Five layers, one timing budget.

A line scan camera with the vision algorithm on the FPGA and the ejection interface on the board. Two Gigabit Ethernet ports carry the image and the valve command, so a defect decision never leaves the imaging chain.

• CMOS sensor at 2048 px resolution with a 14 µm pixel pitch
• 20 kHz line rate for full-width streams on a colour sorter
• FPGA-integrated vision algorithms: colour, Mono, colour difference and intelligent SVM
• Two Gigabit Ethernet ports, ejection valve interface integrated on the camera board
• Domestic core components, NIR ready, custom HD lenses for tighter imaging
• Deep learning extension once the camera is connected to an AI edge box

An embedded Linux box built for the sorter cabinet: 8-core ARM CPU, 5 or 9 network ports and enough compute for four or eight cameras. Deep learning runs on the machine, without an external GPU server.

• 12T and 48T compute tiers for different machine widths
• 8-core ARM CPU on an embedded Linux platform
• 4 or 8 cameras computing on a single box at the same time
• 5-port and 9-port Gigabit configurations for cabinet layout
• Image inference latency down to <1 ms
• Lower cost per channel than a separate industrial PC plus GPU

Valve boards developed for the colour sorter industry, paired with ejection strategy algorithms. Hardware driving changes shorten the open and close timing, which is where Purity Ratio and carry-over are decided.

• Custom valve boards for colour sorter platforms
• Valve open at 0.3 ms and valve close at 0.2 ms
• Ejection strategy algorithms developed together with the board
• Designed for 8+ years of continuous operation
• Ejection timing tuned against Purity Ratio and the good:bad ratio
• Compatible with SMC, MATRIX and MAC valves already used in the industry

An ARM-based industrial panel PC running the Linux HMI, with a custom GUI control set for sorter operators. One-click operation on the machine, remote control, monitoring and upgrades from the service side.

• Linux human machine interface with custom GUI controls
• Industrial thermal range of -20 °C to 70 °C
• ARM architecture for lower power draw
• 10-inch and 15-inch touch panels
• Two Gigabit Ethernet ports for camera and plant network
• One-click smart operation, remote control, monitoring and upgrading over IoT

An image annotation and training platform built in-house. A user uploads sample data, trains on the platform and downloads the model for the edge box, so a material change does not wait on an external algorithm team.

• Image annotation workflow for sorting samples
• Training runs on the platform after samples are submitted
• Trained model downloaded and deployed to the edge box
• Training method work keeps accuracy while inference latency stays low
• Model updates follow material and seasonal changes
• No external GPU cluster needed in the customer plant