lighting
  • reproducible viewing angle

During the project we put particular emphasis on installing the camera exactly at the point where the operator previously inspected the packages visually.

This ensured that the AI system sees the same decision situation as the human did before.

Learning AI model and data collection

The video sampling was dense, but the processing was not based on a continuous stream. The relevant frames were saved and then served as the basis of the training database.

Because the AI model is built on a learning architecture, it was necessary to:

  • collect images taken under the final lighting conditions
  • label them accurately by scrap type
  • create a structured training dataset

Fast and efficient labelling – "Tinder"-style annotation

To speed up image classification, we developed a simple, gesture-based annotation app accessible from a mobile phone.

The logic of operation:

  • swipe right → defect-free product
  • swipe left → label by scrap type

This intuitive solution significantly sped up the creation of the training dataset and minimised the administrative burden.

The images classified this way formed the basis for training the AI model.

Edge-based execution – in an industrial environment

We run the trained AI model on an industrial mini computer placed close to the packaging line.

This ensures:

  • low latency
  • internet-independent operation
  • stable industrial operation
  • instant scrap detection

Detected scrap types and capabilities

The AI model distinguishes several scrap types, including:

  • Recognising vacuum packaging damage
  • Packaging integrity check
  • Estimating the product type in the package

The latter capability – the AI-based estimate of the type of product in the package – makes it possible, through integration with the production scheduling module, to compare the production plan with the actual packaging.

This takes quality control to a new level: it can detect not only defective packages but also mismatched product packaging.

Real-time scrap feedback at operator level

After the AI model went live, the system did not just collect data in the background but also provided immediate feedback to operators.

The result of the scrap detection appeared on a tablet placed next to the packaging line, where:

  • for each product, it decided whether it was defect-free or belonged to a specific scrap type
  • the probability value (%) belonging to the decision was also shown
  • the analysed image was also visible

This ensured that the operator:

  • could see the basis of the system's decision
  • could verify how the AI was working
  • could gradually build trust in the system

The system therefore does not operate as a "black box" but provides transparent decision support.

Real-time OEE display at management and operator level

Once the technical and quality data of the packaging line was being collected in a structured way, the REACH system made real-time OEE calculation possible.

The calculated metrics – Availability, Performance, Quality – appeared on a large TV display set up next to the packaging line.

The display is:

  • real-time
  • automatically calculated
  • independent of manual data entry
  • accurate and auditable

As a result:

  • the shift could see the current efficiency level immediately
  • deviations could be detected immediately
  • intervention took place during operation, not after the fact

Results

After the introduction of the system, the operation of the packaging line became noticeably more stable and transparent. AI-based quality control provides consistent and objective scrap detection, which significantly reduces the room for error caused by fluctuations in human attention. The recognition of vacuum packaging damage became more reliable, and defective products are filtered out according to a uniform logic, so the quality level is more balanced.

Thanks to the real-time feedback, reaction time was shortened considerably. Whereas scrap trends and efficiency deviations previously only became visible after the fact, from paper-based summaries, the new system provides immediate visualisation. The continuous updating of the Quality component allows shift management to intervene during operation, so errors do not accumulate and do not become visible only at the end of the shift.

The operator workload also decreased. Instead of manual scrap counting and administration, the system collects and processes the data automatically while providing visual decision support. This allows the operator to put more emphasis on supervising the process and on actual intervention, rather than on administrative tasks.

Real-time OEE calculation and display created an objective basis for shift evaluation. Instead of paper-based processing after the fact, the efficiency metrics are updated continuously and automatically, giving a more accurate picture of the packaging line's performance. For management, the relationship between machine stoppages, error codes and quality deviations has become transparent, which provides a stable basis for further optimisation steps.

Overall, as a result of the project, the operation of the packaging line has become data-driven: quality control is automated, efficiency can be measured in real time, and decisions are based not on after-the-fact estimates but on objective, auditable data.

Any questions?

Facing a similar challenge?

Tell us about the problem, and after a short assessment we'll show you how REACH would solve it in your plant.

Let's talk