2015 15th International Conference on Control, Automation and Systems (ICCAS 2015) Oct. 13-16, 2015 in BEXCO, Busan, Korea
Artificial Neural Network-Based Model for Quality Estimation of Refined Palm Oil Nurul Sulaiha Sulaiman1 and Khairiyah Mohd Yusof2* 1
Department of Chemical Engineering, Faculty of Chemical Engineering, Universiti Teknologi Malaysia, Johor Bahru, 81310, Malaysia ([email protected]
) 2Centre for Engineering Education, Universiti Teknologi Malaysia, Johor Bahru, 81310, Malaysia ([email protected]
Abstract: The goal of this study is to develop an accurate artificial neural network (ANN)-based model to predict significant quality of refined palm oil which is Free Fatty Acid (FFA) content. The variables; FFA content, Iodine Value (IV), moisture content, bleaching earth and citric acid dosage as well as the pressure and temperature of the deodorizer is used to build the ANN prediction model. A feed forward neural network was designed using a back-propagation training algorithm. Comparison of ANN predicted result with industrial data was made. It is proven in this study that ANN can be used to estimate the quality of refined palm oil. Therefore, the model can be further implemented in palm oil refinery plant as the prediction system of the refined oil quality. Keywords: Artificial Neural Network, Prediction Model, Product Quality System, Palm Oil
2. PROCESS DESCRIPTION
There is a tremendous growth for ANN-based research and application over the past few years. ANN is a computational modeling tool that had been used in many areas for modeling complex real-world problems . ANN is getting popular among researchers due to its ability to learn from training data and engage what had been learned to categorize pattern from unseen test data . Many applications of ANN have appeared over the years for prediction and estimation. Arora and Srivastava  applied ANN in predicting cholesterol in market available edible oils. Morad et al.  explored the function of ANN to improve operating procedure of combined degumming and bleaching processes in palm oil refining. Palm oil is edible oil which has become a common ingredient in many consumer products. The crude palm oil need to undergo refining process to remove impurities and thus the major concern in the palm oil refining process is the refined oil quality . In current practice, the refined oil quality is manually monitored through lab testing which is time consuming. There is no system that can estimate the quality of the refined oil. There is no research on quality estimation in palm oil refining based on our literature survey. The use of ANN for predicting the refined oil quality was explored in this paper. A feed forward neural network was designed using a Levenberg-Marquardt (LM) training algorithm.
Refining process is vital in removing impurities like FFA, coloring pigments (chlorophyll and carotene), trace metal and other contaminants in the crude oil . During degumming and bleaching process, citric acid and bleaching earth were added to the hot crude oil. Citric acid coagulates impurities like phosphatides, chlorophyll, carotenoid, trace metal ions and oxidative product. Bleaching earth absorbed the coagulated impurities and later it removed by filter. The oil is then undergo deodorization process which is steam distillation process at 250–265oC under vacuum condition. Volatile components such as Free Fatty Acids (FFAs) and odoriferous pigments are distilled and removed as Palm Fatty Acid Distillate (PFAD). The three processes produced Refined, Bleached and Deodorized (RBD) palm oil . Figure 1 shows the overall process of palm oil refining. A good quality of RBD palm oil depends on the organoleptic parameters such as taste, odor and color . It also determined based on the quality of the crude oil . The RBD palm oil that does not meet the requirement will be labeled as off–specification product and need to be refined again. This proved that quality of the refined oil is very crucial in a palm oil refinery industry. In current industry, the most crucial refined oil qualities are FFA content and Lovibond color of the oil. However, this paper will focus on FFA quality only.
Bleaching earth Crude oil
vacuum Bleached oil
Bleached Oil tank 1 E-160
Pressure Leaf Filter
vacuum Bleached oil
recycle oil Polishing Filter 1
Bleached Oil Drop tank
Cyclone Crude oil Drop tank
Bleached Oil tank 2
Cold Fatty Acid
Fatty Acid scrubber
Anti-oxidant Citric acid Polishing Filter 2
Deodorizer RBD oil
Polishing Filter 3 RBD oil
Fig.1. Overall Palm Oil Refining Process In this network, data flow in a direct forward direction which is from the input layer to the output layer. The inputs, hidden and output layers were connected by weights (w) and biases (b).
3. MODEL DEVELOPMENT 3.1 Artificial Neural Network ANN is derived from human brain. ANN consists of some neurons called nodes which are connected via weights. Each nodes arranged in parallel layer receives data from behind nodes, adds it and passes the data through a nonlinear function, and then propagates data to the next nodes . ANN modeling follows these steps: 1. Database collection Selection of the input and the output Analysis and pre–processing of the data 2. Training the neural network Selection of architecture, training functions, training algorithms and parameters of the network 3. Testing the goodness of the model 4. Comparing both predicted output and actual output. The model developed here has implemented these steps.
3.3 Data Preprocessing and Generalization A total number of 66 data sets were collected during industrial attachment at one palm oil refinery in Pasir Gudang, Malaysia. The data were pre-processed. The input and the output data were normalized within a uniform range of [0–1] according to the below equation: (1) Where x is variable, xmax is the maximum value and xmin is the minimum value. Over-fitting is one of the common problems that occur during training of a neural network. An overtrained network shows large error on unseen data. To overcome this problem, the early stopping technique was implemented. The data were divided into three parts; training, testing and validation data set. This technique continuously monitors the validation error and stops the training if validation error begins to rise .
3.2 Input and Output Data Selection The network consists of an input layer, one hidden layer and an output layer. The inputs to the network are FFA content in crude oil, moisture, IV, bleaching earth dosage, citric acid dosage, deodorizer pressure and temperature. Output is FFA content in RBD palm oil.
3.4 Training and Testing Neural Network One of the critical tasks in developing neural network is determining the network architecture. This is because the complexity of the relationship between input
and output is relative to the number of hidden neurons . To tackle the problem, the training of the network is started with the smallest possible network with one hidden neuron in the hidden layer. Then the number of neuron was increased to improve the performance. The number of neurons had varied from 1 to 20. The best number of hidden nodes for the model is determined by the smallest MSE. There are numerous types of training algorithms and back-propagation (BP) method is one the most employed classes of training algorithms for feed-forward neural network. By using BP algorithm, the weights and biases are iteratively updated using LM algorithm until the convergence to the certain value is achieved . In this study, logsig was used as a transfer function between input and hidden layer, while purelin was used as a transfer function between hidden and output layer. To calculate the performance of the network, error function based on the predicted output and actual output is carried on . The commonly used error function the mean squared error (MSE) was employed in this work which is defined as follow: ∑
Fig.2. ANN architecture for FFA prediction In Fig. 3, the evaluation of the MSE during training, validation and testing phase is plotted. The best epoch for validation check was found at 19.
Where Yt is the target output, YN is the predicted output and N is the number of points.
4. RESULT AND DISCUSSION The network architecture or also known as topology gave essential influences on the predicted results. The number of input and output nodes is equal to the number of input and output data respectively. By training several ANN topologies and varied the number of hidden nodes, the optimal network topology is determined based on the minimum MSE of the network. Therefore, the optimal topology of the ANN developed in this study is a feed forward neural network with 8 inputs, one hidden layer with 4 nodes and one output layer including single node. This feed forward network topology can be noted as ANN (8:4:1). The optimal network topology of the ANN model in this study is shown in Fig. 2.
Fig.3. MSE of training, validation and testing phase The summarized MSE and determination coefficient (R2) for training, validation, testing and all data sets are presented in Table 1 while Table 2 listed the predicted and actual data for some training, validation and testing data from all three phases. The ANN is trained by using LM algorithm. This algorithm is very well fitted to ANN and it has advantages of faster convergency .
Table 1 Statistical measures and performance of the ANN model for training, testing, validation and all data The best architecture 8:4:1
R2 Testing Validation 0.7637 0.8277
Training 0. 009557
ANN Predicted of FFA %
The value of MSE for all three phases is very small which are 0.009557, 0.0083811 and 0.005526 respectively. Although the R2 for all data is 0.6578, the R2 for all the three phases is in a satisfactory range. Table 2 Predicted and Actual FFA % No.
FFA % in RBD Palm Oil Predicted Actual 1a 0. 024 0. 025 2a 0. 022 0. 023 3a 0. 028 0. 029 4a 0. 020 0. 020 5a 0. 019 0. 016 6a 0. 019 0. 015 7a 0. 019 0. 018 8a 0. 019 0. 016 9a 0. 018 0. 019 10a 0. 033 0. 033 11a 0. 023 0. 023 12a 0. 029 0. 035 13b 0. 028 0. 028 14b 0. 029 0. 025 15b 0. 018 0. 016 16b 0. 029 0. 029 17c 0. 029 0. 028 18c 0. 019 0. 019 19c 0. 019 0. 017 20c 0. 019 0. 023 a Training data b Validation data c Testing data
MSE Testing Validation 0. 0083811 0. 005526
0.080 0.070 0.060 0.050 0.040 0.030 0.020 0.010 0.000 0.000
All 0. 009084
R² = 0.9553
0.020 0.040 0.060 Actual FFA %
Fig. 4 Scatter plot of the ANN predicted versus actual FFA%
5. CONCLUSION In this study, a three layer ANN model was developed. Logsig and purelin transfer functions were used at the hidden and output layer respectively. The ANN was trained by using LM algorithm and the result showed that ANN with 4 nodes in the hidden layer had the best performance. Therefore, the optimal topology of the network is 8 inputs, one hidden layer with 4 nodes and one output layer (8:4:1). The prediction results were fairly good and high degree of accuracy was obtained.
From Table 2 above, the results were fairly good. The developed ANN-based model is able to give an approximately accurate prediction of the FFA % in the refined oil. The network is capable to recognize the pattern of unseen data from testing phase. In Fig. 4, the predicted data versus actual data for training, validation and test data are plotted. This figure shows the goodness prediction of the network. The predicted and actual data nearly fit each other which mean a high degree of accuracy was obtained.
The authors would like to thank MyBrain15 Malaysia, Universiti Teknologi Malaysia and Ministry of Education Malaysia for funding and supporting this research. This work was also supported in part by the Flagship grant under grant no. Q.J130000.2409.02G57.
A. K Bhatt., D. Pant and R. Singh, “An analysis of the performance of Artificial Neural Network technique for apple classification”, AI & Soc, Vol. 29, pp. 103–111, 2014. F. Ahmad, N. A. Mat-Isa, Z. Hussain, R. Boudville, and M. K. Osman, “Genetic Algorithm - Artificial Neural Network (GAANN) Hybrid Intelligence for Cancer Diagnosis”, Second International Conference on Computational Intelligence, Communication Systems and Networks, pp. 78-83, 2010.
J. K. Arora and M. M. Srivastava, “Artificial Neural Network Modeling & Standardization of HPTLC Method for the Estimation of Cholesterol in Edible Oils”, Proceeding of the World Congress on Engineeing 2011, 2011. N. A. Morad, R. Mohd-Zin, K. Mohd-Yusof, and M. K. Abdul-Aziz, “Process Modelling of Combined Degumming and Bleaching in Palm Oil Refining Using Artificial Neural Network”, JAOCS Journal of the American Oil Chemists’ Society, Vol. 87, No. 11, pp. 1381-1388. P. T. Gee, “Analytical Characteristic of Crude and Refined Palm Oil and Fractions”, Eur. J. Lipid Sci. Technol., Vol. 19, pp. 373–379, 2007. H. Haslenda and M. Z. Jamaludin, “Industry to Industry By-products Exchange Network Towards Zero Waste in Palm Oil Refining Processes”, Resources, Conservation and erecycling, Vol. 55, No. 7, pp. 713-718. V. Gibon , W. De-Greyt, and M. Kellens, “Palm oil refining”, European Journal of Lipid Science and Technology, Vol. 109, No. 4, pp. 315-335, 2007. W. F. De Greyt, M. J. Kellens and A. D. Huyghebaert, “Effect of Physical Refining on Selected Minor Components in Vegetable Oils”, Lipid Fett, Vol 10, No. 11, pp. 428–432, 1999. M. Rossi, M. Gianazza, C. Alamprese, and F. Stanga, “The Effect of Bleaching and Physical Refining on Color and Minor Components of Palm Oil”, JAOCS, Vol. 78, No. 10, pp. 10511055, 2001. Akay M.F, Mohommoud–Zayid E. I., Aktürk E. and George J. D., “Artificial neural networkbased model for predicting VO2 max from a submaximal exercise test”, Expert Systems with Applications, Vol. 38, pp. 2007–2010, 2011. Y. Yasin, Fa. H-Ahmad, M. GhaffariMoghaddam, M. Khajeh, “Application of a hybrid artificial neural network–genetic algorithm approach to optimize the lead ions removal from aqueous solutions using intercalated tartrate-Mg–Al layered double hydroxides”, Environmental Nanotechnology, Monitoring & Management, Vol. 1, No. 2, pp. 2– 7, 2014. N. Qazi and H. Yeung, “Modeling of gas–liquid separation through stacked neural network”, Asia-Pacific Journal of Chemical Engineering, Vol. 9, pp. 490–497, 2014. M. Khayet and C. Cojocaru, “Artificial neural network modeling and optimization of desalination by air gap membrane distillation”, Separation and Purification Technology, Vol.
86, pp. 171–182, 2012.  A.R. Khataee and M.B. Kasiri, “Artificial neural networks modeling of contaminated water treatment processes by homogeneous and heterogeneous nanocatalysis”, Journal of Molecular Catalysis A: Chemical, Vol. 331, pp. 86–100, 2010.  B. Zhao and Y. Su, “Artificial neural networkbased modeling of pressure drop coefficient for cyclone separators” chemical engineering research and design, Vol. 88, pp. 606–613, 2010.