Reliability analysis of electronic boards based on field data

The rapid developments in electronics, especially in the last decade, have elevated the importance of electronics reliability More complex electronic systems and their widespread usage, spanning almost all large industrial fields, requires high reliability. This demand reveals a need for an accurate and early reliability prediction to give feedback for the design and warranty precautions. There are many suggested methods to predict reliability of electronics in the literature such as using accelerated life tests, component based numerical and probabilistic simulations, and statistical field data based methods. Conventional accelerated reliability test do not meet the demands of today's very rapid electronic product cycles; they are time consuming and expensive . Using simulations for components and systems is another option, which is time saving, but simulation test data never reflects the real-world performance of the product and results in accuracy problems for various failure mechanisms . Therefore, laboratory and simulation data based predictions can be deceptive for many applications. This underlines the importance of using the product's field return data for reliability analysis that is relatively accurate, cheap and time saving. If one assumes that the return data is error free then it is not difficult to claim that it is the most trustworthy data in analyzing reliability of a product compared to accelerated test or simulation based data. However, in reality field return data contains errors corresponding to improper, correlated, incomplete, and poorly collected data that result in misleading reliability predictions. In this study, a systematic approach based on Weibull analysis to eliminate errors from field return data is proposed. In the elimination process, two types of errors are defined : obvious and hidden errors. Obvious errors are the errors that can be easily determined with applying one-by-one data check. Changing data orderings make this process even faster. Examples for such errors are unknown assembly dates, invalid time-to-failure values, and quality related errors. In contrary to the obvious errors, hidden errors and their probable sources cannot be directly determined from the data. Examples for hidden errors are significant changes in product manufacturing process, data loss, and inappropriate data recording in services. Hidden errors are fatal and overwhelm obvious errors in terms the number of occurrences. This obligates us to eliminate hidden errors before starting statistical analysis. After, debuging field data, prediction model is based on the Weibull-exponential piecewise hazard rate scheme considering early failure and useful life phases of the boards. This scheme is preferred for products not getting into the wear-out region during their warranty times or anticipated lives. Electronic boards, studied in this work, fall into these product groups. Electronic boards are expected to work at least 10-15 years with a warranty period of at most 5 years. Using a piecewise hazard rate function for reliability modelling is a well-studied subject in the reliability and statistics literature. This includes the determination of the turning/change time often called as the change point problem .Different approaches are proposed to solve this problem . Studies exploiting parametric change point analysis of non-monotonic hazard rate functions consider the change point as a parameter and propose statistical estimation methods including maximum likelihood, least squares, and Bayesian methods Additionally, non-parametric methods are studied widely. Regarding these problems, a practical yet accurate change point detection method is given by performing forward and backward data analysis with left and right truncated data. The given reliability prediction model has an input of either short- term or long-term field data corresponding to early failure and useful life phases of the boards, respectively. The output of the model is the reliability prediction covering the full warranty period which is 36 months in this study. In case of having long-term data, proposed practical change point method is directly applied to depict a hazard rate curve with Weibull and Exponential distributions using maximum likelihood and rank regression methods. In case of having short-term data, first it is investigated that how the Weibull shape parameter β varies with different time intervals of the field data. For this purpose it is used 36-month (full warranty period) of field data belonging to previous and/or current (if available) versions of the board. Then a mathematical function of beta is given in terms of time and board dependent parameters. Additionally, empirical findings and iteration methods are used to estimate other distribution parameters. Finally prediction model based on the developed Weibull-exponential scheme is constructed . This study contributes to an early reliability prediction methodology. A 36-month reliability prediction of a board can be made with its field data as short as 3 months. Conventional field data based reliability prediction methods, either parametric or non-parametric, have an important constraint .They can only predict very near future reliability of a product. In non-parametric analysis, it is not possible to accurately predict large time spans. In conventional parametric analysis, it is assumed that the failure mechanism of a system will be the same in contrary to real cases. Indeed, a product, consisting of many components, has different failure mechanisms for different operating time intervals corresponding to early failure, useful life, and wear-out regions. For example, with conventional methods one cannot predict the products reliability performance for the useful life period by using the data from the early failure period. With the proposed methodology, this problem is overcome considerably . Four different data sets belonging to four different boards Board-B, Board-E, Board- F, and Board-K are used. The field data contains assembly and return dates for each warranty call. Warranty period of the boards is 3 years. The maintenance policy for electronic boards under this study is to replace the board with a new one in any suspected case such as stopping from time to time or breaking down entirely in the field. Therefore the analyzed data has no record of repaired boards. The proposed methodology, illustrated, has an input of a new board's field data classified as long-term or short-term that is determined by applying the proposed forward and backward data analysis. If the analysis results in a change point that separates early failure and useful life regions then it is called the input data as the long-term data that is modelled with Weibull and exponential distributions. In this case, reliability prediction is achieved by using the exponential distribution that is expected to keep its validity until the end of the warranty period. If the analysis results in a single distribution, no change point found, then it is called the input data as the short-term data. In this study a framework that connects the prior knowledge based on an old board with a new one is used. The evaluation of the field data of an old board goes in forward time spans in order to obtain time dependent behavior of the Weibull β parameter. Ideally, a specific component or a material is expected have its own constant β that is time independent. However, similarly for most of the electronic applications, electronic boards consist of hundreds of different components/elements. Some of these elements such as integrated chips are almost perfectly reliable throughout the early failure phase that results in β >1. On the other hand, the components, especially used in power supply and mechanical control blocks which face much more stress, have serious early reliability issues that results in β <1. Therefore it is expected to see changes for β along with time in the field. This is the main motivation to derive a time dependent function of β. In order to estimate other parameters in the Weibull- exponential scheme, empirical findings and iteration methods are used. As a result, a full warranty forecasting of a new electronic board is made with given its field data as short as 3 months.