PRODUCT BASED SEQUENCE EVALUATION FOR AUTOMATED ASSEMBLY PLANNING

Assembly time estimation is a key factor in evaluating the performance of the assembly process. The overall goal of this research is to develop a method to automatically estimate product assembly time based on the tessellated model. This paper proposes a way to divide an assembly operation into four actions which consist of a) part movement, b) part installation, c) secure operations, and d) subassembly rotations. Four predictive models are built for estimating these action times with the input features that can be obtained from the tessellated model automatically. In order to estimate the four operation times, a design of experiments is applied to collect experimental data, based on the physical assembly experiments performed on products that are representative of common assembly processes. The Box-Behnken design (BBD) is an experiment design to support response surface methodology to interpret and estimate a prediction model for the four operations. With the experimental data, a stepwise regression method is used to estimate the predictive mean time. After that, a Gaussian Process (GP) model is applied with the basis mean regression function for more accurate prediction. Also, the confidence of the prediction can be given by the GP predictive confidence interval (CI). Hence, the uncertainty of the prediction can be quantified by the designer and can be used for evaluating the risk of an assembly sequence design. An Artificial Neural Network (ANN) is also used in this study to compare with the regression and GP model. A case study of a pump assembly time estimation is conducted. The predictive times from the regression, GP and ANN models are compared with the Design for assembly (DFA) method to ensure the reasonable predictions. The results show our proposed method has good prediction for most of the assembly tasks. Although errors exist in some of the task, the accuracy of the models can be improved with more user feedback since the model quality depends upon the amount of the training data.