BIION-D Docs

Training

Practical training modules and end-to-end industrial inspection scenarios.

Welcome to the BIION-D hands-on training section. This section provides three progressive, production-grade training exercises and practical assessments built around industrial plastic discs quality control.

Instead of providing pre-built solutions, each case presents an industrial scenario, technical specifications, and evaluation criteria, supplemented with practical engineering tips and guidance to help trainees design, configure, and validate their own inspection pipelines independently.

Training Exercises Overview

CaseArchitectureCore TechniquesLink
Case 0YOLO Counting & Face DetectionCamera, YOLO Object Detection, Plugin per-face counting.Case 0: Plastic Discs — YOLO Counting & Face Detection
Case 1Case 0 with Dashboard & PLCReuse YOLO counting, add Webhook telemetry and OPC-UA handshaking.Case 1: Plastic Discs — Case 0 with Dashboard & PLC
Case 2Single Disc, Dual-Pipeline Front/Back with YOLO Crop & Anomaly2 pipelines, YOLO cropping, Anomaly Detection, is_pretrained during anomaly training.Case 2: Plastic Discs — Dual-Pipeline Front & Back with YOLO Crop & Anomaly

Inspection Concepts Covered

1. Optical & Lighting Strategy

Inspecting semi-reflective or matte plastic discs requires uniform, diffuse illumination: glare can obscure the face features detected in Cases 0 and 1, while shadows can trigger false anomalies in Case 2.

2. Anomaly vs. Object Detection

  • Anomaly Detection (Unsupervised): In Case 2, learns what “normal” plastic discs look like from defect-free cropped images. Highlights scratches, dents, burns, or contamination without requiring prior defect labels.
  • Object Detection (Supervised): Identifies discrete objects, counts items, classifies orientation/face (e.g. Face A with embossed logo vs Face B flat), and provides bounding boxes for localization and cropping.

3. Edge Integration & Handshaking

  • Industrial PLC (OPC-UA): Reads machine triggers and writes real-time compliance results, defect flags, and piece counts directly to PLC registers.
  • Custom Dashboards & MES (Webhook): Dispatches structured JSON payloads asynchronously to factory dashboards, production lines, or quality assurance logging systems.

4. Advanced Dynamic Cropping & is_pretrained

When parts move freely on conveyors, combining YOLO object detection with a Plugin cropping script ensures the anomaly model always receives a normalized, tightly bounded image. Setting is_pretrained: true on the Object Detection module guarantees that YOLO executes live inference even while collecting anomaly training datasets in Acquisition mode.

  1. Start with Case 0 to train YOLO for face identification and count discs.
  2. Extend that pipeline in Case 1 to send counts to a custom dashboard and PLC.
  3. Complete Case 2 to inspect one disc on both sides with two YOLO-crop-and-anomaly pipelines, using is_pretrained on Object Detection during anomaly training.