Int J Performability Eng ›› 2006, Vol. 2 ›› Issue (2): 149-161.doi: 10.23940/ijpe.06.2.p149.mag

• Original articles • Previous Articles     Next Articles

Methods for Binning and Density Estimation of Load Parameters for Prognostic Health Monitoring


  1. CALCE Electronic Products and Systems Center,
    University of Maryland,
    College Park, MD 20742, USA


Environmental and usage loads experienced by a product in the field can be remotely monitored in-situ using autonomous sensor systems. This data is useful for assessing the product degradation, predicting remaining useful life, developing load histories for future product designs, and hence minimizing the life cycle cost. One of the major challenges in such a load monitoring activity is the reduction in power and memory consumption of the sensor system for enabling long uninterrupted monitoring. This necessitates reducing the monitored data and storing it in a condense form (in-situ) without sacrificing the load information required for subsequent damage and life assessments.
This paper assesses non-parametric density estimation methods such as histograms and kernel estimators for use in-situ load monitoring. An experiment was conducted where-in an electronics printed circuit board (PCB) was exposed to field temperature conditions. The temperatures on the PCB were measured in-situ using sensor module with embedded processor and limited memory. The raw sensor data was pre-processed to extract cyclic mean temperatures. During monitoring the extracted cyclic mean temperature values were stored in bins with pre-calculated optimal bin widths, based on estimates of standard deviation and sample size. Assessment of density estimation techniques was conducted based on the comparison of pdf obtained from the binned data versus the complete data set. Sensitivity of the derived load parameter pdf to variations in the estimated and actual standard deviation and sample size were studied and a new method to account for these variations was demonstrated. Compared to using the complete data set, kernel functions resulted in more than 78% data reduction per day with an accurate estimate of density of the monitored parameter. The histogram provided more than 85% data reduction but a less accurate density estimate.
Received on September 23, 2005
References: 29