Machine Learning Based Data Validation for Signal Integrity and Power Integrity Using Supervised Time Series Classification

A novel approach for the validation of data in signal integrity and power integrity using machine learning is proposed. This approach presents an alternative to the feature selective validation method outlined in the IEEE Standard 1597.1 for the validation of computational electromagnetics, computer modeling and simulations. The proposed approach focuses on replicating the human visual assessment by using data collected and labeled by expert engineers to train time series classification networks that predict the degree of agreement between two curves. The trained networks are then used for the systematic and automated validation of 1-D datasets. The performance and suitability of this approach for systematic data validation is evaluated and discussed. The trained network surpasses the single human subjects in predicting the expert opinion with an accuracy higher than 70%.

I. Introduction

In The domain of electromagnetic compatibility (EMC) engineering, signal integrity (SI) and power integrity (PI) play an important role in the design of printed circuit boards (PCBs). The use of computational electromagnetics (CEM) and simulation tools is part of the design process. The investigation of EMC related problems relies on physics-based (PB) modeling using measurements or simulations and full-wave (FW) solvers. Since the problems faced are complex, it is important for the engineer to ensure that the used models deliver results accurate enough for the faced problems. The validation of data is the process of examining the quality and accuracy of collected data. For the SI/PI engineer, it is often about ensuring a certain confidence in the modeling techniques by comparing the simulation results with each other or with measurements and accepting some variation between the two. The knowledge and experience of the engineers involved is the major factor in deciding whether the comparison is acceptable for the use-case and field [1]. This process is conventionally subjective in nature, see Fig. 1. For instance, the SI/PI-Database, freely accessible at the homepage of the Institut für Theoretische Elektrotechnik at the Hamburg University of Technology (TUHH) (https://www.tet.tuhh.de/en/si-pi-database/), contains structures simulated with a PB model [2]. The data are mainly collected for machine learning (ML) applications. The generation of data is partially automated to speed up the process [3]. To ensure conformity, the engineer proceeds to validate the generated data against a FW solver before the upload, which can prove to be tedious and time-consuming.