0728_ANN simulation of humic substance and membrane filtration



Title: Artificial neural network simulation of combined humic substance coagulation and membrane filtration 
Journal: Chemical Engineering
Authors: Mohammed Al-Abria and Nidal Hilala
Corresponding author: Nidal Hilal
aCentre for Clean Water Technologies, School of Chemical and Environmental Engineering, University of Nottingham, Nottingham NG7 2RD, UK

The original and creativity of paper: Membrane retention and fouling were investigated using backpropagation artificial neural network (BPNN) as a tool which provide very good result of prediction.

Artificial neural network (ANN) was success to predict membrane performance. The developed BPNN produced high reliability; R-value exceeded 0.95 in all results. The researchers found that a high number of training dataset caused depress the generalizing abilities of the ANN through over-fitting or memorization of the training data set, reducing the networks predictability. So, less number of training dataset was suggested to apply. In addition, number of neurons in hidden layers need to be chosen carefully to obtain a reliable network. 

Why artificial neural network (ANN) was selected in this research?
- Even it is a black box model but it capable to model highly complex and non-linear system with many interrelation parameters.
- It does not require detail parameter.
- It uses available data to predict the relationship between input and output parameters.

Types of ANN
1.    Feedback networks
-    It allows signal to travel in both direction by introducing loops in the network enabling them to route back to previous neural.
-    It is able used for classification problem with binary pattern vectors.
2.    Feed-forward networks
-    It allows signal to travel in one direction from input to output.
-    It tends to be straightforward networks that associate inputs with output.
-    Normally the network used with error correction algorithm such as backpropagation.

What is Backpropagation?
-    It relies on a search technique such as gradient descent.
-    Its learning process relies on the iterative weight.

How does the BPNN work?
1.    The working process start with comparison of  output and target values
2.    Then, modify weight values according to a specific learning algorithm to reduce overall error.
3.    After that, the modified weights are propagated backwards into the system.
4.    Step one to three are carried out for each set of training pattern used to compute the global error. There process will be repeated until the difference between predicted output and target values reach an accepted range.
Fig. 1 shows a typical backpropagation artificial neural network (BPNN)

Fig. 1. Architecture of a typical backpropagation artificial neural network (BPNN)

How does the ANN work?
1.    Assembly of dataset.
2.    ANN will calculate input using neurons in the hidden and output layers. In this case, it calculates by performing a weighted sum of the output they received from previous layer.
3.    Then, it will calculate output by transforming their inputs using a transfer function.
4.    Deciding the network architecture.
5.    Training (network learning)
6.    Simulating the network response to new input.

Most widely used transfer functions (Fig. 2)
1.    Log-sigmoid (logsig): this function produces output in the range of 0 to 1.
2.    Tan-sigmoid (tansig): this function produces output in the range of -1 to +1.
3.    Linear (purelin): this function produces output in the range of -∞ to +∞.

Fig. 2. Typical transfer functions used in BPNN: (A) logsig, (B) tansig and (C) purelin.

Application & further study: BPANN can be applied to the dataset that would like to see the interaction between input and output and also it is the suggested tool by several researchers for prediction reverse osmosis membrane performance.

By Monruedee Moonkhum
Email: moon@gist.ac.kr

첨부 (1)
133.6KB / Download 7