Estimation of balance uncertainty using Direct Monte Carlo Simulation (DSMC) on a CPU-GPU architecture

Abstract

The estimation of balance uncertainty using conventional statistical and error propagation methods has been found to be both approximate and laborious to the point of being untenable. Direct Simulation by Monte Carlo (DSMC) has been shown to be an effective alternative. The long simulation times of DSMC, when applied using conventional sequential codes, has been addressed by re-formulating the code to run on a multiple CPU-GPU, platform. Simulation times spanning minutes have replaced those spanning several hours. Application of uncertainty analysis using DSMC has led to an improvement in both balance calibration-quality and calibration-time, and promises to add understanding and insight not only to balance performance, calibration systems, and balance uncertainty, but also to an improved understanding of commonly quoted statistical data. This paper extends the introductory paper presented in 2013 to provide an overview of the current CPU-GPU system. Data obtained from a six component internal balance is used to show how the relatively large quantity of data generated using DSMC can be managed through the mechanism of data modelling. These models may be used to generate equivalent data related to balance loads as would be generated by a balance when performing a dead-weight roll-polar. This is considered to be not only an effective analysis approach, it also provides a practical link between data generated by a DSMC simulation and data that can be physically generated by a balance. Balance data from a roll-polar is used to show that balance uncertainties arising from the mathematical calibration model, the calibration-loading, and the balance itself, can be separated and quantified. This leads to the recommendation that verification of correct balance installation and the determination of installed uncertainty be obtained by performing a dead-weight roll-polar. This provides confidence in the balance installation, balance uncertainty data, as well as minimisation of installation roll-offset error.