Presentations

Abstract

Introduction Understanding the effect of resource on turbine loading and expected life is essential for planning, strategy and development. This project quantifies the impact of environmental conditions on expected fatigue life using novel methodology. Generally, sensitivity is assessed using computationally complex and data-intensive methods, such as Sobol’s Global sensitivity analysis, or simplistic univariate methods which require fewer data but trade efficacy for efficiency. One potential compromise is the Cox Proportional Hazards (Cox-PH) method: a statistical optimization technique enabling accurate and efficient estimation of the effect of input parameters on a component’s failure behaviour. This method is frequently used in the field of reliability but has limited exposure in the wind industry. One reason for this is that traditionally, the baseline hazard function is calculated using the Breslow estimator, which fails to capture the complexity of time-dependent failure behaviour. When calculated in this way, the relationship between hazard and failure expectation is not intuitive or easily calculable. This project proposes an adaptation, combining Cox-PH with failure-data distribution-fitting to provide a method for quantification of change in fatigue life from change in environmental conditions. This enables mapping of bounded wind conditions to fatigue life with associated confidence.   Methodology The Cox-PH method assumes there exists some underlying hazard function, , defining the failure behaviour of a component, which is then scaled proportionally by a factor determined by its covariate values. Once exponentiated, the hazard ratios may be interpreted as the multiplicative change in the failure rate from increasing that input by 1. This project uses Cox-PH in tandem with distribution-fitting, to quantify the relationship between environmental parameters and fatigue life The hazard ratios are first computed using fatigue life estimated through component S-N curves and the rainflow counting algorithm. The underlying ‘risk’ is then calculated by fitting an exponential distribution to the full population. Finally, mean time till failure is taken as the inverse of the exponential parameter, gamma, scaled by the inverse of the exponentiated hazard ratios. Since this method assumes hazard ratios are linearly independent, we can focus on the effect of changing a single variable to understand the sensitivity or combine changes to predict fatigue life under a different set of conditions. This method is fully analytical, allowing for highly efficient computation even on large datasets. Once trained, the results allow inference of change in expected life from change in input parameters. This allows for uncertainty quantification and construction of confidence intervals for fatigue life with known wind condition uncertainty, as well as prediction of life of new assets based on their individual conditions.   Results and conclusions Life at blade-root, tower-bottom, hub and shaft positions have been estimated for just under 15,000 Saltelli-sampled covariate sets. The normalized sensitivities suggest the variables with highest impact on blade-root fatigue life are the Weibull k parameter and reference turbulence intensity. The C-index amongst this population was 0.94. Wald tests were carried out to validate the estimated hazard ratios, with all showing no statistical evidence against the estimated parameters. Thus, initial study supports these results.