Engineering Change Management using Bayesian Network

Abstract

An engineering change is defined as the changes in forms, fits, materi-als, dimensions or functions of a product or component. As a product evolves into a complex system, engineering changes become major driv-ing force for determining schedules, costs, and quality of product de-velopment process. Change management is defined as companys effort of making changes to a product in a planned or systematic fashion. Several efforts have been tried from both academia and industry

these includes implementing computer aided systems to streamline change implementation process, establishing guidelines of product design, and developing tools and methodologies for analyzing engineering changes in advance One major challenge in change management is change propagation. In a complex system where each part is associated with several other parts, a change presented into a part may influence other parts. As a re-sult, a change in a component may propagate throughout entire system resulting in avalanches of changes. To cope with the risk of change propagation, a number of strategies are proposed to avoid changes as much as possible. These strategies includes for example to make chang-es as early as possible, or to create design buffers. However, eliminating entire changes during the product development process is neither desir-able nor possible. In order to tackle these difficulties, this paper proposes a systematic method to predict and analyze engineering changes in advance. Bayesi-an network, which is a main modeling language of this paper, can effec-tively address both of the complex and uncertain aspects underlying engineering changes with the support of mathematical rigor. Based on a probability theory, one can anticipate risk of engineering changes in the form of conditional probability distribution. The resulting probability distribution can also be utilized for guiding optimal design decisions that reduce the risk of unnecessary changes. This thesis consists of three parts. The first part focuses on change prediction. In order to avoid unnecessary changes, likelihood and con-sequence of design changes should be analyzed in advance. In this sec-tion, Bayesian network is utilized for encoding probabilistic relationship among components and generating conditional probability distribution of each component with respect to various change scenarios. Due to its ability to model and reasoning of uncertain domains, Bayesian-network-based approaches shows significant advantages over tradition-al approaches in terms of modeling, analysis, and data learning aspects. The second part focuses on controlling change propagation. Espe-cially, this part focuses on derivation of a design freeze sequence. De-sign freeze is the end point of design phase at which a technical product description is handed over to production. One way to mitigate the risk of change propagation is to impose a design freeze on components at some point prior to completion of the design process. In this part, a Dy-namic Bayesian Network was used to represent the change propagation process within a system. According to the model, when a freeze deci-sion is made with respect to a component, a probabilistic inference algo-rithm within the BN updates the uncertain state of each component. Based on this mechanism, a set of algorithm was developed to derive optimal freeze sequence. The third part focuses on learning change prediction model from engineering change log database. When a product becomes complex, identifying complex relationship among components complex based on expert elicitation would become almost impossible. One alternative is to automatically mine change propagation network from a collection of previous change records. As a modeling language, dependency network, a graphical model for representing probabilistic relationships among random variables, was utilized. As a result, a complete joint probability distribution that can predict the probability of change propagation was obtained. A case study on Azureus, an open-source software project, was conducted to validate the proposed approach.