Innovative AI visual system built to train in defect recognition for industrial QA using image recognition and deep learning

CLIENT NAME
Large European automotive manufacturer
SECTOR
Automative
MISSION
Achieving high accuracy for recognition of defects when deep learning training has reached a valid plateau depending on defect task and environment.

ABOUT THE CLIENT

The Client is a European division of Japanese automobile manufacturer specialising in the manufacturing and distribution in the UK market. The Client is the 2nd largest UK car manufacturer producing over 500 000 cars early and employing 7000 staff directly just in the UK.

KEY ACHIEVEMENTS

  • Reaching 98% accuracy for recognition of defects when deep learning training has reached a valid plateau
  • Achieving high percentage of accuracy on clearer subjects such as holes and missing badges or where conditions are more easily optimized to reduce false positives – e.g., low light conditions
  • Our client has actioned plans to roll out the AI Visual System factory-wide, encompassing a range of QA processes in different environments

THE CHALLENGE

  • The major challenge / innovation has been in refining to adapt to real industrial environments, ironing out false positives and environment variables, with the objective of producing a stable, scalable, and performant system. Challenges include working within processing parameters (power required for images is huge, optimisation key), generating sufficient test samples for training.
  • Software currently being adapted to enhance visuals to show where defects are on vehicle (can be difficult to spot exactly where a dent is on otherwise featureless panels)
  • Bottleneck in image capture-to-data pre-processing, to be addressed with plans for point-and-click automation.

THE SOLUTION

The success of the project has been in the development of software that works to facilitate both principles to build a stable, scalable system. To reach our goal to identify rejects and remediate requirements early in the manufacture process to reduce costs we have taken actions to:

  • Adapt and fit industrial environment (factory stations) for QA monitoring and provisioning of test images
  • Build for creation, training, and refinement of defect models
  • Apply automation process to live environment
  • Automate QA to detect and flag defects to 98% accuracy
  • Deliver an ability to establish and train new defect models and adapt to different environments
  • Industrialize and deploy factory-wide

THE TECHNOLOGIES

  • Microsoft Visual Software
  • Microsoft SQL Server (DB engine)
  • C# coding
  • Google TensorFlow (open source)
  • HTML — CSS — Vanilla JS
  • Mosquitto (MQTT) transmission protocol transmits between separate layers, providing a lightweight method of carrying out messaging using a publish/subscribe model
  • Desktop application interface
  • Deep learning software

THE RESOURCES

  • 1 Project Manager
  • 2 AI Data Scientists
  • 1 Business Lead
  • 6 Business Pilots
  • 1 Front End/Back End Developer

THE RESULTS

The thorough audit performed by the Infotel UK team gave the Client an objective perspective on their architecture and the maturity of their approach to the implementation of innovative AI system. In addition to enabling the successful implementation of new approach, the provided recommendations allowed the Client to make several other optimisations to their solution.

Mapping processes to IT resources allowed for more detailed auditing and reporting, which contributed to greater visibility and accountability. The final phase of the roadmap established multiple QA processes across the plant to guarantee long-term high performance and business continuity.

The AI Visual System involved the collaborative work of Infotel developers, AI data scientists, business managers and QA experts to capture and refine banks of images for training in the automated recognition of defects – from dents & holes to overpainting & missing elements. The software has been developed to allow for adaptation to pretty much any industrial process, building in variables to:

  • Fine-tune to different environments and station sensors
  • Encompass a range of defect types
  • Apply alternative deep learning models
  • Enable the future potential to extend from object-detection to anomaly detection and optical character definition

The level of Artificial Intelligence and deep machine learning expertise provided by Infotel UK allowed the Client to perform the adaptation of new system seamlessly and with a strong focus on the automation of defect recognition for industrial QA using image recognition and deep learning.