Milarex: 99.9% of contaminants caught on X-ray, on lines running 66.6% faster

The challenge

Milarex’s previous X-ray system caught 62% of contaminants such as metal, glass and ceramics, and it slowed production down. False alarms ran at about 20% and stopped the lines for almost two hours a day.

The new system had to bring false alarms down to 5% at most and run at least as fast as the old one, so every stage of production stayed in step. It also had to use less energy, because X-ray lamps are costly to replace, and stay flexible as Milarex added new products.

What we built

• Six months of classifying a large image database gave us the training material for deep learning models.
• Two neural networks work side by side: one trained on clean products, one on contaminated ones.
• The models run next to the line, on GPU, CPU or VPU, with the X-ray generators and line-scan detectors already in place.
• Operators mark false alarms and confirmed detections in the scanner software, so the models keep learning. A central server manages every machine and collects the detection statistics.
• Cyclic scanning came first. Continuous scanning followed in 2021, so a contaminated tray no longer stops the line.

How it works

Two models score every tray, and Firecrux X-Ray QC reads the two scores together:

• Clean tray: the model trained on clean products scores highest, and the tray moves on.
• Contaminated tray: the scores reverse, and the tray is rejected.
• Anything unusual, such as a contaminant outside the tested range or deliberate sabotage: the two scores come out close, and the tray goes to a person to check.

Results

• Contaminant detection: 62% → 99.9%, up 61.1%.
• False alarms: from about 20% to under 1%, against a target of 5%.
• Line speed: 1.2 → 2 m/s, 66.6% faster, which moved the bottleneck of the whole system to another part of the process.
• Over 330 million salmon trays scanned since 2019.

Timeline

• 2017: work on false alarms starts, with real contamination such as broken knife tips, plasters and glass.
• 2019: the first machine goes live, and 13 more follow by 2022.
• November 2022: licensing and maintenance agreements signed. Since then we tune the classification models to keep detection accurate.
• By February 2023: 23.5 million scan cycles, 107,917 detections and about 40,000 trays a day.
• 2025: the inspection stations move to new industrial computers.