Efficient inspection of loaded pallets: reliably detect leaks in cardboard boxes

The project

A customer from the pharmaceutical industry is faced with the challenge of making the storage of pallets loaded with multiple layers of brown cardboard boxes more efficient. The cartons contain plastic bags with liquids that can leak. These leaks can be recognized by dark discolorations on the outside of the boxes. When such defective pallets are stored in the high-bay warehouse, significant problems arise, including

• Operational disruptions with sometimes lengthy downtimes
• Damage, including stored pallets
• Complex manual re-sorting, often in great amounts

The customer was looking for a solution to detect these discolorations automatically and reliably. As there is no commercially available standard solution for this application, AUTFORCE was commissioned to carry out a comprehensive feasibility study to check the technical viability.

Keyfacts

• Objective: Automated detection of leaked liquids on cardboard surfaces
• Cycle time: One pallet every 10 seconds
• Pallet loading: 6 boxes per layer in 6 layers (36 boxes per pallet)
• Pallets are wrapped in a transparent transport film

Setting up the test setups

• Three different test setups were created in order to establish optimum conditions for image acquisition.
• Various combinations of camera systems, exposures and filters were tested.
• Tests were carried out with both smart camera systems and PC-based systems in order to find the best solution for detecting the leaks.
• Particular attention was paid to minimizing reflections from the transport film and reducing the effects of extraneous light.

Analyzing the image data

• Various AI models were considered for the automated evaluation of the captured images. The following models were examined for their suitability in the course of this feasibility study:
  • Deep Learning: Application of Anomaly Detection and Semantic Segmentation to analyze complex defects.
  • Edge Learning: Local processing on the device with pre-trained algorithms, ideal for smaller data volumes and fast training cycles.

Leak detection with deep learning

Leak detection with edge learning

Findings from the feasibility study

• The results of the study showed that leaked liquids can be reliably detected using machine vision.
• Clear recommendations were given for the selection of suitable hardware and evaluation methods.

Conclusion

The feasibility study demonstrated the technical feasibility of a system for the visual inspection of loaded pallets. The findings from the test series create a solid basis for the development of a customized solution and underline the efficiency of AI-supported technologies in automated quality control.