Journal of Loss Prevention in the Process Industries 24 (2011) 563e567
Contents lists available at ScienceDirect
Journal of Loss Prevention in the Process Industries journal homepage: www.elsevier.com/locate/jlp
An alternative work measurement method and its application to a manufacturing industry deviren a, Ergün Eraslan b, *, Fatih V. Çelebi c Metin Dag a
Department of Industrial Engineering, Gazi University, Maltepe 06570, Ankara, Turkey Department of Industrial Engineering, Bas¸kent University, Baglica Campus, Eskisehir Yolu 20.km, Etimesgut 06810, Ankara, Turkey c Department of Computer Engineering, Ankara University, Bes¸evler 06500, Ankara, Turkey b
a r t i c l e i n f o
a b s t r a c t
Article history: Received 24 December 2009 Received in revised form 8 June 2010 Accepted 21 June 2010
Difficulties in determining the standard time justify the need to develop alternative methods to direct measurement procedures. The indirect methods which are comparison and prediction, standard data and formulation, predefined movement-time systems have several deficiencies in time measurement procedures. In this study, an alternative indirect work measurement method based on artificial neural networks (ANNs) is presented which is simple and inexpensive. For the application of the proposed method, the products that have similar production processes are selected among the whole product family produced in a manufacturing company. The standard times of the sampled products that are previously measured are used and the standard times of the remaining several products and semiproducts are predicted by the proposed method. The model results show that the proposed method can be applied accurately in companies which produce similar products. Ó 2010 Elsevier Ltd. All rights reserved.
Keywords: Artificial neural network Work measurement Time study
1. Introduction Nowadays, companies have great demand to measure the standard time of the products that they produce to compete effectively. It is not possible to be consistent and efficient primarily in preparing manufacturing plans and programs, short and long term forecasts, cost control, pricing and the other technical and managerial activities in a company with a time estimation that is not based on the standard time. On the other hand, despite many condensation studies, it is clear that there is not a cost effective method or tool that can measure the standard time. It is possible to say that the above mentioned reality is the core difficulty in the application of time studies. Time study (TS) records the process time and levels of a predetermined work using specified conditions. The collected data is analyzed and used to identify the time required to finish the work with a defined process speed. Unfortunately, TS is cost ineffective and can be applied only under some specific conditions. Moreover, it depends on the experience of the person performing the TS. The problems encountered in determining the standard time brings the need for alternative work measurement methods in addition to the direct measurement techniques such as TS. These alternative methods are classified to be indirect work measurement * Corresponding author. Tel.: þ90 312 234 10 10/1311; fax: þ90 312 234 10 51. deviren),
[email protected] E-mail addresses:
[email protected] (M. Dag (E. Eraslan),
[email protected] (F.V. Çelebi). 0950-4230/$ e see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.jlp.2010.06.017
methods and can be named as methods of comparison and prediction method (CPM), standard data and formulation (SDF) and pre-determined motion-time systems (MTM) (Niebel & Freivalds, 2003). It is obvious that indirect work measurement methods cannot be used in all companies to measure the standard time of every product or semi-product. The work measurement methods to calculate standard time are used in some recent studies. Koelling and Ramsey (1996) studied the effects of multimedia in developing and applying a work measurement method. Cohen, Bidanda, and Billo (1998) examined successful integration of automatic speech recognition (ASR) into industrial systems and by using the ASR, work analyst can develop more reliable time estimates and supply 70% time reduction for manual tasks. Freivalds, Konz, Yurgec, and Goldberg (2000) tested the effect of work measurement and design systems in customer satisfaction to future engineers using 100 industrial engineering programs in US. In industry applications, Priore, Fuente, Puente, and Parreño (2006) used these algorithms for dynamic scheduling in FMS and Akyol (2000) used it for heuristic scheduling. Artificial neural networks (ANNs) have been recognized as a fast and flexible tool for modeling, analysis and design purposes. These are parallel computational models comprised of densely interconnected adaptive processing units (Hassoun, 1995). Ability to learn, processing in parallel, self organization, adaptability, real time operation, fault tolerance properties make ANNs well suited in many applications such as function approximation, optimization,
M. Dagdeviren et al. / Journal of Loss Prevention in the Process Industries 24 (2011) 563e567
564
control, clustering, prediction, nonlinear system modeling. In this study, an alternative indirect work measurement methodology is proposed which is based on ANNs. The method is specifically developed for a company and multilayer perceptrons (MLPs) network architecture is used. Despite its limited complexity, it is one of the most extensively used ANN architectures because of its wellknown general approximation capabilities. The factors that affect the time of different processes are determined and the standard time is estimated by using MLP ANNs based on these factors. The research studies of MLP ANNs in the last decade are especially stated in chemistry (Baawain, El-Din, & Smith, 2007; Bhattacharya & Solomatine, 2006; Chetouani, 2008), energy and power systems (Mori & Yuihara, 2001), medicine (Zhang, 2007), agricultural industry (Mandavgane, Pandharipande, & Subramanian, 2007), hydrology (Ochoa-Rivera, Andreu, & Garcia-Bartual, 2007; Wang, Van Gelder, & Vrijling, 2006), imaging (Lasch, Diem, & Hansch, 2006), measurement (Danisman, Dalkiran, & Celebi, 2006) and semiconductor laser areas (Celebi, 2005a, 2005b, 2006). The proposed MLP ANN model is reviewed in Section 2. The accuracy of the model is tested by comparing the results that is described in Section 3. Finally, the research results and conclusion remarks are summarized in Section 4.
2. ANNs and developed ANN model ANNs are biologically inspired computer programs designed to simulate the way in which the human brain processes information. ANNs gather their knowledge by detecting the patterns and relationships in data and learn (or are trained) through experience, not from programming. The power of neural computations comes from weight connection in a network. Each neuron has weighted inputs, transfer function and one output. The behavior of a neural network is determined by the transfer functions of its neurons, by the learning rule, and by the architecture itself. Applications of ANNs can be categorized into classification or pattern recognition, prediction and modeling (Charniak & Mcdermott, 1985; Ham & Kostanic, 2002; Schalko, 1997). There are many types of neural networks for various applications available in the literature. Radial Basis Function Networks and MLPs are examples of feed-forward networks and both of them are universal approximators. In spite of being different networks in several important respects, these two neural networks are capable of mimicking each other accurately. MLPs (Lippmann, 1989; Rumelhart, Hinton, & Williams, 1986), which is used in this study, are the simplest and most commonly used ANN architectures which are shown in Fig. 1. As shown in the figure, an MLP consists of three layers: an input layer, an output layer and one or more hidden layers with previously defined number of neurons. The neurons in the input layer only act as buffers for distributing the input signals xi to neurons in the hidden layer. Each neuron j in the hidden layer sums up its input signals xi, after weighing them with the strengths of the respective connections wji from the input layer and computes its output yj as a function f of the sum, namely:
yj ¼ f
X
wji xi
(1)
where f is one of the activation functions used in ANN architectures. The weights are optimized by attempting to minimize the sum of squared differences between the desired and actual outputs namely:
E ¼ 1=2
X
Ydj Yj
2
(2)
where Ydj is the desired output of the system and Yj is the actual output of that neuron.
Calculation Direction (backward)
y1
y2
Outputs
yn
Output Layer
Hidden Layer (one or more)
w1m w11
w12 Input Layer
x1
x2
Inputs
xm
Calculation Direction (forward)
Fig. 1. General form of multilayer perceptrons.
There are many learning algorithms available in literature (Ham & Kostanic, 2002; Lippmann, 1989). The learning algorithms mentioned below give the best results among other learning algorithms used in the analysis. FletchereReeves (CGF) algorithm is a second-order method, which restricts each step direction to be conjugated to all previous step directions. This restriction simplifies the computation because it is no longer necessary to store or calculate the Hessian or its inverse (Hagan, Demuth, & Beale, 1996). BroydoneFletchereGoldfarbeShanno (BFG) algorithm is considered to be the best form of quasi-Newton methods. It uses an update formula derived from the quasi-Newton update of the Hessian (Dennis & Schnabel, 1983). LevenbergeMarquardt (LM) represents the simplified versions of Newton’s method applied to the problem of training MLP ANNs. Also it is a well-established numerical optimization technique with a quadratic speed of convergence (Ham & Kostanic, 2002). In this study, an alternative method that can estimate the standard time by MLP ANNs is developed (Eraslan, 2009). The basics of ANN method are the machine estimation under the circumstances that are studied after learning the past data. Machine does not use a determined statistical method during this estimation. It solves the unknown problems of the related events after learning the current data and documents. It acts as if the human learning method and completes the learning systematic. It estimates the time while learning the real relationships between the qualitative and quantitative factors that affect the production and time which is the result of these factors. Qualitative factors have already satisfied this term. Conversely, quantitative factors are expressed based on the predefined numerical scales. The proposed method can be used in companies for deducing standard time easily. Many companies can estimate the standard time by MLP ANN method with the benefits of lower cost, shorter time, and higher accuracy. Consequently, there is no need to use complex methods because the results are satisfactory for a planning production (Eraslan, 2009). The steps of the proposed method can be expressed as: (i) Firstly, a production system which manufactures numerous finished products should be selected with similar manufacturing processes. (Or, this study can be applied to a semi-product that is being produced in a production system.) (ii) The factors that affect the production time have to be determined. These factors are the ones that are possibly affecting
M. Dagdeviren et al. / Journal of Loss Prevention in the Process Industries 24 (2011) 563e567
the time and production system. The expected results could be more accurate for the estimation process when the factors are determined from actual production system of the companies. In addition to that, it is necessary to express these factors numerically. Besides quantitative factors such as length, weight or area etc, there are also qualitative factors that cannot be measured. These factors can be numerically expressed using the Likert scale, e.g., if a painted material that is used in production, “2” should be assigned due to the time effect. Otherwise “1” is assigned to the unpainted material, i.e., production time is not affected due to the unpainted material. All these factors can be digitized as stated above. (iii) Next, the production time of some of these products should be measured with TS. The factors that affect the time need to be clearly defined and numerically expressed. If necessary, an interview can be made with the worker in order to introduce different factors. By the end of the process, the necessary data for ANN process will have been obtained and the determined factors are the inputs and the production time will be the output of the system. (iv) In the last step the proposed network is trained and tested by changing the network variables. If the final network is capable of estimating the time in desired error band, the same network can be used to estimate the time of finished products, semiproducts or new products which have not been measured previously with TS.
3. An application in a manufacturing company The managers of the company demand that they are in difficulty in determining the standard time of the products or semi-products in manufacturing processes. The TS application takes a long time and needed to endure high cost. On the other hand, the environmental conditions are affecting the team of TS poorly in measurement procedures. The proposed ANN model is advised to predict the standard time in similar processes. The most important step to activate the above mentioned ANN is the determination of the factors. While determining the factors, some facts need to be taken into consideration. These are the ones that affect the production time, have the ease of measurement and are not affecting the time linearly, i.e., it is not known how much it affects the time or if it is the only one that affects the time. The factors can be divided into two as firstly, the ones that affect the product specifications and secondly, others that are based on production method (work flow). This means, some of the factors are based on product or semi-product specifications and some other factors are based on the method of production. The proposed alternative work measurement method is applied in Turkey’s biggest heavy-duty truck and bus manufacturing company. Many steel components are welded in the company. The manufacturing processes of many products of different dimensions and materials are similar to each other. Following this study, there exist five possible factors that affect the production time by the group of managers and experts of the area. These factors are listed below: The number of pieces: The products are composed of different number of steel component combinations. Preparation time can be visualized as taking the components from the related shelves and ranging them according to their patterns. The most important factor that affects the preparation time is the number of components to be welded. The number of welding operations: Welding is the fundamental process performed on the products. The time required to accomplish these welding tasks is determined according to the
565
number of welding operations on that product. At first, it was thought that, there was a relationship between the number of components and the number of welding operation. Unfortunately, studies have shown that there is not a full correlation between these factors. Therefore, it was decided to investigate this important factor separately. Product’s surface area factor (composed of width and length): The differing dimensions of the products from each other demonstrate that this factor can have an effect on time. In this aspect, the volume of the product is not taken into consideration since the depth factor directly affects the ergonomics of the study. It is better to investigate the volume factor as an individual one that affects the time. It is seen that the surface area factor affects the number of workers that can work on the product and reachable dimension limits. In the first two criteria, the coefficients are numerically determined using the ones in direct design level. Number 4 is used for the products that have a surface area more than 15 m2, 3 for 10e15 m2, 2 for 5e10 m2 and finally 1 for 5 m2. Difficulty/working environment factor: This factor, including the depth factor, is concerned with the ergonomics of the pattern location, the complexity of the product and the effects of welding problems caused by the product design. This factor is divided into 3 classes and values of 1, 2 and 3 are entered for each product. The difficulty class of products can be easily decided according to product types. This numerical distinctions come out as; because of much reaching and bending, 3 for ceiling and side wall stations’ ventilating problems and large patterns, 2 for combined (chassis, front wall, back wall) sections, narrow places, compact patterns, unsuitability of the product with respect to ergonomics and 1 for spants which are lower chassis parts that are easier to produce. Number of metal forming process: Welding operation can generate stress/strain on the product due to hot metal forming. This is valid for all products. Relative to the structure form, the limits of allowable deformation can be exceeded in some products. If this occurs, metal forming is applied to that product to recover the physical dimensions. This operation is called rectification and also takes time. In this aspect, the products are classified according to the existence of sheet metal forming and 1 and 0 values are assigned to the pieces. The standard time of the related product will be determined with these 5 factors. In the study, TS is applied to 71 products. 55 of them are used to train the MLP ANNs and the remaining 16 of them (randomly selected) are used to test if the proposed model is generating correct results. After several trials with different learning algorithms and with different network configurations in order to obtain a better performance with simpler structure, it is observed that the most suitable network configuration is 5 5 2 1 using the CGF algorithm. This means that the number of inputs is 5 and the number of outputs is 1. The number of neurons is 5 for the first hidden layer and 2 for the second hidden layer, respectively. The input and output layers have the linear activation functions and the hidden layers have the hyperbolic tangent sigmoid activation function. The number of epochs is 300 for training. Before training, the input and the output data tuples are normalized between 0.0 and 1.0 in order to ensure the learning performance since the normalization is an essential step to improve the training process of ANNs. After completing the training process successfully for 5 input parameters, test process is carried out for one output parameter which is the standard time of the work. The CAD model based on MLP ANN in order to find the standard time is shown in Fig. 2. The determined input factors are applied to estimate the output which is the standard time. Table 1 summarizes the comparison between results of test inputs and real values. The required level of error corresponds to
M. Dagdeviren et al. / Journal of Loss Prevention in the Process Industries 24 (2011) 563e567
566
Nr of Pieces
Nr of Weld
Surface Area
Standard Time
Difficulty/ working environment
Rectification
Fig. 2. The developed CAD model.
Table 1 Comparison of the outputs of network and real values. No.
Measured values
1 32.86 2 70.33 3 19.40 4 92.74 5 37.69 6 145.96 7 36.23 8 36.23 9 70.33 10 16.00 11 17.96 12 24.79 13 32.67 14 36.23 15 18.17 16 18.17 Train error Test error
Generated values by CGF
Absolute difference
Generated values by BFG
Absolute difference
Generated values by LM
Absolute difference
Generated values by regression
Absolute difference
34.38 69.18 17.32 89.46 35.06 146.68 40.96 40.96 69.18 16.70 16.70 22.06 29.02 40.96 17.62 17.62 6.65E04
1.52 1.15 2.08 3.29 2.63 0.72 4.73 4.73 1.15 0.70 1.26 2.73 3.65 4.73 0.55 0.55
32.69 73.85 18.46 99.55 35.68 143.47 41.34 41.34 73.85 18.69 18.69 24.43 30.47 41.34 18.31 18.31 1.72E04
0.17 3.52 0.94 6.81 2.01 2.50 5.11 5.11 3.52 2.69 0.73 0.36 2.20 5.11 0.14 0.14
34.30 70.33 21.22 94.81 31.62 143.62 40.96 40.96 70.33 21.37 21.37 27.03 32.39 40.96 21.14 21.14 2.30E005
1.44 0.00 1.82 2.06 6.07 2.34 4.73 4.73 0.00 5.37 3.41 2.24 0.28 4.73 2.97 2.97
45.15 80.01 22.54 126.77 29.70 134.63 51.22 51.22 80.01 22.14 22.14 30.56 39.37 51.22 22.73 22.73 e
12.29 9.68 3.14 34.03 7.99 11.33 14.99 14.99 9.68 6.14 4.18 5.77 6.70 14.99 4.56 4.56
1.27E05
1.86E05
the total difference between the outputs of the network to train inputs and the real values of the outputs. As seen in Table 1, the developed model can generate very accurate results with respect to real values. The maximum difference between real values and network outputs is observed to be 6.81 min. Since the difference will not affect the planning activities, these results can be used in planning. It is clear that when the cost of TS is considered, the results are quite useful. Besides, increasing the number of predetermined standard times, this difference could be reducible. In order to search for an alternative method to compare the results of MLP ANN method, regression analysis is determined to be the best choice. When evaluating the results of the regression analysis, it seems that there is a significant difference according to real values. The computed maximum overshoot value in the regression analysis is found to be 34.03 min. When the product that has maximum overshoot value is examined, it is observed that it is produced in 92 min but surprisingly estimated as 126. That is, when it is planned to produce 3 of that product, the number of real production can be 4. This would be a big problem in planning studies. 4. Conclusion In this study, the proposed MLP ANN based method can estimate the time in similar processes using the mentioned factors which are number of pieces, number of welding operations, product’s surface
1.99E005
4.265
area, difficulty/working environment, and number of metal forming process. It is shown that the selected factors are relevant and valid for this study since the computed results are highly satisfactory that are converged to the pre-measured real time values. The factors are determined by both the related manager group and experts of the company in this area. Several network architectures are applied for different learning algorithms and the most suitable network configuration has found to be 5 5 2 1 as explained before. Among the learning algorithms used in the analysis, best results have been obtained using CGF, BFG and LM learning algorithms, respectively. While training the network, previously obtained standard time values are used. With the proposed model, the minimum total test error of standard time estimation is calculated as 1.27E05 for CGF algorithm, which is highly accurate. In order to search for an alternative method, regression analysis is determined. When using this method, it seems that the computed maximum overshoot value is not satisfactory and cannot be used for planning activities. In the proposed MLP ANN based time estimation method, many factors effecting the time are taken into consideration and it is possible to deduce the fractions of these factors during time measurements that lead us more accurate results. Moreover, when this method is compared with others, the benefits to the industry listed below can be obtained:
M. Dagdeviren et al. / Journal of Loss Prevention in the Process Industries 24 (2011) 563e567
It does not require long term experience with a product to estimate time or expert for work measurement, It minimizes the amount of TS to avoid time, effort consumptions and environmental difficulties, It gives highly accurate results with in a very limited error band to plan the future activities, It considers all the factors that affect the time by previous time studies and gives more objective estimations, It allows tracing which factors are more effective on time and helps to focus on the points of improvement. Besides these advantages, it can be seen that the method is economically applicable in many companies. The proposed method has found appropriate by the company managers and they thought it is applicable in this area. According to the managers of the selected manufacturing company, using these qualitative and quantitative factors in the estimation process, standard times of more than 2000 pieces of products or semi-products belong to several product families can be estimated in a short time. Furthermore, company saves significant time and cost in planning production and managerial activities in the future horizons. The effects of endurable environmental difficulties during the application of TS (noise, vibration, chemical gas or dust, insufficient lightning, etc.) are prevented.
5. Discussion The proposed MLP ANN based model is a single, simple, new and an accurate one. The model can be easily used for the estimation of standard time. A distinct advantage of the model is that, after proper training the model completely bypasses the repeated use of complex iterative processes for the new cases presented to it. Thus, the model is very fast after the training phase and can be used effectively in real time applications. It should also be emphasized that better results may be obtained either by choosing different training and test data sets from the ones used this study or by supplying more input data set values for training. Companies that do not know the standard time of their products due to the measurement difficulties and other economic factors can easily obtain these time values more accurately in order to aid their critical management decisions. In literature, when using CPM, the situation of the product is the most important factor. The method requires a value assigned according to the experience of the one who has applied the TS. For this reason, CPM can be highly subjective. Furthermore, it is a method where all the time effecting factors are not taken into consideration. SDF is not able to produce accurate results in every case. When the companies in this area develop same kind of similar models and use minimal practical software, they can easily calculate the standard time of the products which are to be designed or produced. On the other hand, it is not easy to apply for every product, semi-product or manufacturing processes. This can be counted as a disadvantage. However, the method is very efficient in similar manufacturing processes of various products.
567
References Akyol, D. (2000). Application of neural networks to heuristic scheduling algorithms. Computers and Industrial Engineering, 46, 679e696. Baawain, M. S., El-Din, M. G., & Smith, D. W. (2007). Artificial neural networks modeling of ozone bubble columns: mass transfer coefficient, gas hold-up, and bubble size. Ozone-Science and Engineering, 29-5, 343e352. Bhattacharya, B., & Solomatine, D. P. (2006). Machine learning in sedimentation modelling. Neural Networks, 19(2), 208e214. Celebi, F. V. (2005a). A proposed CAD model based on amplified spontaneous emission spectroscopy. Journal of Optoelectronics and Advanced Materials, 7(3), 1573e1579. Celebi, F. V. (2005b). A different approach to gain computation in laser diodes with respect to different number of quantum-wells. International Journal for Light and Electron Optics, 116(8), 375e378. Celebi, F. V. (2006). Modeling of the linewidth enhancement factors of the narrow and wide GaAs well semiconductor lasers. J. Fac. Eng. Arch. Gazi Univ., 21(1), 161e166. Charniak, E., & Mcdermott, D. (1985). Introduction to artificial intelligence. USA: Addison Wesley Publishing Company. Chetouani, Y. (2008). Using ARX and NARX approaches for modeling and prediction of the process behavior: application to a reactoreexchanger. Asia-Pacific Journal of Chemical Engineering, 3(6), 597e605. Cohen, Y., Bidanda, B., & Billo, R. E. (1998). Accelerating the generation of work measurement standards through automatic speech recognition: a laboratory study. International Journal of Production Research, 36(10), 2701e2715. Danisman, K., Dalkiran, I., & Celebi, F. V. (2006). Design of a high precision temperature measurement system based on artificial neural network for different thermocouple types. Measurement, 39(8), 695e700. Dennis, J. E., & Schnabel, R. B. (1983). Numerical methods for unconstrained optimization and nonlinear equations. Englewood Cliffs, USA: Prentice-Hall. Eraslan, E. (2009). The estimation of product standard time by artificial neural networks in the molding industry. Mathematical Problems in Engineering1e12. doi:10.1155/2009/527452, ID:527452. Freivalds, A., Konz, S., Yurgec, A., & Goldberg, J. H. (2000). Methods, work measurement and work design: are we satisfying customer needs? International Journal of Industrial Engineering: Theory Applications and Practice, 7(2), 108e114. Hagan, M. T., Demuth, H. B., & Beale, M. H. (1996). Neural network design. Boston, MA: PWS Publishing. Ham, F. M., & Kostanic, I. (2002). Principles of neurocomputing for science and engineering. Singapore: McGraw-Hill. Hassoun, M. H. (1995). Fundamentals of artificial neural networks. Cambridge: The MIT Press. Koelling, C. P., & Ramsey, T. D. (1996). Multimedia in work measurement and methods engineering. Computers and Industrial Engineering, 31(1e2), 49e52. Lasch, P., Diem, M., & Hansch, W. (2006). Artificial neural networks as supervised techniques for FT-IR microspectroscopic imaging. Journal of Chemometrics, 20(5), 209e220. Lippmann, R. P. (1989). Pattern classification using neural networks. IEEE Communication Magazine 47e64. Mandavgane, S. A., Pandharipande, S. L., & Subramanian, D. (2007). Modeling of desilication of agro based black liquor using artificial neural networks. Journal of Scientific and Industrial Research, 66(7), 517e521. Mori, H., & Yuihara, H. (2001). Deterministic annealing clustering for ANN-based short-term load forecasting. IEEE Transactions on Power Systems, 16(3), 545e551. Niebel, B., & Freivalds, A. (2003). Methods, standards, and work design (11th ed.). USA: McGraw-Hill. Ochoa-Rivera, J. C., Andreu, J., & Garcia-Bartual, R. (2007). Influence of inflows modeling on management simulation of water resources system. Journal of Water Resources Planning and Management e ASCE, 133(2), 106e116. Priore, P., Fuente, D., Puente, J., & Parreño, J. (2006). A comparison of machinelearning algorithms for dynamic scheduling of flexible manufacturing systems. Engineering Applications of Artificial Intelligence, 19, 247e255. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323, 533e536. Schalko, R. J. (1997). Artificial neural networks. USA: McGraw-Hill. Wang, W., Van Gelder, P. H. A. J. M., & Vrijling, J. K. (2006). Forecasting daily streamflow using hybrid ANN models. Journal of Hydrology, 324(1e4), 383e399. Zhang, Y. X. (2007). Artificial neural networks based on principal component analysis input selection for clinical pattern recognition analysis. Talanta, 73(1), 68e75.