RESEARCH Development You will find the figures mentioned in this article in the German issue of MTZ 3/2004 beginning on Methods page 220.
Prototypeinsatz evolutionärer Algorithmen in der Motorenentwicklung bei Volkswagen
The Use of Evolutionary Algorithms in Prototypes in Engine Development at Volkswagen AG
How is a new product or a new solution developed? The ever-faster rates at which car models are changed mean that the product development process has to be continuously improved. In a joint research project between Volkswagen AG and the University of Magdeburg, a procedure for adaptation design based on Autogenetic Design Theory (ADT) has been developed. The Chair of Information Technologies in Mechanical Engineering has developed a design system prototype that has already been used to solve different tasks in engine development.
1 Motivation
By Steffen Clement, André Jordan, Ridwan Sartiono, Sándor Vajna and Peter Kellner
MTZ worldwide 3/2004 Volume 65
In engine development at Volkswagen AG, a product developer optimises various kinds of products (e.g. a coolant jacket or an intermediate exhaust pipe) by using the conventional “trial and error” procedure. Due to the large number of possible solu-
tions available, there is no guarantee that the product developer will find the very best one. It is a well-known fact that at least 90 % of all products in engine development are adaptation designs. In order to reduce the extensive development times, costs and risks that are caused by conventional approaches, a procedure to automate
27
RESEARCH
Development Methods
the computer-based adaptation design process was developed. In this procedure, an existing solution is modified until it fulfils the new requirements as ideally as possible. The benefit of this solution is that the desired (“better”) characteristics of the initial solution can be inherited by the succeeding solution. The most well-known design process model presented in VDI guidelines 2221 and 2222 describe product development as a process with different phases. In each phase, various alternatives of a solution are developed and compared, and these alternatives then compete with each other. Solutions that fulfil the requirements sufficiently are selected for further treatment, combination and development. This approach is characterised by searching for, accepting and learning from existing solutions and by testing, evaluation, selection and combination. These are in fact typical processes of natural evolution. Autogenetic Design Theory (ADT) employs this analogy to natural evolution in order to describe and support product development processes. It provides an evolution-oriented approach for modelling and supporting design activity as the substantial activity within the product development process [1, 2, 3]. ADT points out that evolutionary operators and the driving forces of evolution are important for improving both products and the development process. In all phases and at all organisational levels of the evolving new product, it is possible to develop new solutions (called “individuals”) by evolutionary procedures. In order to develop new solutions, the evolutionary procedure (trial and error) dominates compared to the deductive approach of conventional design methodology. In accordance with the principle of evolution, the best solutions (i.e. those that fulfil the requirements better) are selected from the preceding solutions (“parents”) by selection pressure. Selection pressure is formed by the requirements as well as by the driving forces, initial conditions and boundary conditions of the solution space, which may also change during the development process. This so-called autogenetic behaviour is recognisable in the development of any (partial) solution, because each developed solution passes through this process of development [3]. Adaptation is continuous and targeted, and can therefore be understood as an optimisation process [1]. For the computer-based implementation of ADT, so-called evolutionary algorithms are used. These algorithms have been used for many years in the field of artificial intelligence and are able to simulate
28
the processes of natural evolution. Furthermore, efficient approaches and methods are well known from cognitive science and from Chaos Theory. These methods can be transferred by analogy to the product development process. Due to these analogies, it is possible to perform an expanded solution search and to support creative thinking processes in early stages of the development process [3]. 2 Adaptation Design Using ADT at Volkswagen AG
The aim of an adaptation design it is to change an existing solution in such a way that it fulfils the new requirements. ADT is used to obtain an inheritance of “good” characteristics (this also includes the fulfilment of the new requirements) of the initial solution to the succeeding solutions. 2.1 General Procedure
Development can be defined as an activity in which all life cycle characteristics of a product are specified by the product developer based on given geometric-material requirements. From the viewpoint of ADT, the described procedure is a multi-criteria optimisation process (pareto-optimal procedure). The process takes into account the contradictory objectives of preliminary development, the building of prototypes, testing, manufacturing, assembly and suppliers, etc. During this iterative process, the product developer sets up solutions, evaluates them, and, on the basis of the evaluation, creates a more detailed solution, Figure 1. In the engine development department of Volkswagen AG, the product development process is increasingly supported by integrated CAx tools that perform the modelling step “creation of solutions”. Computation and simulation programmes are necessary for the evaluation of the solution. Linking these tools to form a suitable optimisation algorithm simplifies and shortens the iterative procedure of solution identification. The three components modelling, evaluation and optimisation are used to construct a design system that is flexible in use and is able to run almost fully automatically. The design system is called NOA (Natural Optimisation Algorithm) and it supports the product developer in searching for a solution. Success in the product development process requires close co-operation between all departments. The following example shows how co-operation between the computation engineer and product developer can be organised and supported in detail. The focus is on adaptation design.
At Volkswagen AG, the product developer creates a CAD model of the product. For the purpose of co-operation with different departments, the product developer uses an already existing CAD model. On the basis of his experience, the product developer modifies and details the existing model in such a way that it fulfils the new requirements. After this, he transfers the model to the computation department, where the future characteristics of the product are computed and/or simulated by special programmes (CFD, MBA, FEA, analytical computations, etc.). The results of the computation usually show that the product does not fulfil the requirements sufficiently. In this case, the CAD model has to be changed by the product developer and a new computation has to be performed. This process is repeated until the product properly fulfils the requirements. One way to shorten development time is to reduce the number of iterations that are necessary, even though iterations cannot be completely avoided. In addition to the changes that become necessary because of the computation results, other reasons for changes (e.g. changing boundary conditions) need to be considered. These changes may lead to an adjustment of the CAD model and further computations. Usually, computations take a long time, especially if they are numerical flow computations. Additionally, the time required for re-working errors or a loss or data (for example because of the need for conversion into other formats when transferring the CAD model to the computation department) must be considered. In some cases, the evaluation of a CAD model can take weeks. This time can be shortened by the use of computation modules that are integrated into the CAD system. Such computation modules usually presuppose less expert knowledge, which means that the product developer can use these programmes directly, and can obtain computation results much faster. Due to the close link to the CAD system, conversion problems do not arise. Software manufactures continuously extend the functionality of their systems by integrating new and better modules. Nevertheless, an integrated computation module cannot replace high-quality computation software. The advantages of the approaches described here can be combined by linking design (CAD system), computation and optimisation to produce an automated evolutionary adaptation design.
MTZ worldwide 3/2004 Volume 65
RESEARCH
2.2 Implementation of the Design System
By linking geometric variation, flow computation and optimisation in an automatically operating design system for an adaptation design, development time can be shortened and the quality of the solution can be increased. The precondition is that the applied systems have to work without user interaction. The CAD system Pro/Engineer and the CFD system Vectis, which are used at Volkswagen AG, fulfil this requirement. The two systems and an optimisation component that includes an evolutionary algorithm are linked together to form the developed design system. During each session, Pro/Engineer produces a so-called “trail file”, which can be used as a batch file to control this system. For the flow computation process in the modular CFD system Vectis, a script controls all necessary phases (cross-linking, computation, and evaluation). The two systems are connected in a control loop, and are used to evaluate the solutions (individuals) that are produced by the optimisation algorithm, Figure 2. By using the trail file, Pro/Engineer is instructed to read the parameter set produced by the optimisation algorithm. Pro/Engineer creates the new geometry. The geometry is stored in a file format that can be read by Vectis. Typically, the following steps are necessary for a numerical flow computation process[4]: ■ modelling the geometry model (CAD data) ■ mesh generation (generating a structured mesh with direct reference of the elements to the global co-ordinate system) ■ discretisation in space (using the finite volume method, transforming the volume integrals of each volume into six surface integrals) ■ time discretisation (approximation procedure; a typical time step procedure is the Runge Kutta procedure). During the optimisation process, the phase to define the boundary conditions in Vectis can be skipped by a suitable transformation of the geometry data from Pro/Engineer. For the preparation of the optimisation, it is necessary to define different template files with all boundary conditions (inlet/outlet, mesh definition, physical data, etc.). These files are used in each computation. For the optimisation with the design system, it is important that the mesh definition in Vectis (i.e. number and size of elements) is valid for all geometry files produced. In the case of large variations in the model geometry, it is possible that the mesh definition no longer fits the current geometry. In such a case, several
MTZ worldwide 3/2004 Volume 65
variants of the two-dimensional basis mesh may be prepared manually. The design system then chooses the suitable variant during optimisation. The results of the flow computation are stored by Vectis in special files, which have to be evaluated by the design system. The design system uses various tools for the evaluation. These tools filter the desired measured variables from the Vectis files, determine the target criteria and compute the quality value of the solution with the aid of the target function. Using a central script, the design system calls up Pro/Engineer, Vectis and other tools. The optimisation module of the design system only creates the parameters of the solution that is to be evaluated. After termination of the script, the optimisation module expects the value that describes the quality of the solution. Due to its modular and script-based structure, the design system is very flexible, and changes can be implemented very quickly. 3 Examples
The following section describes the adaptation design of a converter system and a coolant jacket of a car engine by using the design system NOA (Natural Optimisation Algorithm). The tasks are typical of the engine development process at Volkswagen AG. In the phases of conceptual design and detail design, the basis variants of the products are prepared in co-operation with the product developer responsible. Their experience also provides the basis for deriving the target function from the determined requirements. The integration of all departments involved in the project ensures that the evolutionary computer-aided adaptation design can be realised. 3.1 Catalytic Converter System
The aim of this development task was to adapt the catalytic converter system in order to improve the uniform distribution (Uniformity Index) of the flow over the catalyst surface, and to increase the efficiency of the converter system. This was to be achieved by changing the incoming flow and by varying the inlet funnel geometry. A catalytic converter system consists of pipes, an inlet funnel, a converter (catalyst) and an outlet funnel. The funnels form the transition between the connecting pipes and the converter. The task of the inlet funnel is to distribute the flow over the catalyst surface. The converter houses the actual catalyst. In this example, a CAD model of the catalytic converter system (in particular the converter and the inlet/outlet funnel) was
Development Methods
created and varied using Pro/Engineer. Vectis was used for the computation of the flow (see section 2.3). When modelling the catalytic converter system, nine parameters within the region of the inlet and outlet funnel were to be considered. Space restrictions of the car’s underbody were used as tolerance values to define the parameter sequence (corresponding to the chromosome in biological evolution) for the evolutionary adaptation design. Furthermore, a suitable representation scheme of the optimisation problem was developed. The geometry parameters that vary the converter system were transferred as a vector to the evolutionary algorithm, Figure 3. In addition, it was necessary to formulate the requirements of the product and to derive the target function. The design was performed according to the flow chart shown in Figure 2. Both the evolutionary algorithm and the additional tools for the evaluation of different variants were controlled by the design system NOA. First, NOA generates the initial population, which can either be done randomly or can be created based on the experience of the product developer. The parameters to be modified are available in an input file, which is read by the evolutionary algorithm. Based on this file, the evolutionary algorithm creates different parameter combinations for each individual. Using prepared scripts, NOA starts (in a given order) Pro/Engineer to create the geometry of the solution and Vectis to simulate the flow characteristics of this solution. A special routine is used to evaluate the quality value of the solution from the flow computation data. Its value is given back to the evolutionary algorithm. Using an integrated selection procedure and predefined probabilities for mutation and recombination, the algorithm calculates the parameter combinations of the next generation. These parameter combinations are used for the next run. The result of the evolutionary adaptation design of the catalytic converter system was a new distribution of the incoming flow over the catalyst surface, which results in better efficiency of the catalyst, Figure 4. The task was solved and discussed together with the product developer. This solution opened up a completely new perspective on the problem of catalytic converter system development. Formerly, such variants were rejected due to the expected increased pressure loss. Now, with the possibilities offered by ADT, the product developer was guided towards new creative ideas, because the expectation of pressure loss turned out to be unfounded.
29
RESEARCH
Development Methods
3.2 Cylinder Coolant Jacket
The aim of this task was to optimise a coolant jacket by increasing the flow around the single cylinders in order to achieve better cooling. As a constraint, an even heat flow to cool down each cylinder had to be assured. The flow around the cylinder can be varied by modifying the openings in the cylinder head seals, which were therefore optimised. For geometry modelling, the female shape of the cylinder head seals was used. The restrictions were structured with the aim of having flow openings in the range of a circular or triangular surface area. For control purposes, one lead parameter per opening was sufficient. As parameters for the adaptation design, eight flow openings were selected to vary the flow between the cylinder feet and cylinder head, Figure 5. In the next step, the requirements to be fulfilled by the product were formulated and integrated into the target function. Also integrated into the target function were a uniform flow in the cylinder head and especially in the lands between the cylinders as well as the corresponding pressure loss. The procedure was performed according to the flow chart shown in Figure 2. Similar to the previous example, the algorithm produces the first population. The parameters to be modified were provided by an input file. Accordingly, NOA started Pro/Engineer and Vectis and the additional tools to evaluate the flow computation and to compute the quality value of the solution, which is then returned to the evolutionary algorithm. The result of the development was the achievement of a more uniform flow in the cylinder head, as shown in Figure 6. The task was solved and discussed together with the product developer. Due to this positive experience, NOA is being prepared for series application for the optimisation of cylinder coolant jackets.
4 Summary and Outlook
Autogenetic Design Theory uses analogies from natural evolution to describe and organise the product development process. It shows that evolutionary operators and driving forces of evolution are of significant importance in improving existing products and developing new ones. Methods and procedures of ADT provide an expanded solution search as well as the very early support of the creative thinking processes [3]. The application of suitable computerbased tools in all design phases allows the evolutionary support of the product development to be achieved. In addition to wellknown computer-aided systems (such as CAx systems), methods of Artificial Intelligence, approaches from Chaos Theory and cognitive sciences can be used to form the bridge between product development and natural evolution. The optimisation results from the two examples show that ADT procedures can already be used in practice for adaptation design. At present, a prototype of the design system NOA is being applied in engine development at Volkswagen AG. After a pilot phase of six months, a decision will be taken on whether NOA will be developed further in order to implement it as a standard tool for broad application.
References [1]
[2]
[3]
3.3 Summary of the Results
The two examples of an automated evolutionary adaptation design show the following results: ■ In the case of the development of a converter system, an equivalent result was achieved in a shorter time compared to the manual procedure. ■ The results of the improved flow distribution in the coolant jacket were confirmed in practice. ■ The procedure applied to the development of these products can be transferred to the same or similar problem definitions in the design process. For the phases of conceptual design, embodiment design and detailing, an automated computer-based procedure was implemented in the context of evolutionary adaptation design.
30
[10] Verein Deutscher Ingenieure: Richtlinie VDI 2221, Methodik zum Entwickeln und Konstruieren technischer Systeme und Produkte. Düsseldorf, VDI-Verlag 1993 [11] Verein Deutscher Ingenieure: VDI-Richtlinie 2222, Konstruktionsmethodik, Blatt 1 (Entwurf). Düsseldorf, VDI-Verlag 1993
[4] [5] [6] [7]
[8]
[9]
Bercsey, T.; Vajna, S.: Ein autogenetischer Ansatz für die Konstruktionstheorie. In: CADCAM Report 13 (1994) 2, S. 66–71 und 14 (1994) 3, S. 98–105 Wegner, B.: Autogenetische Konstruktionstheorie – ein Beitrag für eine erweiterte Konstruktionstheorie auf der Basis evolutionärer Algorithmen. Magdeburg. Otto-von-Guericke-Universität, Dissertation, 1999 Clement, St.; Jordan, A.; Vajna, S.: The Autogenetic Design Theory – an Evolutionary View of the Design Process. In: Proceedings of ICED03, Stockholm 2003 Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning. Reading MA, Addison-Wesley, 1989 Goldberg, D.E.; Smith R.E.; Earickson J.A.: SGA-C: A C-language Implementation of a Simple Genetic Algorithm. TCGA Report, 1994 Oertel, H.; Laurien, E.: Numerische Strömungsmechanik. Springer-Verlag Berlin, Heidelberg, 1995 Rechenberg, I.: Evolutionsstrategie – Optimierung technischer Systeme nach Prinzipien der biologischen Evolution. Stuttgart, Friedrich Frommann Verlag 1973 Vajna, S.; Weber, C.: Sequenzarme Konstruktion mit Teilmodellen – ein Beitrag zur Evolution des Konstruktionsprozesses. In: Konstruktion 52 (2000) 5, S. 35–38 Vajna, S.; Bercsey, T.; Clement, S.; Mack, P.: Autogenetic Design Theory – a Contribution To An Extended Design Theory. In: Proceedings of the 2000 ASME Design Engineering Technical Conferences. Baltimore, September 10 - 13, 2000
MTZ worldwide 3/2004 Volume 65