Reliability prediction of electronic boards by analyzing field return data

In this study, we perform field return data analysis of electronic boards. We cooperate with one of the Europe’s largest manufacturers and use their well-maintained data with over 1000 electronic board failures. We follow two steps that are filtering the return data and modeling the filtered data with probabilistic distribution functions. In the first step of filtering we propose a new technique to eliminate improper data, correlated to incomplete and poorly collected data, from the whole field return data. In the second step of modeling, we use the filtered data to develop our reliability model. Rather than conventionally using a single distribution for all service times that does not accurately model the substantial changes of the electronic boards reliability performance over time, we use different distributions for different service time intervals. In order to determine the distributions we propose a technique that deals with forward and backward time analysis of the data. well accepted to assume that the hazard rate function of a product shows a bathtub characteristic. In general hazard rate models depict an entire bathtub curve as a combine of two Weibull distributions for early failure and wear out periods (Dhillon, 1999; Kececioglu, 2002). This is illustrated in Figure 1. Although the models using all regions of the bathtub curve are theoretically correct, they are not practically applicable to some industrial product groups with relatively long lives (Kleyner & Sandborn, 2004). 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. This is illustrated in Figure 1. In this study, we propose a model regarding early failure and useful life periods of electronic boards. Our model is based on Exponential and Weibull distributions among many other distribution options regarding the optimum curve fitting. Rather than conventionally using a single distribution for all timeto-failures that does not accurately model the substantial changes of the board’s reliability performance over time, we use different distributions for different service time intervals. For this purpose we propose a new technique that deals with forward and backward time analysis of the data. In the fitting process we use “rank regression” and “maximum likelihood” methods. The data used in this study (filtering and modeling analysis) contains assembly and return dates for each warranty call. Warranty period of the board is three 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. Throughout our analysis we benefit from Weibull++ by ReliaSoft Corporation. The structure of this paper is as follows. In Section 2-Filtering, we describe the filtering procedure of field return data. In Section 3-Modeling, we work with the filtered data to develop our statistical model. Finally, Section 4 reports the conclusion of the work. 2 FILTERING In order to guarantee accuracy of the analysis and eliminate errors, field return data must be filtered. For this purpose, we use a step by step procedure. In the first step, we filter obvious errors from the whole data. We filter data having unknown assembly date, data with quality failure records that result in zero time to failure (TTF), and data with negative TTF or unreasonable TTF. In the second step which is the main target of this section we are dedicated to find and eliminate hidden errors. As it is mentioned before, obvious errors can be easily found by applying one-by-one data check. But, we cannot find hidden errors as direct as we find obvious errors. Here, a systematic approach is needed. By using Weibull++ ReliaSoft Corporation software we survey the consistency of the data and systematically investigate if there are hidden errors or not. We use 2 parameter Weibull distribution for our analysis because of being mathematically more tractable than other distributions (Reliasoft, 2014 ; Babington et al. 2007). Also using Weibull distribution to model reliability has long been approved in the literature. Maximum likelihood method (MLE) is selected for parameter estimation since regression methods generally work best for large data sets (O’Connor & Kleyner, 2011). In analysis, we deal with six month assembly time intervals. Selecting 6 month intervals is quite reasonable because we have 54 month return data. The proposed methodology targeting hidden errors is as follows. We first perform forward analysis for 1-6, 1-12, 1-18, 1-24, 1-30, 1-36, 1-42, 1-48 and 1-54 month time intervals. This is illustrated in Figure 2. For example, 1-6 time interval represents products assembled in the first six months. Similarly, 1-42 time interval represents the whole data excluding the ones assembled in the last six months. In forward analysis we expand time window from left to right where the left edge is fixed. We then perform backward analysis for time intervals of 48-54, 42-54, 36-54,,,,,,12-54, and 1-54. This is illustrated in Figure 3. Here, we expand time window from right to left where the right edge is fixed. Finally we perform analysis for seperate 6 month time intervals of 1-6, 7-12, 13-18, 19-24, 2530, 31-36, 37-42, 43-48, 49-54. This is illustrated in Figure 4. Note that X-axes in figures represent assembly times (not TTF). As a result, using parameters of Weibull distributions of these three analysis, we filter improper time intervals corresponding to hidden errors. We develop our filtering systematic mainly on a Weibull parameter β that explains the hazard rate function’s behavior. If β <1, it indicates a decreasing hazard rate and is usually associated with the early Figure 1. A bathtub curve hazard rate function over time.