Prognostics and Remaining Useful Life Estimation of Electronic Assemblies under Shock and Combined Temperature-Vibration Experiments

Abstract

The electronics components experience different high-stress loading environments such as drop, shock, and vibration at various life-cycle stages, i.e., manufacturing, transportation, and deployment in the field. In this research work, resistance and contact strain (strain gauge) based measurements from different solders materials and package architectures are used to predict the failure and also develop methods and algorithms for prognostic health monitoring, damage diagnostics, remaining useful life estimation, and failure mode classifications. By utilizing the data generated from sensors in real-time, the system aims to detect signals for potential failures before they occur by using statistical data processing, machine learning, and deep learning methods. Consequently, it helps address the issues by notifying operators early such that preventive actions can be taken prior to the failure. In this study, the Printed Circuit Board (PCB) with solder material consisting of SAC305, SAC105, and Tin-Lead daisy chain CABGA packages are subjected to different temperatures (25oC, 55oC, 100oC, and 155oC), vibration acceleration levels (5g and 10g) and drop/shock (1500g to 10,000g) loads. The test boards consist of a multilayer FR4 configuration with JEDEC standard dimensions. Strain signals and resistance measurements are acquired from different PCB locations during the experiments. The strain signals are processed in both the time and frequency domain, and the various characteristics of the same are compared for before and after failure signals. The high dimensional data is statistically analyzed using principal component analysis (PCA) as the data reduction technique to understand the patterns produced. The frequency-based study of the strain signals involved different Time-Frequency analysis (TFA) and Joint Time-Frequency analysis (JTFA) in conjunction with pattern recognition techniques to predict the failure of the package on the PCB. The failure prediction is modeled using logistic regression with principal components from frequency and instantaneous frequency matrices. The variations in the feature vectors for different conditions of temperature and loads and the correlation of the same are studied to understand the changes in the feature vectors with various operating conditions. The two major feature vectors that can predict the failure include frequency and spectral content from 500 Hz to 2000 Hz of the strain signal and the instantaneous frequency of the whole signal. The residual of the autocorrelation function of frequency and instantaneous frequency matrices shows behavior that can predict the packages' failure on the PCB irrespective of different conditions of temperature and acceleration levels during vibration analysis. A multivariate time series deep learning technique called as long short-term deep learning technique was used to predict the remaining useful life of the packages under different and varying vibration conditions. The reliability analysis and health monitoring of PCBs subjected to drop and shock loads were carried out by statistical analysis on the strain signal's time and frequency domain components. The strain signals acquired from the corner position of the packages on the backside of the PCB were able to show patterns that can predict failure. Various statistical approaches on the time and frequency domain components were used to identify the changes in the package's damping characteristics. The feature vectors in predicting failure were selected based on the variation in the time-frequency data and the logarithmic decrement of the strain signals. The feature vectors from the single component test board were verified for the multi-component board, and the effectiveness in predicting the package's failure was also estimated. The patterns produced from the frequency components, including and excluding the test board's natural frequency, gave distinctive separation of before and after failure data points. The feature vectors of the multi-component board were able to identify the package's failure corresponding to the strain signal and neighboring packages as well. The residual of the cross-correlation function of frequency matrices shows behavior that can predict the packages' failure on the PCB irrespective of different conditions of acceleration levels during shock experiments. The feature vectors were verified at different acceleration levels of shock loads (1500, 3000, 5000, 7500, and 10,000) g for solder materials of SAC305, SAC105, and Tin Lead solder PCBs. The prognostics and health monitoring (PHM) decision framework for electronics at the component or system level can address various health monitoring issues in systems with varying degrees of complexity. The techniques developed in this work are scalable to system-level reliability. These techniques are data-driven in nature, hence providing the capability to tailor the need for reliability depending on the application. The framework developed in this work can help proactive schedule maintenance and prevent a catastrophic shutdown of the system. Prognostics and health monitoring (PHM) objectives are to make critical systems cost-effective, safe, and reliable. This can be achieved by enabling pre-planned maintenance, forestalling failure, and making systems operationally more reliable. The work presented in this thesis is an advancement toward achieving these objectives.