Nonconventional Technologies Review Romania, June, 2013 2013 Romanian Association of Nonconventional Technologies NEURAL MODEL FOR ABRASIVE WATER JET CUTTING MACHINE Ciupan Cornel 1, Ciupan Emilia 2 and Ferenţ-Pipaş Silvia 3 1 Technical University of Cluj-Napoca 1, cornel.ciupan@muri.utcluj.ro 2 Technical University of Cluj-Napoca 1, emilia.ciupan@mis.utcluj.ro 3 Technical University of Cluj-Napoca 1, silvia.ferent@muri.utcluj.ro ABSTRACT: The paper presents a neural model used to control an abrasive water jet process. The material features, the orifice diameter and the abrasive consumption are considered to be the input parameters. The output parameters are the feed rate, the energy consumption and the water consumption. A neural model with back propagation algorithm was used.a set of data obtained by Waterjet Web Reference Calculator was used to model the process. The training and the validation data were calculated based on the values presented by the water jet cutting machines manufacturers. In another paper [3] the authors have presented a neural model for control the speed of cutting and abrasive consumption. KEY WORDS: waterjet, processing, neural, network, model 1. INTRODUCTION Abrasive waterjet machining is one of the nonconventional technologies used in the industry for processing of different types of materials. The main advantages of the waterjet technology are: high machining versatility, high flexibility, small cutting forces and absence of thermal distortions [5]. Water jet machines can manufacture parts to very good tolerances. Today the modern water jet cutting machines can create parts with a tolerance as small as 0.05 mm, although it is usually easier to obtain tolerances under 0.1 mm. To achieve these tolerances it is necessary to understand the factors that affect the precision of water jet machining. The productivity and costs of processing are other important factors which determine the parameters of the abrasive jet processing and sometimes they may be even critical in choosing this process. Abrasive jet machining is complex and many difficulties arise in fully understanding the physical processes taking place. This in turn leads to difficulties in mathematical modeling. The analytical models used to study the process of waterjet cutting [1,2] do not correspond to the requirements imposed by process control due to their high degree of particularity depending on input material and technological parameters. Experimental research [1, 5] shows that the mass of eroded material varies linearly with transport speed of the abrasive particles in the point of impact with the material to be cut. Numerous models and approaches are known in this direction, including design of experiments, regression modeling, ANOVA analysis, fuzzy logicsand artificial neural networks. Some of these studies are based on mathematical equations developed for predicting the process parameters [4]. The purpose of this paper is to develop a neural model for determining the process parameters for the waterjet cutting machines. Besides the feed rate, other parameters that can be used in cost determination were sought. For this reason, the neural model was built to also provide the energy and water consumption, components that are necessary to calculate the processing cost. The energy consumption used in the training data was calculated based on the pump energy and the time required to cut a length of one meter of varying thickness. Water consumption was calculated the same way taking into account the link between power, pressure and flow and the cutting time. 2. STRUCTURE OF THE NEURAL NETWORK A neural network with back propagation learning algorithmis considered. Figure 1 shows the network topology with three input nodes which represent the material thickness, the orifice diameterand the abrasive consumption and three output nodes that represent the feed rate, the energy consumption and the water consumption. Figure 1. Neural network structure 25
25 neurons were chosen for the hidden layer. The activation function for the hidden layer is the sigmoid function and the output one uses the pure line function. Neural network modelling involves two phases: the training and the validation phase. The network training was accomplished using data from field literature [5].Table 1 presents the training data set. In table 1, the first three columns show the input parameters (material thickness, orifice diameter and abrasive flow rate) and the next three columns show the outputs parameters (feed rate, energy consumption and water consumption). The output parameters were calculated by a waterjet web reference calculator [www.waterjets.org]. The network training was done using a set of 106 inputoutput pairs of values. Table 1. Training data set HP pump (pressure 4000 bar), one cutting head, 0.25 mm orifice diameter, 0.508 mm abrasive nozzle diameter. A part of the training data from table 1 were normalized by scaling, as follows: thickness, divided by 10 feed rate, divided by 100. Both, the training and validation of the neural network was done with normalized data. The simulated results obtained in the operational phase were scaled using the normalization factors. It was found that through data normalization we obtained an improvement of the results for the same training data set. The evolution of the mean square error obtained during the training of the network is shown in figure 2. The process parameters for cutting a mild steel and an average quality processing on a machine which has the following features were chosen as input: 50 Figure 2. Error evolution during the training process The validation of the neural model was realised in the next two stages: - Stage 1: input data belonging to the training data set in which only one parameter (orifice diameter) was changed. This situation is met in the real process due to the nozzle wear. - Stage 2: data in which the value of the abrasive feed rate belongs to the training data and the values of other two parameters was modified: the material thickness and the orifice diameter. The validation results in stage 1 are presented in table 2. Errors in the range of ±10%for the feed rate and ±40% for the energy and water consumption were obtained when the outputs were calculated. In this table, the first three columns show the input parameters and the next three columns show the 26
outputs parameters, like desired outputs. These desired outputs were calculated in the same way as the outputs parameters from table 1. Three columns with simulated outputs are provided of neural model. The last three columns show the relative errors between simulated and desired responses. Table 2. First validation results The errors between simulated and desired responses are shown in figure 3. Analyzing figure 3 shows that the neural model gives better results for feed rates compared to energy and abrasive consumption. We 27
also concluded that the error of the feed rate increases with the material thickness. When the feed rate error is low, the other two parameters have large errors. One explanation for the good results for feed rate could be related to the higher ratio between maximum and minimum values of this parameter in relation to the other two parameters (energy and abrasive consumption). Table 3. Second stage data validation set Figure 3. The first stage outputs parameters errors The second stage validation of the neural model, based on the data set from table 3, is shown in figures 4-7. Figure 4 shows simulated feed rate (blue line) and desired feed rate. The desired feed rate values (when known) are represented by a circle. The results provided by the neural model are situated between known values. Figure 4. Feed rate evolution Figure 5 shows simulated and desired output for the energy consumption. The known energy consumption is represented by a circle. It was found that the results provided by the neural model are well placed among the desired values for this case. Figure 5. Energy consumption 28
Similar to previous figures, figure 6 shows simulated and desired water consumption. The known water consumption is represented by a circle. As in the previous examples the results provided by the neural model are well placed among the desired values. Figure 6. Water consumption The errors of the output of the neural model in relation to the desired values are shown in figure 4. It is found that the maximum error is less than 15%. Figure 7. The second stage outputs parameters errors 3. CONCLUSIONS Of particular importance is the development of experimental research to determine a large number of training data. This research shows that a neural computing model can be used to control AWJ process parameters. It is important to determine a large number of training data for the development of an experimental program in order to obtain a model that provides good results. The experimental program should take into account the material properties, the cutting machine features and must contain many experiments, with a smooth variation of the input parameters. The following steps need to be followed in order to develop a neural model that accounts for all the technological parameters of a waterjet cutting machine: selecting the input parameters; sorting the parameters into categories; determining the number of input variables; determining the output parameters; creating separate models for different types of machined materials; preparing the training data and the validation data; choosing the network structure and begin training; validating the model and using it commercially. It is necessary to train the network with a set of data as large as possible in order to obtain an acceptable error. It is also recommended that the domain covered by training data is greater than the domain covered by normal work data. Another way to improve the results is to normalize the training data using different scaling factors. 4. REFERENCES 1. Andreas Momber. Principles of Abrasive Water Jet Machining. Springer, (1998). 2. Cornel Ciupan, Liviu Morar., Aurel Pop. Innovative system with abrasive water jet. Annals of DAAAM for 2008 & Proceedings. 19th International DAAAM Symposium Inteligent manufacturing and automation, pp. 273-274, Tranava, Slovakia (2008) 3. Emilia Ciupan, Cornel Ciupan., Rareş Petruş. The control of the abrasive water jet processing using a neuronal network model.proceedings of 2012 International Conference of Hydraulics and Pneumatics HERVEX, pp. 315-319, Calimanesti-Caciulata, Romania, (2012). 4. Farhad Kolahan, A. HamidKhajavi. Modeling and Optimization of Abrasive Waterjet Parameters using Regression Analysis. International Journal of Aerospace and Mechanical Engineering, pp. 248-253 (2011). 5. Aurel Pop. Study of Computer Control Strategy for Jet Cutting Integrated Systems, Ph.D. Thesis, Technical University of Cluj-Napoca (2004). 6. Radovan Kovacevic. Modeling of the influence of the abrasive waterjet cutting parameters on the depth of cut based on fuzzy rules. International Journal of Machine Tools and Manufacture,vvol. 34,no.1, pp.55 72, (1994). 29