Welding Anomaly Detection – PatchCore PoC

This repository presents an industry-oriented Proof of Concept (PoC) for automated weld inspection using unsupervised computer vision techniques. The goal is to improve process reliability and reduce manual inspection effort in manufacturing environments.

Problem Statement

Manual weld inspection is time-consuming, subjective, and difficult to scale. In real manufacturing settings, labeled defect data is scarce and new defect types frequently emerge, making supervised approaches hard to maintain.

Solution Overview

This PoC uses an unsupervised anomaly detection approach:

• Train PatchCore only on normal (good) weld images
• Detect deviations from learned normal patterns as anomalies
• Produce both anomaly scores and pixel-level explainable heatmaps

This design allows detection of unseen defects without relying on explicit defect labels.

System Pipeline

Weld Image → Preprocessing → PatchCore Model → Anomaly Score + Heatmap → Decision

Model & Methodology

PatchCore learns feature embeddings from normal weld images and builds a compact memory representation. During inference, weld images are scored based on their deviation from learned normal patterns, while heatmaps highlight localized defect regions for explainability.

Tech Stack

• PyTorch
• Anomalib (PatchCore)
• OpenCV
• NumPy

Current Status

• Model training completed
• Inference pipeline validated
• Threshold-based decision logic implemented
• Demo-ready qualitative results generated

Demo Results

Normal Weld (Clean)

Defective Weld (High Confidence Anomaly)

Future Scope

• Real-time camera or video-based inspection
• Adaptive threshold calibration per weld type
• MLOps integration for monitoring and retraining
• Dashboard and BI integration for quality monitori