| The traditional MLE method for analyzing Accelerated Degradation Test often en- counters convergence problem because no closed form is available for most integrals. In this thesis, our target is to develop a Bayesian model to avoid the integral problem and effectively analyze the Accelerated Degradation Test data. Then we validated the model with simulation ADT data and compared with traditional and approximated MLE method's result. In comparison, the Bayesian model showed a significant advantage, as it doesn't need too much prior information, and convergences better than the MLE methods. We then further expanded the Bayesian model, combined more than one Accelerated Factors and Non-accelerated Factors. For these complex models, Bayesian method performs even better compared with MLE methods, suggestings that the Bayesian method has the ability of dealing with more complex models. An R function is developed for this model, which enables the automation of this procedure in the future. In the last part, we developed a method to find the best experimental design for a given ADT plan. This method is based on the Bayesian Network random walk model, the node with highest through frequency will be considered as the best plan, several alternative plans can also be selected from the network path based on the frequency. This model is first tested on a simple one Accelerated Factor case, and com- pared that design with current available software's result. It provided reasonable experimental plans similar to the traditional method's result for the simple case. We then applied our design method on a more practical example, and provided a detailed plan result for the example. This method has shown better accuracy than traditional balanced design in the simulated example. |
