Data Augmentation in Convolutional Neural Networks for Channel Operating Margin Classification

The application of Deep Neural Networks has exploded in different fields. However, they require large amounts of data to be generated and curated. This process is time consuming; this is especially true in signal integrity applications where 3D-simulations are computationally intensive due to the use of electromagnetic software solvers for complex geometries. For example, signal integrity time-domain analysis of channels can take considerable amount. Although there are now signal and power integrity public databases, there are few databases specifically tailored for channel operating margins, where eye diagrams are extensively used. In a prior paper, we proposed a method using deep neural networks to determine when a channel passes the channel operating margin (COM) standards. While initial results were promising, however, to achieve better performance a larger amount of data is needed. In this paper, more data was obtained for the COM database through more simulations as well as increasing the volume and diversity by using data augmentation techniques. In addition, a newer CNN, Resnet 101, was used that provided better performance. Results show a more resilient and more stable method to obtain channels that pass the COM standards.