- Email: [email protected]
Optimization of process parameters in feed manufacturing using artificial neural network
Computers and Electronics in Agriculture 120 (2016) 1–6
Contents lists available at ScienceDirect
Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag
Original papers
Optimization of process parameters in feed manufacturing using artificial neural network L. Sudha a, R. Dillibabu b, S. Srivatsa Srinivas b,⇑, A. Annamalai b a b
Department of Management Studies, College of Engineering, Guindy, Anna University, Chennai 600 025, India Department of Industrial Engineering, College of Engineering, Guindy, Anna University, Chennai 600 025, India
a r t i c l e
i n f o
Article history: Received 14 May 2015 Received in revised form 2 November 2015 Accepted 3 November 2015
Keywords: Feed manufacturing Process parameters Optimization Artificial neural network
a b s t r a c t Feed manufacturing faces enormous challenges and with the demand for good quality feed increasing gradually, it becomes essential to improve the processes in a feed mill. This article provides a brief overview of the different processes in feed manufacturing and identifies the critical process parameters. Five critical parameters are identified where the production rate is the output parameter. Mash feed size, steam temperature, conditioning time and feed rate are the input parameters. Artificial neural network is the methodology which is used to optimize the process parameters. Root mean squared error and coefficient of determination and computation time are used as performance measures and it is observed that Polak–Ribiere conjugate gradient backpropagation training function with log sigmoid – pure linear transfer function combination provided good results among the different available alternatives. The process parameters are then optimized using the appropriate ideal settings of neural network parameters. This model is extremely useful for the prediction of production rate for 1 specific recipe in a feed mill. Ó 2015 Elsevier B.V. All rights reserved.
1. Introduction Process parameters optimization has been an active area of research for the past few decades. Different industries face the problem of optimizing multiple, conflicting input parameters to achieve the desired value of the objective function. The objective function may be maximization (e.g. profits) or minimization (e.g. costs incurred). The various techniques of the optimization of process variables have been applied in different contexts and have showed impressive results in the respective areas. The machine setting up tasks in the feed industry are cumbersome. As such, there is no standard operating procedure for setting up of optimum levels of process variables as far as feed mills are concerned. The operations in a feed mill are sequential and require the optimization of process parameters to achieve maximum productivity. Without the knowledge of ideal levels of input parameters, recurrent problems arise day-to-day in the machine setup which ultimately affects the feed mill performance. Thus, the optimization of process parameters in feed manufacturing would help in overcoming these issues and achieve better performance. The optimization of process parameters is a problem that has been actively studied as far as food processing industries are
⇑ Corresponding author. E-mail address: [email protected] (S. Srivatsa Srinivas). http://dx.doi.org/10.1016/j.compag.2015.11.004 0168-1699/Ó 2015 Elsevier B.V. All rights reserved.
concerned. Various tools and techniques have been used in this regard to achieve good quality solutions. Ramakrishnan et al. (2004) used simulation as a tool for the prediction of process variables for the intelligent control of tunnel freezers. Arora et al. (2007) present a classic example of optimization of process parameters for milling of enzymatically pretreated basmati rice using response surface methodology. Lamberts et al. (2007) studied the effect of milling on color and nutritional properties of rice through statistical analysis. In another study, Oztop et al. (2007) used Taguchi technique for the optimization of microwave frying of potato slices. Roy et al. (2008) investigated the influence of different operating conditions on overall energy usage and rice quality. Feed processing is not an exception and requires a detailed study and optimization of process variables to achieve good results. Different procedures have been used earlier for the optimization of process parameters in rice milling and allied food processing branches. Due to its successful implementation as evident in the literature, artificial neural networks are utilized to optimize parameters in this case. Agatonovic-Kustrin and Beresford (2000) defined an artificial neural network as a biologically stimulated computational model consisting of hundreds of single units, artificial neurons, connected with weights which constitute the neural structure. Cus and Zuperl (2006) illustrated that a neural network approach can produce significant gains in productivity in the optimization of cutting conditions for materials. Chegini et al. (2008) developed
2
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
a supervised artificial neural network to predict seven performance indices based on three input parameters in the prediction of process and product parameters for an orange juice spray dryer. Hamzaoui et al. (2015) made use of artificial neural network and its inverse to optimize operating conditions in a steam turbine. This model provided a new artificial intelligence approach to optimize process parameters and showed good results. These studies clearly demonstrate the usefulness of neural networks in solving optimization problems. In one of the earliest studies in the area of feed manufacturing, Behnke (1996) presented the various problems in the feed manufacturing sector. A series of research findings during the late 1990s exhibited the various aspects of animal feed production. Thomas and Van der Poel (1996) put forward the criteria for pellet quality. According to them, animal feed in the form of pellet has many advantages. They are: The better flow properties of pellets when compared to the mash, makes transfer in conveyors smooth and damage free. Pellets possess high bulk density and so, more tonnage can be carried by truck. The formulation of pellets helps to maintain the moisture content and nutrition. Thomas et al. (1997) emphasized the need to understand the processes in a feed mill better and quantify the parameters so that high performance levels can be achieved with low cost. Also, they discussed the impacts of variations in process variables on pellet quality which is measured in terms of hardness and durability. The pelleting process considered in this respect was the combination of conditioning, pelleting and cooling. In the following year, Thomas et al. (1998) studied the effect of feedstuff components on the quality of animal feed. Pathumnakul et al. (2009) consider the application of neural networks for the selection of appropriate feed mix in the feed processing industry. Though the study by Pathumnakul et al. (2009) contributed greatly by applying neural networks to the feed industry, it was not successful in predicting production rate using process parameters because of the consideration of numerous recipes. Hence, the physical–chemical complexity of the interactions among feed ingredients made it not quite possible to obtain the production rate using simple neural network and input factors stated. Our study has only one recipe as stated in Table 1, hence, the complexities have tremendously been reduced. Therefore, our case is possible for recipe specific and easily adoptable. In addition, this work focusses on the study of processes in a feed mill and the critical operations in a feed mill which includes conditioning and pelleting (Thomas et al., 1997). This article provides an overview of the processes in a feed mill and the significant operating parameters are identified which is explained in Section 2. The methodology to solve the given problem, artificial neural network is discussed in Section 3. For the appropriate settings of the neural network, the optimized process parameter levels are elaborated in Section 4. Finally, Section 5 provides the conclusions of the research.
Table 1 Feed formulation. Raw material type
Inclusion (%)
Maize Soya Other ingredients Total
60 20 20 100
2. Material and methods 2.1. Operations in a feed mill There are different types of feed manufacturing facilities which possess certain unique characteristics. The components in the feed mill, the type of arrangement and the general layout are the primary causes of variation in the feed mills. However, the feed manufacturing consists of eight major operations. The feed mill considered in this case possesses vertical layout where the movement of materials is from top to bottom during the course of the operations. This system effectively utilizes the gravity phenomenon, which reduces unnecessary costs for movement. 2.1.1. Raw ingredient receiving Feed mills typically receive incoming ingredients by both rail and truck. The primary raw materials, maize and soya are received separately using the rail and truck systems. Secondary raw materials are transferred through bulk trucks and pneumatically moved to the storage bins. 2.1.2. Raw ingredient distribution and storage Screw conveyor is generally used in feed mills as it offers many advantages over other types of conveyors. Screw conveyors not only transport the materials very fast, but also additional functionality of accurate measurement of the ingredients during the transfer process. This leads to significant savings in the production time in the long run. If the raw materials are greater than 50 kg, bins are used for the purpose of storage. Otherwise, bins are not used. 2.1.3. Grinding Grinding is a process used to ground irregularly shaped raw material particles to fine powdered substances. The advantage of grinding operation is that the finer particles produced in this stage helps in healthy mixing of the ingredients, thereby leading to the production of pellets with accurate formulation. Grinding mills are generally positioned under whole grain storage bins, in a separate apartment within the mill facility. Hammer mills are the most common type of milling equipment. 2.1.4. Batching In order to produce specific feed combinations, certain materials must be transferred through batches to the mixer. This process is referred to as batching. This is an important operation before the mixing stage and requires all the raw materials to be batched in accurate proportions to manufacture good quality feed. Screw conveyors, which possess very good measurement control and scale hoppers, which are mounted above the mixers are used for this purpose. Bulk bag and hand dump stations are also used to add materials to the feed mix. Proper venting must be provided between the mixer and scale hoppers to maintain steady flow of materials to the mixers. 2.1.5. Mixing Mixing is one of the important operations in a feed processing plant. The gap between the body of the mixer and the blades must be maintained as small as possible to produce uniform mixtures. Another factor to be considered in this process is the mixing time. Mixing times of 3–5 min are most commonly used to ensure even mixing. Mixing times do not hold much significance in our study as they are already optimal. Mash resides in storage until needed for mixing operation. Ribbon mixers vary in size and are widely used. They operate at a speed of approximately 40 rpm. At the end of the mixing operation, the output is mash. The mash feed size plays a significant role in the quality of feed and production time.
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
2.1.6. Conditioning Conditioning and Pelleting are the critical operations as far as feed manufacturing is concerned. The mixture from the previous stage is transported to the conditioner. In the conditioner, the mash is mixed with steam at certain temperature for a certain period of time. This process ensures that the moisture content in the final feed is maximized. Otherwise, the pelleted feed might have very little moisture content which affects the nutritious value. In addition, the production rate of pellets also drops rapidly.
2.1.7. Pelleting Pelleting is an operation that is used to increase the density of the feed, which ultimately helps in storage, shipping and handling. The improvement in feed efficiency and palatability is also significant. The conditioned feed materials are forced through die openings in this process with the help of rotating rollers. The feeder frequency or feed rate must also be maintained at an optimum value to ensure continuous production of pellets at a faster rate. The pellets are cooled so that pellet temperature is brought to surrounding temperature to avoid wastes. They are also screened to remove broken pellets from the mix and then packed.
2.3. Data and methodology The study was conducted in a feed mill in India. This case study involved a thorough understanding of the processes in feed manufacturing to find the appropriate process parameters for optimization. In this case, only poultry feed manufacturing is considered. The raw materials for the production of cattle and poultry feed are almost similar, with minor variations in the quantity of the contents. Generally, the primary raw materials are maize and soya. The secondary raw materials which include fat, vitamins, minerals and medicines are added during mixing stage for various purposes. These liquid contents also play an additional role of maintaining the moisture level in the pelleted feed to an optimum level which
2.1.8. Final product storage and load out Bulk feed materials are stored in bins whereas, the packed feed materials are sent to warehouse for dispatch. Production capacity, space available, and utilization of the facility are some of the factors which has an effect on the type and arrangement of warehouse. Reversible screw conveyors and weigh lorry systems are generally used in the load out systems. Adequate clearance and platforms need to be provided in the both the systems for quicker service.
2.2. Process parameters After a thorough literature survey and understanding of the operations in a feed mill, the process parameters of significance are identified. It is found out that conditioning and pelleting are important operations in feed manufacturing and have high impact on the quality of the pelleted animal feed (Thomas et al., 1997). Thus, the identified process parameters are relevant to this process. The input parameters are:
Mash feed size. Steam temperature. Conditioning time. Feed rate.
While the mash feed size is decided during the mixing stage before conditioning operation, 2 input parameters – the steam temperature and conditioning time are set and processed at the conditioning stage. Finally, the feed rate is maintained at the pelleting phase to ensure good productivity in feed manufacturing. The only output process parameter considered is Production rate. Production rate is different from the feed rate and refers to the rate at which feed is produced at the outlet. However, these input parameters have conflicting effects on the production rate which needs to be addressed. The collected data is shown in Table A1. From the collected data, it can be observed that feed rate and steam temperature are directly proportional to the production rate. There are no conclusive evidences to show that mash feed size and conditioning time are directly proportional to the output variable. As all the input parameters do not have the same effect on output, it becomes necessary to optimize the process parameters to achieve maximum production rate.
3
Fig. 1. Schematic methodology.
4
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
Fig. 2. The artificial neural network.
Table 2 Comparison of training functions for log sigmoid – pure linear combination. Function
RMSE
R2
Best linear equation
Seconds/ iteration
Runs/ iteration
Trainlm Traingd Traingdm Traincgp Trainscg Trainbfg Traincgb Trainoss Traincgf Traingdx
0.273 0.099 0.385 0.057 0.365 0.227 0.164 0.193 0.350 0.161
0.968 0.974 0.978 0.986 0.976 0.947 0.980 0.980 0.976 0.978
Y = 1.000X 0.330 Y = 0.980X + 0.130 Y = 0.990X + 0.073 Y = 0.990X + 0.150 Y = 1.000X + 0.018 Y = 1.000X + 0.390 Y = 1.000X + 0.072 Y = 1.000X 0.072 Y = 0.970X + 0.250 Y = 1.000X 0.055
3 6 6 13 9 8 2 14 17 8
102 1000 1000 794 1000 527 45 810 1000 715
Table 3 Comparison of training functions for tan sigmoid – pure linear combination. Function
RMSE
R2
Best linear equation
Seconds/ iteration
Runs/ iteration
Trainlm Traingd Traingdm Traincgp Trainscg Trainbfg Traincgb Trainoss Traincgf Traingdx
0.336 0.415 0.416 0.271 0.207 0.319 0.206 0.239 0.187 0.296
0.943 0.974 0.966 0.958 0.972 0.962 0.972 0.964 0.978 0.960
Y = 0.940X + 0.400 Y = 0.960X + 0.360 Y = 0.960X + 0.230 Y = 0.980X + 0.080 Y = 0.970X + 0.150 Y = 1.000X 0.052 Y = 0.980X + 0.300 Y = 0.950X + 0.460 Y = 0.990X + 0.140 Y = 0.970X 0.320
3 8 8 27 17 6 12 17 22 11
102 1000 1000 1000 1000 293 688 1000 1000 878
functions, hidden layers and neurons. The speed of this process varies according to the specifications of the system. There are different types of neural network available. However, we chose to solve this particular problem using feedforward backpropagation network. The network flow is only in one direction. There is no feedback. Input and Output data is fed into the network as this network uses supervised learning. In case of unsupervised learning, only input is given and linear equations are built such that maximum correlation coefficient is achieved. Signals flow forward and errors are propagated backward to ensure that errors are reduced. Random weights are initialized and change in each run of each iteration. The objective is to make errors minimal. The data is normalized to improve the performance of the network. The different parameters considered in the neural network include training functions, performance measures, number of layers and neurons, and transfer function. The schematic methodology is presented in Fig. 1. 3. Optimization methodology 3.1. Training function Training function involves changing weights of each arc in the neural network. 10 different training backpropagation algorithms are compared to find the best algorithm which effectively changes weights to improve the performance measure and leads to optimal weights. Ten different training functions are compared based on various performance measures to find the best fitting training function for the given data. 3.2. Performance measures
helps in the manufacture of high quality and nutritious feed. The feed formulation is presented in Table 1. Other ingredients denote the secondary raw materials. This analysis is carried out in MATLAB 7.12.0, Windows 7 operating system, Intel R CPU 2.53 GHz, 4.00 GB of RAM. The input data and target data was the data collected which are fed into the network for training. The specifications of operating system are provided because we are comparing different training and transfer
The different performance measures considered are root mean squared error (RMSE), coefficient of determination (R2), number of iterations, time per iteration and runs per iteration. The first two performance measures are used to compare neural network parameters based on errors. The other performance measures are based on computational performance. The equations for RMSE and R2 are presented in Eqs. (1) and (2).
5
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6 Table 4 Weights and bias for the ideal ANN settings. IW(s,k)
W0(s) b1(s) b2
1.192 1.261 0.424 1.681 1.136 0.613 1.972 1.588 1.212 0.648 0.590 2.489 0
0.351 0.608 1.438 0.457 1.125 1.776 0.406 1.082 0.840 1.573 0.173 1.936
1.030 2.323 0.886 1.530 1.483 1.766 1.157 0.686 1.604 2.210 0.623 1.383
0.001 1.329 0.008 1.076 0.456 0.011 0.981 1.380 1.166 0.609 0.139 0.829
0.942 0.276
0.239 0.829
0.171 1.383
0.603 1.936
0.449 2.489
tion is fixed to a certain value. Bias neuron helps in fixing the transfer function to a certain value.
Table 5 Optimum levels for input parameters for maximization of production rate. Input parameters
Output parameter
Mash feed size = 1.5 mm Steam temperature = 80 °C Feed rate = 16 tons/h Conditioning time = 32 s
Production rate = 13.82 tons/h
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 Pn i¼1 xi;net xi;act RMSE ¼ n 2 Pn x i;net xi;act R2 ¼ 1 Pi¼1 2 n i¼1 xi;act xav g
0.177 0.276
ð1Þ ð2Þ
where n is the number of actual observations, xi;net is the predicted output values from the network, xi;act is the actual output value from experiments and xav g is the mean of actual values from experimental observations. 3.3. Number of layers and neurons The number of layers and neurons should be kept at an optimum level to enhance the computational efficiency of the network and error reduction. For this case, 1 hidden layer and 1 output layer are fixed. The number of neurons is also set to 10. Thus, no variation occurs due to the number of layers and neurons for this problem. If the number of layers is increased, a proportionate increase of 10 neurons would take place which would enormously increase the computational effort of the network. So, the fixed number of layers and neurons would reduce the problem complexity to an extent. 3.4. Transfer function Transfer functions are input–output functions and 2 sets of transfer function combinations are compared to check for best results. Transfer functions are present in hidden layer–output layer. The combinations compared in this case are log sigmoid – pure linear and tan sigmoid – pure linear. A comprehensive neural network is shown in Fig. 2. Weights are present in the arcs connecting input–hidden and hidden–output layer. The weights are changed to reduce the error in all the iterations. Bias is the additional input of the neuron, which improves the neuron properties but increases complexity. The weights and the value in the transfer functions can be varied during the learning process of a neural network. This is not practical and would be better if only one of the variables should be varied. Thus, a bias neuron is utilized to solve this problem. The weights can be changed only if the transfer func-
4. Results and discussion Though different variables are present, every function does not fit suitably to the given data. For example, a training function might present with less error for one dataset. However, the same training function may not be precise and accurate for any other dataset. This leads us to a point where we need to compare the different parameters based on training and transfer functions, with the number of layers and neurons fixed. 4.1. Comparison of neural network parameters 10 different training functions are compared for 2 sets of transfer function (hidden–output layer) combinations – log sigmoid – pure linear and tan sigmoid – pure linear. It is shown in Tables 2 and 3. Epochs are set to 1000 and Iterations are set to 5 for all training functions. Two sets of transfer functions and training functions are compared and it is found that Polak–Ribiere conjugate gradient backpropagation training function with log sigmoid – pure linear transfer function combination provided best results with least RMSE and high positive R2 at satisfactory computation time. The weights and bias for the ideal neural network settings are shown in Table 4. These settings are used to simulate and find the optimum parameter levels for maximum production rate. In Table 4, s denotes the number of neurons in hidden layer and k denotes the number of input variables. b1(s) and b2 refer to the biases and IW(s,k) and W0(s) refer to weight of arcs of the neural network. 4.2. Prediction of optimum parameters Sample data is used to find out the levels of input variables which results in maximum production rate. The accuracy of the solution increases as the number of levels of each of these input variables becomes large. The optimum levels of input parameters are shown in Table 5. Thus, maximum production rate of 13.82 tons/h can be achieved in a feed mill by properly maintaining the input process variables. This would lead to significant cost savings as quality is increased and power consumption is decreased. However, it should be noted that the results obtained are recipe specific and these results should be used with caution. 5. Conclusion The optimization of process parameters for one recipe in feed manufacturing is a step in the right direction to improve the productivity of feed processes. The prediction of production rate using the neural networks for a specific recipe acts as a useful guideline
6
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
Table A1 Data on input parameters and the corresponding production rate. S. No.
Mash feed size (mm)
Steam temperature (degrees)
Feed rate (tons/h)
Conditioning time (s)
Production rate (tons/h)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5
80 81 60 81 80 80 81 81 80 42 81 81 81 62 81 81 81 60 80 81 82 82 80 80 81 81 80 48 80 81 65 80 80 80 80 60 80 80 60 80 60 83 83 84 60
16.5 17 5 16.5 16.5 17 17 17 17 5 17.5 17.5 17.5 6 17.5 8.4 8 5 8 8 8 8 8 8 8 8 8 4.5 9.6 9.8 5 14.2 14.1 14.2 14.5 5 14 14 5 14 5 14 13.5 13.2 4.5
33 34 32 32 34 35 35 32 32 34 33 33 32 34 33 32 32 33 34 32 33 34 33 33 32 32 32 33 32 32 32 33 34 33 32 32 32 32 32 33 34 33 33 34 33
12.4 12 4.6 13 13.2 12 13.4 11.8 12.2 3.7 13.6 12 13.2 4 12.2 6.7 7 3.8 6 5 5 6 6 6 6 6.3 6 3.2 7.8 8 3.7 11.6 12 11.8 12.3 4 10.4 10.2 3.6 11.2 4.1 12.4 12.7 12.1 3.2
for adoption in feed mills. Further, the work could be extended by incorporating other parameters that are significant in other operations apart from conditioning and pelleting to achieve a much truer prediction of the production rate. The scope of research in the feed manufacturing sector is not restricted to feed mill operations alone. Pathumnakul et al. (2011) focus on the need for integration
of information among the upstream and downstream members of the animal feed supply chain to reduce costs and produce better performance. Similarly, the study also analyzes the pros and cons of multiple period planning and its effect on costs. Thus, inventory management is one of the crucial management problems that show enough promise in the feed supply chain. Moreover, the current problem’s scope could be expanded by considering multiple objectives and integrating with other issues in the animal feed supply chain. The production rate settings could be synchronized and optimized along with the inventory to meet the required demand and avoid loss of material and money. Appendix A. Data collection See Table A1. References Agatonovic-Kustrin, S., Beresford, R., 2000. Basic concepts of artificial neural network (ANN) modeling and its application in pharmaceutical research. J. Pharm. Biomed. Anal. 22 (5), 717–727. Arora, G., Sehgal, V.K., Arora, M., 2007. Optimization of process parameters for milling of enzymatically pretreated Basmati rice. J. Food Eng. 82 (2), 153–159. Behnke, K.C., 1996. Feed manufacturing technology: current issues and challenges. Animal Feed Sci. Technol. 62 (1), 49–57. Chegini, G.R., Khazaei, J., Ghobadian, B., Goudarzi, A.M., 2008. Prediction of process and product parameters in an orange juice spray dryer using artificial neural networks. J. Food Eng. 84 (4), 534–543. Cus, F., Zuperl, U., 2006. Approach to optimization of cutting conditions by using artificial neural networks. J. Mater. Process. Technol. 173 (3), 281–290. Hamzaoui, Y.E., Rodríguez, J.A., Hernández, J.A., Salazar, V., 2015. Optimization of operating conditions for steam turbine using an artificial neural network inverse. Appl. Therm. Eng. 75, 648–657. Lamberts, L., De Bie, E., Vandeputte, G.E., Veraverbeke, W.S., Derycke, V., De Man, W., Delcour, J.A., 2007. Effect of milling on colour and nutritional properties of rice. Food Chem. 100 (4), 1496–1503. Oztop, M.H., Sahin, S., Sumnu, G., 2007. Optimization of microwave frying of potato slices by using Taguchi technique. J. Food Eng. 79 (1), 83–91. Pathumnakul, S., Piewthongngam, K., Apichottanakul, A., 2009. A neural network approach to the selection of feed mix in the feed industry. Comput. Electron. Agric. 68 (1), 18–24. Pathumnakul, S., Ittiphalin, M., Piewthongngam, K., Rujikietkumjorn, S., 2011. Should feed mills go beyond traditional least cost formulation? Comput. Electron. Agric. 75 (2), 243–249. Ramakrishnan, S., Wysk, R.A., Prabhu, V.V., 2004. Prediction of process parameters for intelligent control of tunnel freezers using simulation. J. Food Eng. 65 (1), 23–31. Roy, P., Ijiri, T., Okadome, H., Nei, D., Orikasa, T., Nakamura, N., Shiina, T., 2008. Effect of processing conditions on overall energy consumption and quality of rice (Oryza sativa L.). J. Food Eng. 89 (3), 343–348. Thomas, M.A.F.B., Van der Poel, A.F.B., 1996. Physical quality of pelleted animal feed 1. Criteria for pellet quality. Anim. Feed Sci. Technol. 61 (1), 89–112. Thomas, M., Van Zuilichem, D.J., Van der Poel, A.F.B., 1997. Physical quality of pelleted animal feed. 2. Contribution of processes and its conditions. Anim. Feed Sci. Technol. 64 (2), 173–192. Thomas, M., Van Vliet, T., Van der Poel, A.F.B., 1998. Physical quality of pelleted animal feed 3. Contribution of feedstuff components. Anim. Feed Sci. Technol. 70 (1), 59–78.
Contents lists available at ScienceDirect
Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag
Original papers
Optimization of process parameters in feed manufacturing using artificial neural network L. Sudha a, R. Dillibabu b, S. Srivatsa Srinivas b,⇑, A. Annamalai b a b
Department of Management Studies, College of Engineering, Guindy, Anna University, Chennai 600 025, India Department of Industrial Engineering, College of Engineering, Guindy, Anna University, Chennai 600 025, India
a r t i c l e
i n f o
Article history: Received 14 May 2015 Received in revised form 2 November 2015 Accepted 3 November 2015
Keywords: Feed manufacturing Process parameters Optimization Artificial neural network
a b s t r a c t Feed manufacturing faces enormous challenges and with the demand for good quality feed increasing gradually, it becomes essential to improve the processes in a feed mill. This article provides a brief overview of the different processes in feed manufacturing and identifies the critical process parameters. Five critical parameters are identified where the production rate is the output parameter. Mash feed size, steam temperature, conditioning time and feed rate are the input parameters. Artificial neural network is the methodology which is used to optimize the process parameters. Root mean squared error and coefficient of determination and computation time are used as performance measures and it is observed that Polak–Ribiere conjugate gradient backpropagation training function with log sigmoid – pure linear transfer function combination provided good results among the different available alternatives. The process parameters are then optimized using the appropriate ideal settings of neural network parameters. This model is extremely useful for the prediction of production rate for 1 specific recipe in a feed mill. Ó 2015 Elsevier B.V. All rights reserved.
1. Introduction Process parameters optimization has been an active area of research for the past few decades. Different industries face the problem of optimizing multiple, conflicting input parameters to achieve the desired value of the objective function. The objective function may be maximization (e.g. profits) or minimization (e.g. costs incurred). The various techniques of the optimization of process variables have been applied in different contexts and have showed impressive results in the respective areas. The machine setting up tasks in the feed industry are cumbersome. As such, there is no standard operating procedure for setting up of optimum levels of process variables as far as feed mills are concerned. The operations in a feed mill are sequential and require the optimization of process parameters to achieve maximum productivity. Without the knowledge of ideal levels of input parameters, recurrent problems arise day-to-day in the machine setup which ultimately affects the feed mill performance. Thus, the optimization of process parameters in feed manufacturing would help in overcoming these issues and achieve better performance. The optimization of process parameters is a problem that has been actively studied as far as food processing industries are
⇑ Corresponding author. E-mail address: [email protected] (S. Srivatsa Srinivas). http://dx.doi.org/10.1016/j.compag.2015.11.004 0168-1699/Ó 2015 Elsevier B.V. All rights reserved.
concerned. Various tools and techniques have been used in this regard to achieve good quality solutions. Ramakrishnan et al. (2004) used simulation as a tool for the prediction of process variables for the intelligent control of tunnel freezers. Arora et al. (2007) present a classic example of optimization of process parameters for milling of enzymatically pretreated basmati rice using response surface methodology. Lamberts et al. (2007) studied the effect of milling on color and nutritional properties of rice through statistical analysis. In another study, Oztop et al. (2007) used Taguchi technique for the optimization of microwave frying of potato slices. Roy et al. (2008) investigated the influence of different operating conditions on overall energy usage and rice quality. Feed processing is not an exception and requires a detailed study and optimization of process variables to achieve good results. Different procedures have been used earlier for the optimization of process parameters in rice milling and allied food processing branches. Due to its successful implementation as evident in the literature, artificial neural networks are utilized to optimize parameters in this case. Agatonovic-Kustrin and Beresford (2000) defined an artificial neural network as a biologically stimulated computational model consisting of hundreds of single units, artificial neurons, connected with weights which constitute the neural structure. Cus and Zuperl (2006) illustrated that a neural network approach can produce significant gains in productivity in the optimization of cutting conditions for materials. Chegini et al. (2008) developed
2
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
a supervised artificial neural network to predict seven performance indices based on three input parameters in the prediction of process and product parameters for an orange juice spray dryer. Hamzaoui et al. (2015) made use of artificial neural network and its inverse to optimize operating conditions in a steam turbine. This model provided a new artificial intelligence approach to optimize process parameters and showed good results. These studies clearly demonstrate the usefulness of neural networks in solving optimization problems. In one of the earliest studies in the area of feed manufacturing, Behnke (1996) presented the various problems in the feed manufacturing sector. A series of research findings during the late 1990s exhibited the various aspects of animal feed production. Thomas and Van der Poel (1996) put forward the criteria for pellet quality. According to them, animal feed in the form of pellet has many advantages. They are: The better flow properties of pellets when compared to the mash, makes transfer in conveyors smooth and damage free. Pellets possess high bulk density and so, more tonnage can be carried by truck. The formulation of pellets helps to maintain the moisture content and nutrition. Thomas et al. (1997) emphasized the need to understand the processes in a feed mill better and quantify the parameters so that high performance levels can be achieved with low cost. Also, they discussed the impacts of variations in process variables on pellet quality which is measured in terms of hardness and durability. The pelleting process considered in this respect was the combination of conditioning, pelleting and cooling. In the following year, Thomas et al. (1998) studied the effect of feedstuff components on the quality of animal feed. Pathumnakul et al. (2009) consider the application of neural networks for the selection of appropriate feed mix in the feed processing industry. Though the study by Pathumnakul et al. (2009) contributed greatly by applying neural networks to the feed industry, it was not successful in predicting production rate using process parameters because of the consideration of numerous recipes. Hence, the physical–chemical complexity of the interactions among feed ingredients made it not quite possible to obtain the production rate using simple neural network and input factors stated. Our study has only one recipe as stated in Table 1, hence, the complexities have tremendously been reduced. Therefore, our case is possible for recipe specific and easily adoptable. In addition, this work focusses on the study of processes in a feed mill and the critical operations in a feed mill which includes conditioning and pelleting (Thomas et al., 1997). This article provides an overview of the processes in a feed mill and the significant operating parameters are identified which is explained in Section 2. The methodology to solve the given problem, artificial neural network is discussed in Section 3. For the appropriate settings of the neural network, the optimized process parameter levels are elaborated in Section 4. Finally, Section 5 provides the conclusions of the research.
Table 1 Feed formulation. Raw material type
Inclusion (%)
Maize Soya Other ingredients Total
60 20 20 100
2. Material and methods 2.1. Operations in a feed mill There are different types of feed manufacturing facilities which possess certain unique characteristics. The components in the feed mill, the type of arrangement and the general layout are the primary causes of variation in the feed mills. However, the feed manufacturing consists of eight major operations. The feed mill considered in this case possesses vertical layout where the movement of materials is from top to bottom during the course of the operations. This system effectively utilizes the gravity phenomenon, which reduces unnecessary costs for movement. 2.1.1. Raw ingredient receiving Feed mills typically receive incoming ingredients by both rail and truck. The primary raw materials, maize and soya are received separately using the rail and truck systems. Secondary raw materials are transferred through bulk trucks and pneumatically moved to the storage bins. 2.1.2. Raw ingredient distribution and storage Screw conveyor is generally used in feed mills as it offers many advantages over other types of conveyors. Screw conveyors not only transport the materials very fast, but also additional functionality of accurate measurement of the ingredients during the transfer process. This leads to significant savings in the production time in the long run. If the raw materials are greater than 50 kg, bins are used for the purpose of storage. Otherwise, bins are not used. 2.1.3. Grinding Grinding is a process used to ground irregularly shaped raw material particles to fine powdered substances. The advantage of grinding operation is that the finer particles produced in this stage helps in healthy mixing of the ingredients, thereby leading to the production of pellets with accurate formulation. Grinding mills are generally positioned under whole grain storage bins, in a separate apartment within the mill facility. Hammer mills are the most common type of milling equipment. 2.1.4. Batching In order to produce specific feed combinations, certain materials must be transferred through batches to the mixer. This process is referred to as batching. This is an important operation before the mixing stage and requires all the raw materials to be batched in accurate proportions to manufacture good quality feed. Screw conveyors, which possess very good measurement control and scale hoppers, which are mounted above the mixers are used for this purpose. Bulk bag and hand dump stations are also used to add materials to the feed mix. Proper venting must be provided between the mixer and scale hoppers to maintain steady flow of materials to the mixers. 2.1.5. Mixing Mixing is one of the important operations in a feed processing plant. The gap between the body of the mixer and the blades must be maintained as small as possible to produce uniform mixtures. Another factor to be considered in this process is the mixing time. Mixing times of 3–5 min are most commonly used to ensure even mixing. Mixing times do not hold much significance in our study as they are already optimal. Mash resides in storage until needed for mixing operation. Ribbon mixers vary in size and are widely used. They operate at a speed of approximately 40 rpm. At the end of the mixing operation, the output is mash. The mash feed size plays a significant role in the quality of feed and production time.
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
2.1.6. Conditioning Conditioning and Pelleting are the critical operations as far as feed manufacturing is concerned. The mixture from the previous stage is transported to the conditioner. In the conditioner, the mash is mixed with steam at certain temperature for a certain period of time. This process ensures that the moisture content in the final feed is maximized. Otherwise, the pelleted feed might have very little moisture content which affects the nutritious value. In addition, the production rate of pellets also drops rapidly.
2.1.7. Pelleting Pelleting is an operation that is used to increase the density of the feed, which ultimately helps in storage, shipping and handling. The improvement in feed efficiency and palatability is also significant. The conditioned feed materials are forced through die openings in this process with the help of rotating rollers. The feeder frequency or feed rate must also be maintained at an optimum value to ensure continuous production of pellets at a faster rate. The pellets are cooled so that pellet temperature is brought to surrounding temperature to avoid wastes. They are also screened to remove broken pellets from the mix and then packed.
2.3. Data and methodology The study was conducted in a feed mill in India. This case study involved a thorough understanding of the processes in feed manufacturing to find the appropriate process parameters for optimization. In this case, only poultry feed manufacturing is considered. The raw materials for the production of cattle and poultry feed are almost similar, with minor variations in the quantity of the contents. Generally, the primary raw materials are maize and soya. The secondary raw materials which include fat, vitamins, minerals and medicines are added during mixing stage for various purposes. These liquid contents also play an additional role of maintaining the moisture level in the pelleted feed to an optimum level which
2.1.8. Final product storage and load out Bulk feed materials are stored in bins whereas, the packed feed materials are sent to warehouse for dispatch. Production capacity, space available, and utilization of the facility are some of the factors which has an effect on the type and arrangement of warehouse. Reversible screw conveyors and weigh lorry systems are generally used in the load out systems. Adequate clearance and platforms need to be provided in the both the systems for quicker service.
2.2. Process parameters After a thorough literature survey and understanding of the operations in a feed mill, the process parameters of significance are identified. It is found out that conditioning and pelleting are important operations in feed manufacturing and have high impact on the quality of the pelleted animal feed (Thomas et al., 1997). Thus, the identified process parameters are relevant to this process. The input parameters are:
Mash feed size. Steam temperature. Conditioning time. Feed rate.
While the mash feed size is decided during the mixing stage before conditioning operation, 2 input parameters – the steam temperature and conditioning time are set and processed at the conditioning stage. Finally, the feed rate is maintained at the pelleting phase to ensure good productivity in feed manufacturing. The only output process parameter considered is Production rate. Production rate is different from the feed rate and refers to the rate at which feed is produced at the outlet. However, these input parameters have conflicting effects on the production rate which needs to be addressed. The collected data is shown in Table A1. From the collected data, it can be observed that feed rate and steam temperature are directly proportional to the production rate. There are no conclusive evidences to show that mash feed size and conditioning time are directly proportional to the output variable. As all the input parameters do not have the same effect on output, it becomes necessary to optimize the process parameters to achieve maximum production rate.
3
Fig. 1. Schematic methodology.
4
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
Fig. 2. The artificial neural network.
Table 2 Comparison of training functions for log sigmoid – pure linear combination. Function
RMSE
R2
Best linear equation
Seconds/ iteration
Runs/ iteration
Trainlm Traingd Traingdm Traincgp Trainscg Trainbfg Traincgb Trainoss Traincgf Traingdx
0.273 0.099 0.385 0.057 0.365 0.227 0.164 0.193 0.350 0.161
0.968 0.974 0.978 0.986 0.976 0.947 0.980 0.980 0.976 0.978
Y = 1.000X 0.330 Y = 0.980X + 0.130 Y = 0.990X + 0.073 Y = 0.990X + 0.150 Y = 1.000X + 0.018 Y = 1.000X + 0.390 Y = 1.000X + 0.072 Y = 1.000X 0.072 Y = 0.970X + 0.250 Y = 1.000X 0.055
3 6 6 13 9 8 2 14 17 8
102 1000 1000 794 1000 527 45 810 1000 715
Table 3 Comparison of training functions for tan sigmoid – pure linear combination. Function
RMSE
R2
Best linear equation
Seconds/ iteration
Runs/ iteration
Trainlm Traingd Traingdm Traincgp Trainscg Trainbfg Traincgb Trainoss Traincgf Traingdx
0.336 0.415 0.416 0.271 0.207 0.319 0.206 0.239 0.187 0.296
0.943 0.974 0.966 0.958 0.972 0.962 0.972 0.964 0.978 0.960
Y = 0.940X + 0.400 Y = 0.960X + 0.360 Y = 0.960X + 0.230 Y = 0.980X + 0.080 Y = 0.970X + 0.150 Y = 1.000X 0.052 Y = 0.980X + 0.300 Y = 0.950X + 0.460 Y = 0.990X + 0.140 Y = 0.970X 0.320
3 8 8 27 17 6 12 17 22 11
102 1000 1000 1000 1000 293 688 1000 1000 878
functions, hidden layers and neurons. The speed of this process varies according to the specifications of the system. There are different types of neural network available. However, we chose to solve this particular problem using feedforward backpropagation network. The network flow is only in one direction. There is no feedback. Input and Output data is fed into the network as this network uses supervised learning. In case of unsupervised learning, only input is given and linear equations are built such that maximum correlation coefficient is achieved. Signals flow forward and errors are propagated backward to ensure that errors are reduced. Random weights are initialized and change in each run of each iteration. The objective is to make errors minimal. The data is normalized to improve the performance of the network. The different parameters considered in the neural network include training functions, performance measures, number of layers and neurons, and transfer function. The schematic methodology is presented in Fig. 1. 3. Optimization methodology 3.1. Training function Training function involves changing weights of each arc in the neural network. 10 different training backpropagation algorithms are compared to find the best algorithm which effectively changes weights to improve the performance measure and leads to optimal weights. Ten different training functions are compared based on various performance measures to find the best fitting training function for the given data. 3.2. Performance measures
helps in the manufacture of high quality and nutritious feed. The feed formulation is presented in Table 1. Other ingredients denote the secondary raw materials. This analysis is carried out in MATLAB 7.12.0, Windows 7 operating system, Intel R CPU 2.53 GHz, 4.00 GB of RAM. The input data and target data was the data collected which are fed into the network for training. The specifications of operating system are provided because we are comparing different training and transfer
The different performance measures considered are root mean squared error (RMSE), coefficient of determination (R2), number of iterations, time per iteration and runs per iteration. The first two performance measures are used to compare neural network parameters based on errors. The other performance measures are based on computational performance. The equations for RMSE and R2 are presented in Eqs. (1) and (2).
5
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6 Table 4 Weights and bias for the ideal ANN settings. IW(s,k)
W0(s) b1(s) b2
1.192 1.261 0.424 1.681 1.136 0.613 1.972 1.588 1.212 0.648 0.590 2.489 0
0.351 0.608 1.438 0.457 1.125 1.776 0.406 1.082 0.840 1.573 0.173 1.936
1.030 2.323 0.886 1.530 1.483 1.766 1.157 0.686 1.604 2.210 0.623 1.383
0.001 1.329 0.008 1.076 0.456 0.011 0.981 1.380 1.166 0.609 0.139 0.829
0.942 0.276
0.239 0.829
0.171 1.383
0.603 1.936
0.449 2.489
tion is fixed to a certain value. Bias neuron helps in fixing the transfer function to a certain value.
Table 5 Optimum levels for input parameters for maximization of production rate. Input parameters
Output parameter
Mash feed size = 1.5 mm Steam temperature = 80 °C Feed rate = 16 tons/h Conditioning time = 32 s
Production rate = 13.82 tons/h
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 Pn i¼1 xi;net xi;act RMSE ¼ n 2 Pn x i;net xi;act R2 ¼ 1 Pi¼1 2 n i¼1 xi;act xav g
0.177 0.276
ð1Þ ð2Þ
where n is the number of actual observations, xi;net is the predicted output values from the network, xi;act is the actual output value from experiments and xav g is the mean of actual values from experimental observations. 3.3. Number of layers and neurons The number of layers and neurons should be kept at an optimum level to enhance the computational efficiency of the network and error reduction. For this case, 1 hidden layer and 1 output layer are fixed. The number of neurons is also set to 10. Thus, no variation occurs due to the number of layers and neurons for this problem. If the number of layers is increased, a proportionate increase of 10 neurons would take place which would enormously increase the computational effort of the network. So, the fixed number of layers and neurons would reduce the problem complexity to an extent. 3.4. Transfer function Transfer functions are input–output functions and 2 sets of transfer function combinations are compared to check for best results. Transfer functions are present in hidden layer–output layer. The combinations compared in this case are log sigmoid – pure linear and tan sigmoid – pure linear. A comprehensive neural network is shown in Fig. 2. Weights are present in the arcs connecting input–hidden and hidden–output layer. The weights are changed to reduce the error in all the iterations. Bias is the additional input of the neuron, which improves the neuron properties but increases complexity. The weights and the value in the transfer functions can be varied during the learning process of a neural network. This is not practical and would be better if only one of the variables should be varied. Thus, a bias neuron is utilized to solve this problem. The weights can be changed only if the transfer func-
4. Results and discussion Though different variables are present, every function does not fit suitably to the given data. For example, a training function might present with less error for one dataset. However, the same training function may not be precise and accurate for any other dataset. This leads us to a point where we need to compare the different parameters based on training and transfer functions, with the number of layers and neurons fixed. 4.1. Comparison of neural network parameters 10 different training functions are compared for 2 sets of transfer function (hidden–output layer) combinations – log sigmoid – pure linear and tan sigmoid – pure linear. It is shown in Tables 2 and 3. Epochs are set to 1000 and Iterations are set to 5 for all training functions. Two sets of transfer functions and training functions are compared and it is found that Polak–Ribiere conjugate gradient backpropagation training function with log sigmoid – pure linear transfer function combination provided best results with least RMSE and high positive R2 at satisfactory computation time. The weights and bias for the ideal neural network settings are shown in Table 4. These settings are used to simulate and find the optimum parameter levels for maximum production rate. In Table 4, s denotes the number of neurons in hidden layer and k denotes the number of input variables. b1(s) and b2 refer to the biases and IW(s,k) and W0(s) refer to weight of arcs of the neural network. 4.2. Prediction of optimum parameters Sample data is used to find out the levels of input variables which results in maximum production rate. The accuracy of the solution increases as the number of levels of each of these input variables becomes large. The optimum levels of input parameters are shown in Table 5. Thus, maximum production rate of 13.82 tons/h can be achieved in a feed mill by properly maintaining the input process variables. This would lead to significant cost savings as quality is increased and power consumption is decreased. However, it should be noted that the results obtained are recipe specific and these results should be used with caution. 5. Conclusion The optimization of process parameters for one recipe in feed manufacturing is a step in the right direction to improve the productivity of feed processes. The prediction of production rate using the neural networks for a specific recipe acts as a useful guideline
6
L. Sudha et al. / Computers and Electronics in Agriculture 120 (2016) 1–6
Table A1 Data on input parameters and the corresponding production rate. S. No.
Mash feed size (mm)
Steam temperature (degrees)
Feed rate (tons/h)
Conditioning time (s)
Production rate (tons/h)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5
80 81 60 81 80 80 81 81 80 42 81 81 81 62 81 81 81 60 80 81 82 82 80 80 81 81 80 48 80 81 65 80 80 80 80 60 80 80 60 80 60 83 83 84 60
16.5 17 5 16.5 16.5 17 17 17 17 5 17.5 17.5 17.5 6 17.5 8.4 8 5 8 8 8 8 8 8 8 8 8 4.5 9.6 9.8 5 14.2 14.1 14.2 14.5 5 14 14 5 14 5 14 13.5 13.2 4.5
33 34 32 32 34 35 35 32 32 34 33 33 32 34 33 32 32 33 34 32 33 34 33 33 32 32 32 33 32 32 32 33 34 33 32 32 32 32 32 33 34 33 33 34 33
12.4 12 4.6 13 13.2 12 13.4 11.8 12.2 3.7 13.6 12 13.2 4 12.2 6.7 7 3.8 6 5 5 6 6 6 6 6.3 6 3.2 7.8 8 3.7 11.6 12 11.8 12.3 4 10.4 10.2 3.6 11.2 4.1 12.4 12.7 12.1 3.2
for adoption in feed mills. Further, the work could be extended by incorporating other parameters that are significant in other operations apart from conditioning and pelleting to achieve a much truer prediction of the production rate. The scope of research in the feed manufacturing sector is not restricted to feed mill operations alone. Pathumnakul et al. (2011) focus on the need for integration
of information among the upstream and downstream members of the animal feed supply chain to reduce costs and produce better performance. Similarly, the study also analyzes the pros and cons of multiple period planning and its effect on costs. Thus, inventory management is one of the crucial management problems that show enough promise in the feed supply chain. Moreover, the current problem’s scope could be expanded by considering multiple objectives and integrating with other issues in the animal feed supply chain. The production rate settings could be synchronized and optimized along with the inventory to meet the required demand and avoid loss of material and money. Appendix A. Data collection See Table A1. References Agatonovic-Kustrin, S., Beresford, R., 2000. Basic concepts of artificial neural network (ANN) modeling and its application in pharmaceutical research. J. Pharm. Biomed. Anal. 22 (5), 717–727. Arora, G., Sehgal, V.K., Arora, M., 2007. Optimization of process parameters for milling of enzymatically pretreated Basmati rice. J. Food Eng. 82 (2), 153–159. Behnke, K.C., 1996. Feed manufacturing technology: current issues and challenges. Animal Feed Sci. Technol. 62 (1), 49–57. Chegini, G.R., Khazaei, J., Ghobadian, B., Goudarzi, A.M., 2008. Prediction of process and product parameters in an orange juice spray dryer using artificial neural networks. J. Food Eng. 84 (4), 534–543. Cus, F., Zuperl, U., 2006. Approach to optimization of cutting conditions by using artificial neural networks. J. Mater. Process. Technol. 173 (3), 281–290. Hamzaoui, Y.E., Rodríguez, J.A., Hernández, J.A., Salazar, V., 2015. Optimization of operating conditions for steam turbine using an artificial neural network inverse. Appl. Therm. Eng. 75, 648–657. Lamberts, L., De Bie, E., Vandeputte, G.E., Veraverbeke, W.S., Derycke, V., De Man, W., Delcour, J.A., 2007. Effect of milling on colour and nutritional properties of rice. Food Chem. 100 (4), 1496–1503. Oztop, M.H., Sahin, S., Sumnu, G., 2007. Optimization of microwave frying of potato slices by using Taguchi technique. J. Food Eng. 79 (1), 83–91. Pathumnakul, S., Piewthongngam, K., Apichottanakul, A., 2009. A neural network approach to the selection of feed mix in the feed industry. Comput. Electron. Agric. 68 (1), 18–24. Pathumnakul, S., Ittiphalin, M., Piewthongngam, K., Rujikietkumjorn, S., 2011. Should feed mills go beyond traditional least cost formulation? Comput. Electron. Agric. 75 (2), 243–249. Ramakrishnan, S., Wysk, R.A., Prabhu, V.V., 2004. Prediction of process parameters for intelligent control of tunnel freezers using simulation. J. Food Eng. 65 (1), 23–31. Roy, P., Ijiri, T., Okadome, H., Nei, D., Orikasa, T., Nakamura, N., Shiina, T., 2008. Effect of processing conditions on overall energy consumption and quality of rice (Oryza sativa L.). J. Food Eng. 89 (3), 343–348. Thomas, M.A.F.B., Van der Poel, A.F.B., 1996. Physical quality of pelleted animal feed 1. Criteria for pellet quality. Anim. Feed Sci. Technol. 61 (1), 89–112. Thomas, M., Van Zuilichem, D.J., Van der Poel, A.F.B., 1997. Physical quality of pelleted animal feed. 2. Contribution of processes and its conditions. Anim. Feed Sci. Technol. 64 (2), 173–192. Thomas, M., Van Vliet, T., Van der Poel, A.F.B., 1998. Physical quality of pelleted animal feed 3. Contribution of feedstuff components. Anim. Feed Sci. Technol. 70 (1), 59–78.