Fatigue behavior of welded joints Part 1 - Statistical methods for fatigue life prediction

Statistical models for fatigue life prediction for welded joints are discussed and fitted to experimental data for fillet-welded steel joints where cracks emanate from the weld toe. The models are based on an S-N approach where the number of cycles N to failure is assumed to be directly correlated to the applied nominal stress range AS. The models assume the existence of a fatigue limit given as a stress range below which no failure. will take place. Emphasis is laid on the modeling of the fatigue life close to this limit where the service stresses for welded details often occur. Experimental data in this stress regime are sparse and do not fit the knee point of the conventional bi-linear S-N curve found in rules and regulations. Consequently, an alternative model where both the fatigue life and the fatigue limit are simultaneously treated as random variables is investigated. The model parameters for this random fatigue-limit model (RFLM) are determined by the maximum likelihood method, and confidence intervals are obtained by the profile likelihood method. The advantage of the model is that it takes into consideration the variation in fatigue limit found from specimen to specimen and that run-out results are easily included. The median S-N curve obtained from the model coincides with the conventional bi-linear curves in the high-stress regime (stress ranges higher than 110 MPa), but predicts longer lives as the stress range decreases below 100 MPa. The model gives a nonlinear S-N curve for a log-log scale in the fatigue-limit area; the fatigue life is gradually increasing and is approaching a horizontal line asymptotically instead of the abrupt knee point of the bilinear curve. The nonlinear curve is more in accordance with experimental data. At stress ranges below 100 MPa, the predicted fatigue lives are between 2 to 10 times longer than predictions made by the bilinear F-class curve. The conclusion is that the rule-based S-N curves may be overly pessimistic in the stress regime where service stresses frequently occur. A more correct statistical model based on a random fatigue-limit model results in S-N curves that give decreased dimensions for a given fatigue design factor under constant amplitude loading.