An Integrated Knowledge Based System

Published Date: 02 Nov 2017

Noureddine Ben Yahia, Romdhane Ben Khalifa, Ali Zghal

Abstract:

This paper describes an automated deep drawing process using a knowledge-base with Artificial Neural Network (ANN) system. The proposed system is organized in three knowledge-base modules. First module forms the geometrical base for cylindrical deep drawing part, second module is a technological deep drawing base, and third module is an intelligent automatic deep drawing system. The development of this system is based on the experiments and the knowledge of specialists in this field. Indeed in this research we begin with a theoretical study concerned the influence of deep drawing parameters and causes of the principal defects in manufacturing operations. However this work deals only with cylindrical parts and several typical examples of processes validated with industrialists.

Consequently we focus only in deep drawing knowledge base using ANN structure. We define different ANN input/output in the purpose to give industrial optimal solution. The proposal method can substantially reduce the time needed to design process planning and the results are enough of consistent and sufficiently promising

Keywords: Computer Aided Manufacturing, Knowledge-base, Deep drawing process, Artificial Neural Network

1 Introduction

Currently the process of deep drawing passes by several stages in many companies. Indeed the research is oriented toward the automation of the manufactures, as the automatic range in machining, in folding, in carving, as well as in deep drawing. The automation of deep drawing forms the global object of this paper.

Recently many researchers have worked to develop Computer Aided Process planning and deep drawing die to ease the difficulty of die designers and process planners and to reduce manufacturing lead time. Cheok et al. developed some aspects of a knowledge-based approach for automating progressive metal stamping die design [1]. Pilani et al. developed IA based expert system for forming die using Artificial Neural Network [2]. Fang et al developed a Rule-Based Deep-Drawing Process Planning for Complex Circular Shells [3]. Lee et al. [4], proposed an axi-symmetric shell element for the multi-step inverse analysis for more accurate prediction of design variables from initial blank shape, strain distribution, to intermediate shapes. This approach is more accurate and the punch increment for each step is much larger for the conventional incremental analysis. Zhang et al. proposed a Computer Aided Process Planning (CAPP) using in multi-stage, non axi-symmetric sheet metal deep drawing by a Case Based Reasoning (CBR) approach [5]. Several other approaches proposed aided systems for decision in sheet metal, as well as many models of optimization of the costs in deep drawing [6],[7],[8],[9].

Otherwise, we contribute in this domain by the development of a new modeling system that permits design and calculation in deep drawing, and then we show the process and the results of development of knowledge base of know-how first, as well as the development of the system for the automatic choice of processes.

2. Development of Knowledge-base

Manufacturing companies that develop complex products (such as automotive, aviation, electronics, computer,... ) by the sheet metal process, produce and use a great deal of information since the first step of development until the end step of the product life cycle. Generally these informations are gathered to elaborate a knowledge base which is based on the know-how of the specialists of the domain. Does the deep drawing specialist look for Indeed how to produce in a minimum delay with a best quality and a least cost?

Otherwise, specialists in deep drawing process had behind them a long experience acquired in industrial environment and know how to integrate their knowledge intelligently without have resort to computerized means. Today where the change of individual is more and more fast, it is going to be difficult to find preparers in offices of methods experienced and remaining to their station a prolonged period.

In this case we must try therefore to develop new techniques of process planning, essentially based on resources with Artificial Neural Network. However, the correct choice of the processes during a deep drawing operation can be realized from the analysis of information of process planner or equivalent in mechanical manufacture and sheet metal domain.

2.1. Concepts of Investigation support

The first step for the development of an investigation support is to fix people representative group has contact. It is necessary, in this step, to determine who would participate in this investigation, the subject which would deal with and the way that enable us to obtain necessary information, to pass then to the writing of question. The next step is to concept which serves for the collection of Knowledge. Making sure that the same question is proposed in the same way to each answerer.

In this work of appraisal and extraction of information we used a methodology of knowledge description of the designer in deep drawing. This method is called ETED developed by the research and studies center on Qualifications (CEREQ). This center has for mission to conduct studies and investigations on training - job relations for expertise and use. These works aim the acquirement and the certification of expertise, as well as to the professional motilities all along the professional course. In this context we use an approach to lead the investigation bound to technician and engineer specialist in deep drawing. The collection of information with ETED method reposes on:

- The compilation of the documentation concerning the data of justification of the sector of activities and its evolution, it is about collecting a maximum of documentation concerning the organization, their assignments, their context, their primary activities, their secondary activities, the types of referential and used data bases, etc.

- The interviews with the hierarchy concerned jobs, but also with the preparers and the other people resources.

- The interviews with the engineers and the designer in deep drawing that contributed to the development of the global knowledge base.

During the realization of this investigation we contacted several expertises in this domain, engineers and designers in manufacturing companies for oil filter and diesel filter of automotive. We also made some interviews with the expertise of the societies of manufacture of pans and steel cuts. The goal it is to look for the good practices of deep drawing, the shortcomings of shape, the quality of surface finished and essentially the optimization of number of operation process. Ten investigators answered our questions correctly.

Among the open questions proposed on the cylindrical deep drawing:

- What types of shortcoming generated during cylindrical deep drawing?

- What geometrical constraints are used?

- What dimensional constraints are used?

- What kind of CAD software you use?

- How you classify the parameters of manufacture?

- How you determine the number of the intermediate operations (from blank to deep drawing)?

- How to choose the type of punch and die for intermediate operations?

- How does make to determine the elastic return (figure 1-b) of sides that influences on measurements of final part?

- How avoid the rupture of blank in deep drawing operation?

Answers to these questions help us to elaborate a reliable knowledge base to determine the intermediate deep drawing operations.

Also the goal in this part it is to master the failings to avoid them in the development of aided planning system for deep drawing (figure 1).

2.2. Development of deep drawing rules

Main idea in this research is to define automatically deep drawing process planning, based on a strategy of scheduling under constraints [15]. These constraints result from the analysis of the geometrical and technological specifications definite between the different features and capacity of deep drawing machine. Actually each manufacturer or planner asks the question - to avoid manufacturing defects - why we proposed several rules in cylindrical deep drawing part?

During the operation of deep drawing it’s necessary to center the blank between the punch and the die, if this condition not released, several problems will be generated among which: wrinkling on the level of the radial surface of the final parts (figure 1-a). This problem is explained by the bad distribution of metal between die and punch. The choice of the interval between die and punch is very important.

apply suitable lubrication with the materials used to improve the quality of surface finished (to avoid curling)

Ascending approach ensures the good process control. We start with the shape of the final parts to calculate the processes parameters and the intermediate stages.

Select the material of parts according to the customer requirements and the complexity for each step of deep drawing and the final shape.

Number of operations in the forming process depends on the properties of materials, of product geometry and thickness feature.

In the general case of deep drawing, it's necessary to begin the interior form always toward the outside. If a hole exists in drawing feature, it’s necessary to reverse the sense of the drawing and begin from outside toward the interior to keep exacted measurements of the finished shape (figure 2-3).

Take account of the material behaviors’ in each step to choose well the passage between two operations in order to choose a suitable die and punch dimensions.

4. Development of geometrical Base with design by feature.

Features 3D are also the expression of basis to programming manufacturing process.

Deep drawing feature is chosen while respecting two conditions:

1. Topological condition of process deep drawing existence,

2. Condition of independence between operations and processes.

The design by feature is based on an original method able to study all cases presented in this section for obtaining of a composed piece by two, three, four or five simple elements under the condition to be practically feasible in the industrial environment (Table 1).

4.1. Design by feature approach

Several definitions of manufacturing feature have been studied in the literature. We propose for example: feature of deep drawing is a geometric shape and a set of specifications (material, quality of surface) for which a process is defined. This process is almost independent of other features processes. Therefore feature is a developed shape from initial blank toward the finished drawing shape.

In figure 2 we present the decompositions of deep drawing part simple elements. This decomposition facilitates the calculation of blank diameter. We used the analytic calculation to determinate the initial blank. This method is based on a conservation of volume between the final parts and the blank. To simplify calculation, approximate equations that assume constant thickness is implemented our system. Relation between initial blank D0 and deep drawing part is:

(1)

For example to calculate the blank diameter for (F3, F9) in figure 2a:

Do = (2)

Do = (3)

D1, H, D2, R1 and R2 represented geometrical parameters for F3 and F9 (Table 1)

In figure 2 of drawing shape (F3, F9, F1) we add to the formula (2) the section of the blank:

(4)

Equation (3) becomes therefore:

Do = (5)

combinations by this matrix (table...) of intersection that presents two features solely in the first colone. The superior line represents the thirteen bases feature to add.

Inside of this matrix we find three types of possible intersections:

01: First element is added to the part

10: Last element is added to the part

11: Element can added in first or last place to the part

4.2. Codification for geometrical parameters

The measurements of drawing part are coded by several variables Hi (heights), D1, D2 (Diameters)., R1, R2, (Radius) α (Conical angle), E (thickness).

The results of the investigation and the bibliographic research allow us to share all measurements in gap. The transition from a gap to other gap permits to modify the total number of the operations. For example the variation of the height permits to increase the number of the operations of drawing. Indeed to pass the height from 20 mm to 40 mm (Table ) we have to add a supplementary operation of drawing.

5. Development of technological knowledge Base

The families of knowledge base represent the types of deep drawing features in the same way characteristic that don't influence on the choice of process. For example if we change the nuance of sheet metal stainless steel we will have the same number of stages for deep drawing process. The choice of the geometrical deep drawing part will be distributed on beaches of size. The variation of sheet metal thickness is expressed in interval in the following table.

6. Implementation of knowledge base in intelligent deep drawing system

The automatic system of the choice of deep drawing process which we used in this study is based on multi- layer artificial neural networks. They have the advantage to permit with a certain number of tests to select the appropriate process characteristics of proposed deep drawing process. The model of multi-layer neural networks is based on a simple representation of the biological neurons in form of a function of several variables. For this sort of networks, the activity of a neuron is modeled by a real number and the synapses by coefficients. As their name indicates it, the multi-layer neural networks are divided into layers; the first layer is a layer of inputs because it receives the inputs vector, reciprocally the last layer is a layer of outputs, it produces the results. The intermediate layers are called hidden layers, because states of neurons that they contain are not observable. The proposed neural networks are a self-adapting structure, it internally modified until attaining the desired result following the phase of training and generalization. Indeed, the training is a development phase of neural networks during which the behavior of the networks is modified until obtaining the desired behavior. It is done in the context of task or a behavior to be learned.

Figure 5 : Structure for automatic drawing process

Stage 1: Choice of the training rate μ and the moment α.

μ: Training rate, it is a constant ranging between 0 to 1 which fixes the training speed of the network.

α Moment coefficient, generally takes a value ranging between 0.1 to 1; used to accelerate the convergence of algorithm.

Stage2: Randomly initialization of the weights w (i).

Stage3: Choice of the inputs sample and propagation of the calculation through the network.

Stage4: Calculation of the outputs for all the neurons leaving the inputs layer towards the outputs layer using the equations (2) and (3).

Stage5: Measure TMSE by deference between real output and desired output. It is defined by the equation (1).

Stage6: Stop of the algorithm: if calculated TMSE is lower than a threshold value of beforehand definite convergence, or if the iteration time is high.

Stage7: Calculation of the contribution of one neuron to the error starting from the output and determination of the weight modification sign.

Stage8: Correction of the neurons weights in order to decrease the error.

Stage9: Repetition of calculation from Stage 3.

The process starts again from the choice of the example in input, until it that a minimal error rate is reached. The training of the network consists in making thus learn with a pattern who represents the combinations of a random inputs, then we tests the answer of our network, therefore according to error values’ of training and generalization TMSE and GMSE (Generalization Means Square Error), we can evaluate the training of the network.

6.1. Training Base

The training of ANN consists in making thus learn with a pattern who represents the combinations of a random inputs, then we tests the answer of network, therefore according to error values of training and validation RMSE, we can evaluate the global performance of network. In figure 7, we illustrate the training algorithm for ANN structure.

Training base contains several deep drawing features cases for training and other different cases that latter for validation and test. The outputs of ANN are the deep drawing process plans (P1, P2,…,Pn) (Figure 6).

Figure 6 : ANN architecture

6.2. Evaluation of ANN performance

To select the best architecture of the network it is necessary to test many models. Unluckily, this part is usually a trial and error procedure and various situations are necessary to be attempted. There are several algorithms and training suitable function to train an ANN model. In our work, a feed-forward neural network is used with back propagation algorithm which is the most common structure is developed to predict the deep drawing process.

The training of the network consists in making to learn with a pattern who represents the combinations of a random inputs, then we tests the answer of network, therefore according to error values of training and validation RMSE (figure 7), we can evaluate the global performance of network .

Characteristics of back propagation network are:

Training method: supervised training

Training function: log-sigmoid function

Training function: levenberg-Marquardf

Learning function: gradient descent

6.3. Industrial validation using ANN

A complex cylindrical part is considered to test proposed ANN as example show how the neuronal system works (figure 7). The examples are also useful for understanding the feature design method and the identification of deep drawing parameters used in the previous modules. In this case of validation ANN proposes results linked to the possible processes. For example so P1=01 (table 7) that means the choice of process is possible and optimum, otherwise it is impossible or difficult deep drawing process.

In this case study, the basis of learning for each family type stamping parts is divided into seven ANN systems using a binary coding. The following table provides an example of one input / output. The input vector is composed of ten basic parameters (1 material, 8 dimensions, 1 angle). Outputs are the possible processes for the proposed part. For example the first line (Table 8) ANN output = [01 01 00 01 01 00 00 00] shows that processes P1, P2, P4 and P5 are practically possible and can give good quality results to obtain required feature by ANN. Other processes 00 such as output P3, P6, P7 and P8 are not practically possible or very difficult to perform and the results can cause deficiencies on products.

The first dialog box created can select the type of drawing developed based on proposed features. In the present example form No. 5 of total 7 is proposed. Indeed, the neural network specific to this form will be loaded in the same dialog box by programming Visual Basic and Matlab (Figure 8). The user must enter the geometrical and technological parameters. The validation of this section is to display all input parameters and propose possible process which is optimal scheduling deep drawing operations (Figure 9).