Articles & News Stories

ANN modelling to optimize manufacturing processes: the case of laser welding

In this work, an artificial neural network (ANN) was implemented to investigate the main effects of process parameters on the laser welding process quality. A high brightness Yb fiber laser was used to carry out the analysis. Full penetration
of 6
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  IFAC-PapersOnLine 49-12 (2016) 378–383 ScienceDirect   Available online at 2405-8963 ©  2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.Peer review under responsibility of International Federation of Automatic Control.10.1016/j.ifacol.2016.07.634 ©  2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.   * Department of Mechanics, Mathematics and Management (DMMM), Politecnico di Bari, Bari, 70126 ITALY (Tel: 0039-080-5962769; e-mail: Abstract: In this work, an artificial neural network (ANN) was implemented to investigate the main effects of process parameters on the laser welding process quality. A high brightness Yb fiber laser was used to carry out the analysis. Full penetration autogenous welding of 6 mm thick AA5754 aluminum alloy sheets was performed in butt configuration. The welding speed and the shielding gas varied in the experimental plan. The process quality was analyzed by visually inspecting the bead appearance. The ANN modeling code was built by  Neural Tools (Excel add-in) – Palisade Corporation ® . The statistical estimation revealed the relationship of the process parameters with the weld geometry, which provides a deeper understanding of the welding process. Eventually, the usefulness of ANN modeling for optimizing the quality of manufacturing processes was demonstrated. G. Casalino*, F. Facchini*, M. Mortello*, G. Mummolo* ANN modelling to optimize manufacturing processes: the case of laser welding  Keywords:  Modeling, simulation, control and monitoring of manufacturing processes, simulation technologies 1. INTRODUCTION In recent times, the firms operating in the global market are expected to meet increasingly higher quality and productivity standards, which guarantees high level of competitiveness. Innovative technical solutions, novel hybrid technological  processes and customized products have been developed in numerous industrial and civil fields. Moreover, costs reduction and preservation of resources are basic requirements for both environmental and economic purposes.  Nowadays, the optimization of manufacturing processes has  prompted the interest of many researchers, since it represents a critical issue for the production engineering. In particular, a wide range of manufacturing cycles includes joining and welding operations. Laser welding presents several unarguable advantages, since it ensures process efficiency and high product quality. On one hand, laser welded joints dispaly narrow bead, deep penetration, high precision and low distortion, on the other, this process exhibits high  productivity, feasibility, efficiency and versatility for a wide range of working configurations and geometries. The advantages of laser sources were discussed by many authors, like Akman et al. (2009), Suder et al. (2012), katayama et al., (2013), Schneider et al. (2014). Currently, these aspects have encouraged the demand for laser assembled lightweight metal structures. In particular, Al alloys, thanks to their high strength-to-weight ratio, low density, forming properties, reduced costs and recyclability, are widely employed in many automotive and aerospace applications. Assuncao et al. (2010) explained the advantages and the great potential of these alloys. However, Al alloys are difficult to be fusion welded. High light reflectivity and thermal conductivity, loss of alloying elements, pore formation and solidification cracking can compromise the joint quality. Different authors investigated the laser welding of Al alloys. El-Batahgy et al. (2009) studied the effects of process parameters and joint  preparation on the quality of joints. Haboudou et al. (2003) estimated the benefits of surface preparation on porosity rate. Paleocrassas et al. (2010) investigated the causes of instability at low welding speed values. Finally, Casalino et al. (2013) improved the geometry, mechanical and microstructural characteristics of Al assemblies by coupling the laser with an electric arc source. A deeper understanding of the process-related dynamics and the adequate control of all significant variables needs more than experimental trials. So, industries encouraged the development and integration of informative technology for enhancing the quality of manufacturing systems. The implementation of numerical and analytical models can reduce time and cost for experiment and analysis by means of quantitative solutions. Artificial Neural Networks (ANNs) can help to unveil the relation between process parameters and quality of weld. Moreover, they furnish a response to optimisation of process  parameters. This statistical method was inspired by the  behaviour of biological neurons. Weighed interconnections link nodes called “neurons”, which are allocated in layers. A complex relationship between input and output of the network can be estimated. According to Facchini et. al. (2013), an artificial neural network can predict one or more output parameters after learning from a training data set. Such networks connect a number of individual elements, each of which take a set of inputs and produce a single real number. The learning algorithm determines the numeric weights to the link between neurons that produce a robust and correct output. One of the main advantage of this technique is that it can produce good results, even when supplied data are noisy or incomplete.  Neural Networks can be used to develop models that express the interrelationship between the input and the output of very complex systems. The key benefit of ANNs in the domain of engineering design and group technology is in their ability to store a large set of parameter patterns as memories for the system that can later be recalled. Neural network software   G. Casalino et al. / IFAC-PapersOnLine 49-12 (2016) 378–383 379    packages are very common among scientists and manufacturing researchers. In particular, their applications in the field of welding have showed good success. Nagesh and Datta (2002) applied the back-propagation neural network to  predict the bead geometry and penetration in shielded metal-arc welding without considering the structure of the neural network. An ANN was designed to calculate penetration-to-fused zone widths for different laser powers (LPs), welding speeds (W), and focal positions (FPs). This algorithm permits to relate the input of the process parameters to the output of the weld quality features. The error of the ANN output rarely exceeded 20%. Thus, the implemented network resulted highly accurate. Yilmaz and Ertunc (2007) developed a generalized regression neural network model to predict the tensile strength. The predicted values of the tensile strength were found to be in good agreement with the actual experimental values. Ates (2007) introduced a new technique based on ANNs for the prediction of gas metal arc welding parameters. Input  parameters of the model consisted of gas mixtures, whereas, outputs of the ANN model included several mechanical  properties, i.e. tensile strength, elongation and weld metal hardness. ANN controller was trained with the extended delta-bar-delta learning algorithm. The results showed that the calculated results were in good agreement with the measured data. Khorram et al. (2010) built an ANN to demonstrate the significance of several process parameters on bead geometry of laser welded titanium alloy sheets. Casalino et al. (2005)  processed the data coming from the laser welding experimental trials by ANN. The aim was to interpolate the database in order to form a suitable database for the analysis of the variance (ANOVA) and Taguchi analysis. Olabi et al. (2006) applied ANN to find out the optimum levels of the welding speed, the laser power and the focal  position for CO 2  keyhole laser welding of medium carbon steel butt weld. Therefore, it is very clear that a very common problem in industry is the difficulty to measure process quality. There are several approaches to this issue but these techniques are restricted either in applicability (a priori model), accuracy (envelop curves) or scalability (look up tables). They are also limited in their capability to adapt to changes. In particular as far as concern the laser welding process, in scientific literature there are only few papers that discuss the modelling of this welding processes by a neural network. In this study, an ANN model was built to predict the geometric parameters of the weld bead of an aluminium alloy that was fabricated by laser welding. The process consisted of the welding of an aluminium alloy adopting a laser device. 6 mm thick laser welded sheets were fabricated by autogenous welding in butt configuration. Different welding conditions were tested. Process parameters were optimized according to  both experimental trials and computational data. The target of the proposed ANN architecture allowed  predicting, integrating and controlling the laser welding  process. Therefore, this approach promises to address key industry needs relating to both the calibration and optimization of laser welding processes. The rest of this paper was organized as follows: the experimental and model implementation procedures were described in Section 2, while, the empirical results are  presented in Section 3. Eventually, conclusions s were  provided in Section 4. 2. PROCEDURE 2.1 Experimental procedure and set-up In this work, 6 mm thick 5754 Aluminum alloy sheets were autogenously welded in butt configuration. The chemical composition and physical properties of the as-received material are shown in tables 1 and 2, respectively. Table 1. Chemical composition of AA5754 alloy Chemical element Weigth % Silicon (Si) 0.40 Iron (Fe) 0.40 Cooper (Cu) 0.10 Manganese (Mn) 0.50 Magnesium (Mg) 2.6÷3.6 Chromium (Cr) 0.30 Zinc (Zn) 0.20 Titanium (Ti) <0.15 Aluminium (Al) balance Table 2. Physical properties of AA5754 alloy Thermal conductivity K (W/(m*K)) Melting point Tm (k) Density ρ  (g/ cm 3 ) 147 670 2.66 Prior to welding, sheets were prepared by milling at low speed, wire brushing and cleaning with acetone, with the aim to both improve the surface contact and reduce asperities. A high brightness Yb fiber laser was focused in continuous wave regime on the top surface. The active gain medium consisted of a 200 μ m diameter optical fiber. The focusing system was made up of a focal lens and a collimating lens, which focal lens were 250 mm and 120 mm, respectively. The work-piece was fixed to the support by four clamps. Since aluminium alloys are highly reflective, the beam was inclined (40° to the horizontal) to prevent the integrity of the components of the machine. The focus spot was about 400 μ m diameter. The incident light had a near-to-Gaussian distribution. Preliminary test were conducted to find the suitable welding conditions. Then, the laser power was kept constant at the maximum value (4 kW). The welding speed ranged between 2 and 3  380  G. Casalino et al. / IFAC-PapersOnLine 49-12 (2016) 378–383   m/min while the type of shielding gas varied at 3 levels (table 3). Figure 1. Laser welding equipment. Table 3: Shielding gases adopted. Sample Shielding gas Level 1 Ar Level 2 He Level 3 N After welding, the joints were visually inspected, cross sectioned perpendicularly to the bead, polished with standard grinding procedures and chemically etched by Keller’s reagents solution (1% HF, 1.5% HCl, 2.5% HNO3, 95% H20). The bead appearance was observed by capturing the images detected from the cross sections. Figure 2 shows the main welding defects detected to evaluate the soundness of the joints. Figure 2. Main defects detected 2.2 ANN prediction model An ANN was proposed to establish a relationship between  bead geometry and technical parameters of the welding  process, by using  Neural Tools (Excel add-in) – Palisade Corporation ®  software. The input parameters of the ANN were the welding speed (S) and type of shielding gas (G). The value of input parameters are identified according to the experimental plan (as shown in table 3). On the other hand four responses (output nodes) are identified to weld bead geometry as: fusion zone (FZ), bead width (W), depth of penetration (D), and pore area (A  p ). The relationship between the output  y(t)  and the input  x(t)  is given by the following equation:                                                    (1) where w(i,j)  ( i =0,1,2,…,  p ;  j =1,2,…, u ) and w(j)  (  j =0,1,2,… u ) are the synaptic weights, v  is the number of input nodes, u  is the number of hidden nodes, and  f   is a non-linear activation function that enables the system to learn non-linear features. The most widely used activation functions for the output layer are the sigmoid and hyperbolic functions. In this case, the sigmoid transfer function, given by equation (2), it was employed. The adoption of this type of the function depends  by data fitting.             (2) 2.2.1 Data set and ANN architecture In order to design the architecture of ANN, one of the most difficult tasks, consisting of identifying the number of hidden layers and the number of neurons for each layer. Too many neurons can lead to memorization of the training sets with the danger of losing the ANN’s ability to generalise. On the other hand, a lack of neurons can inhibit appropriate pattern classification. In this work, a number of tests are performed varying the number of hidden layers and the number of neurons in the hidden layer. The ‘best’ number of each  parameters, it is identified varying the number of neurons from 5 to 20 and adopting a number of hidden layers included between 1 and 4 layers. For every network trained, a validation set of data is obtained in order to evaluate its  prediction ability. In these cases the best accuracy is achieved adopting an ANN with three hidden layers, characterized by 6, 12 and 11 neurons, respectively. For the other layers of the ANN are adopted two nodes as input layer and four nodes as output layers (fig. 3). A supervised learning mechanism was adapted to training of the ANN, hence the input parameters must be linked to their respective targets. Considering the overfitting phenomenon, the total data set (200 inputs/targets pairs) is randomly split into three subsets: 1.   Training subset  : the group of data constituted by a sample of 150 inputs/targets pairs for training the ANN. In this phase, the synaptic weights (each link  between neurons has a synaptic weight attached to it) are repeatedly updated in order to reduce the error  between experimental outputs and respective targets; 2.   Validation subset  : the data group include a sample of 25 inputs/targets, this phase consisting of   G. Casalino et al. / IFAC-PapersOnLine 49-12 (2016) 378–383 381   identifying the underlying trend of the training data subset, avoiding the overfitting phenomenon. In case of the error measured, compared to validation subset, begins to increase, the training is stopped. This procedure runs together with the training  procedure. In this case, in order to minimise the overfitting problem, the training phase is stopped when the mean square error (MSE) assumes values lower than 0.01. 3.   Testing subset: the data group include a sample of 25 inputs/targets pairs given to the network during the learning phase, in this phase the error is evaluated in order to update the threshold values and the synaptic weights. Figure 3: Back-propagation neural network used to forecast the bead geometry parameters 2.2.2 Training algorithm The network was trained using the back-propagation routine. This method allows to minimize the squares of the differences (  E  ) between the desirable output, identified as  y d  (t),  and the predicted output  y  p (t) . ‘E’   is given by the follow equation:                     (3) Back-propagation routine was trained with the steepest descent algorithm (eq. 4), where  Δ w(k)  is the vector of weights,  g(k)  is the current gradient, α (k)  is the learning rate, and m  is the momentum parameter, which prevents that the algorithm converges to a local minimum or to a saddle point. Moreover, it avoids the risk of minimum overshooting, which can cause instability of the network. The learning rate and the momentum parameter are arbitrarily set to 0.001 and 0.999, respectively.                           (4) The training of the network was stopped when the mean square error (MSE) assumed values lower than 0.01. 3. RESULTS 3.1 Experimental results Figures 4 and 5 show the cross sections of samples 2 and 4, which process parameters are shown in table 4. Although the slight difference in welding speeds, the seam morphology of the two cross sections appeared widely different. Sample 2 exhibits a high quality with not significant defects, while the defectiveness of sample 4 cannot be neglected. Table 4. Process parameters of samples 2 and 4 Sample Shielding gas Welding speed (mm/min) Sample 2 Ar 2300 Sample 4 Ar 2650 Figure 4. Cross section of sample 2 Figure 5. Cross section of sample 4. Discernible zones can be observed in the pictures. The fusion zone (FZ) appeared restricted. On one hand, sample 2  presents a slight inclination of the bead, because of the inclination of the laser head. On the other hand sample 4 excessive penetration, undercut and porosity. Both the S G FZ W D A P  INPUT LAYER HIDDEN LAYERS OUTPUT LAYER  382  G. Casalino et al. / IFAC-PapersOnLine 49-12 (2016) 378–383   samples presented a significant excess weld metal (also called overfill). This is caused by various contributes which influence the weld pool motion, such as excessive gas flow rate and longitudinal metal liquid flows. The analysis of the cross sections of all samples revealed that a slight difference of process parameters can involve very different seam morphology. This derived from some aspects concerning the  process dynamics. The keyhole behaviour, which stability mainly derives from the balance between recoil pressure and surface tension of the surrounding liquid walls, strictly influences the successfully outcome of the joining. In fact, the keyhole collapse involves geometric defectiveness and  porosity formation. This last is due to gas bubbles suppression. Thus, the optimization of technological  parameters is a basic requirement for increasing the seam quality, since it promotes a stable and defects-free process. 3.2 Model validation In order to evaluate the reliability of the geometric  parameters predicted by modelling, the features of the weld  bead in output by ANN, was compared to experimental results. For this scope, the Mean Absolute Percentage Error (MAPE) see equation 5, and the p-value parameter (according student's t-tests) were calculated.                  (5) where  x i  parameters are the results obtained by experimental  plan, while     are the output parameters by ANN. Table 5: Mean absolute percentage error computed for each geometric parameter of the welding bead for AA5754 aluminum alloy sheets Output parameters MAPE (%) p-value Fusion Zone (FZ) 3.64 0.95 Bead Width (W) 3.70 0.93 Depth of penetration (D) 3.65 0.94 Pore Area (A P ) 3.21 0.93 The results obtained (Tab.5) demonstrated that the predicted  parameters, generated by the ANN, ensure a higher level of reliability. Therefore, the neural network was able to predict with significant accuracy the weld bead geometry under given set of welding conditions. It is very interesting noted that the adoption of the ANN could allows to identify the optimal setting process  parameters in order to achieve the desired welding quality. For this scope the ANN could be interrogated according to ‘reverse process’ approach, where the quality parameters of the welding are configured as input of the model, and the  process parameters will be the output of the prediction algorithm. In this case the reverse process model, according analytical approach, might fail because the transformation matrix might not be invertible. This limitation of the problem could be effectively handled by using a heuristic approach. In  particular a software routine can be realized (by means Visual Basic for Application, C++, or other language programs) in order to perform a series of automated iterative sessions focused on interrogation of the ANN according a plan of simulation preventively defined. In this phase a sensitivity analysis will be of paramount importance in order to identify the most significant process parameters for the laser welding  process. 4. CONCLUSIONS The difficulty to predict the geometrical features of the bead in the laser welding process of the Aluminium alloy was satisfactorily solved by adopting an ANN that allowed to establishing the relationship between process parameters and  bead characteristics. The welding process were monitored and the features of the welding bead were forecasted. ANN allows improving the  performance and optimizing the quality of the welding bead through identification and control of the process parameters. The increase of the data set used for the training of the ANN would enhance the reliability of the forecast. The inclusions of new experiments would improve the control of the overfitting phenomenon during training. The adoption of different mathematical routines (such as Bayesan method, Markov chain, Monte Carlo simulation, etc.) has brought to the identification of the number of neurons in the ANN hidden layers. Moreover, this work suggests that the full integration of analysis, prediction, control, and continuous learning into a single framework holds great promise not only in laser welding process but also in the prospect of other manufacturing technologies. Therefore, it is possible to claim that this type of approach can be transfer to a broader range of industrial joining systems. REFERENCES Akman, E., Demir, A., Canel, T., S  nmazcelik ,T. (2009). Laser welding of Ti6Al4V titanium alloys. Journal of materials processing technology, 209, 3705–3713. Assuncao, E., Quintino, L., Miranda, R. (2010). Comparative study of laser welding in tailor blanks for the automotive industry. Int J Adv Manuf Technol, 49, 123-131. Ates H. (2007). Prediction of gas metal arc welding  parameters based on artificial neural networks.  Materials and Design, 28,   2015–2023. Casalino, G., Mortello, M. , Leo, P., Benyounis, K.Y., Olabi, A.G. (2014). Study on arc and laser powers in the hybrid welding of AA5754 Al-alloy. Materials & Design, 61, 191-198. Casalino, G., Curcio, F., Memola Capece Minutolo, F. (2005). Investigation on Ti6Al4V laser welding using
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!