中華大學博士論文

(1)

中華大學博士論文

題目：整合性組裝順序規劃之 KBE 系統

A Knowledge-Based Engineering System for Integral Assembly Sequence Planning

系所別：科技管理研究所學號姓名： D09403015 戴培豪指導教授：陳文欽博士指導教授：謝玲芬博士

中華民國九十八年八月

(2)

中華大學

(博碩士論文授權書)

本授權書所授權之論文為本人在中華大學

科技管理學系(所) 97 學年度第 2 學期取得博士學位之論文。

論文題目：整合性組裝順序規劃之KBE系統指導教授：陳文欽博士、謝玲芬博士研究生姓名：戴培豪

授權事項：

一、博碩士紙本論文著作權授權

本人同意將本著作，以非專屬、無償授權中華大學與國家圖書館，基於推動讀者間「資源共享、互惠合作」之理念與回饋社會與學術研究之目的，

中華大學圖書館及國家圖書館得以紙本收錄、重製與利用；於著作權法合理使用範圍內，讀者得進行閱覽或列印。

二、博碩士論文全文電子檔上網授權

本人茲同意將授權人擁有著作之上列紙本論文全文(含摘要)，以非專屬、無償授權中華大學圖書館、國家圖書館，不限地域、時間與次數，

以微縮、光碟或其他各種數位化方式，將上列論文重製，並得將數位化之上列論文及論文電子檔以上傳網路方式，提供讀者基於個人非營利性質之線上檢索、閱覽、下載或列印。

指導教授：陳文欽博士、謝玲芬博士授權人：戴培豪

簽名：______________________ (請親筆正楷簽名)

中華民國年月日備註：本頁授權書請列印三份，並親筆簽名，一份裝訂於論文紙本，二份於畢業辦理離校手續時繳交至圖書館。

(3)

中華大學博士班研究生論文指導教授推薦書

科技管理研究所博士班戴培豪君所提之論文整合性組裝順序規劃之 KBE 系統，係由本人等指導撰述，同意提付審查。

指導教授 ^(簽章) ^(簽章)

中華民國九十八年八月

(4)

中華大學博士班研究生論文口試委員會審定書

科技管理研究所博士班戴培豪君所提之論文整合性組裝順序規劃之 KBE 系統，經本委員會審議，符合博士資格標準。

論文口試委員會召集人

^(簽章)

委員

^(簽章)

(簽章)

所長

^(簽章)

中華民國九十八年七月二十九日

(5)

(6)

謝辭

時光荏苒，日月如梭！回顧四年博士班求學的過程，誠為人生難得且難忘的歷程。記得當初步入中華大學博士班入學口試時，就曾誓言要認真求學，戮力研究，希望能有番作為。很慶幸也很感恩能有兩位學養豐富、治學嚴謹的指導老師陳文欽教授、謝玲芬教授，從旁提攜及悉心教導，從他們的專業知識及言行典範薰陶下，讓學生在射出成形製程參數最佳化、類神經網路應用技術及決策分析、績效評估與作業研究管理領域中，獲致長足的進步，也發覺學海無涯、唯勤是岸的真諦。

本篇論文的完成，感謝盛中德教授、雷鵬魁教授、吳鴻輝教授、蔡志弘教授、王珉玟副教授、徐永源副教授，以及指導老師陳文欽教授、謝玲芬教授等八位口試委員，於百忙之中，不辭辛勞，撥冗指導，提供諸多寶貴意見，使得本論文得以更臻完善，在此誠摯表達對教授們的謝意！尤其是徐永源副教授，

藉由他知識工程(KBE)實驗室的全力贊助，讓本論文組裝順序規劃的KBE系統得以具體實現，也讓學生得以將碩士所學延續，並應用於知識工程領域中。

感謝鄧維兆老師在「科盛科技股份有限公司產學合作案」射出成形的熱心指導，感謝「竹永成股份有限公司」丁進興先生與杜鳳榜先生的射出成形相關技術的經驗交流與建議，所長李友錚老師在實驗設計課程的教誨，馬恆老師在類神經網路的啟蒙，李欣怡老師在會計學的應用及林淑萍老師在統計學觀念建立，在此一併致謝。感謝中華大學自動化實驗室學長陳振臺、學弟妹傅公良、

賴東燦、范揚志、王伯元、林芷郁、鄭仁豪、周裕泰、周書全、曾文柏、陳柏睿、陳希平在各類實驗及相關事務的協助。再則也感謝台達電子股份有限公司品保處處長鄭子建先生對我工作的包容與支持。

感謝軍中退役學長朱宗緯先生及黃華國先生對我的種種鼓勵及建議。最後，我要特別感謝我的妻子－林惠美女士照顧我日常生活及三個可愛的女兒－

維儀、維萱及維頡，讓我可以無後顧之憂、全力以赴，完成博士學位。

(7)

中華大學博士論文

題目：整合性組裝順序規劃之 KBE 系統

A Knowledge-Based Engineering System for Integral Assembly Sequence Planning

系所別：科技管理研究所學號姓名： D09403015 戴培豪指導教授：陳文欽博士指導教授：謝玲芬博士

中華民國九十八年八月

(8)

i

ABSTRACT

Research in assembly planning can be categorized into three types of approach:

graph-based, knowledge-based and artificial intelligence approaches. The main drawbacks of the above approaches are as follows: the first is time-consuming; the second approach is difficult to find the optimal solution; and the third approach requires a high computing efficiency.

To tackle these problems, this study develops a three-stage approach (i.e., firstly create a correct CAD-oriented explosion graph and then find a graph-based assembly sequence using Above graph, relational model graph and assembly precedence diagram; at last, generating a feasible assembly sequence) integrated with robust back-propagation neural network (BPNN) engines via Taguchi method and design of experiment (DOE), and a knowledge-based engineering (KBE) system to assist the assembly engineers in promptly predicting a near-optimal assembly sequence for mechanical or plastic products. The research focuses on building a novel KBE system for assembly sequence planning (ASP), which joins BPNN predictor and Siemens NX/KF second development module together to create feasible assembly sequences. System user can easily access the volume, weight and feature number through Unigraphics NX system interface, and input the related parameters such as contact relationship number and total penalty value, and predict the feasible assembly sequence via a robust BPNN engine. Furthermore, the proposed system can demonstrate the explosion views and vivid assembly simulations, save the entire assembly information, and setup a consolidate knowledge base.

Finally, three real-world examples- the toy car model as a learning (training) sample, toy motorbike model and a brushless DC fan as verifying (testing) samples, are dedicated to evaluating the feasibility of the proposed KBE system in terms of the differences in assembly sequences. The results show that the proposed model can efficiently generate robust BPNN engines, facilitate feasible assembly sequences and allow the designers to recognize the contact relationships, assembly difficulties and assembly constraints of three-dimensional (3D) components in a virtual environment type.

Keywords: assembly planning; assembly precedence diagrams; neural networks;

design of experiment; Taguchi method

(9)

ii

摘要

產品組裝規劃研究一般可分成圖形導向、知識導向及人工智慧等三類，而其三種研究的各自缺點在於圖形導向需耗費大量時間，知識導向很困難去發掘最佳解，人工智慧需要高度計算效率。為了克服上述問題，本研究發展一個整合三階段圖形導向啟發式作業規則，並以田口及實驗設計建構穩健倒傳遞神經網路預測器與智慧型物件導向的知識工程(KBE)系統。

首先協助工程師依據產品上位圖規則去產生一正確CAD 爆炸圖，並依據圖

形導向啟發式作業規則，建構關係模型圖及組裝優先順序圖，進而產生圖形解之最佳組裝順序-當作倒傳遞類神經網路學習訓練樣本；再則透過倒傳遞類神經

網路穩健預測器及物件導向的KBE 系統，產生可行的組裝順序。本研究所建立

之組裝順序規劃KBE 系統，係利用 Siemens NX/KF 二次開發軟體模組與結合倒

傳遞神經網路預測器，預測並獲得較佳的機構件或塑膠件產品的組裝順序；其

中在使用者UI 介面可獲取 CAD 系統組合件的體積、重量、特徵數量等資訊，

並輸入相關參數(接觸關係值與總懲罰值)，以倒傳遞類神經網路訓練樣本，建構穩健的組裝順序預測引擎，預測可行的組裝順序；本系統不但可顯示其爆炸圖及動態組裝模擬，並可將所有組裝資訊儲存，建立完善的知識庫。

最後應用三種實體樣品在組裝順序規劃KBE 系統上，以玩具模型車當為學

習(訓練)樣本，玩具摩托車模型與無刷直流風扇當為測試樣本，去評估及驗證上述方法的可行性。結果顯示所提的研究模式，可有效率地建立穩健倒傳遞神經

網路引擎及快速預測可行的組裝順序，並讓 R&D 設計者深入了解組裝接觸關

係、組裝困難度與虛擬3D 實體組裝限制因素。

關鍵字：組裝規劃、組裝優先順序圖、類神經網路、實驗設計、田口方法

(10)

iii

LIST OF TABLES

Table 1 An example of penalty index... 30

Table 2 Information on the factors’ assumed settings via Taguchi method ... 34

Table 3 Information on the factors’ assumed settings via DOE... 35

Table 4 The optimal assembly sequence of a toy car. ... 43

Table 5 The optimal assembly sequence of a toy motorbike. ... 44

Table 6 The optimal assembly sequence of a toy boat... 44

Table 7 The optimal assembly sequence of a DC brushless fan. ... 44

Table 8 Comparisons of BPNN performance between Taguchi method and DOE ... 45

(13)

vi

LIST OF FIGURES

Figure 1. DFX in the Concurrent Engineering environment framework. ... 1

Figure 2. Working concepts and procedures. ... 14

Figure 3. KBE model for assembly sequence optimization... 15

Figure 4. The construction flowchart of ANN-based ASP engine. ... 16

Figure 5. The ASP flowchart applied on KBE system... 17

Figure 6. An APD tree used for assembly planning. ... 30

Figure 7. Max- and min-dimensions of a part in deployment direction... 31

Figure 8. Nine Above Graph sets. ... 32

Figure 9. Symbol definition for the combined parts... 33

Figure 10. Main effects plots of BPNN’s factors... 35

Figure 11. The RSM response optimization of BPNN’s parameters. ... 36

Figure 12. The part parameter of UI-Style Push Button. ... 37

Figure 13. The BPNN parameters and UI-Style AI and TPV interface ... 37

Figure 14. The explosion graph generation per UI-Style interface... 38

Figure 15. The assembly sequence generation per UI-Style interface... 38

Figure 16. The parts list and exploded view of a toy car. ... 39

Figure 17. The complete RMG and APD of a toy car... 40

Figure 18. The parts list and exploded view of a motorbike... 41

Figure 19. The complete RMG and APD of a motorbike ... 41

Figure 20. The parts list and exploded view of a toy boat. ... 42

Figure 21. The complete RMG and APD of a toy boat... 42

Figure 22. The parts list and exploded view of a DC brushless fan... 43

Figure 23. The complete RMG and APD of a DC brushless fan. ... 43

Figure 24. An assembly sequence prediction via BPNN engine 1... 45

Figure 25. An assembly sequence prediction via BPNN engine 2... 46

Figure 26. The realization procedure of KBE system... 46

Figure 27. Build the components information of a DC-brushless fan. ... 47

Figure 28. Build the components information of of a toy motorbike. ... 48

Figure 29. Input the contact relationship number of a DC-brushless fan. ... 48

Figure 30. Input the contact relationship number of a toy motorbike... 48

Figure 31. Input the total penalty value of a DC-brushless fan. ... 49

(14)

vii

Figure 32. Input the total penalty value of a toy motorbike... 49

Figure 33. Constructing the BPNN engine and testing ASP of a DC-brushless fan. . 50

Figure 34. Constructing the BPNN engine and testing ASP of a toy motorbike. ... 50

Figure 35. The parts list and exploded view of a brushless DC fan... 51

Figure 36. ASP generation of a brushless DC fan via KF interface... 51

Figure 37. The exploded view of a toy motorbike via KF interface. ... 52

Figure 38. ASP generation of a toy motorbike via KF interface... 52

(15)

1

CHAPTER 1 Introduction

Assembly sequence planning (ASP) is a critical technology which achieves product design and realization. Product design is an important determinant of business’ competitiveness. It has been determined amount to 80% of the costs of product lifecycle during the initial design stage (Mascle & Zhao, 2008). DFX, or

“Design for eXcellence”, is a major concept that incorporates “Design for Assembly (DFA)”, “Design for Disassembly (DFD)”, “Design for Manufacturing (DFM)”,

“Design for Test (DFT)”, “Design for Quality (DFQ)”, “Design for Cost”, “Design for Environment” and all other design aspects that you can come up with. DFX was developed in the late 1970s (Kuo et al., 2001), and it is widely used in the development of new products (Huang & Mak, 1997). DFX is a general term; ‘X’ can represent assembling, manufacturing, quality, safety, and so on. The exact nature of the variable

‘X’ in any instance defines the focus of a DFX tool. ‘X’ has two parts X = x+bility; x represents life-cycle processes in product development as can be seen in Fig. 1 and the suffix ‘-bility’ corresponds to the performance matrices (Huang, 1996).

Figure 1. DFX in the Concurrent Engineering environment framework.

Design for X (DFX) philosophy suggests that a design be continually reviewed from the beginning to the end of product lifecycle to find ways to enhance production quality and other non-functional aspects. These rules are nothing new, they are just common sense items written down, but they can be a good guide through the design

(16)

2

process. Advantages of DFX techniques are fewer production steps, shorter production times, smaller parts inventory, simpler designs that are more likely to be robust, more standardized parts, proven to be very successful over decades of application, and can be used as a very substantial part of “concurrent engineering”

(Huang & Mak, 1997). DFX modeling and simulation systems will provide a robust environment for designing, planning, and driving seamless assembly/ disassembly/

reassembly processes that enable engineering operations.

A feasible assembly plan ensures the standard of RMA (Reliability, Maintainability and Availability) during a product’s life cycle; this can increase production efficiency and reduce the cost of a product. Moreover, effective assembly planning facilitates concurrent engineering in product development, operations and system analysis, to enhance the quality of a particular system (Lim et al., 1995).

However, achieving an optimal assembly plan requires a significant amount of time to analyze the combination of relationships of each component or part (Lotter, 1989) and assembly process may include welding, soldering, adhesive bonding, wiring, press fitting, shrink fitting, brazing, riveting, and other mechanical or electric fastening, in which a series of standard sequences is required to accomplishing the process without regard to the product configuration, material, or production quantity (Crowson, 2006).

Assembly-related researches can fall into three areas: (1) the evaluation and optimization of the assembly structures such as DFA, DFD, and DFM (2) compter-aided assembly process planning (CAAPP) in view of assembly sequence, assembly tools, assembly line layout, fixtures, and assembly operation (3) assembly system design, assembly line balancing and optimization (Su, 2009).

Section 1 Literature Review and Hierarchical ASP

In general, assembly involves the integration of components and parts to create a product or system through computer-aided design and manufacturing (CAD/CAM) systems. Assembly planning is a crucial design step for generating a feasible assembly sequence (Wang et al., 2005). Traditional assembly planning is manual and based on the experience and knowledge of industrial engineers; however, manual analysis does not allow the feasibility of assembly sequences to be easily verified. In the electronics industry, the approximate 40%- 60% of total wages was paid to assembly labors, and assembly cost accounts for 10%- 30% of the total labor cost and amounts to 50% of

(17)

3

product manufacturing cost (Kalpakjian, 1992; Smith & Smith, 2002). The implementation of design for assembly (DFA) and design for manufacturing (DFM) resulted in enormous benefits, including the simplification of products, reduction of assembly product costs, improvement of quality, and shrinkage of time to market (Kuo et al., 2001). Good assembly sequence planning (ASP) has been recognised as a practical way of reducing operational difficulties, the number of tools and the working time (Lai & Huang, 2004).

Design for assembly (DFA) focuses on product related factors such as size, weight, symmetry, orientation and form features, and other assembly processes:

handling, gripping and insertion (Kai and Basem, 2003). Automated design for assembly (ADFA) tools will enable product and process designers to rapidly compose, adjust, and decompose “micro-models” of constituent materials, parts, components, subassemblies, and assemblies. The above system will enable them to determine the fastest and most effective assembly sequence, evaluate and select the best joining/

attachment methods, ensure correct fitting and tolerance, and optimize the total assembly process to create the most effective use of manual and automated assembly resources from all assets available to the enterprise-including those of partners, suppliers, vendors, and remote facilities as well as the local factory.

De Fazio and Whitney (1987) adopted the exhaustive concept of Bourjault (1984) to obtain a complete set of assembly sequences. They generated sequences in two stages – creating the precedence relations between liaisons (i.e., physical contacts) or logical combinations of liaisons in a product and verifying the liaison sequence in terms of graph search theory and simplified approach. However, some components cannot be assembled successfully due to the hidden geometric constraints. Homen de Mello and Sanderson (1991a) made a representation of the directed AND/OR graphs and disassembly concerns to create feasible assembly sequences. In addition, Kroll (1994) used directed graph-based procedures with conventional representations to reduce the number of sorting operations required. He then extended his previous approach from uniaxial assemblies to triaxial assemblies and presented a set of rules for resolving conflicts between multiple parents and multiple offspring. However, these direct approaches only apply on the orthogonal assembly structure involved six orthogonal directions.

Assembly planning is commonly a tedious design procedure requiring a

(18)

4

significant amount of manpower to analyze the combination of relationships of each component or part (Tai, 1997). Rationalization of an assembly, including the development of new materials, time-and-motion studies, methods analysis and improvement, product development and design, etc., must be dedicated to assembly planning for improving the quality of the assembled products and for reducing its cost (Nof et al., 1997; Amen, 2001); The whole product assembly planning entails assembly sequence, the assembly tools and fixtures, the assembly line layout, and assembly line balancing, which are highly related to assembly sequence and precedence graphs (Henrioud et al., 2003; Su, 2009).

In practice, most assembly companies use semi-automatic systems to generate an assembly plan and employ 2D cross-sectional views to represent their heuristic models (Lin & Chang, 1993). Moreover, CAD-directed assembly sequence planning is a graph-based heuristic approach using four critical stages to generate assembly sequences: creates connective graphs, decomposes an assembly into subassemblies, generates the disassembly sequence for each subassembly, merges the disassembly sequences into a complete disassembly sequence, and converts the complete disassembly sequence into the final assembly sequence (Gu & Yan, 1995).

Automated assembly planners generally use the above artificial intelligence techniques or graph- searching methods to search for better assembly plans in view of product geometric and physical constraints, minimizing assembly reorientations, tool changes or reducing fixture device (Smith & Smith, 2002). However, there exist two kinds of potentially applicable problems: a large number of generated assembly sequences and the computational complexity derived from the geometrical and topological features (Yin et al., 2004).

Assembly planning is also regarded as “assembly by disassembling,” i.e., an assembly sequence results from systematically disassembling the final product and reversing the disassembling sequence (Lee, 1989). This approach usually employs the contact-based feature to represent the precedence relationships of the product. A designer can successively assign the assembly relations to form the assembly plan based on the precedence diagram. However, the contact-based precedence diagram cannot effectively express the complexity of the assigned assembly relations.

There are two ways to disassembly a part by reverse assembly or brute force. A critical aspect of almost all objects or devices existing more than one part is the

(19)

5

nature of fasteners used to join their parts together. Fastening methods can be employed in dictating the level of difficulties during assembly/disassembly process as follows (Prasad, 1997):

1. Adhesive 2. Heat Staking 3. Induction Welding 4. Inserts

5. Screws

6. Snap-fit Latches 7. Thermal Methods

An effective assembly plan must include other graphs, such as explosion graph, relational model graph, incidence matrix, assembly precedence diagram (APD), etc. In reality, few experts or engineers know exactly how to derive a correct explosion graph, draw a complete relational model graph or incidence matrix among the components, or determine a complete APD to generate a global optimal assembly sequence (Chen et al., 2004b; Chen et al., 2008). Besides, the graph-based approach based on the assembly relationship and the precedence constraints can find a global optimal solution in theory (Lu et al., 2006).

The current essential technology, knowledge-based engineering (KBE), allows an engineer to create a product model based on rules and the powerful CAD/CAM applications that used to design, configure and assemble products, examples of which include the so-called expert systems, web-based knowledge bases involving the engineering knowledge with a knowledge-based language such as Knowledge Fusion, capturing both geometry and non-geometry attributes of a given part or assembly, writing the rules and essentially giving us the ability to capture intelligence and engineering know-how, and becoming an critical part of business strategy (Kulon et al., 2006; Homen de Mello & Sanderson, 1991b).

In addition, numerous researchers have employed an artificial intelligence (AI) tree search or graph search methodology to generate an assembly sequence (Chen et al., 2004a). Unfortunately, the search space increases explosively when the number of components in a design grows. To relieve this computational and combinational complexity, heuristic rules, ant colony optimization, algorithm of Self-Guided Ants (ASGA) and genetic algorithms (GAs) have been used in the searching process

(20)

6

(Marian et al., 2003; Chen et al., 2004a; Wang et al., 2005; Tripathi et al., 2009).

Other studies have used the Hopfield and BPNN as the means to generate optimum or near-optimum assembly sequences (Chen, 1990; Hong & Cho, 1993; Sinanog˘lu, 2006). An AI system has three key components: representation, reasoning and learning, and must be able to achieve three things: store knowledge, apply the knowledge stored to solve problems and, obtain new knowledge through experiences.

An optimal assembly plan must simultaneously consider crucial factors, such as physical and geometrical constraints, the similarity of assembly operations, the frequency of tool replacement, etc. Besides, the optimization problem of assembly sequence, like traveling salesperson problem (TSP), knapsack problem (KP), minimized spanning tree (MST), constrained shortest path problem (SP), as well as many well-known decision problems, belongs to one type of typical NP-complete/

NP-hard optimization problems with the complexity of precedence graphs, which cannot be solved problem in polynomial time (Ramos et al., 1998; Levitin, 2007).

Integration of process planning and scheduling (IPPS) is a critical research issue to achieve manufacturing and assembly planning optimization using particle swarm optimization (PSO), genetic algorithms (GAs) and simulated annealing (SA) as a combinatorial model to tackle significant barriers (i.e., vast search spaces, complex technical constraints and operational accessibility) and avoid being trapped into local solution of optimization (Guo et al., 2009). Besides, the application of soft computing (SC) including core methodologies of Fuzzy Logic (FL), Artificial Neural Networks (ANN) and GAs providing an efficient way of solving multi-variables and non-linear system has been reported in civil engineering, human-machine workstation design, and the optimization of clamping forces (Saridakis & Dentsoras, 2008).

Systematic Optimization (SO) is usually fulfilled using two stages. The first stage is to create a feasible model describing the behavior of the system, whereas the second stage is dedicated to figuring out the best parameter values (input variables) for achieving the optimal response (output characteristics) of the system (Anjum et al., 1997). Mostly one can develop a regression model, in the first stage, and fit it the given data to find the best parameter values by trial and error. However, the process of which could be very time consuming and very misleading for highly non-linear systems.

(21)

7

Section 2 Back-Propagation Neural Network

An artificial intelligence (AI) system, including knowledge-based and ANN system, has three key components: representation, reasoning and learning, and must be able to achieve three things: store knowledge, apply the knowledge stored to solve problems and, obtain new knowledge through experiences (Sage, 1990). Artificial neural networks (ANN) have widely applied in signal processing, pattern recognition, medical diagnosis, speech recognition, identification of geophysical features and mortgage evaluation (Haupt, 2004). In much of the ANN literature, back-propagation neural networks (BPNNs) have been adopted because they have the advantages of a fast response and high learning accuracy (Maier & Dandy, 1998; Liu et al., 2001; Lee et al., 2001; Yao et al., 2005; Chen & Hsu, 2007). The superiority of a network’s functional approach depends on the network architecture and parameters, as well as the problem complexity. If inappropriate network architecture or parameters are selected, undesirable results may be obtained. Conversely, the results will be much more significant if good network architecture and parameters are selected. The BPNN consists of an input layer, hidden layer, and output layer. The parameters for the BPNN include the number of hidden layers, number of hidden neurons, learning rate, momentum, etc. All of these parameters can significantly impact the performance of the neural network.

Fogel (1991) proposed a final information statistical (FIS) process based on Akaike’s information criterion (AIC) to determine the number of hidden layers and neurons. One hidden layer is sufficient to compute arbitrary decision boundaries and quite adequate to model nonlinearity in most cases of the BPNN (Khaw et al., 1995;

Anjum et al., 1997). The limitation of Fogel’s research is that the process can only perform simple binary classifications. Murata and Yoshizawa (1994) and Onoda (1995) respectively proposed methods to improve AIC. These methods, called the network information criterion (NIC) and neural network information criterion (NNIC), use statistical probabilities together with an error energy function to determine the number of hidden neurons.

In this research, the steepest-descent method was used to find the weight change and to minimize the error energy function. The activation function is a hyperbolic sigmoid function. According to past studies (Hush & Horne, 1993; Cheng & Tseng, 1995), there are a few conditions for network learning termination: (1) when the root

(22)

8

mean square error (RMSE) between the expected value and network output value is reduced to a preset value; (2) when the preset number of learning cycles has been reached; and (3) when cross-validation takes place between the training samples and test data. The first two methods are related to the preset values. This research adopts the first and second approaches by gradually increasing the network training time to gradually decrease the RMSE until it is stable and acceptable. The RMSE is defined as follows:

( )

∑

=

−

= ^N

i

i y

N d RMSE

1

1 2 ; (1) where N, di, and yi are the number of training samples, the actual value for training sample i, and the predicted value of the neural network for training sample i, respectively.

In network learning, input information and output results are used to adjust the weighting values of the network. The more detailed the input training classification and the greater the amount of learning information which are provided, the better the output will conform to the expected result. Since the learning and verification data for the BPNN are limited by the functional values, the data must be normalized by the following equation:

(

max min

)

min

min max

min D D D

P P

P

PN P × − +

−

= − ; (2)

where PN is the normalized data, P is the original data, Pmax is the maximum value of the original data, Pmin is the minimum value of the original data, Dmax is the expected maximum value of the normalized data, and Dmin is the expected minimum value of the normalized data.

When applying the neural network to the system, the input and output values of the neural network fall in the range of [0.1, 0.9].

The learning (training) process and algorithm of BPNN can be stated below.

Step 1. Setup the network parameters of learning rate (η) - the smaller the changes to the synaptic weights from one iteration to the next, and the smoother will be the trajectory in weight space. If the learning-rate is too large in order to speed it up, the resulting large changes may become unstable (say, oscillatory).

Step 2. Setup the synaptic weight (W_xh, W_hy) and bia (θ_h, θ_y), where h is the amount of neurons: neuron is a fundamental information- processing unit to neural

(23)

9 networks;

Step 3. Input training samples of input vector (Xi) and desired (target) output vector (Tn), where i is the amount of input signals, and n denotes discrete time of an iterative process involved in adjusting the synaptic weights of neuron h.

Step 4. Calculate the hidden-layer output Hh and actual output Yj, where j is the amount of neurons of output-layer.

Calculate the hidden-layer output Hh

Vh= _h _ _ih _i _ _h

i

net =

∑

W xh •X −θ h (3)

Hh= 1

( )

1 exp ^h

h net

f net = ₋

+ (4) Vh, Vy are activation potentials or induced local fields generated in hidden-layer and output-layer. Say, f is an activation (threshold) function for limiting (squashing) the amplitude of the output of a neuron.

Calculate the actual output Y

Vy= __j _hj _h _ _j

i

net =

∑

W hy •H −θ y (5)

Yj= 1

( )

1 exp ^j

j net

f net = ₋

+ (6) Step 5. Calculate the local gradient δ

Calculate the output-layer local gradient δj

δ_j=Y_j• −(1 Y_j) (• Tn_j−Y_j) (7) Calculate the hidden-layer local gradient δh

δ_h= _h (1 _h) _ _hj _j

j

H • −H

∑

W hy δ (8) Step 6. Calculate the synaptic weight correction ΔW and bias correction Δθ

Calculate the output-layer synaptic weight correction ΔW_hy and bias correction Δθ_y.

_ _hj _j _h

W hy ηδ H

Δ = (9) _y_j _j

θ ηδ

Δ = − (10) Calculate the hidden-layer synaptic weight correction ΔW_xh and bias

(24)

10 correction Δθ_h

_ _ih _h _i

W xh ηδ X

Δ = (11) _h_h _h

θ ηδ

Δ = − (12) To avoid the danger of instability, the delta rule can be modified by including a momentum (α)- controlling the feedback loop acting around ΔW, as shown by (Rumelhart et al., 1996).

i hX )

n

( ηδ

αΔ − +

=

ΔW_xh_ih W_xh_ih 1 (13) Step 7. Update the weight matrix W and bias vector θ

Update the output-layer weight matrix W_hy and bias vector θ_y

_ _hj _ _hj _ _hj

W hy =W hy + ΔW hy

(14) _y_j _y_j _y_j

θ =θ + Δθ (15) Update the hidden-layer weight matrix W_xh and bias vector θ_h

_ _ih _ _ih _ _ih

W xh =W xh + ΔW xh (16) _h_h _h_h _h_h

θ =θ + Δθ (17) Step 8. Repeat 3~7 until reach learning convergence (i.e., ej value amounts to constant) or a limited (pre-set) learn cycle (epoch).

tor.

output vec the

of element jth

the vector;

reponse

desired the

of element jth

the is signal;

error the is

j j j

Y Tn e

Y Tn

e = − ;

Step 9. Determine the RMSE (Root mean square error) while the learning (training) is convergence.

2

1

( ) /

n

j j

j

RMSE Tn Y n

=

∑

− ⁼

∑

= n

1 j

2

j /n

e (18)

Section 3 Taguchi Method

Taguchi’s robust parameter design is a systematic method which normally selects an appropriate formulation of the S/N ratio and calculates the S/N ratio for each treatment. There are three types of S/N ratios: nominal the best, the larger the better, and the smaller the better. Most engineers choose the highest S/N ratio treatment as the preliminary optimal initial process parameter setting. Taguchi method has also been used to design the parameters for neural networks in previous research (Khaw et al., 1995; Santos & Ludermir, 1999). Khaw et al. (1995) applied Taguchi method to design

(25)

11

⎟⎟⎠

⎜⎜ ⎞

⎝

× ⎛

=10 log ₂² /

S N y

S

⎟⎟⎠

⎜⎜ ⎞

⎝

× ⎛

−

=

∑

= n

i yi

N n S

1 2

1 log 1

10 /

] [

log 10 1 -

log 10

/ ² ²

1

2 y S

n y N

S

n

i

i ⎟= × +

⎠

⎜ ⎞

⎝

× ⎛

−

=

∑

=

yi i n

y yi _S

yi

the parameters and verified that the method could rapidly and robustly design the optimal parameters. Santos and Ludermir (1999) applied a factorial design to assist the design and implementation of a neural network. The formulae of the three types of S/N ratios are given as follows:

nominal the best: , (19)

the larger the better: , and (20)

the smaller the better: ; (21)

where is the response value of a specific treatment under replications, is the number of replications, is the average of all values, and is the standard deviation of all values.

A crucial procedure using the Taguchi method can be conducted as follows- using analysis of variance (ANOVA) of the S/N ratio, determine factors having a significant effect on the S/N ratio; then identify levels of these factors to maximize the overall S/N ratio through the main effect graphs of S/N ratio. In case interactions are significant, information obtained from the plots of interactions is employed to determine the optimal setting for the corresponding factors.

Section 4 Response Surface Methodology

Response surface methodology (RSM) is a specialized technique for design of experiment (DOE), which is also a combination of mathematical, statistical and optimization techniques useful for analyzing the problems and applying to create model and optimize designs (Anjum et al., 1997; McDonald et al., 2007). The role of RSM focuses on design improvement and optimization. It is crucial to find the precise transfer functions between variables and output responses, and desirable to conduct a low-resolution and two-level factorial experiment called “screening experiment”. The tasks of screening experiment can be summarized below:

1. Eliminating insignificant variables from the further assessment and analysis.

2. Determining optimal parameter settings of many discrete variables.

3. Identifying a small number of key variables for further investigation.

(26)

12

The response surface methodology (RSM) is designed to search for an optimal solution after a screening experiment where the output response is related to a few number of important continuous variables. In practice, design improvement and optimization problems are more than one output response, and mathematically called multi-objective optimization. There are three kinds of optimization criteria, in RSM, which are maximum (the larger, the better), minimum (the smaller, the better) and target (nominal the better).

Section 5 Research Motivation and Objectives

Traditionally, a larger number of components will produce more constraints to an assembly product, and increase the complexity of solving assembly problem. The motives of this study firstly assist engineers to develop a correct CAD-oriented explosion graph and find a feasible assembly sequence using practical graph-based heuristic rules such as Above graph, relational model graph and assembly precedence diagram; and consequently create a feasible assembly sequence through an BPNN-based robust predictor and a Siemens NX/KF-based KBE system.

The purpose of this research is described below.

1. To achieve a feasible assembly sequence via three-stage graph-based reasoning and systematic planning.

2. To evaluate the difficulties of assembly with some mating-criteria and assembly constraints, validate the available assembly sequences based on CAD-based Knowledge Fusion (KF) programming language and robust BPNN engines under knowledge-based database.

3. To promptly find a near-optimal assembly sequences via an intelligent KBE system.

4. Some kinds of real-world products are utilized to evaluate the feasibility of the proposed KBE system.

Section 6 A guide to This Dissertation

The remainder of this dissertation is denoted as follows: Chapter 2 specifies the working scheme and process. Chapter 3 and Chapter 4 describe feature-based assembly modeling and knowledge-based assembly planning. Chapter 5 explains how to create a feasible assembly sequence planning through a three-stage procedure.

Chapter 6 illustrates the novel KBE system for assembly sequence planning.

Expected achievements and suggestions are finally made in Chapter 7.

(27)

13

CHAPTER 2 Working Scheme and Processes

Generally, assembly sequence planning consists of feature-based assembly modeling (i.e., all assembly information are modeled), knowledge-based assembly sequence generation and interactive assembly planning systems demonstration. This study develops a three-stage integrated approach with some heuristic working rules to assist the planner in generating a best and most effective assembly sequence. In the first stage, Above Graph and transforming rule are used to create a correct explosion graph of the assembly models. In the second stage, a three-level relational model graph, with geometric constraints and assembly precedence diagrams (APDs), is generated to create a complete relational model graph, an incidence matrix and a feasible assembly sequence. In the third stage, the back-propagation neural network (BPNN) engine via parameter optimization using Taguchi method and design of experiment (DOE) is employed to predict the available assembly sequences. The afore-mentioned BPNN engine, created by the toy car model- as a learning (training) sample, and toy motorbike model and a brushless DC fan as testing samples, is dedicated to evaluating the feasibility of the proposed model in terms of the differences in assembly sequences. As conducts the above three-stage jobs, a knowledge-based engineering (KBE) system can be installed to assist the assembly engineers in promptly predicting a near-optimal assembly sequence.

The working concepts and procedures fall into two portions: the first is to construct a graph-based assembly sequence planning and the second is to implement a systematic optimization through a robust BPNN engine and KBE system. The proposed flow chart is shown in Fig. 2. Initially, detailed data is input from a 2D engineering drawing and related assembly information into a CAD assembly package.

Then, the correct explosion graph is developed using the transforming rules. Finally, the relational models are generalized to represent the assembly precedence relations, and an evaluating mechanism is then employed to find a global feasible solution. The planning process is recursive until the defined criteria are satisfied. The main outputs of the integrated graphs and BPNN-based assembly planning are the complete RMG, APD, and BPNN engines. In addition, Fig. 3 represents the knowledge-based engineering (KBE) system rendering a NX/KF-based operational interface to access

(28)

14

the potential graphs and BPNN-related details via different types of databases, and a robust BPNN engine (see Fig. 4) dedicated to promptly generating a near-optimal assembly sequence.

Figure 2. The flow chart of this proposed study.

(29)

15

Figure 3. KBE model for assembly sequence optimization.

Section 1 Constructing an ANN-based ASP Prediction Engine

Step 1: Obtaining a feasible ASP using a graph-based methodology and manual

operation.

Step 2: Catching CAD-oriented ASP knowledge from a NX/KF CAD system.

Step 3: Building up a robust BPNN-based prediction engine using Taguchi method and design of experiment.

A most adaptable ANN-based prediction engine for assembly sequence planning using Taguchi method and design of experiment is created as shown in Fig. 4.

(30)

16 Collect related assembly information

(Section graph, Above graph, Part name, Assembly code ...)

Optimal ASP database

Optimal?

End Start

Generate the correct Explosion graph

& the relational model garph

Generate the assembly precedence diagram(APD)

No

Yes

CAD-oriented ASP model

Conduct BPNN training and test ( using NeuroSolution package)

CAD-oriented ASP Knowledge (Features, Weight, Volume,

Total penalty value, AI)

ANN-based ASP prediction engine database Manual procedure

Neural networks

Catch CAD-oriented ASP Knowledge

(CAD potential data:

Features, Weight, Volume, and, assembly relation data:

Total penalty value, AI) CAD system

Figure 4. The construction flowchart of ANN-based ASP engine.

Section 2 Creating Near-Optimal ASP via KBE System

The KBE database representation and flowchart of creating knowledge-based ASP report can be seen in Fig. 5.

(31)

17

Explosion Graph

Assembly part Part Feature CAD model ...

CAD- oriented ASP Knowledge

(Features, Weight, Volume, Total penalty value,Contact number) UG

(CAD/CAM/CAE)

ANN-Based ASP engine

ASP KBE System

User

Simulation animation & Certification Knowledge Database

(Knowledge & Case storage)

Knowledge report

D e v e lo p a ro b u st A N N -B a se d A S P e n g in e

C re a tin g n e a r- o p tim a l A S P C e rtific a tio n ?

E n d S ta rt

G e n e ra te a n ex p lo sio n g ra p h & a fe a sib le a sse m b ly se q u e n c e

C o n d u c t S im u la tio n (A n im a tio n ) C a tc h C A D -o rie n te d A S P K n o w le d g e

(C A D m o d e l, A sse m b ly re la tio n sh ip , F e a tu re s, W e ig h t, V o lu m e , T o ta l p e n a lty v a lu e , C o n ta c t n u m b e r)

N o

Y e s C a se S to ra g e

A S P

K n o w le d g e D a ta b a se

Figure 5. The ASP flowchart applied on KBE system.

(32)

18

CHAPTER 3 Feature-Based Assembly Modeling

In recent years, features, combining geometric and functional information, have been introduced in modeling and planning for product assembly and components manufacturing. Assembly features represents assembly information used to handle a component and connections information between components (Holland &

Bronsvoort, 2000). Analysis and modeling of the final mechanical or plastic products can be one of the most essential steps in assembly planning. The product model ought to contain all information required such as geometrical information, component information, topological information, technological information, and final assembly information. The fundamental assembly modeling strategy is based on the mating features of its components. The mating features are the faces in contact including low level polygon faces and higher level features for the above products (Eng et al., 1997).

Section 1 Feature Fundamentals

1. Definition of a feature

Features represent physical constituent, generic shape, predictable properties and the engineering meanings of parts or assembly geometry. The shape of a feature can be expressed in terms of dimension parameters, enumeration and interactive relationships of topological/ geometric entities or geometry construction procedure.

The engineering meaning may include the function formalization the feature serves, the way of production and actions to be taken. Feature attributes may be shape, dimensions or size tolerances, and feature to feature relation attributes may involve positioning, geometric constraints and compatibility. Assembly attributes may encompass such information as mating surface, fits/ clearances, depth of insertion and relative orientation vectors. Modeling is an activity where the models or representations of how devices or objects of interest can be made or considered such as words, drawing or sketches, computer programs, or mathematical formulas (Clive

& Patrick, 2009). The modeling activity using those features is called feature-based modeling, and popular features supported by most feature-based modeling are manufacturing features such as chamfer, hole created by drilling, fillet, slot, pocket created by milling, and so on.

(33)

19 2. Feature types and properties

Features are used to include a wide variety of entities which need to distinguish between the following types (Andrew, 1992).

1. Form features: portions of nominal geometry or recurring shapes.

2. Precision features: deviations from nominal form, size and location.

3. Material features: material composition, treatment and condition.

4. Technological features: non-geometric (i.e., engineering and manufacturing rules) parameters with respect to function, performance, etc.

5. Assembly features: part relative orientations, interaction surfaces, fits and kinematic relations.

The shape and engineering significance properties of a feature can be separated into generic property, regardless of its size and location, and specific property. The list of feature properties has been represented below: Generic shape, dimension parameters, location, orientation, tolerances, construction procedure for geometric model, recognition algorithm, inherited parameters, validation rules and procedures, and non-geometric attributes.

3. Feature relationships

The major feature-based function is to create association between entities in a product definition database. This association of entities makes it possible to encompass design or geometric constraints and do geometric judgments. Some generic relationships or dependencies are the following:

1. Direct parameter Inheritance- One or more dimensions of a feature must be calculated from parameters of others.

2. Orientation dependencies- The orientation of a feature is fixed by its instant parent.

3. Positioning dependencies features- The location of a feature is fixed by the surface, edge, or corner of another feature.

4. Feature face dependencies- One or more geometric entities of a feature are constrained to lie on a specified face or faces of other features.

Section 2 Geometrical Information

Geometrical Information determines the shape and dimensions of components as well as their relative positions within the final assembly. As the parts are modeled in a CAD system, this information can be extracted directly from the corresponding

(34)

20

geometrical database. The components consist of one or several objects, built from shell- describing a polyhedral surface with facets, facet- represented by several loops, loop- defined by the vertices of the polygon (i.e., a loop is a list of half edges forming a closed circuit, and thus any face is bounded by one peripheral loop), and vertices- the set of points represented by three vectors of real numbers X(i), Y(i), Z(i) with i the index of part’s boundary, and X, Y, Z the coordinates of this point.

Geometry level assembly models are based on geometrical information to specify and model assemblies in terms of mating conditions (relationships between geometric entities). Some examples of such relations are: against (two faces against each other), align coplanar (two faces aligned to lie in the same surface), co-axial (two axes aligned to lie in the same straight line) and coincident (two points constrained to be coincident).

Geometric modeling systems are normally classified as wireframe modeling systems, surface modeling systems and solid modeling systems. And the data structures needed to describe a solid can fall into three types: the first type is called constructive solid geometry (CSG) tree applying Boolean operation on the primitives.

The second type is called boundary representation (B-Rep) providing connectivity information in terms of the boundary information (i.e., vertices, edges, and faces) for a solid. The third type is called a decomposition model storing a solid as an aggregate of simple solids with a set of surfaces and the mathematical description of a shape (Lee, 1999).

Section 3 Component Information

Component Information characterizes the component’s roles and supplies the physical features of the components, specified independently, which constitute the assembly. These features are exploited during generation and selection of assembly sequence.

1. Standard categories

Some components have been designed to play particular roles such as fastening, sealing and the possibility of relative motion in any types of assembly; more often, they require special treatments, categorized by the screw, the bolts, the pins, the clips, the bearings, the springs, the o-rings and the threaded parts, during assembly.

2. Physical features

(1) Base part: In an assembly line, a part generally presenting minimum

(35)

21

handle-ability can be regarded as a base component. More generally, the operator may determine several parts combined into a subassembly which will thus be present at the beginning of the assembly plan.

(2) Handle-ability: The handle-ability is taken into account in determining which component is to remain fixed during assembly. Screws and bolts are always fastened by the robots, and will offer maximum handle-ability. The operator may specify how easy it is to handle a part. There are five options given to the operator: very easy, easy, normal, difficult and very difficult to manipulate.

(3) The man-machine interfaces (MMIs): The aim is to show how the user interacts with the system in order to introduce the component information. The MMIs are composed of a read/write window which contains all the mandatory fields to be filled.

Section 4 Assembly Information

1. Possible assembly directions

Assembly information characterizes the features of the whole assembly, which is the only one global feature: the possible assembly directions. When an assembly presents axial symmetry, generally only two directions need to be considered: the positive and negative directions along the axis of symmetry.

2. The man-machine interfaces

The man-machine interfaces (MMIs) having developed relating to the final assembly information are similar to the component information, and used to log on the assembly directions.

Section 5 Topological Information

Topological Information describes relations between the various parts of the assembly, and it includes two classes of data: parts preventing removal of other parts and the product graph.

1. Parts preventing removal of other parts

The type of data used very often by the various assembly planning modules, which makes a list of component preventing removal, in each of the six assembly directions, of each part in the final assembly.

If a part is an obstacle to another one in a given direction, then the second part is an obstacle to the first one in the opposite direction. Here are three steps involved in obstacle detection:

(36)

22

(1) Intersection of projections of parallelepiped envelopes parallel to the coordinate axes and completely enclosing the considered part;

(2) Intersection of apparent outlines (of a part) which is the polygon (including points and edges) external to the projection of the part in a plane perpendicular to a given direction;

(3) Search for obstacle facets: this search is activated only if the two previous tests result in suspicion of a collision. The test of superposition using the projections themselves is the classical problem of the intersection of the two polygons.

2. Product graph representation

The product graph is identified in four steps:

(1) Definition of part clusters- it is best to define clusters of parts having similar features: standard fasteners of the same type, size, orientation and attached to a same part; parts with the same dimensions and similar relative positions with respect to another component.

(2) Identification of contacts between parts- for each part in the assembly and each direction parallel to the axes of the global reference frame, all parts can be listed in physical contact with the part considered, and preventing removal of the examined part; Once a contact between two components is found, a connection is created between the two components.

(3) Determination of contact types- contacts belong to three major categories:

fixations (no significant motion is possible between parts forming the connections, such as screwing, insertion of a pin, clipping and o-ring insertion), insertions (generally connections between cylindrical surfaces) and placements (two components of the connection are not actual connected where the contacts are usually planar);

(4) Definition of additional fixation means- it is possible for the operator to define additional fixation means: pressure fit and gluing, which stabilize the connection and eliminate any remaining type of connection.

Section 6 Technological Information

The assembly sequence can be used as soon as product design is completed.

General knowledge of cell layout will enable someone to choose the most efficient assembly sequences. Fixture features constitute an important part of such knowledge,

(37)

23

which is why left the user frees to specify certain features of these fixtures in the form of constraints.

Assembly directions parallel to the coordinate axes and geometrical concepts can be adapted and affected to a spatial transformation, namely a rotation. Axis X denotes assembly direction (oblique) and the Y and Z axes define a plane perpendicular to the assembly direction.

Section 7 Processing Complex Assembly Directions

In several assemblies, it is not always possible to assemble the parts in straight line motions. The basic problem is to find a path to disassemble two components or subassemblies from their final assembly position. The majority of mechanical assemblies can be treated as combinations of straight-line motions. Hoffman has proposed an algorithm solving this problem as follows.

1. Detection of complex assembly directions- during the generation of geometrical precedence constraints, the program can check if it is not possible to disassemble them considering combinations of straight line motions.

2. Elaboration of the solution tree- each node of this tree represents the current relative position of the two parts (or subassemblies), and the root node is the final assembly position. A straight motion is described by its direction and by its length. If this length becomes infinite, then a possible path will be found to disassemble the two parts.

3. Computation of other positions from the current one- moving the first part A until a collision with a facet of the other part B is discovered, next moving part A around facets of part B until A does not touch B any more. Only motions along an assembly direction are permitted.

4. Choice of a solution in the tree- A solution consists of a sequence of positions taken by the part during the disassembly phase. The choice of these positions uses techniques of graph search and heuristics.

Section 8 Feature-Based Assembly Planning

Most products belong to assemblies of discrete components where a critical manufacturing stage is the physical assembly of the components. Assembly planning can be roughly divided into three phases: selection of assembly method- to recognize the most suitable assembly method and consider the type of the assembly system which will be used; assembly sequence (route) planning- to recognize a sequence of

(38)

24

assembly operations with the chosen assembly method; assembly operations planning- to focus on the assembly directions, mating movements and application of fasteners (Gu & Norrie, 2006).

Assembly modeling is to provide a logical structure for grouping and organizing parts into assemblies and sub-assemblies with parametric constraint relationships between parts. The structure enables a designer to identify individual parts, keep track of associated part data, and maintain relationship data among parts and sub-assemblies. Mating conditions between parts in the assembly are crucial relationship data which identify how the part is connected to others. Data on fit, position, and orientation, which are derived from mating conditions, can specify how parts are joined in the assembly and often include allowable tolerances. Exploded views created from an assembly model are helpful in clearly showing the physical relation and describing their assembly instructions of all the parts in complex assemblies (Lee, 1999).

Assembly process problems (poor quality, poor efficiency, higher cost, etc.,) have resulted in the design for assembly (DFA) approach where the most economical production process is already selected during the design stage. Features can be applied for the data retrieval and geometric reasoning tasks required for the above approach. Obviously, the actual physical arrangement of components is fundamental for the assembly and their sequencing. Assembly sequence planning is closely related to the path planning problems of robotics and inspection planning; its task can be defined as assembly sequence generation.

For robotic assembly, the assembly operation planning involves the design of a valid sequence of rigid motions (translations and rotations) that will bring a component into its final position. Various tools for assembly must also be designed, such as grippers, jigs, fixtures, pallets, and component feeders. Assembly path planning includes the possible access directions for positioning the component and a geometric search (i.e., the obstacle polygon method- a popular method for translational path planning) for free space.

中 華 大 學 博 士 論 文

中 華 大 學 博 士 論 文

題目：整合性組裝順序規劃之 KBE 系統

A Knowledge-Based Engineering System for Integral Assembly Sequence Planning

系 所 別：科 技 管 理 研 究 所 學號姓名： D09403015 戴 培 豪 指導教授：陳 文 欽 博 士 指導教授：謝 玲 芬 博 士

中 華 民 國 九 十 八 年 八 月

中 華 大 學

(博碩士論文授權書)

中 華 大 學 博 士 班 研 究 生 論 文 指 導 教 授 推 薦 書

科技管理研究所博士班戴培豪君所提之論文 整合性組裝順序規劃之 KBE 系統，係由本人等 指導撰述，同意提付審查。

指導教授 (簽章) (簽章)

中華民國九十八年 八 月

中 華 大 學 博 士 班 研 究 生 論 文 口 試 委 員 會 審 定 書

科技管理研究所博士班戴培豪君所提之論文 整合性組裝順序規劃之 KBE 系統，經本委員會 審議，符合博士資格標準。

論文口試委員會 召集人

委 員

所 長

中華民國九十八年 七 月 二十九 日

謝辭

中 華 大 學 博 士 論 文

題目：整合性組裝順序規劃之 KBE 系統

A Knowledge-Based Engineering System for Integral Assembly Sequence Planning

系 所 別：科 技 管 理 研 究 所 學號姓名： D09403015 戴 培 豪 指導教授：陳 文 欽 博 士 指導教授：謝 玲 芬 博 士

中 華 民 國 九 十 八 年 八 月

ABSTRACT

摘 要

CONTENTS

LIST OF TABLES

LIST OF FIGURES

CHAPTER 1 Introduction

Section 1 Literature Review and Hierarchical ASP

Section 2 Back-Propagation Neural Network

( )

∑

(

)

∑

∑

∑

θ ηδ

∑

∑

Section 3 Taguchi Method

∑

∑

Section 4 Response Surface Methodology

Section 5 Research Motivation and Objectives

Section 6 A guide to This Dissertation

CHAPTER 2

Working Scheme and Processes

Section 1 Constructing an ANN-based ASP Prediction Engine

Section 2 Creating Near-Optimal ASP via KBE System

CHAPTER 3

Feature-Based Assembly Modeling

Section 1 Feature Fundamentals

Section 2 Geometrical Information

Section 3 Component Information

Section 4 Assembly Information

Section 5 Topological Information

Section 6 Technological Information

Section 7 Processing Complex Assembly Directions

Section 8 Feature-Based Assembly Planning

中華大學博士論文

中華大學博士論文

系所別：科技管理研究所學號姓名： D09403015 戴培豪指導教授：陳文欽博士指導教授：謝玲芬博士

中華民國九十八年八月

中華大學

中華大學博士班研究生論文指導教授推薦書

科技管理研究所博士班戴培豪君所提之論文整合性組裝順序規劃之 KBE 系統，係由本人等指導撰述，同意提付審查。

指導教授 ^(簽章) ^(簽章)

中華民國九十八年八月

中華大學博士班研究生論文口試委員會審定書

科技管理研究所博士班戴培豪君所提之論文整合性組裝順序規劃之 KBE 系統，經本委員會審議，符合博士資格標準。

論文口試委員會召集人

委員

所長

中華民國九十八年七月二十九日

中華大學博士論文

系所別：科技管理研究所學號姓名： D09403015 戴培豪指導教授：陳文欽博士指導教授：謝玲芬博士

中華民國九十八年八月

摘要