Delta Electronics Is Redefining Production Lines and Industrial Inspection With Digital Twins and Synthetic Data

The process of creating a digital twin begins with aggregating and visualizing 3D assets from the environment. Delta uses multiple 3D applications, including Autodesk 3ds Max, FlexSim, and Visual Components to model and simulate their production line. Historically, it would be nearly impossible to bring data from these modeling and simulation tools into one environment without time-consuming data transfer, model decimation, and interoperability hurdles. Even then, once the data is aggregated, if any changes are made to models or processes in the source 3D applications, the process of aggregating starts again to update any changes.

With Omniverse, Delta can connect their various applications and data together using a unified asset pipeline, allowing their teams to visualize and collaborate on the entire output in a single environment. This is all made possible by OpenUSD and additional connectors and extensions that allow third-party tools to feed live data into a USD stage through a Nucleus server. 

Once all the assets are aggregated in the digital twin, the Delta team can run simulations and identify issues early in the design and reconfiguration process, well before making changes to their physical production lines. This is key to avoiding costly downtime and change-order requests.

"When apps connect to Omniverse, this allows all our files and data to be synced simultaneously, so team members can see updates in the USD Stage and collaborate in real time," said Ares Chen, PSBG General Manager at Delta Electronics. "USD enables seamless collaboration with each team member, so we can design a production line faster and more efficiently than before."

Delta uses computer vision for automated inspection in the final assembly, checking for issues like missing components or misaligned screws. AI-assisted Automated Optical Inspection (AOI) significantly accelerates the inspection process, enabling Delta to detect defects early and minimize the need for manual intervention.

Delta faced a challenge in training AI models because collecting and labeling data manually took a lot of time–up to two days for 1000 images. To speed things up and cut costs, they switched to using synthetic data with Omniverse Replicator. Now, in just 10 minutes, they can generate the same amount of labeled images that would take two days manually.

The team also achieves a 90% accuracy with synthetic data, similar to real data. The big win is that they reached this accuracy level in just a tiny fraction of the time it usually takes to collect real data—now done in one-hundredth of the time. This not only accelerates AI training, but also makes the process more cost-effective and efficient for computer vision applications.

Delta also uses NVIDIA Isaac Sim—a scalable robotics simulation application—to accurately simulate the performance of the model and detect the position of a tray in the production line.

By creating the digital twins in Omniverse, Delta can minimize downtime and reduce risks while reoptimizing production lines.