Degradation Prediction of Electronic Packages using Machine Learning

Abstract

The piezoresistive silicon based stress sensor has the potential to detect a precursor for the Prognostics and Health Management (PHM) implementation in automotive electronics. One solution to enforce reliability in automotive electronics is the use of Machine Learning (ML). One or more physical parameters are being monitored, and algorithms are used to illustrate the health state and predict remaining useful life based on the current and past health information. Piezo-resistive stress sensors are employed to measure the internal stresses of electronic packages, an Acquisition Unit (AU) to read out sensor data and a Raspberry Pi as PHM server to perform evaluation. Accelerated tests in air thermal chamber are performed to get time series data of the stress sensor signals, with which we can know better about how delamination develops inside the package. In this study stress measurements are performed in several electronic packages during the delamination. The delamination is detected by the stress sensor due to the continuous change of the stiffness and the local boundary conditions causing the stresses to change. Moreover, the stress change in multiple cells can give more information regarding the delamination such as the location and its state. Data preprocessing methods to remove outliers and filter raw measurement results, and feature extraction methods to capture only meaningful information by reducing the data are chosen and applied to raw data. A logical assumption is made regarding the data behavior and delamination state, based on data analytics and with Scanning Acoustic Microscope (SAM) images confirmed the delaminated area. FEM simulation are used to provide a qualitatively physical explanation of the stress change due to the delamination. A prognostic model using neural network is trained to estimate the degradation grade. Back Propagation Neural Networks are chosen to provide a fast and quick training for the mechanical stress data.