Using Higher-Level Models to Predict Survival and How It Compares to Weibull Models
Aug 13, 2015 6:29 AM
Anshuman Singh, Mechanical Engineer, GE Power & Water John Korsedal, Consulting Engineer, GE Power & Water Brad Foulkes, Technical Leader, Life Cycle Reliability Engineering, GE Power & Water
The life of a machine is comprised of hundreds of components. Each component is influenced by a variety of operational conditions, many of which are not considered as part of the design process. Because of this, the actual life of a component is not a single point, but a distribution that is affected by operational data. Additionally, the exact time to failure can sometimes not be determined, as failure of a component may not cause failure of the entire machine. These sources of variation make survival modeling increasingly difficult, but including operational data is one way to reduce the impact of this variation. JMP Pro is used to understand the operational data. With the help of artificial neural networks and bootstrap modeling, we created a parallel approach to traditional survival models using unbalanced discrete outcomes to identify the drivers of component life. With this parallel approach and using JMP Pro visualization, vital X’s are identified and added to a survival model. A coherent use of these methods leads to a better survival model for the component life. This presentation will demonstrate the modeling technique used to identify factors, leading to more accurate models.