AI Defect Detection in Manufacturing: reTerminal DM as the Vision System HMI
An inspection solution provider uses reTerminal DM as the HMI of its AI vision system for mass-production defect detection and assembly verification.
Industrial Visual Inspection
Millisecond edge inspection on-station, direct to PLC
Fanless edge IPCs provide real-time inspection verdicts locally, while high catch rates require retraining with line defect samples.
10ms Hardware Reject
Isolated DO triggers pneumatic kick-off cylinders in under 10ms
Real-Time Edge AI
Orin compute inspects locally; images never leave the station
Native Bus & MES
Modbus TCP interfaces PLC while MQTT uploads records to MES
Scenario details
Covers three core industrial inspection tasks: surface defect detection, missing-part and dimension checks, and label OCR.
Pain Points & Challenges
Manual visual inspection suffers from worker fatigue and inconsistent subjective criteria, causing high defect escape rates on high-speed continuous lines; micro-scratches and pits slip into downstream stages, triggering costly mass recalls.
Edge AI Solution
Deploy a fanless industrial edge host beside the line, triggering high-speed industrial cameras with strobe illumination to run INT8 defect models at 30–60 FPS, firing isolated optocoupler DO within 10 ms to actuate pneumatic rejectors.
Core Advantages
- YOLOv8 INT8 inference runs at 35 to 60 FPS per stream
- Isolated DO triggers kick-off cylinder reject within 10ms
- Retrain YOLO models with real factory defect samples
Live Verdict Panel
A screen by the line shows live video, defect boxes and the verdict.
Location & Class
Defects marked by location and class; NG reasons reviewable.
Trial Run First
Record line images first and set threshold and false-alarm rate on your data.
Pain Points & Challenges
Multi-component assembly stations relying on manual inspection frequently suffer from missing washers, loose screws, or reverse orientations; offline coordinate measuring machines only allow batch sampling rather than 100% in-line error proofing.
Edge AI Solution
Mount high-resolution industrial cameras above workstations, leveraging optical calibration and template matching to verify dozens of assembly points in a single frame, measuring dimensional tolerances and interlocking with PLC via Modbus.
Core Advantages
- Verifies missing and misaligned parts in a single frame
- Sub-millimeter dimension inspection with calibration targets
- Connects up to 8 industrial cameras per edge host
Size in Tolerance
Marker-based mm conversion; pass within nominal ± tolerance.
Out of Tolerance = NG
Oversize or undersize is NG, with the measured size in mm.
PCBA Board Defects
Marks the position and class of opens, shorts, spurs and other board defects.
Pain Points & Challenges
Handheld barcode scanners require manual alignment per item, heavily bottlenecking packaging throughput; metallic label glare, scratched dot-matrix characters, and ink bleeding cause rule-based vision systems to fail and falsely halt production.
Edge AI Solution
Position all-in-one edge AI cameras along packaging conveyors running deep-learning OCR on-device, reading alphanumeric batch codes and QR codes from moving workpieces in milliseconds and verifying against MES databases.
Core Advantages
- On-device OCR for Chinese and English batch codes
- Matches work orders to prevent printing omissions
- Streams OCR text to line systems for reject gating
Field Reading
Reads batch codes, production dates and model numbers in Chinese or English.
Order Check
The line system checks results against the work order; misprinted parts are rejected.
On-Camera OCR
Recognition runs on the AI camera itself; images stay on the local network.
Deployment & Selection
Left to right along the diagram: set camera interfaces, select fanless wide-temp edge IPCs, and configure PLC & MES outputs.
Not sure which to choose? See the full selection guide →
Station Camera
Which camera is at the station? The camera connection decides which checks are available.
Cameras link to edge IPCs, sending millisecond verdicts to the PLC
The deployment architecture consists of station cameras, edge industrial compute, and verdict outputs. The next three tabs detail camera interfaces, edge host selection, and line verdict outputs.
Core Advantages
- Two routes: industrial camera plus IPC, or smart AI camera
- Inference runs on-station; raw images never leave
- Isolated DO triggers reject while Modbus talks to PLC
| Camera Type | Interface | Inspection Tasks | Hardware Sync |
|---|---|---|---|
| GigE Vision Camera | Gigabit PoE to host | Surface defects, coil, precision assembly | Optocoupled DI triggers strobe |
| USB3 Vision Camera | USB 3.0 port to host | Stepped assembly, tolerance checks | Software or external photo trigger |
| RTSP IP Camera (Existing) | Ethernet RTSP stream | Packaging inspection, presence check | Continuous per-frame stream |
| reCamera Pro All-in-One | Built-in 8MP sensor | PCB defects; no missing-part/size checks | Onboard exposure & edge inference |
Gigabit industrial cameras with hardware sensor strobe sync
Supports GigE Vision, USB3 Vision, and RTSP industrial cameras across diverse fields of view and resolutions. Isolated DI inputs sync with photoelectric sensors and strobe lighting to prevent motion blur.
Core Advantages
- Compatible with standard GigE Vision and USB3 cameras
- Hardware sync strobe and sensor to eliminate blur
- Calibration targets enable millimeter size conversion
| Image | Product | Role | Field I/O | Checks | Camera Input |
|---|---|---|---|---|---|
 | reComputer Industrial J3011 | Fanless host (Orin Nano 8GB) | CAN, DI/DO, Gigabit Ethernet | Surface, missing part, size | RTSP IP camera |
 | reComputer Industrial J4012 | Fanless host (Orin NX 16GB) | RS-485, CAN, DI/DO, Gigabit Ethernet | Surface, missing part, size | RTSP IP camera |
 | reComputer Industrial R2035-12 | Industrial host + Hailo-8 | RS-485, RS-232, DI/DO, Gigabit + Fast Ethernet | Surface, missing part, size | RTSP IP camera |
 | reCamera Pro | All-in-one AI camera, no host | Gigabit Ethernet | PCB defect detection | Built-in 8MP sensor |
 | reComputer R1025-10 | Line-side gateway, no inspection | RS-485, Ethernet | Runs no inspection model | No camera input |
Match fanless wide-temp IPCs to camera stream count and cycle times
Multi-camera high-framerate inspection is best served by the fanless J4012 IPC, while single lightweight stations run on J3011, and the R1000 gateway bridges plant data and legacy PLCs. The Orin series leverages INT8 hardware acceleration to meet high-speed line cycle times.
Core Advantages
- Fanless wide-temp design with DIN-rail and wall mounting
- 4x PoE Gigabit ports power and connect multiple cameras
- Orin platform accelerates YOLOv8 INT8 inference
| Output Method | Physical Interface | Target System | Verdict Message & Action |
|---|---|---|---|
| Isolated DO (Hardwired) | Optocoupler terminal | Blow-off valve / reject cylinder | Pulses within 10ms to trigger pneumatic reject mechanisms |
| Modbus TCP (Port 502) | RJ45 Gigabit Ethernet | Line-control PLC | Coil 0=NG, Coil 1=OK; holding registers store defect codes and sizes |
| MQTT | Ethernet / Wi-Fi | Plant MES / SCADA | Publishes JSON events with timestamp, station ID, cause, and latency |
| HTTP / Web | Local network | Station HMI / Dashboard | Streams live video with defect bounding boxes and reject snapshots |
Isolated DO triggers pneumatic reject, fieldbus connects PLC
Verdict outputs support dual channels: isolated DO triggers pneumatic cylinders within 10ms, Modbus TCP reports defect categories to the PLC, and MQTT publishes traceable inspection events to the MES.
Core Advantages
- Isolated DO triggers pneumatic reject in under 10ms
- Modbus TCP coils and registers link directly to PLC
- Asynchronous MQTT events stream to MES for trace
Success Cases
FAQ
Yes. Any GigE Vision, USB3 Vision, or RTSP camera connects to the inspection host. Mounting must be rigid, lighting must be stable, and the field of view must cover the part.
Yes. Dual links are supported: isolated DO directly triggers pneumatic blow-off or cylinders within 10ms, while Modbus TCP coils and registers stream defect classes and dimensions to the PLC.
Retrain with your line samples. The baseline model provides out-of-the-box evaluation; fine-tune YOLO models with real defect images before production. Assembly checks allow drawing-based on-site calibration.
No. All image acquisition, YOLO inference, dimension calculations, and OK/NG verdicts run locally on fanless edge IPCs. Only structured verdict records and events are transmitted outbound.
No. Vision measurement uses calibration targets on the plane for sub-millimeter conversion to prevent assembly defects. It is not a legally calibrated gauge and cannot replace CMM certified metrology.
Yes. Seeed edge IPCs and cameras support OEM/ODM customization including colors, custom logo silk-screening, packaging, and pre-installed firmware. See ODM/OEM Customization or contact our team.
Didn't find your answer?
Contact Us
We Are Glad to Be Your Hardware Partner !