ScrapVision — Reject Station Intelligence

ScrapVision: Real-Time Reject Chute Intelligence for Discrete Manufacturing

ScrapVision transforms the factory reject chute from a passive waste bin into a high-value production sensor. In modern discrete manufacturing—spanning Tier-1 automotive components, precision plastic injection molding, and electronics assembly—most scrap tracking is still recorded manually on paper clipboards at the end of a shift. By the time quality metrics are compiled, the underlying machine drift or tooling root cause has vanished, and the financial cost of poor quality (COPQ) remains invisible until monthly scrap accounting.

ScrapVision positions ruggedized IP67 camera sensors directly at the reject chute to photograph every rejected part in isolation. Using high-speed multimodal AI vision models (Google Gemini), it classifies each defect against your plant's custom, closed-set defect catalogue in real time, computes the exact euro cost of the scrap event, and pushes actionable quality metrics to your MES and ERP dashboards via MQTT and OPC UA.

---

🚀 Key Capabilities & Architecture

1. Automated Reject Chute Visual Classification

• Chute-Mounted Inspection: Captures high-resolution imagery of rejected parts in isolation as they enter the chute, eliminating background factory clutter.
• Multimodal Defect Classification: Classifies defects against your specific defect catalogue (e.g. Flash, Sink Marks, Warpage, Short Shots, Burn Marks) with explicit natural-language visual rationales.
• Closed-Set Catalogue Enforcement: Defect types are strictly validated against pre-approved plant taxonomies. Unknown or hallucinated codes are rejected outright to ensure pristine data integrity.

2. Deterministic Cost-of-Poor-Quality (COPQ) Engine

• Financial Grounding: Costs are computed strictly from configured defect unit costs (basis: defect_type_unit_cost). The AI model classifies defect geometry; it never invents pricing figures.
• Live Financial Visibility: Real-time financial loss tracking segmented by production line, defect class, machine tool, and operating shift.

3. Confidence-Gated Human Review Queue

• Active Quality Routing: Predetermined confidence thresholds (e.g. 85%) automatically route ambiguous classifications to quality supervisors rather than quietly corrupting production metrics.
• Auditable Quality Preservation: Reviewer corrections update defect classifications to ai_corrected while immutably preserving the original prediction, creating a continuous feedback loop for active model retraining.

4. Privacy & Operator Safety by Design

• Chute-Restricted Field of View: The camera is optically and physically constrained to the reject station.
• Zero Facial Recognition: No employee biometric surveillance, facial recognition, or operator tracking features—fully compliant with GDPR and European Works Council (BetrVG) privacy mandates.

5. Factory-Optimized Interface & Industrial Integration

• Light-Mode Only Industrial UI: Engineered with high contrast specifically for plant-floor tablets, overhead monitors, and mezzanine workstations under high-ambient factory lighting.
• Industrial Telemetry Sync: Real-time integration with PLCs, MQTT brokers, and OPC UA servers for synchronized machine telemetry and MES dispatching.

---

🏭 Target Manufacturing Applications

• Automotive Tier-1 Suppliers: Stamped metal parts, plastic trim, welded subassemblies, and precision chassis components.
• Plastics & Injection Molding: Rapid identification of flash, short shots, sink marks, jetting, and thermal burn marks.
• Electronics & Electromechanical Assembly: Solder bridging, misaligned connectors, damaged housings, and missing fasteners.
• Continuous Quality Improvement (CI): Process engineers correlating defect timestamps with injection pressure, mold temperature, and cycle times.

---