Study Spray Draying Modeling by Artifitual Neural Network

Study Spray Draying Modeling by Artifitual Neural Network

A. S. Ghasemi , A. Siami, M. Lotfi

Dept. Chemistry, Payame Noor University Payame Noor University

Tehran, Iran

Abstractconsidering the large Caspian Sea water source have caused to propose a plan for the salt production from salt water by spray dryer. 300 ml salt water with %8 concentrate has been used during 56 experiments. The experiments have done with

%35 efficiency for water product and %75 of that for the humid salt. Also, the data have been used for Artifitual Nerual Network (ANN) modeling and has been found in well agreement between the experiments result and predicted one by ANN.

Keywords spray dryer; moisture; salt water; Artifitual Nerual Network

1. INTRODUCTION

   Drying is the oldest method of preserving food [1]. The main purpose of drying is to allow longer periods of storage, minimize packaging requirements and reduce shipping weights [2]. Open-air sun drying has been used since ancient times to dry grain, vegetables, fruits and other agricultural products. However, this method of drying is not always suited to large-scale production. There are some problems, the most important ones are: lack of ability to control the drying operation properly, the length of the drying time, weather uncertainties, high labor costs, large area requirement, insect infestation, mixing with dust and other foreign materials and so on [3].The industrial drying processes have been established to overcome these problems [4]. Drying is defined as a process of moisture removal due to simultaneous heat and mass transfer [5]. This complicated process depends on different factors such as air temperature and velocity, relative humidity of air, air flow rate, physical nature and initial moisture content of the drying material, exposed area and pressure [6]. Knowledge of drying behavior is important in the design, simulation and optimization of drying process [7]. Drying behavior of different natural materials has been studied by several investigators [5,8]. Spray dryers are used to obtain a dry powder from a liquid feed. Although the process equipment is very bulky and operation is expensive, it is an ideal process for drying heat sensitive materials. Spray dryers have been used for nearly a century now, but it is still very difficult to model the performance of this type of process equipment, especially with respect to the quality of the dried product [9]. Some years ago, Bahu (1992) made the question of whether enough modeling of spray dryers had actually been done or not [10]. The rapid expansion of interest in spray

   drying research appears to indicate that there is still much to do [11]. At first sight, it might appear that the spray drying process is a simple one, with a liquid solution or slurry being fed through an atomizing device into a hot gas stream, with dried particles then being separated from the gas. However, the complexity of the process is that it includes many aspects of fluid mechanics, heat and mass-transfer, particle technology, reaction engineering, and materials science, in other words a large part of chemical, mechanical and process engineering [12]. The analysis of the process may also include many aspects of numerical analysis and mathematical modeling [13]. Spray drying consists of atomizing a solution into liquid drops in a hot air flow to get dry solid particles after solvent evaporation. The convective drying at the drop surface leads to a very fast evolution of temperature and water content due to initial high differences of temperature and water vapor pressure between the drop surface and the drying air [14]. One of the problems often encountered in the operation of spray dryers is wall deposition, which potentially causes fire hazards, reduced product quality and lower production efficiency. The characteristics of airflow significantly affect the drying process of atomized droplets, where recirculation of the droplets affects their drying times and histories. Understanding the features of the airflow inside a spray dryer is essential for the investigation of the resulting complicated droplet trajectories [15]. Several researchers have also developed simulation models for drying processes [16- 18]. Artificial neural networks (ANN) models have been successfully used in the prediction of problems in bio- processing and chemical engineering [19]. Artificial neural network (ANN) is a general non-linear model based on a simplified model of human brain function and this technique is particularly useful when a phenomenological model of a process is not available or would be too far complex. Several studies have been reported ANN on modeling of drying [20- 24]. ANN technique has also been applied for modeling of water sorption isotherms of black tea [25] and corn starch [26] and the ANN models were found to be better than other mathematical models [27].

2. MATERIAL AND METHODS

   1. Experiments with Spray drying

      We used an experimental spray dryer for the separation. In the dryer, the feed flow through pumped to the reduced (low) atomizer and then a chamber dryer is driven. A magnetic stirrer and a magnet for feed to remain consistent and not precipitated salt in a measuring cylinder is used during testing. With the help of air compressor inlet air flow ranging Debye 1.3 to 4 m3/min, which is directed into the dryer chamber. Dryer in a room with a temperature of about 25 ° C was constant environmental conditions. In this study, the salt water (use the unrefined salt) at a concentration of 8% and no additives are used.56 tests in three Debye different feed flow and different temperature of inlet air flow is done, that the table (1) is given. During experiments, a magnetic stirrer is used to maintain uniform concentration of the feed brine. In each experiment, 300 ml of brine was

      used, it can be stated that an average of about 25% feed in each test, of liquid products were collected. Also, an average of 75% of salt in the input feed was seen as a solid product. Part of the water contained in the feed as moisture with salt output out of the system and much of it was evacuated with exhaust air from the dryer, with the placing a cooling metal plate on the front of the vapor output, we were able to collect some of the water vapor condenses and in each experiment, approximately 35% of the feed liquid product was collected.

      In some experiments, the salt stick to the walls of the dryer due to high temperature or high Debye air flow, no salt product but the amount of liquid water output increased. Parts of the ultrafine particles of salt with exhaust air were removed from the tank. In each test sample of salt with polymeric tubes insulator heat and moisture collected and hygrometry were tested in the laboratory.

      The amount of salt water just to gather information for the simulation of neural network was measured. The results are visible in Table (2).

      TABLE (1): Details the conditions tested

      Number of tests	Temperature (oC)	feed Debye (m3/min)	air Debye (m3/min)	Number of tests	Temperature (oC)	feed Debye (m3/min)	air Debye (m3/min)
      1	150	8.64	1.4	29	180	13.95	3
      2	150	8.64	1.8	30	210	13.95	1.4
      3	150	8.64	2.2	31	210	13.95	1.8
      4	150	8.64	2.6	32	210	13.95	2.2
      5	150	8.64	3	33	210	13.95	2.6
      6	150	8.64	3.4	34	210	13.95	3
      7	180	8.64	1.4	35	240	13.95	1.4
      8	180	8.64	1.8	36	240	13.95	1.8
      9	180	8.64	2.2	37	240	13.95	2.2
      10	180	8.64	2.6	38	240	13.95	2.6
      11	180	8.64	3	39	240	13.95	3
      12	210	8.64	1.4	40	150	13.95	1.4
      13	210	8.64	1.8	41	150	22.1	3
      14	210	8.64	2.2	42	180	22.1	1.4
      15	210	8.64	2.6	43	180	22.1	1.8
      16	210	8.64	3	44	180	22.1	2.2
      17	240	8.64	1.4	45	180	22.1	2.6
      18	240	8.64	1.8	46	180	22.1	3
      19	240	8.64	2.2	47	210	22.1	1.4
      20	240	8.64	2.6	48	210	22.1	1.8
      21	240	8.64	3	49	210	22.1	2.2
      22	150	13.95	1.4	50	210	22.1	2.6
      23	150	13.95	2.6	51	210	22.1	3
      24	150	13.95	3	52	240	22.1	1.4
      25	180	13.95	1.4	53	240	22.1	1.8
      26	180	13.95	1.8	54	240	22.1	2.2
      27	180	13.95	2.2	55	240	22.1	2.6
      28	180	13.95	2.6	56	240	22.1	3

      TABLE (2): Results of measuring moisture content of salt in each test

      Number of tests	Percent of Moisture	Number of tests	Percent of Moisture	Number of tests	Percent of Moisture	Number of tests	Percent of Moisture
      1	2.15	15	1.14	29	2.07	43	3.01
      2	2.31	16	1.05	30	3	44	3.46
      3	1.34	17	1.43	31	2.56	45	6.54
      4	1.27	18	2.19	32	2.24	46	12.06
      5	1.93	19	1.26	33	1.9	47	10.22
      6	1.11	20	1.49	34	2.39	48	5.26
      7	1.36	21	1.34	35	1.39	49	2.35
      8	3.12	22	4.79	36	1.49	50	1.60
      9	1.85	23	9.29	37	1.51	51	1.35
      10	2.11	24	6.2	38	3.27	52	2.64
      11	1.96	25	2.82	39	5.61	53	4.05
      12	0.99	26	3.26	40	4.18	54	2.83
      13	2.9	27	4.53	41	7.03	55	2.13
      14	1.41	28	3.13	42	1.93	56	1.90

   2. The characteristics of neural network and input data.

      56 tests were performed and in total 224 data for four sets were collected. Initially, all data that including parameters: air temperature, air Debye, feed Debye and the moisture in the form of dimensionless numbers were changed. Sample of dimensionless at the temperature of 210

      ° C is expressed in the following calculations. Thus, all parameters were in the range of zero to one. This has increased the speed of network convergence.

      Maximum temperature interval is 90 ° C.

      Minimum temperature is 150 ° C.

      Dimensionless temperature at 210 ° C. (210-150)/90= 0.666

      Experimental data were divided randomly into two groups. one group for training and another group for testing. Overall, 70% of data were used for training. In fact, the number of neurons depends on the number of independent variables in the input, and dependent variables in the output. In these experiments, the dependent variable (the amount of salt water output) and three independent variables (air temperature, air Debye and brine inlet Debye) and the one- and three neurons were assigned to the input and output layers. The number of hidden layers and their neurons are dependent on the complexity of the problem. Leading (Ahead) network and its input data is descried in the following tables.

      TABLE (3) Input data and dimensionless of them

      Number of Tests	Temperature (oC)	Dimensionless temperature	feed Debye (m3/min)	Dimensionless feed Debye	air Debye (m3/min)	Dimensionless air Debye	Percent Of Moisture	Dimensionless Moisture
      1	150	0	8.64	0	1.4	0	2.15	0.104
      2	150	0	8.64	0	1.8	0.2	2.31	0.119
      3	150	0	8.64	0	2.2	0.4	1.34	0.031
      4	150	0	8.64	0	2.6	0.6	1.27	0.025
      5	150	0	8.64	0	3	0.8	1.93	0.084
      6	150	0	8.64	0	3.4	1	1.11	0.010
      7	180	0.333	8.64	0	1.4	0	1.36	0.033
      8	180	0.333	8.64	0	1.8	0.2	3.12	0.192
      9	180	0.333	8.64	0	2.2	0.4	1.85	0.077
      10	180	0.333	8.64	0	2.6	0.6	2.11	0.010
      11	180	0.333	8.64	0	3	0.8	1.96	0.087
      12	210	0.666	8.64	0	1.4	0	0.99	0
      13	210	0.666	8.64	0	1.8	0.2	2.9	0.172
      14	210	0.666	8.64	0	2.2	0.4	1.41	0.037
      15	210	0.666	8.64	0	2.6	0.6	1.14	0.013
      16	210	0.666	8.64	0	3	0.8	1.05	0.005
      17	240	1	8.64	0	1.4	0	1.43	0.039
      18	240	1	8.64	0	1.8	0.2	2.19	0.108
      19	240	1	8.64	0	2.2	0.4	1.26	0.024
      20	240	1	8.64	0	2.6	0.6	1.49	0.045

      21	240	1	8.64	0	3	0.8	1.34	0.031
      22	150	0	13.95	0.394	1.4	0	4.79	0.343
      23	150	0	13.95	0.394	2.6	0.6	9.29	0.749
      24	150	0	13.95	0.394	3	0.8	6.2	0.470
      25	180	0.333	13.95	0.394	1.4	0	2.82	0.165
      26	180	0.333	13.95	0.394	1.8	0.2	3.26	0.205
      27	180	0.333	13.95	0.394	2.2	0.4	4.53	0.319
      28	180	0.333	13.95	0.394	2.6	0.6	3.13	0.193
      29	180	0.333	13.95	0.394	3	0.8	2.07	0.097
      30	210	0.666	13.95	0.394	1.4	0	3	0.181
      31	210	0.666	13.95	0.394	1.8	0.2	2.56	0.141
      32	210	0.666	13.95	0.394	2.2	0.4	2.24	0.112
      33	210	0.666	13.95	0.394	2.6	0.6	1.9	0.082
      34	210	0.666	13.95	0.394	3	0.8	2.39	0.126
      35	240	1	13.95	0.394	1.4	0	1.39	0.036
      36	240	1	13.95	0.394	1.8	0.2	1.49	0.045
      37	240	1	13.95	0.394	2.2	0.4	1.51	0.046
      38	240	1	13.95	0.394	2.6	0.6	3.27	0.205
      39	240	1	13.95	0.394	3	0.8	5.61	0.417
      40	150	0	13.95	0.394	1.4	0	4.18	0.288
      41	150	0	22.1	1	3	0.8	7.03	0.545
      42	180	0.333	22.1	1	1.4	0	1.93	0.084
      43	180	0.333	22.1	1	1.8	0.2	3.01	0.182
      44	180	0.333	22.1	1	2.2	0.4	3.46	0.223
      45	180	0.333	22.1	1	2.6	0.6	6.54	0.501
      46	180	0.333	22.1	1	3	0.8	12.06	1
      47	210	0.666	22.1	1	1.4	00	10.22	0.8330
      48	210	0.666	22.1	1	1.8	0.2	5.26	0.385
      49	210	0.666	22.1	1	2.2	0.4	2.35	0.122
      50	210	0.666	22.1	1	2.6	0.6	1.6	0.055
      51	210	0.666	22.1	1	3	0.8	1.35	0.032
      52	240	1	22.1	1	1.4	0	2.64	0.149
      53	240	1	22.1	1	1.8	0.2	4.05	0.276
      54	240	1	22.1	1	2.2	0.4	2.83	0.166
      55	240	1	22.1	1	2.6	0.6	2.13	0.102
      56	240	1	22.1	1	3	0.8	1.9	0.082

      TABLE (4): Percent of training data

      Train	70%
      Valid	15%
      Test	15%

   3. Simulation output data.

   In this study, many hidden layers were examined, and finally, a hidden layer of 18 neurons was suggested as the

   best mode. To find the best state, at each step of the regression and least square error was evaluated. The case that the higher regression and least-square error is less, Conditions is desirable.

   MSE	0.00482
   Regression	0.94451

   TABLE (5): least squared error and the final regression

   TABLE (6): humidity was calculated by the software

   Num	ANN result for humidity	Num	ANN result for humidity	Num	ANN result for humidity	Num	ANN result for humidity
   1	0.053	15	0.031	29	0.186	43	0.229
   2	0.030	16	0.005	30	0.099	44	0.279
   3	0.061	17	0.015	31	0.128	45	0.414
   4	0.071	18	0.023	32	0.104	46	0.625
   5	0.059	19	0.024	33	0.102	47	0.559
   6	0.059	20	0.023	34	0.112	48	0.336
   7	0.003	21	0.032	35	0.079	49	0.175
   8	0.105	22	0.294	36	0.069	50	0.112
   9	0.095	23	0.565	37	0.098	51	0.167
   10	0.047	24	0.6	38	0.186	52	0.205
   11	0.037	25	0.172	39	0.294	53	0.201
   12	0.052	26	0.219	40	0.294	54	0.159
   13	0.084	27	0.229	41	0.570	55	0.112
   14	0.061	28	0.195	42	0.167	56	0.150

3. Discussion and Conclusion

During the experiments, in areas where moisture was lower, more salt and more liquid product was collected. Namely, at higher Debye due to the shorter contact time feed with hot air, less salt and more humid were obtained. According to tests done in four temperature ranges, to compare the efficiency of the dryer, considering the low moisture in the solid salt product is more desirable, moisture curves are plotted (graphs (1) to (4)).

The average moisture content of the product, graphs (3) and

1. are very similar and in both graphs, about 87% experiments show moisture below 20%.

   Sudden maximum points in each of the graphs are due to an increase in both the feed Debye and hot air Debye, or are only due to an increase in feed Debye.

   It can be stated that the increase in feed Debye is due to liquid surface contact time with the hot air and thus reduces mass transfer and heat and increases the moisture content.

   Graph (1): Moisture content at 150 ° C.

   Graph (2): Moisture content at 180 ° C.

   Graph (3): Moisture content at 210 ° C.

   Graph (4): Moisture content at 240 ° C.

   The results for each of the three areas identified for feed Debye are shown as diagrams (5) to (7).

   Graph: (5) Moisture in Debye 64/8 (m3/min).

   Graph: (6) Moisture in Debye 13.95 (m3/min).

   Graph (7): Moisture in Debye 22.1 (m3/min).

   In the diagram (8) measurement of moisture content during 56 tests were compared with the values predicted by the software, as it was shown, there is a good agreement.

   Graph (8): Measuring Moisture in the 56 experiments.

   Experiments 46, 23 and 41 respectively show the greatest amount of salt moisture, and are related to conditions that there are almost simultaneously the high feed Debye and the air Debye.

   The best state is diagram (5), that there are humidity under 20% to almost 100% of the samples, and test 12 with low feed Debye and air Debye in 210 ° C temperature, has the

   best conditions during testing. In this case, there is sufficient time for heat and mass transfer.

   The lower simulation lines than experimental lines in graph

2. is partly due to the openness of the system, low feed Debye and also measurement errors.

This networks solve problems that simulation is difficult through logical, analytical techniques and advanced systems.

REFERENCES

[1]- Lewicki, P.P., 2005. Design of hot air drying for better foods. Trends Food Sci. Technol. XX, 111.

[2] -Simal, S., Femen´a, A., Llull, P., Rossell´o, C., 2000. Dehydration of aloe vera: simulation of drying curves and evaluation of functional properties. J. Food Eng. 43, 109114.

[3]- Togrul, I.T., Pehlivan, D., 2004. Modeling of thin layer drying kinetics of some fruits under open-air sun drying process. J. Food Eng. 65, 413425.

[4]- Kamyar Movagharnejad, Maryam Nikzad, 2007. Modeling of tomato drying using artificial neural network.J. Computers and Electronics in Agriculture .59, 7885.

[5]- G¨og¨us, F., 1994. The effect of movement of solutes on millard reaction during drying. Ph.D. Thesis. Leeds University, Leeds.

[6]- Akpinar, E.K., Bicer, Y., Yildiz, C., 2003. Thin layer drying of red pepper. J. Food Eng. 59, 99104.

[7]- Senadeera, W., Bhandari, B.R., Young, G., Wijesinghe, B., 2003. Influence of shapes of selected vegetable materials on drying kinetics during fluidized bed drying. J. Food Eng. 58, 277283.

[8]- Midilli, A., Olgun, H., Ayhan, T., 1999. Experimental studies of mushroom and pollen drying. Int. J. Energy Res. 23, 11431152.

[9]- F. G. Kieviet, J. Vanraaij, P. P. E. A. De moor and P. J. A. M. Kerkhof, 1997. Measurement and modelling of the air flow pattern in a pilot-plant spray dryer.J. Institution of Chemical Engineers.75,321- 328.

[10]- Bahu, R.E., 1992. Spray drying maturity or opportunities? In: Mujumdar, A.S. (Ed.),

Drying92, Proceedings of the Eighth International Drying Symposium, IDS92, 25 August, vol. A. Elsevier, Amsterdam, pp. 7491.

[11]-Mujumdar, A.S., Huang, L.X., 2007. Global R&D needs in drying.J. Drying Technology

25 (4), 647658.

[12]-Chen, X.D., 2004. Heat-mass transfer and structure formation during drying of

single food droplets.J. Drying Technology 22 (1&2), 179190.

[13]-T.A.G. Langrish, (2009). Multi-scale mathematical modelling of spray dryers. Journal of Food Engineering 93, 218228.

[14]- C. Turchiuli, A. Gianfrancesco, S. Palzer, E. Dumoulin, (2011). Evolution of particle properties during spray drying in relation with stickiness and agglomeration control.J. Powder Technology 208, 433440.

[15]- B. Guo, T. A. G. Langrish and D. F. Fletcher,( 2003). Simulation of gas flow instability in a spray dryer. J. Trans IChemE, Part A,81,631- 638.

[16]- Diamante, L.M., Munro, P.A., 1991. Mathematical modeling of hot air drying of sweet potato slices. Int. J. Food Sci. Technol. 26, 99 109.

[17]-Diamante, L.M., Munro, P.A., 1993. Mathematical modeling of thin layer solar drying of sweet potato slices. Sol. Energy 51, 271276.

[18]- Ratti, C., Mujumdar, A.S., 1997. Solar drying of foods: modeling and numerical simulation. Sol. Energy 60, 151157.

[19]- Baughman, D.R., Liu, Y.A., 1995. Neural Networks in Bio- Processing and Chemical Engineering. Academic Press, New York.

[20]- Huang, B., Mujumdar, A.S., 1993. Use of neural network to predict industrial dryer performance. Drying Technology 11, 525541.

[21]- Bala, B.K., Ashraf, M.A., Uddin, M.A., Janjai, S., 2005.

Experimental and neural network

prediction of the performance of the solar tunnel drier for drying of jackfruit bulbs and jackfruit leather. Journal of Food Process Engineering 28, 552566.

[22]- Poonnoy, P., Tansakul, A., Chinnan, M., 2007. Artificial neural network modeling for temperature and moisture content prediction in tomato slices undergoing microwave-vacuum drying. Journal of Food Science 72 (1), E42E47.

[23]- Lertworasirikul, S., Tipsuwan, Y., 2008. Moisture content and water activity prediction of semi-finished cassava crackers from drying process with artificial neural network. Journal of Food Engineering 84, 6574.

[24]- Sablani, S.S., Ramaswamy, H.S., Parasher, S.O.A., 1995. Network approach for thermal processing applications. Journal of Food Processing and Preservation 19, 283301.

[25]- Panchariya, P.C., Popovic, D., Sharma, A.L., 2002. Desertion isotherm modeling of black

tea using artificial neural networks. Drying Technology 20, 351362. [26]- Peng, G., Chen, X., Wu, W., Jiang, X., 2007. Modeling of water

sorption isotherm for corn

starch. Journal of Food Engineering 80, 562567.

[27]- S. Janjai, P. Intaweea, K. Tohsinga, B. Mahayotheeb, B.K. Balac,

M.A. Ashraf d, J. Müllere , 2009. Neural network modeling of sorption isotherms of longan(Dimocarpus longan Lour.).J. Computers and Electronics in Agriculture66, 209214.