Failure Mode Classification for Prognostics and Health Monitoring of Electronic-Systems under Mechanical Shock
Type of Degreedissertation
MetadataShow full item record
Electronics have become an integral part of most systems and subsystems used in various fields such as avionics, defense, space exploration, manufacturing, household appliances, health care (implantable biological devices such as pacemakers, defibrillators), portable electronics (laptops, PDA’s, cell phones) etc. The electronics experience accidental drop and shock at various stages of their life-cycle i.e. manufacturing, transportation, and deployment in field. In this research work, strain based techniques are developed and implemented on test vehicles with ball grid array packages (BGA’s) to address various aspects of health monitoring such as damage detection, diagnostics, study of damage trends, fault mode classifications and isolation. The methodologies developed in present work are completely independent of continuous monitoring of daisy chain resistance. The most common damage quantification techniques currently in use in electronics are, continuous monitoring of daisy chain resistance, and use of built in self-test (BIST) which are purely based on reactive failure. Auxiliary devices such as fuses and canaries are also used to detect damage in electronic systems. Though these techniques are significantly efficient in damage diagnostics, they do not provide any prior knowledge on damage initiation, damage progression, failure mode identification and isolation of dominant fault modes. This can lead to catastrophic failure and shutdown of the system under consideration. Prognostics and health monitoring framework developed in this research work focuses on addressing leading indicators of failure in pre-failure space, hence prior knowledge about various aspects of system state and its health can be known and monitored in real time. Damage detection and study of damage trends is implemented on feature vectors derived from spectral analysis, and joint time-frequency analysis of transient strain histories obtained from test assemblies under drop and shock. Statistical pattern recognition techniques are used for quantification of damage initiation. Autoregressive models are used for studying damage trends on feature vectors derived from the above mentioned domains. Once damage is detected in the electronic assemblies, various data driven techniques for fault mode classification and isolation are developed. The focus is on developing methodologies which can address damage initiation and draw inference on its progression as well as identification of dominant fault modes. Before fault mode classification is addressed, experimental and simulation data sets are procured from test assemblies under study. Explicit finite element simulations of pristine state of the system, and the system with various dominant failure modes such as inter-connect cracking, complete inter-connect failure, die cracking, chip de-lamination and part fall off etc. are performed. Experimental data sets are procured from mechanical drop tests on assemblies according to JEDEC drop standards. The time domain data sets are used for extracting shock features in time-frequency domain. De-correlation of the feature space from time-frequency analysis of the data sets is performed using statistical classifier: Karhunen-Loeve transform. An artificial intelligence based framework for fault monitoring is also developed for test assemblies. Unsupervised neural nets based on a self-organization algorithm are used for detection and isolation of failure modes. Supervised neural trainers in conjunction with nonlinear mapping technique, such as sammon’s mapping, are designed for real time monitoring of the system state. Hard parity between different failure modes in the feature space is achieved. Early classification of failure modes in assemblies under drop and shock is novel. The results of classification of various dominant fault modes are statistically validated using a number of statistical tests: multi-variate analysis of variance, box M-test, and Hotelling’s T-square along with paired t-test and principal component similarity factor. Failure analysis of the test assemblies is also performed by studying experimental cross-sections of the failed packages. Currently there is no prognostics and health monitoring (PHM) decision framework for electronics at component or system level, hence there is a need for developing a platform of techniques which can address various health monitoring issues in systems with varying degree of complexity. The techniques developed in this work are scalable to system level reliability. These techniques are data driven in nature, hence provide capability to tailor need for reliability depending on the application. The framework developed in this work can help schedule proactive maintenance and prevent catastrophic shutdown of the system. The objective of prognostics and health monitoring (PHM) is 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 advancement towards achieving these objectives.