Automating Metal Corrosion Detection with AI-Powered Visual Classification

Quick Summary

Challenge

Identifying metal coupon types and corrosion patterns was a manual process that took over a week & delayed treatment decisions in water systems.

Solution

Tatras Data built a mobile-friendly ML pipeline that detects coupon types from images and classifies corrosion types, reducing dependency on lab analysis.

Result

95% classification accuracy
7-day manual workflow → 5 minutes.

Tech Stack

AI: Fine-tuned CNN models | ML: Transfer learning Image clustering for corrosion types Visual embedding clustering | Data & Retrieval: Mobile image ingestion | Dev: TensorFlow Dataiku Jupyter Notebooks | Viz: Classified images with next steps

The Challenge

Monitoring corrosion in industrial reverse osmosis systems is critical, but identifying the exact metal coupon type and corrosion pattern relied on lab-based analysis. The process involved manual tagging, visual inspection, lab testing, and long wait times. For every new corrosion event, technicians had to:

Identify the metal manually
Document surface-level changes
Send it for lab validation
Wait for reports before acting

This delayed treatment adjustments, increasing both risk and cost.

A Day in the Life: Before Our Solution

Every week, engineers collected metal coupons from installed RO systems.Back at HQ, someone would manually examine them under a microscope. If corrosion was present, the coupon was sent for lab analysis. The lab would return the results in 7 days turnaround, often longer. In the meantime, the field team waited. The client waited. The corrosion often worsened. There were no fast insights. No clear next steps. Just slow decisions in high-stakes systems.

Pain Points:

Lab analysis created 7+ day delay in detecting corrosion issues
Coupon types had to be manually identified from visual clues
No scalable way to flag patterns across different metal types
Field teams lacked real-time decision support
Scaling this workflow globally was impossible

Solution

1. Core Innovation

Tatras Data developed an ML-driven visual classifier that turns image captures into instant decisions:

Mobile-captured images are uploaded via an app.
A fine-tuned CNN model classifies metal coupon types with 95%+ accuracy
Corrosion detection runs as a second layer using unsupervised visual clustering.
Clusters are reviewed and labeled by experts to bootstrap faster future recognition.

2. Key Features

Coupon type detection using transfer learning on image datasets
Visual clustering to group corrosion patterns by similarity
Fine-tuned model that generalizes across lighting and camera noise
Semi-supervised annotation workflow for faster labeling cycles

3. Workflow Integration

Captured images are uploaded to the system via mobile or web. The classifier returns coupon type and corrosion diagnosis in minutes, with explainable visuals and next-step prompts for treatment or escalation.

Outcomes

✅ 95%+ accuracy in coupon classification

⏱️ 7-day manual process reduced to under 5 minutes

⚙️ Deployed as a Dataiku flow for easy reuse and scaling

🧪 Visual cues used to recommend corrosion-specific treatments

🧠 Built-in clustering accelerates expert review