Developer Tutorial: Analyzing Fatigue Testing Data using JMP Fatigue Modeling

JMP 18 offers new Fatigue Model capabilities that implement a new methodology, based on years of research from Prof. William Meeker and a team of researchers, for analyzing fatigue testing data (also known as S-N data). The new methodology respects the nature of censoring in fatigue testing data; unifies the theory of inferring fatigue life distribution and fatigue strength distribution; and establishes robust algorithms to estimate parameters. The analytical results are inputs to other reliability engineering practices to predict the reliability of final structures. 

See how to:

• Understand situations where fatigue modeling can be useful
• Use JMP Fatigue Modeling capabilities to implement the new methodology
  • Create, view results for, and select from twenty-four different model
  • Choose one of six S-N curve types together with one of four distributions to assemble a mode
  • Compare different models, graphically and numerically
  • Get detailed information for individual models, such as parameter estimates and their confidence intervals; model diagnostics; distribution profilers for fatigue life and fatigue strength distributions; quantile profilers for the two distributions; and custom estimations of probabilities and quantiles with both Wald and likelihood intervals.
  • Look for improvements aready planned for V 18.01
    • Better proportions in density curves
    • Distribution and quantile profiles use two frames

Q&A is included at the end of the demo, beginning at time 1:09:00. The sample data used in the demos is available in JMP 18:

JMP 18 sample files used in the demoJMP 18 sample files used in the demo

Questions answered by Peng Liu @peng_liu and Rajneesh Rajneesh @Rajneesh at the live webinar:

Q: Is the slope of the asymptote informed by the tensile strength (0 cycles to failure)?

A: At the moment we fit from a stat perspective, if you have knowledge about the slope you can check to see if it matches.

Q: Does the Cox-Snell residual plot not show the runouts/censored data points?

A: Cox-Snell slope is not data points. they are non-param estimates of the distribution of the residuals. The purpose of Cox-Snell residual plot is to check for something. If curve deviates from the diagonal line (lognormal is fitted and it does not match). Then if we fit Weibell it is much better, so use Weibell.

Q: The construction of the Nishijima and Coffin-Manson Zero Slope models appear similar in their construction (i.e., slope in low-cycle fatigue, asymtote for high-cycle fatigue). Is there a functional difference that would influence your decision to pick one model over the other for a particular dataset?

A: You can use the AIC criterion along with your prior knowledge about the data.

Q: Can JMP's non-linear modeling platform be used to do the same analysis as in the fatigue model platform?

A: You can but you need to know the starting values to give it to nonlinear platform and use the loss function.

Q: If I use Non-linear, which loss function would I use? Log-likelihood? 

A: Yes, express log-likelihood as a loss in the nonlinear platform launch.

Q: If you choose multiple models, will JMP compare them for you and help you choose the best fit? (AIC, BIC)

A: Yes, and you can compare them.

References

• Modern Statistical Models and Methods for Estimating Fatigue-Life and Fatigue-Strength Distributions..., by William Q. Meeker, Luis A. Escobar, Francis G. Pascual, Yili Hong, Peng Liu, Wayne M. Falk, Balajee Ananthasayanam
• JMP 18 Fatigue Model documentation