Artificial Neural Networks for Fault Diagnosis of Milk ... · diagnostic approaches have emerged as...
Transcript of Artificial Neural Networks for Fault Diagnosis of Milk ... · diagnostic approaches have emerged as...
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
Artificial Neural Networks for Fault Diagnosis of Milk
Pasteurization Process - A Comparative Study
Chebira Samia, Bourmada Noureddine, and Boughaba Abdelali
Industrial Safety Department, Hygiene and Industrial Safety Institute
Mustafa Ben Boulaid BATNA 2 University
Batna, Algeria
Abstract
The increasing complexity of most industrial processes always tends to create problems in monitoring and
supervision systems. Detection and early fault diagnosis are the best way to manage and solve these problems.
Artificial neural networks (ANNs), by their ability to learn and store a large volume of information, are tools
particularly suitable for diagnostic support systems. Effectiveness of ANNs for fault diagnosis in milk
pasteurization process is presented in this paper. The initial data base used for fault diagnosis is constructed
using data extracted from FMEA (Failure Modes and Effects Analysis) tables of milk pasteurization process.
Indeed, this analysis makes it possible to establish the links of cause and effect between the faulty components
and the observed symptoms. Three models of ANNs, namely Feed-Forward Back Propagation (FFBP), Radial
Basis Function based Neural Network (RBNN), and Generalized Regression Neural Networks (GRNN) are
developed and compared. The determination coefficient (R2), Root Mean Square Error (RMSE), and Mean
Absolute Error (MAE) statistics were used as evaluation criteria of all the models. The comparison results
indicate that the performances of GRNN model are better than the FFBP and RBNN models. The same neuronal
models can be extended to any technical system by considering appropriate parameters and defects.
Keywords Fault diagnosis, Feed-forward back propagation, Radial basis function based neural network, Generalized
regression neural networks.
1. Introduction
The problem of detecting and diagnosing faults in complex industrial plants is strategically important for its
various implications, e.g., avoiding breakdowns that can lead to major industrial disasters, problems related to
the workers and plants safety, fast and appropriate response to emergency situations and plant maintenance. For
example, the following systems represent only a small part of systems where fault detection and diagnosis are
usually a very difficult but important tasks: chemical and petrochemical plants, refineries, power plants,
airplanes, ships, submarines, space vehicles and space stations, automobiles and household appliances.
Generally, in industrial plants, there is a crucial need for checking and monitoring the equipment condition
precisely since they are mostly subject to hazardous environments, such as severe shocks, vibration, heat,
friction, etc. So fault detection, fault identification and diagnosis of equipments, machineries and systems have
become a vigorous area of work. Due to the broad scope of the process fault diagnosis problem and the
difficulties in its real-time solution, many analytical-based techniques (Isermann ,1997; Leonhardt and Ayoubi,
1997) have been proposed during the past several years for the fault detection of technical plants. The important
aspect of these approaches is the development of a model that describes the ‗cause and effect‘ relationships
between the system variables using state estimation or parameter estimation techniques.
The problem with these mathematical model-based techniques is that under real conditions, no accurate models
of the system of interest can be obtained. In that case, the better strategy is of using knowledge-based techniques
where the knowledge is derived in terms of facts and rules from the description of system structure and behavior
(Rajakarunakaran et al., 2008). Classical expert systems were used for this purpose. The major weakness of this
approach is that binary logical decisions with Boolean operators do not reflect the gradual nature of many real
world problems. Recently, with the development of artificial intelligence, Computational Intelligence (CI)
methods, Neural Networks (NN), Fuzzy Logic (FL), Evolutionary Algorithms (EA), etc., more and more fault
diagnostic approaches have emerged as new techniques for fault diagnostic systems (Venkatasubramanian et al.,
2003; Patton et al., 2000) .
Artificial neural networks, by their ability to learn and store a large volume of information, are tools particularly
suitable for diagnostic support systems. Neural networks are known to approximate any non-linear function,
1393
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
given suitable weighting factors and architecture. NN can generalize when presented with inputs not appearing
in the training data and make intelligent decisions in cases of noisy or corrupted data. However, the NN operates
as a ―black box‖ with no qualitative information available of the model it represents (Patton et al., 1994) .
In this study, effectiveness of ANNs for fault diagnosis in milk pasteurization process is presented. The initial
data base used for fault diagnosis is constructed using data extracted from FMEA tables, of milk pasteurization
process. Indeed, this analysis makes it possible to establish the links of cause and effect between the faulty
components and the observed symptoms. Three models of ANNs, namely Feed-Forward Back Propagation
(FFBP), Radial Basis Function based Neural Network (RBNN), and Generalized Regression Neural Networks
(GRNN) are developed and compared.
This paper is organized as follows: The next section is devoted to present the review of artificial neural
networks. Section 3 describes the milk pasteurization process. Sections 4 presents the fault detection and
diagnosis in milk pasteurization process. Simulation results of fault diagnosis are presented in Section 5. Finally,
in Section 6, conclusions are drawn from the work.
2. Review of Artificial Neural Networks (ANNs)
2.1 Artificial Neural Networks (ANNs)
Artificial neural networks (ANNs) are computational modeling tools that have recently emerged and found
extensive acceptance in many disciplines such as data processing, process analysis and control, fault detection
and diagnosis, pattern recognition, and defining complex and nonlinear relationship and employs number of
input–output training patterns from the experimental data (Hagan et al., 1996; Chen et al., 2015; Yerrabolu et al.,
2013) . Finding a nonlinear algorithm between inputs and outputs is obtained by natural ability of ANNs. They are made
of nodes or neurons, number of simple computing components, which utilized to form respectively an input
layer, one or more hidden layers and an output layer (Hornik et al., 1990). Flexibility of model and accuracy in
prediction are needed for developing a model. Also, neural networks are suitable for prediction of complex
nonlinear functions compared to other literature models (Carrera and Aires-de-Sousa, 2015; Ghaedi et al., 2015).
An artificial ANN consists of some basic elements called neurons. Input variables are processed through
successive layers of neurons. There is always an input layer, with a number of neurons equal to the number of
variables of the problem and an output layer, where the response is made available with a number of neurons
equal to the desired number of quantities computed from the inputs. Layers between the input and output layers
are called hidden layers and may contain a large number of hidden processing units. The ability to effectively
approximate non-linear systems is due to the presence of this hidden layers and non-linear transfer functions in
the hidden layer‘s neurons. The output of each neuron is determined by using an activation function; usually
nonlinear activation functions are used, such as sigmoid or Gaussian. To obtain the desired output for any given
input, the coefficients should be determined by training the network where sets of inputs with the corresponding
outputs are given to the network through a training algorithm. This process should be repeated several times in
order to minimize the output error. Each run of a complete set is called an epoch (Bishop, 1996; Haykin, 1994).
The neural networks used in this work are Feed-Forward Back Propagation (FFBP), Radial Basis Function based
Neural Network (RBNN), and Generalized Regression Neural Networks (GRNN).
2.2 Feed-Forward Back Propagation (FFBP)
An FFBP network structure has one input layer, one output layer, and at least one hidden layer with hidden
neurons.
Figure 1 illustrates a three-layer neural network consisting of layers i, j, and k, with interconnection weights Wij
and Wjk between layers of neurons. The input signals presented to the system in input layer are processed in
forward through to the hidden layer. The summation of weighted input signals is transferred by a nonlinear
activation function.
The response of network is compared with the actual observation results and the network error is calculated
(Sen, 2004) . The error of network is propagated backwards through the system and the weight coefficients are
updated (Firat et al., 2010). After that, a feed-forward process is again formed until a target total error or number
of prescribed iterations is reached (Partal, 2009). The numbers of hidden layer neurons is found using simple
trial–error method in applications. The detailed theoretical information about FFBP can be found in (Haykin,
1999), ( Medhat et al., 2016), and (Bilhan, 2010).
1394
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
2.3 Radial Basis function based Neural Network (RBNN)
Radial-basis function neural networks were proposed by Broomhead and Lowe (1988). An RBNN neural
network is a type of feed-forward network that learns using a supervised training technique, and its output nodes
form a linear combination of the radial basis functions computed by the hidden layer nodes (Hush and Horne,
1993). The RBNN structure consists of an input layer, a single hidden layer, and an output layer as shown in
Figure 2.
The basic functions in the hidden layer produce a significant nonzero response to input stimulus only when the
input falls within a small localized region of the input space. The input-output relationship of this RBNN
network can be described by:
𝑌𝑖 = 𝑊𝑖𝑗𝜑𝑗 𝑥 + 𝑏𝑖𝑁ℎ
𝑗=1 (1)
where φ = the radial basis function of the hidden unit j; x = input data vector; wij represents a weighted
connections between the radial basis function and output layer; Nh = the number of hidden-layer neurons. The
constant term bi in Eq. (1) represents a bias. The hidden neuron of an RBNN has a Gaussian function as its
activation function.
𝜑𝑖 𝑥 = 𝑒𝑥𝑝 − 𝑥−𝑐𝑖
2
2𝜎𝑖2 , 𝑖 = 1,2,𝑁ℎ (2)
Here cis are centers and σi widths (or spreads). ‖. ‖ is the Euclidean distance norm. For simplicity, the centers and
variances are predefined and fixed.
Figure 2: Schematic diagram of RBNN
Outputs Inputs
Output
Layer
Input
Layer
Hidden
Layer
Weights Weights
LLayer LLayer LLayer
Figure 1: Typical feed forward network architecture
1395
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
From a design point of view, the training of RBNN networks involves finding the number of hidden layer nodes
(neurons) Nh and the appropriate parameter set (ci, σi and wij) to map a given input vector to a desired output
scalar efficiently with good accuracy and generalization.
2.4 Generalized regression neural network (GRNN)
The generalized regression neural network, as proposed by Donald Specht (Specht, 1991), falls into the category
of probabilistic neural networks. GRNN is a neural network architecture that can solve any function
approximation problem in the sense of estimating a probability distribution function. GRNN is a universal
approximator (Park and Sandberg, 1991) for smooth functions, allowing it to solve any function approximation
and estimate any continuous variable problem when given enough data.
A schematic of the GRNN is shown in Figure 3. The GRNN consists of four layers (Patterson, 1996): input,
pattern, summation, and output layers.
The number of input units in the first layer is equal to the total number of parameters. The first layer is fully
connected to the second, pattern layer, where each unit represents a training pattern and its output is a measure of
the distance of the input from the stored patterns. Each pattern layer unit is connected to the two neurons in the
summation layer: S-summation neuron and D-summation neuron. The S-summation neuron computes the sum of
the weighted outputs of the pattern layer while the D-summation neuron calculates the unweighted outputs of the
pattern neurons. The connection weight between the ith neuron in the pattern layer and the S-summation neuron
is yi; the target output value corresponding to the ith input pattern. For D-summation neuron, the connection
weight is unity (Yilmaz et al., 2010).
The output layer merely divides the output of each S-summation neuron by that of each D-summation neuron,
yielding the predicted value to an unknown input vector x as
𝑌𝑖 = 𝑦𝑖 .𝑒𝑥𝑝 −𝐷 𝑥 ,𝑥𝑖 𝑛𝑖=1
𝑒𝑥𝑝 −𝐷 𝑥 ,𝑥𝑖 𝑛𝑖=1
(3)
Where n indicate the number of training patterns, and the Gaussian D function in (3) is defined as
𝐷 𝑥, 𝑥𝑖 = 𝑥𝑘−𝑥𝑖𝑘
𝜎 2𝑚
𝑘=1 (4)
yi is the weight connection between the ith neuron in the pattern layer and the S-summation neuron, n is the
number of the training patterns, D is the Gaussian function, m is the number of elements of an input vector, and
xk and xik are the jth element of x and xi, respectively. The σ notation, known as the spread (or width), determines
the generalization performance of the GRNN. In general, a larger σ value may result in better generalization; its
optimal value is determined via trial and error. It should be noted that in conventional GRNN applications, all
units in the pattern layer have the same single spread (Specht, 1991).
Figure 3: Schematic diagram of GRNN
1396
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
2.5 Performance indices
Various statistical measures have been developed and used in the literature. To assess the fitting and predictive
accuracy of the models, the data sets were mathematically evaluated by calculating the following evaluation
criteria: determination coefficient (R2), root mean squared error (RMSE), and mean absolute error (MAE). In
addition, a graphical comparison was performed to illustrate the accuracy of the proposed models. The
computations of R2, RMSE, and MAE are given below:
𝑅2 = 1 − 𝑂𝑖−𝑃𝑖
2𝑁𝑖=1
𝑂𝑖 2𝑁
𝑖=1
(5)
𝑅𝑀𝑆𝐸 = 1
𝑁 𝑂𝑖 − 𝑃𝑖
2𝑁𝑖=1 (6)
𝑀𝐴𝐸 =1
𝑁 𝑂𝑖 − 𝑃𝑖
𝑁1 (7)
where N is the number of the points in the data set, Oi is some measured value, and Pi is the corresponding model
prediction (Willmott, 1981; Willmott, 1982). In addition, a graphical comparison was performed to illustrate the
accuracy of the proposed models.
3. The proposed fault diagnosis methodology
The proposed methodology for fault detection and diagnosis as shown in Figure 4, is based on using three
ANNs, namely FFBP, RBNN, and GRNN to detect and diagnose the failures which can lead to abnormal
operating conditions. The main purpose of selecting ANNs as a tool is ability to capture the non-linear
relationship between the inputs and the outputs, generalization ability and fast real-time operation. The neural
network approach for this application has two phases; training and testing. During the training phase, neural
network is trained to capture the underlying relationship between the chosen inputs and outputs. After training,
the networks are tested with a test data set, which was not used for training. Once the networks are trained and
tested, they are ready for real-time application. Then, a comparison of the training and testing performances
between the FFBP, RBNN, and GRNN models is carried.
For the application of machine learning approaches, it is important to properly select the input variables, as
ANNs are supposed to learn the relationships between input and output variables on the basis of input–output
pairs provided during training. In this work, the starting data are extracted from the FMEA tables and associated
with an initial rule base for establishing cause-and-effect relationships between the failing organs and the
observed symptoms. These data are used as a database of neural networks, and the cause-and-effect links will be
represented in the form of a binary coding constructing the data set, corresponding to 44 vectors.
The data vectors used in the three models ANNs are intervals limited by two values, minimum and maximum.
The symbol '1' represents a normal functioning of the system, and the symbol '0' represents a failure situation.
FMEA cause-and-
effect
FFBP
RBNN
GRNN
Fault
diagnosis
Figure 1: The proposed methodology for fault diagnosis using three models of ANNs
1397
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
4. Study system
In this section, the proposed methodology is applied to milk pasteurization process, the schematic diagram of
which is shown in Figure 5. The process of milk pasteurization consists mainly of five major elements, a control
system, a balancing tank (constant level tank), a plate heat exchanger with four sections, two pumps and valves
which can be interconnected as shown in Figure 5. A detailed explanation of the working of the process is
described in (Djelloul, 2013).
5. Results and discussions
This section presents the details of the training and testing of ANN models for fault diagnosis on the milk
pasteurization process. Three different ANN models were developed for fault diagnosis, FFBP, RBNN, and
GRNN. In all models, 70 % of the data set was randomly assigned as the training set and 30 % was used for
testing the performance of the model predictions. Note that the FFBP, RBNN, and GRNN models employ the
same training and test data sets for appropriate performance comparison. The neural network model is
developed using MATLAB 8.1 Neural Network Toolbox.
The FFBP used here is composed of an input layer and an output layer that respectively contain the effects and
causes of failures obtained from FMEA arrays. The FFBP can have more than one hidden layer; therefore, in this
study, one hidden layer FFBP was used. The tansigmoidal and linear activation functions were used in the input
layer and the output layer, respectively. Moreover, the activation function, in the hidden layers, was chosen as
tansigmoidal function. The number of hidden layer neurons to minimize MSE was found using simple trial and
error with different architectures of all models. The FFBP model, comprising one hidden layer with 10 neurons
has the lowest performance (0.0526) and hence was considered optimal for this study.
In the RBNN model, the key parameter is the spread constant, plays a crucial role in establishing a good ANN
model with high prediction accuracy and stability. Therefore, this parameter needs to be correctly determined on
the basis of the evaluation criteria to optimize prediction performance. The best performance (0.0010) of the
RBNN is obtained by a value of the spread constant that is equal to 0.1.
In this study, different spreads (between 0.01 and 0.8) were tried to find the best value for the GRNN model. The
best testing performance (1.1166 e-62) of the GRNN was obtained when the spread parameter equal to '0,1'.
Tables 1 present the performance during the training, and testing phases of FFBP, RBNN, and GRNN models for
the detection and diagnosis of milk pasteurization process failures, in terms of R2, RMSE, and MAE statistics.
Figure 5: Milk pasteurization process
1- Tank
2- Feed pump
3- Heat exchanger
4- Valve
5- Pipe
1398
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
Table 1 show that the GRNN performed better during training and testing, and it outperforms the FFBP and
RBNN in terms of all the standard statistical measures.
RMSE provides a measure to judge the accuracy of the fit of the models. A lower RMSE indicates a better fit. As
seen from Table 1 that the GRNN model has the smallest RMSE (1.1444e-31) and MAE (1.6601e-32), and the
highest R2 (1) in training period; also in testing period the GRNN has the smallest RMSE (8.0179e-32) and MAE
(8.1494e-33), and the highest R2 (1). According to the training and testing results, the RBNN and FFBP models
provide RMSE, and MAE values close to each other. The RMSE and MAE values of the FFBP model are higher
than the corresponding values of the RBNN model in the training and testing periods, and the values of the
determination coefficient for the two models are identical in both periods.
A very high R2 value (R
2 = 1) for the GRNN model means that it has a better linear relationship (perfect linear
regression) between the observed and calculated failures, compared to the FFBP and RBNN models.
The scatter plots of the observed versus simulated failures of the FFBP, RBNN, and GRNN analyzed herein are
shown in Figures 6, 7, and 8 for the training and testing phases, respectively. It can be obviously seen from the
Figures 6, 7, and 8 that the GRNN simulations are closer to the corresponding observed failures than those of the
RBNN and FFBP. For the training and testing phases, a total coincidence is observed with the adjustment line of
the GRNN model with respect to the RBNN and FFBP models.
The Figures 6, 7, and 8 shows that model performances is generally accurate and that the GRNN model is
consistently superior to the RBNN and FFBP models. Overall, the performance of all models is very satisfying.
The results demonstrate that the GRNN can be applied with better performance, to establish diagnostic models.
6. Conclusion
In this paper, a fault diagnosis system using artificial neural networks was proposed. The FFBP, RBNN, and
GRNN models were developed to detect and diagnose failures of milk pasteurization process. The predictive
performance of each model was assessed using three statistical measures: R2, RMSE, MAE, and a study of the
graphs were used. The results of the statistical measures suggest that GRNN model provides more accurate
results than the FFBP and RBNN models. The high value of the determination coefficient and the low value of
RMSE and MAE in the testing set indicate that the developed models can be used for prediction failures of milk
pasteurization process.
This study has indicated that the GRNN model is the best predictor of failures among three ANN models in
respect to R2, RMSE, and MAE statistics. GRNN model can be successfully employed in fault diagnosis.
RBNN model has lower RMSE and MAE than FFBP model. A lower RMSE means that the accuracy of the
RBNN model is higher than FFBP model. The ranking of prediction was obtained as GRNN, RBNN, and FFBP,
respectively.
Table I: Performances of the FFBP, RBNN, and GRNN models in the training and
testing phases
ANNs
Models
Training Testing
R2
RMSE MAE R2
RMSE MAE
FFBP 0.99978 0.1925 0.0370 0.99958 0.3015 0.0909
RBNN
0.99983 0.0221 9.7752e-04
0.99963 0.0483 0.0047
GRNN 1 1.1444e-31 1.6601e-32 1 8.0179e-32 8.1494e-33
1399
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
0 5 10 15 20 25 30 35 40 45 50
0
5
10
15
20
25
30
35
40
45
50
R² = 0.99978
Observed Failures
Calc
ula
ted
Fail
ure
s
0 5 10 15 20 25 30 35 40 45 500
5
10
15
20
25
30
35
40
45
50
R² = 0.99958
Observed Failures
Calc
ula
ted
Fail
ure
s
Figure 6: Scatter plots of calculated versus observed failures for
the FFBP for (a) training and (b) testing phases
0 5 10 15 20 25 30 35 40 45 500
5
10
15
20
25
30
35
40
45
50
R² = 0.99963
Observed Failures
Calc
ula
ted
Fail
ure
s
0 5 10 15 20 25 30 35 40 45 50
0
5
10
15
20
25
30
35
40
45
50
R² = 0.99983
Observed Failures
Calc
ula
ted
Fail
ure
s
Figure 7: Scatter plots of calculated versus observed failures for
the RBNN for (a) training and (b) testing phases
0 5 10 15 20 25 30 35 40 45 50
0
5
10
15
20
25
30
35
40
45
50
R² = 1
Observed Failures
Calc
ula
ted
Fail
ure
s
0 5 10 15 20 25 30 35 40 45 50
0
5
10
15
20
25
30
35
40
45
50
R² = 1
Observed Failures
Calc
ula
ted
Fail
ure
s
Figure 8: Scatter plots of calculated versus observed failures for
the GRNN for (a) training and (b) testing phases
1400
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
References
Awadalla, M., Yousef, H., Al-Hinai, A., Al-Shidani, A., Prediction of Oil Well Flowing Bottom-hole Pressure in
Petroleum Fields, Proceedings - International Conference on Industrial Engineering and Operations
Management, Kuala Lumpur, Malaysia, March 8-10, 2016, 3007-3017, 2016.
Bilhan, O., Emiroglu, M.E., Kisi, O., Application of two different neural network techniques to lateral outflow
over rectangular side weirs located on a straight channel, Adv Eng Softw, 41:831–837, 2010.
Bishop, C.M., Neural networks for pattern recognition, Clarendon Press, Oxford, 1996.
Broomhead, D.S., Lowe, D., Multivariable functional interpolation and adaptive networks, Complex Syst, 2:321–
355, 1988.
Carrera, G., Aires-de-Sousa, J., Estimation of melting points of pyridinium bromide ionic liquids with decision
trees and neural networks, Green Chem., 7:20–27, 2005.
Chen, G., Luo, X., Zhang, H., Fu, K., Liang, Z., Rongwong, W., Tontiwachwuthikul, P., Idem, R., Artificial
neural network models for the prediction of CO2 solubility in aqueous amine solutions, Int. J.
Greenhouse Gas Control, 39: 174–184, 2015.
Djelloul, I., Souier, M., and Sari, Z., Neuro-fuzzy genetic algorithms for monitoring in a production system,
IEEE International conference on systems and control, 12-17, 2013.
Firat, M., Turan, M.E., Yurdusev, M.A., Comparative analysis of neural network techniques for predicting water
consumption time series, J. Hydrol., 38:446–451, 2010.
Ghaedi, M., Ansari, A., Assefi Nejad, P., Ghaedi, A., Vafaei, A., Habibi, M.H., Artificial neural network and
Bees Algorithm for removal of Eosin B using Cobalt Oxide Nanoparticle-activated carbon: isotherm
and kinetics study, Environ. Prog. Sustainable Energy, 34: 155–168, 2015.
Hagan, M.T., Demuth, H.B., Beale, M.H., Neural Network Design, Pws Pub, Boston, 1996.
Haykin, S., Neural Networks a Comprehensive Foundation, Prentice Hall, Upper Saddle River, 1994.
Haykin, S., Neural networks: a comprehensive foundation, 2nd edn. Prentice-Hall, Englewood Cliffs, 1999.
Hornik, K., Stinchcombe, M., White, H., Universal approximation of an unknown mapping and its derivatives
using multilayer feed-forward networks, Neural Netw., 3: 551–560, 1990.
Hush, D.R., Horne, B.G., Progress in supervised neural networks, IEEE Signal Process Mag, 10(1):8–39, 1993.
Isermann, R., Supervision, fault detection and fault diagnosis methods—an introduction, Int. J. Control Eng.
Pract., 5 (5): 639–652, 1997.
Leonhardt, S., Ayoubi, M., Methods of fault diagnosis, Int. J. Control Eng. Pract., 5 (5): 683–692, 1997.
Park, J., Sandberg, I.W., Universal approximation using radial basis function networks, Neural Comput,
3(2):246–257, 1991.
Partal, T., River flow forecasting using different artificial neural network algorithms and wavelet transform, Can
J Civil Eng, 36:26–39, 2009.
Patterson, D.W., Artificial Neural Networks Theory and Applications, Prentice Hall International Editions,
Korea, p 477, 1996 .
Patton, R. J., Chen, J., and Siew, T. M., Fault Diagnosis in Non-linear dynamic systems via neural-networks,
Proc. IEE Int. Conference Control ’94, Coventry, UK, 2:1346-1351, 1994.
Patton, R. J., Uppal, F. J., and Lopez-Toribio, C. J., Soft Computing Approaches To Fault Diagnosis For
Dynamic Systems: A Survey, Proc. of 4th IFAC Symposium on Fault Detection Supervision and Safety
for Technical Processes, Budapest, 14-16 June 2000, 1:298-311, 2000.
Rajakarunakaran, S., Venkumar, P., Devaraj, D., Surya Prakasa Rao, K., Artificial neural network approach for
fault detection in rotary system, Applied Soft Computing, 8:740–748, 2008.
Sen, Z., Principals of artificial neural networks, Water Foundation Publications, Istanbul (in Turkish), 2004.
Specht, D.F., A general regression neural network, IEEE Trans Neural Netw, 2(6):568–576, 1991.
Venkatasubramanian, V., Rengasamy, R., Yin, K., Kavuri, S.N., A review of process fault detection and
diagnosis, Part I: quantitative model-based methods, Int. J. Comput. Chem. Eng., 27: 293–311, 2003.
Venkatasubramanian, V., Rengasamy, R., Yin, K., Kavuri, S.N., A review of process fault detection and
diagnosis, Part III: process history-based methods, Int. J. Comput. Chem. Eng., 27: 327–346, 2003.
Willmott, C.J., On the validation of models, Physical Geography, 2:184–194, 1981.
Willmott, C.J., Some comments on the evaluation of model performance, Bulletin American Meteorological
Society, 63:1309–1313, 1982.
1401
Proceedings of the International Conference on Industrial Engineering and Operations Management Dubai, UAE, March 10-12, 2020
© IEOM Society International
Yerrabolu, P., Mareddy, L., Bhatt, D., Aggarwal, P., Kumar, A., Devabhaktuni, V., Correction model-based
ANN modeling approach for the estimation of radon concentrations in Ohio, Environ. Prog. Sustainable
Energy, 32: 1223–1233, 2013.
Yilmaz, T., Seckin, G., Yuceer, A., Modeling of effluent COD in UAF reactor treating cyanide containing
wastewater using artificial neural network approaches, Adv Eng Softw, 41:1005–1010, 2010.
Biographies
Chebira Samia is a professor at the Institute of Hygiene and Industrial Safety at Mustafa Ben Boulaid BATNA2
University, Batna, Algeria, and she is member of the Industrial Prevention Research Laboratory (LRPI). She
holds magister in control of industrial risks. She is interested, since the late 1990s, the safe operation and fault
diagnosis of industrial process. His current research themes concern the fault diagnosis using artificial neural
networks.
Institute of Health & Industrial Safety, Mustafa Ben Boulaid BATNA 2 University, Fesdis Road Constantine,
Batna, Algeria.
Bourmada Noureddine is a professor at the Institute of Hygiene and Industrial Safety at Mustafa Ben Boulaid
BATNA2 University, Batna, Algeria, he is doctor of science in spectrochemistry, and he is member of the
Industrial Prevention Research Laboratory (LRPI). He has been director of the Institute of Hygiene and
Industrial Safety since 2001 to 2016. His research focuses on the analysis of environmental risks.
Institute of Health & Industrial Safety, Mustafa Ben Boulaid BATNA 2 University, Fesdis road Constantine,
Batna, Algeria.
Boughaba Abdelali is Professor at the Institute of Hygiene and Industrial Safety at Mustafa Ben Boulaid
BATNA 2University, Batna, Algeria, he is doctor of science in electro-technical, and he is member of the
Industrial Prevention Research Laboratory (LRPI). He was deputy director in charge of post graduation and
currently he deputy director of pedagogy. Since the beginning of the 1990s, he has been interested in the
numerical control of electrical machines and the diagnosis of failures in industrial and particularly electrical
systems. His current research interests concern the control of motors without collectors (BLDCM), the fault
diagnosis using artificial neural networks.
Institute of Health & Industrial Safety, Mustafa Ben Boulaid BATNA 2 University, Fesdis road Constantine,
Batna, Algeria.
1402