استخراج شکل : یک روش داده کاوی جامع برای طراحی و مهندسی

Although the integration of engineering data within the framework of product data management systems has been successful in the recent years, the holistic analysis (from a systems engineering perspective) of multi-disciplinary data or data based on different representations and tools is still not realized in practice. At the same time, the application of advanced data mining techniques to complete designs is very promising and bears a high potential for synergy between different teams in the development process. In this paper, we propose shape mining as a framework to combine and analyze data from engineering design across different tools and disciplines. In the first part of the paper, we introduce unstructured surface meshes as meta-design representations that enable us to apply sensitivity analysis, design concept retrieval and learning as well as methods for interaction analysis to heterogeneous engineering design data. We propose a new measure of relevance to evaluate the utility of a design concept. In the second part of the paper, we apply the formal methods to passenger car design. We combine data from different representations, design tools and methods for a holistic analysis of the resulting shapes. We visualize sensitivities and sensitive cluster centers (after feature reduction) on the car shape. Furthermore, we are able to identify conceptual design rules using tree induction and to create interaction graphs that illustrate the interrelation between spatially decoupled surface areas. Shape data mining in this paper is studied for a multi-criteria aerodynamic problem, i.e. drag force and rear lift, however, the extension to quality criteria from different disciplines is straightforward as long as the meta-design representation is still applicable.

The intensive use of computational engineering tools in the recent years and the transition from an experiment to a simulation based product design process, in particular in the automotive industry, has led to a significant increase of computer-readable design data relating design characteristics1 to the design quality.2 In the context of Product Data Management (PDM) and Product Lifecycle Management (PLM), product related data is maintained and integrated through the whole design process or even through the whole lifetime of the product. Although PDM/PLM frameworks have been successful in managing CAD models and documents as well as in integrating CAD and ERP (Enterprise Resource Planning) systems, PLM solutions still need customization to the actual tools used in the design process [1]. Furthermore, the handling of multi-disciplinary processes, tools and data structures as well as a systems engineering or holistic interpretation of the design process remains to be challenging, e.g. see [1] and [2]. Industrial informatics in the domain of PDM and PLM still has not received the required attention in the literature, e.g. see [3]. As a result the application of data mining techniques to engineering data in practice is still often restricted to single design processes and individual design teams working on a certain CAE task, which we will call a sub-process in the following. The stronger the variation between the CAE tasks is (different representations, different disciplines, different tools and data structures), the more isolated is the data handling. Even though the data might be integrated into an overall PDM framework, it is not available for a holistic data mining approach from a systems engineering perspective. As a simple example different design teams might focus on the aerodynamics of the frontal part of the car, the rear part of the car, the noise generated or the cooling of the front brakes. However, the CAE results as well as the changes the teams proposed to the design are seldom independent from each other, since they are very likely to employ different representations of the design parts. This makes it difficult for data mining techniques to integrate data across teams and disciplines. On the one hand, the decomposition of the overall design problem (known as Simultaneous or Concurrent Engineering) is necessary for an efficient design process. On the other hand, we can expect that new important insight about the design can be gained only when we examine the data holistically and relate previously unrelated parts of the design process to each other. In a more formalized way, the targeted approach is illustrated in Fig. 1. The engineering design process is considered to be best described as a goal oriented iterative decision making process [4]. In each iteration, engineers decide about individual or a sequence of design variations, which lead to a final design configuration fulfilling pre-defined constraints and design goals best. The overall design process is spatiotemporally decomposed into a number of (multi-disciplinary) sub-processes {1,…,i,j,…,P}{1,…,i,j,…,P}. Based on the result of a decision making process (DMP), each sub-process defines design changes that contribute to the synthesis process (SP) for the finally submitted design. The aim of the paper is to propose an approach that allows cross-process design data management (DB) and that enables the analytics process (AP) to integrate knowledge and information gained from all sub-processes. Finally, the results of the holistic analysis can be fed back to the individual sub-processes to improve the individual decision making.Apart from the problem of relating different design representations to each other in the overall design process, in general the application of data mining techniques to engineering data has been less explored than, e.g. to economic data. Literature related to the extraction of human readable knowledge in the field of aerodynamic and structural design is rare. The team of Obayashi [5] and [6] have addressed the extraction of knowledge from a given data set in order to gain insights into the relationship between geometry and multi-criteria performance measurements. The authors applied self-organizing maps (SOM) in order to find groups of similar designs for multi-criteria performance improvements and tradeoffs, and used the analysis of variance technique (ANOVA) to identify the most important design parameters. Their methods have been applied to supersonic wing design. In [7] the use of methods from information theory have been studied to reveal higher order interrelations between design and flow field properties. Their methods have been tested in the domain of turbine blade and passenger car design. In most of the literature, the extracted information is linked to a specific and well-defined representation being used in the design process. Thus, the usability of the extracted information beyond this particular design and optimization process is only possible to a limited extent. Therefore, Graening et al. in [8] started to study the use of data mining techniques on a unified object representation. However, data mining based on such a typically high dimensional representation goes beyond the application of individual modeling technologies. Furthermore, it requires the consideration of other data mining aspects like, feature extraction, feature reduction and post-processing. Wilkinson et al. [9] adopted the basic idea of Graening et al. and utilized unstructured surface meshes as unified object representation for the prediction of the local wind pressure distribution on tall buildings. In this paper, we generalize the concept behind the analytics of design data based on a unified shape representation by introducing the shape mining framework. The remainder of the paper is organized in two parts. In the first part, we discuss different methods for shape mining and embed them into an overall framework. In the second part, the shape mining framework is applied to the analysis of passenger car design data. The first part is divided into four sections. In Section 2, a unified design representation3 is defined together with methods for the evaluation of local design differences. In Sections 3, 4 and 5, methods for sensitivity analysis, for the extraction of design concepts, and for interaction analysis are introduced and discussed. The second part of the paper is organized almost synonymously with the first part. Firstly, elements of different design processes that are the sources for the passenger car design data are described in Section 6. Statistical methods are applied to the meta design representation in Section 7, e.g. to investigate the course of design processes. In Sections 8, 9 and 10 the methods from part one for sensitivity analysis, the extraction of design concepts, and the interaction analysis are applied to the data from the industrial design process. The aim is to model and understand the relation between shape variations of the car and changes in their aerodynamic quality. Whereas part two of the paper is an application specific example, the approach presented in part one is generally applicable to all problems in the area of shape or topology mining. At the same time, some readers might find it useful to see the practical use of algorithms introduced in part one immediately; those readers are invited to read, e.g., Section 8 after Section 3. The paper closes in Section 11 with a conclusion and summary of the work. Part I: shape mining More recently, technologies from computational intelligence and data mining, e.g., see [5] and [6], have been adopted to exploit experimental design data and computational resources for the support of engineers in the decision making process. However, the multi-disciplinary characteristics of complex design processes and the huge variability in computational design representations hinders the analysis of design data beyond individual design configurations and processes. Especially the variation in the computational representations being used makes an efficient knowledge exchange between design processes difficult. The shape mining framework, as illustrated in Fig. 2, targets the integration of technologies for the implementation of a holistic analysis processes. It requires the transformation of designs into a meta-representation, which facilitates the evaluation of design differences on a holistic basis. Just the transformation of the designs into such a unified meta-representation, together with the evaluation of design quality differences, allows a holistic modeling of the design data independently of the originating process. Depending on the stated problem, modeling techniques from data mining and machine learning are applicable to investigate design sensitivities, retrieve abstract design concepts and analyze the interrelations between distinct design parts with the focus to understand the interplay between local design differences and changes in their quality. The resulting knowledge from the analysis of the design data can be utilized to support engineers in decision making and to improve future design and optimization processes.

In this paper, we have motivated the need for holistic data analytics in engineering design and outlined a framework for its realization. In order to be able to combine information from different design processes into one framework, a unified design representation has to be defined. The unified surface data is stored in a database, which is the central element of the framework as depicted in Fig. 4. The adaptation of statistical data mining methods to surface data is preceded by the definition of appropriate distance measures between geometrical objects. Even though the surface differences considered in our research are not large, we have shown that the Geodesic distance is more suitable than the standard Euclidean distance between surface vertices. The sensitivity analysis based on correlation methods and information theoretic approaches applied to surface data constitutes the first step towards shape mining. The application of more sophisticated methods of knowledge formation necessitates to resolve the typical drawback of the universal design representation, i.e., its high dimensionality. Therefore, we apply feature evaluation, reduction and clustering techniques to reduce the representation to a significant subset. Based on this feature set, methods for concept retrieval can be applied. The procedure for the retrieval, description and evaluation of design concepts has been generalized and can be carried out independently of the used modeling technique. A new measure has been introduced to evaluate extracted design concepts based on the estimation of their utility. The new measure allows the ranking of concepts according to the formulation of the engineer’s objectives. The last analytics step in our framework as shown in Fig. 4 is the interaction analysis. The extension of mutual information to more than two random variables is non trivial and its comprehensive statistical treatment is beyond the scope of this paper. Nevertheless, we are able to formulate the statistical interaction between the changes of different surface patches or features and changes in the design objectives. The resulting interaction graphs provide a fast and easy way to visualize rather complex statistical information. The remaining component in the proposed shape mining framework in Fig. 4 is the utilization of the extracted knowledge. Although it is impossible to directly show how the result of shape mining influences a realistic design process involving different tools, engineers and decision making processes, we apply the framework to the practical example of passenger car design. We choose two meaningful objectives and generate data from three different processes involving different shape representations and different strategies for shape space sampling. In the second half of the paper, we go through the different steps of our framework using these three realistic data sets. The findings of the displacement and sensitivity analysis provide the engineer with a concise picture of the direction of the overall design process. The effects of different representations and sampling techniques can be visualized and unexplored regions of the design space can be identified. Whether those ”white spots” on the design landscape are due to constraints or due to shortcomings of the design processes has to be decided by the engineer. Although the statistical methods to extract this information are not complex, it is the comprehensive framework that allows their application to the complete process. In a practical engineering design process for complex products like automobiles the impact of such rather basic information should not be underestimated. The concept retrieval augmented by the new utility measure allows the formulation of simple rules for passenger car design matched to the specific objectives that have been formulated. The key here are the rules of intermediate complexity because they are still readable and are less likely to represent standard engineering knowledge for car design. Furthermore, the algorithmic extraction of design concepts allows the wider distribution of subjective engineering knowledge in a company. Finally, the interaction analysis revealed the joint influence of modifications at the front and the back of the car on the accumulated objective ϕFϕF. This is likely to be new and interesting for the engineer. Whereas experienced engineers are quite confident to judge the interaction between design changes for single objectives, this is usually not the case if objectives are accumulated or when optimal trade-offs between objectives are sought. At the same time, design processes based on single objectives become more and more obsolete and are replaced by multi-disciplinary and many-objective approaches.