DESIGN OF NEURONAL NETWORK TO CONTROL SPIRULINA

AQUACULTURE

Ernesto Ponce, Claudio Ponce and Bernardo Barraza

Mechanical Engineering Department, Electronic Engineering Department, Tarapac´a University

18 de Septiembre 2222, Casilla 6-D, Arica, Chile

Keywords:

Neural Network, Spirulina, Control, Aquaculture, Signal Processing System.

Abstract:

A neural network that was designed to control a Spirulina aquaculture process in a pilot plant in the north of

Chile, is presented in this work. Spirulina is a super food, but is a delicate alga and its culture may be suddenly

lost by rapid changes in the weather that can affect its temperature, salinity or pH. The neural network control

system presented is complex and non linear, and has several variables. The previous automatic control system

for the plant proved unable to cope with large climatic variations. The advantage of this new method is the

improvement in efﬁciency of the process, and a reliable control system that is able to adapt to climatic changes.

The future application of this work is related to the industrial production of food and fuel from micro algae

culture, for the growing world population.

1 INTRODUCTION

Spirulina Platensis is a single cell micro alga that be-

longs to the cianobacteria group. It has a blue-green

colour and spiral shape and reproduces by intracellu-

lar rupture. Its length varies between 20 and 50 mi-

crons. Spirulina can survive at temperatures between

◦

and 33

◦

C and at a pH between 8.5 and 10.5. This

micro alga is 60% all-vegetable protein, rich in beta

carotene, iron, vitamin B-12 and the rare essential

fatty acid, GLA. It is considered the super food of

the future (Henrikson, 1994), (Spirulina.com, 2004),

(Jourdan, 2002).

In 2002 a sudden weather change destroyed the

Spirulina culture in a pilot plant (raceway system) in

Azapa Valley, Arica, Chile, by causing variations in

density, salinity, pH and temperature. The classic au-

tomatic control in place could not manage the changes

and the culture was lost. It was therefore necessary to

design an intelligent control system able to adapt to

adverse weather changes.

In Chilean modern water farms a complete system

that controls Spirulina alga cultivation does not exist.

The control system designed and presented here was

based on a neuronal network. The variables are pH,

temperature, salinity (electrical conductivity, directly

related to the density of the solution) and population

density. The independentvariables of the system were

entered. The exit of the network was forced with the

evolution state of the culture in time. This proce-

dure was repeated several times having changed the

input variables (Hagan et al., 2002). Once the control

system was in place, the neural network continued to

learn and evolve.

2 THEORY

2.1 Neuronal Networks

Neural networks, inspired by biological nervous sys-

tems, are composed of simple elements that operate in

parallel (Figure 1). As in nature, the network function

is determined by the connections between elements.

A neural network can be ”trained” to perform a par-

ticular function by adjusting the values of the connec-

tions between elements (Hunt et al., 2002). Thus neu-

ral networks can be trained to solve problems that are

difﬁcult for conventional computers or human beings.

Moreover, they can incorporate the best techniques

for pattern recognition and tendency analysis.

289

Ponce E., Ponce C. and Barraza B. (2008).

DESIGN OF NEURONAL NETWORK TO CONTROL SPIRULINA AQUACULTURE.

In Proceedings of the Fifth International Conference on Informatics in Control, Automation and Robotics - ICSO, pages 289-292

DOI: 10.5220/0001482002890292

 SciTePress

Figure 1: Basic model of a neuronal network.

For applications that demand networks with fewer

than 100 neurons and little training, the software im-

plementations are sufﬁcient. When the problem re-

quires over 100 neurons and 10000 synapses, is nec-

essary to use hardware.

3 DEVELOPMENT

- Design of neuronal network

- Method

- Network training

- Operation

3.1 Design of the Neuronal Network

The variables of the system are:

• pH: This variable is very important because the

alga survives within a range of pH between 8.5

and 10.5. pH can be measured with a pH meter.

If the pH is low, then a valve connected with a

little tank, is open to give a bicarbonate sodium

solution. If it is high, another valve gives CO2 to

a dome on the surface aquaculture.

• Temperature: Spirulina Platensis can tolerate tem-

peratures between 13

◦

to 33

◦

C. If the tempera-

ture rises over the top limit a fan is connected and

the windows open. If the temperature is low, the

windows are closed (the aquaculture is in a green-

house).

• Flow density: The density of the ﬂow is directly

related to salinity and conductivity. Thus, it can be

controlled by the regulation of the doses of clean

water, marine salt, nitrate and sodium bicarbonate

supplied to the culture. The maximum and mini-

mal density values are 1.05 and 1.20 g/cm

• Population density: When the Spirulina popula-

tion reaches a maximum density, this may result

in sudden death. The optimum time for collection

(harvest) is therefore before maximum density is

reached, at a density of 900 mg/l. Density was

measured with a laser device (Ponce, 2001). At

800 mg/l a pump is connected and the harvest be-

gins.

In order to train the neural network, the independent

variables were measured and the dependent variable

was forced to a desired value. During the training,

time was used as an additional variable to distribute

the learning into discrete cycles (Hagan and Demuth,

1999).

Independent variables:

- pH

- Temperature

- Flow density (salinity and electrical conductivity)

Dependent variable:

- Population density

After training, the system controlled 4 variables:

- pH

- Temperature

- Flow density (salinity and electrical conductivity).

- Population density

3.2 Method

It has 4 inputs (measured variables), and 4 outputs

- 3 of which variables that must be controlled; the

remaining output is for training the neural network.

The best way to train a neuronal network is by means

of variable forced learning (Guti´errez et al., 2004).

Training consists of producing successive cultures,

each one under different constant conditions, max-

imum and minimum values. Training is completed

when the learning margin error approaches zero. For

a different condition the system will computebetween

the 2 extreme data. Offside these maximum and min-

imal data the control does not work.

The extreme conditions are: pH min= 8.5 ; pH

max = 10.5

Temperature: min = 13

◦

C; max = 33

◦

Population density mg/l: min = 100; max = 800

Flow density g/cm

: min = 1.05 ; max = 1.20

The best control scheme for this problem is a

NARMA2 network (Norvig and Russell, 2003), a

neuronal controller that transforms nonlinear system

dynamics into linear dynamics by canceling the non-

linearities during training. This optimizes the perfor-

mance of the hardware. A linear system enables faster

training and control in real time for microprocessors.

The control input is computed to force the plant

output to follow a reference signal. The neural net-

work plant model is trained with static back propaga-

tion and is reasonably fast. It requires minimal online

computation (Rio and Molina, 2002).

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

290

The hardware has a 4x4 parallel pic processor array,

with 4 inputs and 4 outputs. This array emulates 20

layers with 4 ﬁles, (80 neurons). This hardware is an

experimental prototype.

In a NARMA2 structure, each neuron is simulated

as:

u(k+ 1) =

yr(k + d) − f[y(k), ..., y(k − n+ 1), u(k), .. ., u(k − n+ 1)]

g[y(k), .. ., y(k − n+ 1),u(k),... ,u(k− n+ 1)]

(1)

Where y(n) is the system input, and u(n) is the system

output. The system function is a linear combination of

f(y(n− 1),u(m− 1)) and g(y(n− 1),u(m− 1)). Each

neuron can be simulated as shown in Figure 2:

Figure 2: Neurons of NARMA2 model.

3.3 Network Training

The neural network was connected to the culture as

shown in Figure 3. The training error is e

. The inputs

are y

, and the exit variables are u.

Each line is equal to 7 lines. The controller with

80 neurons emulates the system as a linear combina-

tion of g and f (generated in the train).

Figure 3: Connection of the Neural Network to the cultiva-

tion.

The training of the neural network is online, super-

vised and forced. It takes at least 6 weeks (6 cultiva-

tions in different conditions).

Employing 2x3 maximum and minimal values (2 by

each parameter) a simulation with 3 set of variables

was made in MATLAB.

Based on the variation of each input, the neural

network built a linear model system. This enables

each output to be related to the variation of all inputs.

The training simulation is carried out with a forced

reference. The dependent variable automatically ad-

justs the inputs and tries to follow the reference (Fig-

ure 4).

Figure 4: Simulation in MATLAB during the training stage.

Figure 5 shows the output 1 (u1) due to input 1 (y1),

when the system is in operation.

Figure 5: Simulation in MATLAB using 1 variable after

training with 3 variable.

3.3.1 Operation

After training, the neural network has 3 inputs and

3 outputs. During the operation, the neural network

learns new rules, it works as an expert controller.

DESIGN OF NEURONAL NETWORK TO CONTROL SPIRULINA AQUACULTURE

291

4 CONCLUSIONS

When used to control a Spirulina alga aquaculture

process, a classic automatic control system was un-

able to respond adequately to sudden climatic varia-

tions.

This neural network design presented here pro-

vides a good control alternative, one that is able to

adapt to climatic changes and seasons, between max-

imum and minimal operation parameters (previously

known) .

The neural network bestows a reliable, robust con-

trol system which can give a correct response to un-

known situations. If part of the hardware is damaged,

the information can be saved.

Furthermore, implementation is not expensive.

The cost of building the prototype control system was

approximately US $ 500 for a pilot plant race way

system (Area= 3m

, volume = 0.5m

In the future it will be necessary to obtain more

and more foods and fuels. The aquaculture is not an

expensive solution and neither is the intelligent con-

trol proposed. A similar control system could be ap-

plied, with some changes, to another aquaculture fo-

cused to obtain fuel from glycerol (Dunaliella Salina)

or petroleum (Botryococcus).

ACKNOWLEDGEMENTS

With acknowledgements to the Spirulina Control Pro-

cess Project: DIPOG 8741-00, Tarapaca University

(Chile) and Associate Professor Claus Brieba of the

Arthur Prat University, without all of whom, this work

would have been impossible.

REFERENCES

Guti´errez, J., Cano, R., and Coﬁno, A. (2004). Redes Prob-

abil´ısticas y Neuronales. Monograf´ıas del Instituto

Nacional de Meteorolog´ıa. Ministerio de Medio Am-

biente, Madrid, Spain.

Hagan, M. and Demuth, H. (1999). Neural networks for

control. pages 1642–1656, San Diego, CA.

Hagan, M., Jesus, O., and Schultz, R. (2002). Recurrent

Neural Networks: Design and Applications, chap-

ter 12, pages 311–340. CRC Press.

Henrikson, R. (1994). Microalga Spirulina - Superalimento

del futuro. Urano S.A., Barcelona, Spain.

Hunt, K., Sbarbaro, D., Zbibowski, R., and Gawthrop, P.

(2002). Neural networks for control system - a survey.

volume 28, pages 1083–1112.

Jourdan, J. (2002). Cultivez Votre Spiruline. Manuel de Cul-

ture Artisanale de la Spiruline. Le Castanet, Mialet,

30140 Anduz, France.

Norvig, P. and Russell, S. (2003). Artiﬁcial Intelligence: A

Modern Approach. Prentice-Hall, 2nd edition.

Ponce, C. (2001). Implementaci´on de sistemas de medici´on

de poblaci´on de microalgas. Master’s thesis, Tarapaca

University.

Rio, B. D. and Molina, A. (2002). Redes Neuronales Y Sis-

temas Borrosos. RA-MA, 2nd edition.

Spirulina.com (2004). Spirulina. green superfood for life.

http://www.spirulina.com/SPBSpirulina.html.

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

292