Exploring and Implementing the “Monitoring of a Multi-stage Manufacturing Process” using Dynamic Bayesian Networks

1. INTRODUCTION Hewlett-Packard (HP) is one of the world's largest computer companies and the foremost producer of test and measurement instruments. In Corvallis, Oregon, HP manufactures several precision products on high speed, automated assembly lines. The alignment process of a cap to a base part is one of the essential processes in its manufacturing process. This process has several stages with a large number of parts flowing through between stages. Each stage composes of an operation device and a visual sensor device. When a part flows pass the operation device, the visual sensor device then makes an observation on that part. The monitoring of this process has to be efficient because of two important factors. First, the accuracy of the alignment process affects the quality of the product. Second, the more accuracy the process is, the lower manufacturing cost the product has. Dynamic Bayesian networks are used to design the monitoring and diagnosis of the alignment process. Unlike the traditional Bayesian networks that statically show the relationships of the state variables at a particular time, the relationships between state variables in dynamic Bayesian networks are monitored across the time steps. Bayesian networks have several advantages over other diagnostic methods for the following reasons. First, Bayesian networks can lead to rational decisions by using conditional probabilities although there is not enough information. Second, Bayesian networks provide a better way to represent the problem using directed graph. Finally, Bayesian networks use prior knowledge of the causal relationships between variables in the domain, leading to a more accurate initial result. The purpose of this project is to monitor and expeditiously identify a component failure. The component is either an operation device or a visual sensor device. The paper will then describe the application of dynamic Bayesian networks in developing a system for monitoring 2 and diagnosis of the cap alignment process. In addition, the project is a continuation of the former work by Eric Wolbrecht. In his work, a prototype of a real time monitoring and diagnosis system was developed in order to compute the posterior probability of the state variable at any time step. The posterior probability of the state variable tells how well the component is working at that time. However, the existing posterior updating algorithm performs multiplication of the joint probabilities of every part in the system, resulting in a significant speed problem for a large number of parts. This …