Precision Defect Detection Case Study


By Mindtrace
3-minute read

Project Overview

Our client is a leading European automotive company. This OEM organisation produces automotive parts and consequently requires high levels of quality control, focusing on the welding performed on the frame of the chassis.

After discussions with Mindtrace, it was determined that we could support their Defect Inspection process and aim to significantly increase the accuracy of their quality inspection process.


Mindtrace faced the following project challenges.

Minimal amounts of labelled training data

Mindtrace's AI Modelling used an average of 50 images.

Significant variability between welding lines

Requiring multiple camera angles.

Defects are unseen during the training stage

There are no images of major or minor defects available in the training data set.

Mindtrace Process


Mindtrace aimed to create an AI solution using our Brain-Sense™ technology for unsupervised anomaly detection, which could differentiate between serious and minor defects, ignoring minor defects upon request.

Working alongside the client, it was determined that two camera angles would be used to capture 4 different welding lines on the chassis, like the example on the right.

The model underwent a training phase prior to testing, in which an average of 50 training images per inspection point.

A Car Chassis from multiple angles showing defect focus areas


After a day, the AI model was ready for testing. The model processed 177 testing samples for each of the 4 different welding lines.  

Across the 4 Welding Lines, the AI model returned an overall accuracy of 97.19%, with 3 out of the 4 welding lines being 100% accurate. The welding line that returned an accuracy of 90.6% was due to the misclassification of healthy images as defective. However, the model still detected all of the defective images correctly. The reason for the false positives being present was due to a subjective healthy annotation selection.

Welding Line 1 – 100% Accurate. 

Welding Line 2 – 90.6% Accurate. Misclassified 5 healthy images as defective.

Welding Line 3 – 100% Accurate. 

Welding Line 4 – 100% Accurate. 

Once the data is captured in the production line, Mindtrace’s solution generates anomaly detection alerts and reports. The generated reports inform operators to take action and can be stored for further traceability. Mindtrace’s models can be deployed and fine-tuned on-site as SoC solutions with cloud support or specialised inference servers.

Solution Summary

This one-week project successfully created an AI solution for unsupervised anomaly detection, producing effective results of 97%+ accuracy on the first round of testing, and successfully detected 100% of the defects.

The model was also capable of differentiating between major and minor defect. If required, the model could filter out minor defects.

Over time, this model would improve further as additional samples are processed by the AI Brain, producing a comprehensive, unsupervised solution for the client.

results in numbers
Day Project
Average Learning Samples
Testing Data Samples
0 %+

See what our Brain-Sense™ Technology can do for you

Contact Mindtrace using the form below.