Artificial Neural N etwork approach for fault [622221]

Artificial Neural N etwork approach for fault
recognition in a Wastewater Treatment Process

Mihaela Miron , Sergiu Caraman
Faculty of Automation , Computers, Electrical Engineering and Electronics
“Dună rea de Jos ” University of Galaț i
Gala ți, Romania
(Mihaela.Miron , Sergiu.Caraman) @ugal.ro

Abstract— A p attern recognition approach using neural
networks is applied to isolate different types of faults in a
wastewater treatment process (WWTP) . This method was
successfully used to determine the hardware components (sensors
and actuators) that have partial or net faults.
Keywords —wastewater treatment proces s; neural networks ;
fault isolation ; fault recognition; actuator fault; sensor fault ;
proce ss diagnosis;
I. INTRODUCTION
In the recent years, the need for efficient methods in fault
detection of dynamical systems has increased

The paper structure is as follows: the second section
presents the simulation model of a wastewater treatment
process; the third section describe s the neural network
approach in order to recognize the faults ; the fourth section
presents the results obtained and the last section is dedicated
to the conclusions.

II. THE SIMULATION MODEL
A wastewater treatment process , presented in [8], [9] was
chosen to illustrated the fault recognition technique proposed
in this paper. The model is described by the following
equations:

) )) ) ) ) ) ) )
(1)

) )
) ) ) ) ) (2)

) ) )) )
) ) )
)( )) ) (3)

) ) ) ) ) )
) ) ) (4)
) )
) )
) (5)

(6)
where X(t) – biomass concentration, S(t) – substrate
concentration, DO(t) – dissolved oxygen concentration, Xr(t) –
recirculated biomass concentration, ) – specific growth
rate, – maxi mum specific growth rate, D(t) – dilution
rate, W(t) – aeration rate, r – recirculating rate, Sin – influent
substrate concentration, DO in – influent dissolved oxygen
concentration, DO sat – saturation value of dissolved oxygen, Y
– yield coefficient, Ks – saturation constant of the substrate,
KDO – saturation constant of dissolved oxygen, – oxygen
transfer rate, – the rate of the sludge in excess, Fin – influent
flow, V – bioreactor volume, Ds – the dilution rate of the
sludge, Vs – sludge volume.
The initial conditions for the mathematical model are
described in table 1:
Table I Initial conditions of the process parameters []
0.11 [h-1]
D(0) 0.025 [h-1]

W(0) 5 [L∙min-1]
r 1
) 0.8 [g∙L-1]
DO in 2 [mg∙ L-1]
DO sat 8 [mg∙L-1]
Y 0.67
Ks 0.18 [g∙L-1]
KDO 0.2 [g∙L-1]
0.0033 [L-1]
0.2
V 35 [L]
Vs 6 [L]
X(0) 0.5 [g∙ L-1]
S(0) 0.8 [g∙ L-1]
DO(0) 2 [mg∙ L-1]
Xr(0) 0 [g∙ L-1]

III. THE NEURAL NETWORK APPROACH
A. Algorithm presentation
This section describes a fault recognition method using a
feed-forward neural network to isolate the faults which can
occur in WWTP’s. Pattern r ecognition technique uses the
observed symptoms and compares it to a set of known
symptoms for each type of fault in searching the best fit . The
fault pattern can be a symptom vector for each fault
considered . This vector contains values of or ( – normal
behavior, 1 – a symptom observed in a particular type of fault).
All faults patterns are collected into a matrix where each
column corresponds to a specific type of fault and each row
corresponds to a particular type of symptom. When the
recognition algorithm chooses a signature as t he best fit, it is
considered that it is a classifier that practically solves a
classification problem.
According to [ 1], wastewater treatment plants are complex
process es in which sensors and its equipment operate in harsh
conditions, and there are often quite large delays in the
response of th e variables to disturbances. Therefore, the types
of faults analyzed in this process are net and partial and can
occur at the level of measurement and control equipment
(transducers and execution elements).
B. Fault recognition s cheme
The fault recognition scheme is shown in Fig. 1. To
recognize faults , a set of signals which refers to all types of
faults and the normal operation state, is required . The inputs of
the neural network are: input variables ), output variables
) and the ir history over the last N samples. Its output
consists in the result of the recognition approach for each type
of fault considered .
Fig. 1. Fault recognition scheme
C. Designing the neural network

IV. RESULTS AND DISCUSSIO NS

V. CONCLUSIONS
In this paper a n actuator fault detection method.
ACKNOWLEDGMENT
The authors wish to thank the support received from the
program “Partnerships in priority areas – PN II”, implemented
with the support of MECS – UEFISCDI, Project No. 269/2014
– BIOCON.
REFERENCES
[1] G.M. Zeng, X.D. Li, R. Jiang, J.B. Li, and G.H. Huang, “Fault
Diagnosis of WWTP Based on Improved Support Vector Machine”,
Environmental Engineering Science. November 2006, 23(6): 1044 -1054.