Welding Joints Classifications using Attention Networks-Case Study

Client Overview

This Vision AI project was built for a client that develops Safe Electrical Systems for industries, infrastructure, energy, and even commercial/residential projects. They engineer a wide range of electrical products and solutions like Electrical Cabinets, Fasteners, Heat Tracking systems, RCC Steel connections, Thermal Management Systems and Electrical Protection for facilities. Their large-sized business is well recognised by industry-leading brands for quality, reliability and innovation.

The manufacturing process of several products developed by the client required the use of welding to connect different joints and product parts. Welded joints are key to the structural integrity of these products. As a result, quality control for welded joints was a prime area of focus. The client wanted to ensure that the welded joints were free from defects and faults so that they could attain industry-high standards.

Prior to Visionify deployment, the inspection and quality check process for welding joints was completely manual and human-centered. These are the major challenges related to manual inspection and quality control:

• Low accuracy (Increased False Negatives)
• Time-taking process (Delays the entire production mechanism)
• Shortage of skilled resources (Highly-specialized welding joints inspectors are scarce)

Outcome-based requirement

The client wanted to automate the process, increase the accuracy of defect detection, speed up the process and improve product quality.

Problem-statement

To classify welding joints into ‘Bad’ or ‘Good’ types for 9 primary categories which were:

Additionally, the classification had to be determined using images and visual data for the front view, side view, bottom view, and top view of a joint

This classification enabled our client to identify damaged/poor quality joints which were filtered from the production lines before reaching the next stage of manufacturing. Visionify also provided them with the option to integrate small camera-enabled android devices, which could classify these joints as per the requirement without any massive hardware installations.

Approach and Implementation Details

Our team did not implement a Binary Classification to classify an object into simply ‘Good’ and ‘Bad’ categories; instead, our experts experimented with various classifiers, starting with RESNET18. Eventually, we implemented multiclass classification models such as VGG-16 and VGG-19 after obtaining high-accuracy test results. 

The team also utilized Attention Networks in the VGG model to get even higher accuracy. With the help of Attention Networks, our model was able to focus on important areas of an image and ignore irrelevant regions in the background and surroundings of the welding joints.

Dataset details

Initially, when the pilot model was being built, we were provided a limited number of images. To tackle this, the team enriched the dataset to add more diversity, enabling the deep neural networks to extract key visual features and utilize them for final categorization.

Dataset: 180 images

Augmented Dataset: Features: Zoom, Scaling, Translation, Adding variations in the lighting

Utilization of Attention Networks: Deep Learning model that focused on joints while ignoring irrelevant background areas

Training-Testing Ratio: Results-based classification reports of product training and testing with an 80:20 (training:testing) ratio of the images in the dataset.

The results

Even with such a small dataset, Visionify’s solution was able to perform decisively with an accuracy of 91%, which provided 91% recall rates. Even though this experiment was based on a smaller dataset – this provided the foundation for the next step of engagements with the client.

Evaluation parameters

Confusion Matrix

Visualization of Attention Networks

Clients Feedback

Our client was impressed and more than satisfied with the accuracy that we were able to provide. They acclaimed our team’s ability to build a highly-accurate model in such short notice based on a limiting dataset.

What’s Next?

If you want to automate your quality check process, increase its accuracy and improve product quality by eliminating defective products, then opt for Visionify’s factory vision solutions.