Advanced Analytics Office, Western Digital - Leveraging Deep Learning Object-detection Models to Improve the Accuracy of Defect Detection
Sameera Kodagoda Roger Yu
Country: United States
Organization: Advanced Analytics Office (AAO), Western Digital
We provide enterprise-level solutions for big data analysis, machine learning, and artificial intelligence.
Value at Scale
Most Impactful Transformation Story
High-volume manufacturing for consumer electronics depends on automation, and innovation is often needed to improve yield and quality. One of the workhorses of the industry is Automated Optical Inspection (AOI), which uses traditional inspection algorithms and high-speed tooling for defect detection. However, most inspection algorithms from AOI vendors suffer from low accuracy. Manual defect validation incurs large labor costs and is prone to quality escapes due to error/variability in human judgment.
Deep Learning object-detection models offer a robust solution to improve the accuracy of defect detection, both overkill and escape. However, there are big challenges with the MLOps, model revisioning/roll-back, and E2E deployment. When deployed to the production line, downtime due to model serving and model performance drift is not acceptable. A resilient MLOps system Is needed to manage model development, deployment, and maintenance.
With a solution designed for scaling, our factory will be able to support this automation use case, as well as many others to come.
In the Dataiku DSS Platform, we set up an object-detection common template based on code recipes with YoloV4 model and built a comprehensive ML pipeline including CI/CD/CT (Continuous Training), which we refer to as Level#2 ML Ops capability. A user can change the configuration in variables and run the scenario for model training and evaluation. After that, the model can be deployed to the corresponding API node infrastructure with the click of a button.
Dataiku provides the capability for users to easily roll back to a previous model version in case the current version is experiencing model drift. This common template can be replicated to efficiently develop models for any other object detection problem within WDC.
This project enables an E2E object-detection model w/ MLOps capabilities for one use case and provides a common template to automate other manufacturing operations with image inspection. Before this implementation in Dataiku, the typical lead time for object-detection model development and deployment project would take approximately one to two quarters. Even then, the solution is sub-optimal with no model monitoring, retraining, version control, etc.
Now the whole process only takes one to two weeks for a new use case of similar nature. It has also enabled data scientists with limited exposure to full stack ML engineering concepts and IT infrastructure to build and deploy Computer Vision solutions, democratizing AI in the WDC factory environment.
Value Brought by Dataiku:
Reduce model development and deployment lead time from months to weeks.
Enable junior data scientists to build complex applications like a senior data scientist/ ML engineer.
Centralization and standardization of the mode reduces the risk of "reinventing the wheel" across different teams and use cases. Data scientists can collaborate to improve the model in the template project to enhance the model performance.
Better Fail-Over, SLA support compared with the self-deployed model because of the standardization of the K8S workloads.