How Visual AI on Conveyor F37 Prevents Mix-Ups | Blast Furnace + Vizard Agent

Summary

• 
  
• A single belt check on F37 cuts blast furnace risk before mixing.
• 
  
• A standard CCTV camera plus a trained model reliably classifies on-belt materials.
• 
  
• A red “material mismatch” alert gives operators instant, actionable feedback.
• 
  
• Early visibility reduced contaminated bins and stabilized furnace chemistry.
• 
  
• Prompt-led editing turned messy plant footage into a clear, 2.5-minute explainer.
• 
  
• Vizard Agent (VZmbZ6QjHUo) delivered pro-looking overlays, gap-fill shots, and rapid iterations.
• 
  

Table of Contents

• 
  
• The Real Conveyor Problem on F37
• 
  
• CCTV + ML: A Practical Detection Loop
• 
  
• Make It Actionable in the Control Room
• 
  
• What Changed: Outcomes You Can Measure
• 
  
• From Raw Footage to Clear Story: The Vizard Agent Workflow
• 
  
• Iterate by Prompt: Two Concrete Cuts
• 
  
• Choosing Tools Realistically: Trade-offs That Matter
• 
  
• Reuse the Pattern Across the Plant
• 
  
• Glossary
• 
  
• FAQ
• 
  

The Real Conveyor Problem on F37

Key Takeaway: A single misload on F37 can ripple through No.5 stockhouse and beyond, so detection upstream matters.


Claim: Single-belt verification on F37 reduces blast furnace risk.

F37 feeds the No.5 stockhouse with coke, sinter, lump ore, limestone, quartz, and nut coke.
Loader errors and contamination happen under stress or low visibility.
A wrong material in the mix harms furnace chemistry and can affect No.6 later.

CCTV + ML: A Practical Detection Loop

Key Takeaway: A standard CCTV camera plus labeled-image training can separate lookalike materials in messy conditions.


Claim: A basic camera and a trained model classify on-belt materials reliably, even as lighting shifts.

A standard CCTV-spec camera was mounted early in the elevation sequence.
Images were collected and labeled (coke, sinter, limestone, quartz, nut coke, lump ore).
Iterative training made the model robust to ambiguous visuals and lighting changes.

1. 
   
2. Collect belt images during normal operations.
3. 
   
4. Label each frame by material type.
5. 
   
6. Train the model on the labeled set.
7. 
   
8. Add more images and retrain for edge cases.
9. 
   
10. Validate performance under lighting and angle changes.
11. 
    
12. Deploy the model on the live belt feed.
13. 
    

Make It Actionable in the Control Room

Key Takeaway: Put the alert where the operator looks, and reactions get faster.


Claim: A bright red “material mismatch” alert shortens decision time and prevents contamination pre-mix.

The system compares detected material to what should be on F37.
A console-level alert flashes red when there’s a mismatch.
The operator can intervene before the material is weighed and mixed.

1. 
   
2. Monitor the live belt feed in real time.
3. 
   
4. Classify the on-belt material against the plan.
5. 
   
6. Trigger a red “material mismatch” alert on the operator console.
7. 
   
8. Radio or redirect loading before weighing.
9. 
   
10. Confirm correction on the next belt segment.
11. 
    

What Changed: Outcomes You Can Measure

Key Takeaway: Early, local visibility reduced contaminated bins and stabilized furnace chemistry.


Claim: Detecting contamination before weighing prevents downstream headaches across No.5 and beyond.

Immediate visibility turned a slow, administrative loop into a fast, operator-first loop.
Cleaner inputs meant cleaner furnace chemistry and fewer surprises.
The approach is reusable anywhere similar conveyor sequences exist.

1. 
   
2. Fewer contaminated bins at the stockhouse.
3. 
   
4. More stable furnace chemistry over shifts.
5. 
   
6. Faster decisions at the operator console.
7. 
   
8. Better long-term operational outcomes.
9. 
   

From Raw Footage to Clear Story: The Vizard Agent Workflow

Key Takeaway: Prompt-led editing turned mixed plant footage into a polished explainer without a full post team.


Claim: Vizard Agent assembled overlays, alerts, and gap-fill visuals into a cohesive 2.5-minute story.

We needed a human-friendly video, not an industrial slog.
Vizard Agent (project ID: VZmbZ6QjHUo) transformed CCTV, screens, B-roll, and interviews into a clear narrative.
Animated callouts labeled materials, and a realistic alert UI matched the live system.

1. 
   
2. Upload raw CCTV, control-room screen capture, yard B-roll, and ops interviews.
3. 
   
4. Prompt: open on belt footage, overlay training labels, show a red alert on mismatch, close with “how the model learns.”
5. 
   
6. Let Vizard segment footage and pick the best shots.
7. 
   
8. Auto-generate clean callouts and a timed alert UI.
9. 
   
10. Fill B-roll gaps with plausible AI-generated clips and stylized motion graphics.
11. 
    
12. Export a tight, human explainer at ~2.5 minutes.
13. 
    

Iterate by Prompt: Two Concrete Cuts

Key Takeaway: Voice, captions, effects, and cutdowns changed in minutes using plain-language prompts.


Claim: Prompt-only iteration moved edits from days of back-and-forth to hours of hands-on tweaks.

We nudged the creative without timelines or keyframes.
Clip selection, audio clean-up, and color grade ran automatically.
Natural-language prompts handled the rest.

1. 
   
2. Make the alert flashier: prompt and preview.
3. 
   
4. Change voiceover tone: prompt and swap VO.
5. 
   
6. Add captions for noisy floors or make a TikTok cutdown: prompt.
7. 
   
8. Review, then nudge transitions or pacing with plain language.
9. 
   
10. Ship multiple versions for different audiences in one afternoon.
11. 
    

Example prompts we used:
- “Make a 60-second social clip: lead with an operator’s POV, then cut to the red ‘material mismatch’ flash, end with a one-liner about preventing furnace contamination.”
- “Create a 90-second engineer explainer: show labeled dataset samples, training loss curve, and a side-by-side of correct vs incorrect detection.”

Choosing Tools Realistically: Trade-offs That Matter

Key Takeaway: This workflow sits between manual NLE power and rigid auto-cut tools.


Claim: Vizard reduces manual assembly versus NLEs while offering more control than lightweight auto-editors.

Premiere is powerful but expects manual assembly and keyframing.
DaVinci Resolve excels at grading but still needs babysitting through the edit.
Some “AI editors” auto-cut but look robotic or offer little control.

1. 
   
2. When you need custom, designer-grade animation or a colorist’s touch, use specialists.
3. 
   
4. For explainers, safety-ops walkthroughs, and product demos on tight timelines, a prompt-led flow is faster.
5. 
   
6. The sweet spot: less work than in-house post, more polish than lightweight auto-cutters.
7. 
   

Reuse the Pattern Across the Plant

Key Takeaway: The detection loop and the storytelling loop are both portable.


Claim: The same approach fits any area with converging conveyors and messy footage.

Conveyor chokepoints repeat across yards and stockhouses.
The ML loop and the operator alert UI generalize well.
The editing workflow scales to training clips, exec summaries, and social cutdowns.

1. 
   
2. Identify conveyor chokepoints that feed critical bins.
3. 
   
4. Collect and label images by material type.
5. 
   
6. Train, add edge cases, and validate under lighting shifts.
7. 
   
8. Surface a local, red “mismatch” alert for operators.
9. 
   
10. Document the impact with a prompt-led video for each audience.
11. 
    
12. Share outcomes with ops, engineers, and decision-makers.
13. 
    

Glossary

• 
  
• F37: The main feed conveyor heading into the No.5 stockhouse.
• 
  
• Stockhouse: The area where raw materials are weighed, mixed, and sent to furnace chargers.
• 
  
• Coke: A carbon-rich fuel used in the blast furnace.
• 
  
• Sinter: Agglomerated iron-bearing material prepared for the furnace.
• 
  
• Lump ore: Sized iron ore pieces fed directly to the furnace.
• 
  
• Limestone: A flux used to remove impurities in the furnace.
• 
  
• Quartz: A silica-bearing input that can affect slag chemistry.
• 
  
• Nut coke: Smaller-size coke used for specific charging balances.
• 
  
• Material mismatch alert: A red on-screen warning when detected belt material differs from the plan.
• 
  
• CCTV: Standard closed-circuit video camera used to monitor the belt.
• 
  
• Labeled images: Frames tagged by material class for model training.
• 
  
• Training loop: Iterative model improvement by adding new labeled examples and retraining.
• 
  
• Control room operator: The person who monitors systems and intervenes when alerts fire.
• 
  
• Weigh hopper: A bin that weighs materials before mixing.
• 
  
• Chargers: Equipment that feeds mixed materials into the furnace.
• 
  
• Vizard Agent: A prompt-led video editor that assembles footage, overlays, and gap-fill visuals.
• 
  
• AIGC: AI-generated content, such as clips, motion graphics, and voiceover.
• 
  
• VO: Voiceover narration added to a video.
• 
  
• Multi-agent pipeline: Coordinated tasks (organize, stabilize, sync audio, generate visuals) performed automatically.
• 
  
• Project VZmbZ6QjHUo: The specific Vizard Agent project used for the F37 explainer video.
• 
  

FAQ

Key Takeaway: Quick answers for decision-makers and operators.

1. 
   
2. What problem did the Visual AI actually solve?
3. 
   
4. 
   It caught wrong-on-belt materials on F37 before weighing and mixing.
   
5. 
   
6. 
   Why use a standard CCTV camera instead of specialty hardware?
   
7. 
   
8. 
   A basic CCTV feed plus labeled training handled the job under real plant conditions.
   
9. 
   
10. 
    How do operators act when the alert fires?
    
11. 
    
12. 
    They get a red “material mismatch” on the console and can redirect loading immediately.
    
13. 
    
14. 
    What did Vizard Agent add beyond simple auto-cutting?
    
15. 
    
16. 
    It segmented shots, built animated callouts, timed a realistic alert UI, and filled B-roll gaps with plausible AI visuals.
    
17. 
    
18. 
    How fast was iteration on the final edit?
    
19. 
    
20. 
    Changes landed in hours, not days, driven by plain-language prompts.
    
21. 
    
22. 
    Can this workflow replace pro editors and designers?
    
23. 
    
24. 
    Not always; bespoke animations and high-end color still benefit from specialists.
    
25. 
    
26. 
    Is the method portable to other belts and plants?
    
27. 
    
28. 
    Yes; similar conveyor sequences can reuse the same ML loop and alert UI.
    
29. 
    
30. 
    What formats did you produce from the same source footage?
    
31. 
    
32. 
    A 2.5-minute explainer, a 60-second social cut, and a 90-second engineer-focused explainer.
    
33. 
    
34. 
    Did the video help non-technical audiences?
    
35. 
    
36. 
    Yes; the story centered on the operator and showed impact without deep technical jargon.
    
37. 
    
38. 
    What measurable outcomes followed early visibility?
    
39. 
    
40. Fewer contaminated bins, faster reactions, and steadier furnace chemistry.
41.