9th International DAAAM Baltic Conference "INDUSTRIAL ENGINEERING" April 2014, Tallinn, Estonia

Size: px

Start display at page:

Download "9th International DAAAM Baltic Conference "INDUSTRIAL ENGINEERING" April 2014, Tallinn, Estonia"

Alexis Wilcox
5 years ago
Views:

1 9th International DAAAM Baltic Conference "INDUSTRIAL ENGINEERING" April 2014, Tallinn, Estonia DEVELOPMENT OF THE INTELLIGENT FORECASTING MODEL FOR MANUFACTURING COST ESTIMATION IN POLYJET PROCESS Rimašauskas, M.; Rimašauskienė, R.; Balevičius, G. Abstract: In global manufacturing environment, high product quality, manufacturing flexibility and low production cost are the main keys to competitiveness. Additive layer manufacturing (ALM) technologies have very high level of flexibility and could be used for production of precise parts. In other hand these technologies are still very expensive. This paper presents the model of manufacturing cost estimation according to the part geometrical parameters. The manufacturing cost estimation method is based on Artificial Neural Networks (ANN). ANN is helpful in the situations when conventional linear regression does not work. The obtained results have been considered, discussed and appropriate conclusions have been drawn. Key words: additive layer manufacturing, cost forecasting 1. INTRODUCTION In the rapidly changing manufacturing environment of today, it is crucial to quickly and accurately react to the ever changing demands of the environment. Such customer demands as low product price, high quality or short delivery time have become an inseparable part of manufacturing. On the other hand challenges such as these force engineers and scientists to create new technologies and seek optimal solutions. Recently a great deal of work has been done in the area of additive layer manufacturing [ 1 ]. It is necessary to admit that earlier ALM technologies were created with rapid prototyping and custom production in mind however challenges from the environment make these technologies change as well. Currently, the factor of new materials and flexibility is gaining importance. Hence with the constant improvement of technology and especially the increasing emergence of new materials, it becomes a perfect tool not only for prototyping but for serial production too. Moreover ALM technologies in specific areas such as biomedicine could be used [ 2, 3 ]. Such trends give rise to expectations that in the future these technologies will become of increasing importance in the field of manufacturing [ 4 ]. A contributing factor to such speculations is the constant expansion of open source machine supply and demand. On the other hand adoption of these technologies gives rise to a great deal of technical problems. The accuracy of equipment and produced parts as well as repeatability, recycling of materials, definition of optimal work modes, determining of manufacturing cost, these are the areas that currently require more clarity [ 5 ]. Hence the customer is interested not only in the properties of the products, but also delivery periods and product price. These qualities are inseparable from production expenditure and production time evaluation. On the other hand it is worth mentioning that ALM technologies are still fairly expensive, therefore cost forecasting in the early stages of product design becomes very important. Manufacturing cost forecasting can be conducted using such widely known methods as: Parametric cost forecasting; Forecasting by applying artificial intelligence; 175

2 Forecasting based on experts experience; Forecasting by means of knowledgebases; Forecasting by means of classifiers [ 6 ]. Parametric models work great, when a linear dependence can be established between manufacturing cost and the factors that define them. In the meantime models based on artificial intelligence are great for processing non linear dependencies and can be used to solve manufacturing cost and quality forecasting problems [ 6-9 ]. On the other hand, forecasting by means of applying expert knowledge or experience is also suitable. However, lack of information remains a major flaw. Even though all ALM technologies share the same principle layer by layer manufacturing, there still exist differences that influence manufacturing cost. Application of the same practical knowledge on different technologies is a very rare occurrence. The aim of this publication is to collect and analyze information about Poly Jet technology, determine the main factors that have influence on manufacturing cost. 2. DEVELOPMENT OF COST FORECASTING MODEL STRUCTURE A crucial step to model creation is creation of appropriate structure. It consists of selection of input variables, data collection and processing and neural network structure selection. The established network needs to be trained, its validation and testing must be performed. In this case it is not only crucial to have good knowledge of operation principles of neural networks, but also on the technologies under investigation. The publication analyzes factors of Poly Jet technology that may have influence on manufacturing cost. The technology under investigation is one of the newer industrial ALM technologies; it distinguishes itself with accuracy and good repeatability. Photopolymers with different properties are used as production materials, which allow production of products with varying mechanical and visual qualities. Other technological parameters are provided in previous publications [ 10, 11 ]. However, it is necessary to pint out main factors that influence manufacturing cost production time, material consumptions. T = Ts + Te + Tm (1) where, T total production time, T s setup time, T e post processing time, T m machining time, in this case printing time. M = M s + M m (2) where, M total material consumption, M s support material consumption, M m model material consumption. Total material consumption, as in most ALM technologies consists of model material and support material. In the meantime, manufacturing time can be divided into machining time and time meant for supplemental operations. Preparation time is intended to prepare 3D CAD model as well as to prepare the machine for operation. Preparation of 3D model varies according to the qualification of the operator, therefore it is difficult to forecast. However post processing time meant for removal of the finished part from the machine and support material from the surface is also hard to forecast as well. It depends on the experience of the operator as well as the complexity of the part. It is known that highly detailed parts, or parts with deep holes or small construction elements requires longer post processing time. In the meantime, machining time depends on model geometry, positioning, size and machine possibilities. In this case Objet 30 machine with two nozzles one for model material, other for support material was used for experiments. Thickness of the layer being printed is constant and equal to 28 µm. Printing resolution in x and y axes is 600x600 dpi. 176

Maximum dimensions of a model are 294x196x150 mm. The printing speed is 8 mm per hour, as declared by vendor. Achievable printed part accuracy is 0.1 mm.

In order to fully print out one layer of the part, the nozzles must perform four passes. In four passes a 65 mm wide layer is printed.

3 Maximum dimensions of a model are 294x196x150 mm. The printing speed is 8 mm per hour, as declared by vendor. Achievable printed part accuracy is 0.1 mm. However, real printing speed is highly dependent on model, therefore such forecasting method becomes very inaccurate. In order to fully print out one layer of the part, the nozzles must perform four passes. In four passes a 65 mm wide layer is printed. Therefore the positioning of the part in the work plane is of crucial importance to the production time. A few positioning rules are known that help in lowering the production time: First and most important criterion is the height or the part. If possible, the part should positioned in such a way that it s size on the z axis is smallest in comparison to x and y axes. The longest dimension of the part should be parallel to x axis. As mentioned before, printing width is 65 mm therefore; when printing the parts with dimension with respect to y axis is larger it is necessary to perform nozzle displacement with respect to y axis. It is necessary to keep in mind that displacement in y axis direction is relatively slow. cannot be relied on constantly, since when aiming for the lowest production time, material consumptions are sometimes left unevaluated as well as accuracy parameters. Therefore in the first case, the manufacturing time will be 8 hours and 3 min, consumption of model material 178 g, of support material 88 g. After performing manual placement, according to figure 2, material consumptions doesn t change significantly. However production time turns out to be 10 hours 57 min. Fig. 2 Manual placement 1 Hence we see that when the dimension of part or parts in relation to y axis increases, the production time becomes longer. In the third case parts are positioned parallel to the z axis (Fig. 3). Fig. 1 Automatic placement In the figure 1 automatic placement of three parts in in work area is displayed. Automatic placement seeks lowest manufacturing cost, i.e. production time must be smallest. However it is worth mentioning that automatic placement Fig. 3 Manual placement 2 Even though parts used are identical, the production time rises to 25 hours 35 min. However, the most interesting part is that the material consumption increases greatly, consumption of model material 316 g, support material 467 g. Increase in support material consumption can be 177

4 explained, the material is used to fill vertical cavities. However the consumption of model material can only be explained if specifics of the technology are understood. One way in which the technology differs from other ALM technologies is the mixing of support and model materials. On the other hand, model material is mixed in with intent to improve support material and its mechanical characteristics. Also the amount of material mixed in can be changed by altering parameters of the machine. Hence we see that placement of the part not only influences the duration of the production, but material consumptions as well. On the other hand part placement is very important due to before mentioned qualities in other ALM technologies as well. Separate technologies have intellectual models or decision support systems created specifically for them in order to optimize this process [ 12, 13 ]. It is worth mentioning that the shortest production time was 26 minutes, while the longest was 35 hours and 45 minutes. Results on how manufacturing time changes according to part height with respect to z axis are provided in figure 4. This however completely disregards the width of the part with respect to y axis, as mentioned before that is a crucial factor. A clear trend is visible in the provided figure, as the height of the part increases, so does the production time. However overall dependence cannot be considered in this case. Parts with width greater than 63 mm were excluded in the next step. This was meant to eliminate the influence the part width to the production time. As mentioned before, during one pass a 65 mm wide layer is printed. 3. EXPERIMENTAL TESTING OF MODEL STRUCTURE Proper selection of input data is crucial for creation of a well-functioning model. It was determined in the previous chapter, that dimensions and placement of the parts are some of most crucial factors in forecasting of manufacturing cost. Hence, 100 parts of different geometries and sizes were chosen for experimental tests. All parts were produced and the dependence of production time on their dimensions was analyzed. Fig. 4 Dependence between part height in z axis and manufacturing time Fig. 5 Dependence between part height in z axis and manufacturing time (parts width less than 63 mm) In figure 5, results for the performed analysis can be seen. Here it is also worth mentioning, that after discarding parts whose width is over 63 mm the analysis is performed on data from 55 parts. As it can be seen from the graph, the dependence between part height and production time is practically linear when parts with width exceeding 63 mm are excluded. On the other hand it can be seen that there are 5 parts whose manufacturing time greatly exceeds the allowed limits. However such results may be due to length of the part with respect to x axis or other geometrical parameters. Considering previous experiences it can be stated that in this case length in x axis direction has negligible or no influence on the production time. This 178

5 is perfectly illustrated by the further test. In figure 6 are two identical parts placed along x axis. here x 1 x n neuron input values, w 1 w n weight values, b displacement. Mathematical model of a two layer neural network may be expressed as follows: [ W f ( W X )] Y = f , or (5) Y n g f 2 wi f1 w j x j (6) i= 1 j = 1 = Fig. 6 Positioning of the parts along x axis Manufacturing time for both parts is 7 hours 28 minutes, however with one part removed, production time decreases to 6 hours 50 minutes. Such difference shows that part length along x axis has negligible influence to production efficiency. On the other hand this also stands to show that material consumption is not a proper factor to production time. In this case material consumption decreases by exactly two times, however the production time decreases only by 30 minutes. Theoretically, a neural network with one hidden layer containing sufficient neurons of that layer can approximate any continuous function. In practice, neural networks with one or two hidden layers are most frequently used [ 14 ]. This information is used when making the structure of a neural network. A network input layer is created of the following part parameters: height part dimension in direction of z axis, width part dimension in direction of y axis, length part dimension in direction of x axis, material consumptions. Thus, during testing neural network with one hidden layer will be used. Transfer function will be determined experimentally. The one neuron transfer function may be expressed as follows: n y = f wi x i= 1 i + b (3) here X matrix of input values, W 1 matrix of the first layer weight values, W 2 matrix of the second layer weight values, Y matrix of output values, w i weights of the second layer neurons, w j weights of the first layer neurons, x j input values, f 1 hyperbolic tangent function, f 2 linear transfer function. 4. FURTHER RESEARCH Verification of the established neural network structure will be conducted in further research. Currently it is being attempted to collect a sufficient amount of verified data, to be analyzed and used to establish the intelligent model. In further steps of the investigation, application of model to forecast manufacturing cost for other machines is planned. 5. CONCLUSIONS Performed research enables to make following conclusions: 1. During the investigation, most important factors influencing manufacturing cost were determined dimensions of the part and placement in the work area. 2. It has been determined that height of the part in z axis direction is a critical factor to production time. While influence of part width in y axis direction, increases dramatically when part width exceeds 63 mm. 3. A neural network structure that has to be experimentally verified and applied 179

6 for practical use was proposed. 6. REFERENCES 1. Wesley, M., Cunico, M. Development of new rapid prototyping process. Rapid Prototyping J., 2011, 17, Kouhi, E., Masood, S., Morsi Y. Design and fabrication of reconstructive mandibular models using fused deposition modeling. Assembly Automation, 2008, 28, Tek, P., et al. Rapid prototyping for neuroscience and neural engineering. J. of Neuroscience Methods., 2008, 172, Chryssolouris, G., et al. Digital manufacturing: history, perspectives, and outlook. Proc. of the Institution of Mech. E. Part B, J. of Engineering Manufacture, 2009, 223, Brajlih, T., Valentan, B., Balic, J., Drstvensek, I. Speed and accuracy evaluation of additive manufacturing machines. Rapid Prototyping J., 2011, 17, Shehab, E. M., Abdalla, H. S. Manufacturing cost modelling for concurrent product development. Robotics and Computer Integrated Manuf, 2002, 17, Wang, Q., Stockton, D. Cost model development using artificial neural networks. Aircraft engineering and aerospace technology, 2001, 73, Chung, W. W. C., Wong, K.C.M., Soon, P.T.K. An Ann based DSS system for quality assurance in production network. Journal of Manufacturing Technology Management, 2007, 18, Meziane, F., Vadera, S., Kobbacy, K., Proudlove, N. Intelligent systems in manufacturing: current developments and future prospects. Integrated Manufacturing Systems, 2000, 11, Bargelis A., et al. Virtual and distance labs for vocational education training of industrial employees. Mechanika, Proc. of the 18th Int. Conf., 2013, Rimašauskas, M., Balevičius, G. The influence of water absorption to the properties of the photopolymer used in PolyJet process. Mechanika, Proc. of the 18th Int. Conf., 2013, Phatak, A. M., Pande, S. S. Optimum part orientation in rapid prototyping using genetic algorithm. J. of Manufacturing Systems, 2012, 31, Rimašauskas, M., Rimašauskienė, R. Development of decision support system for fused deposition modelling manufacturing cost estimation. Mechanika, 2012, 5, Rimašauskas, M., Bargelis, A. The development of the intelligent forecasting model for productivity index in manufacturing. Mechanika, 2012, 3, ADDITIONAL DATA ABOUT AUTHORS Rimašauskas Marius, Dr. Assoc. Prof. Kaunas University of Technology Kęstučio 27, LT Kaunas, Lithuania Phone: Rūta Rimašauskienė, Dr. Lecturer. Kaunas University of Technology Kęstučio 27, LT Kaunas, Lithuania Gytautas Balevičius Kaunas University of Technology Kęstučio 27, LT Kaunas, Lithuania 180

Cutting tools in finishing operations for CNC rapid manufacturing processes: simulation studies

Loughborough University Institutional Repository Cutting tools in finishing operations for CNC rapid manufacturing processes: simulation studies This item was submitted to Loughborough University's Institutional