Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract—Wireless Mesh Network (WMN) is a wireless com- Artificial Neural Network is provided. Section IV describes the
munication technology that has many advantageous character- proposed approach. Performance analysis results are presented
istics such as dynamic self-organization, self-configuration, self-
in section V and section VI concludes the paper.
healing, high scalability and reliable services coverage. However,
wireless mesh links frequently experience quality variations and
undesirable performance or poor connectivity. In this paper we
II. R ELATED W ORK
propose a fault detection and diagnosis approach for enhancing
performance of WMNs based on Artificial Neural Networks
(ANN). The artificial neural network is trained to detect and Li Xu et al, have proposed a fault diagnosis model for
classify individual or composite faults. We consider node failure, WMNs [1] using shortest path spanning tree algorithm to
link failure and traffic congestion as the target faults. The diagnose faults in the network. In their approach all mesh
approach is implemented in NS3 network simulator and its routers build a fault tree based on shortest path spanning
performance is evaluated considering detection rate, false positive
and false negative. tree and every node in the tree has the shortest path to the
root based on hop count. The authors have stated that they
I. I NTRODUCTION have prevented the overhead of maintaining and repairing the
spanning tree which in turn has made communication more
Wireless Mesh Network (WMN) is a new wireless tech- efficient and reduced complexity.
nology that is suitable to qualify the high demand of con- Li et al, have proposed W-C4.5-RP mechanism which is a
nectivity today. Even though WMNs possess qualities such as fault diagnosis model based on decision tree algorithm [2]. The
dynamic self-management and flexibility, frequent link quality authors have employed rule post-pruning to improve the C4.5
variations and undesirable performance has limited their wider decision tree algorithm. They have compared both algorithms
deployment and application. in regards with accuracy rate and the results that they perform
The existing fault detection approaches for WMNs are not nearly the same. The authors have shown that using W-C4.5-
efficient enough to detect and classify faults. This makes the RP decreases the size of rule set which in turn reduces the
fault recovery process costly and complex specially when the matching time for rules. They have considered DoS attack
network in state of composite faults [5], [6], [7]. To optimize and power down attack in their approach as faults.
the capacity of WMNs it is necessary to detect faults and Moustapha et.al [3] have proposed wireless sensor networks
increase the pace of network’s recovery process to make (WSNs) modeling using modified recurrent neural networks
the network operate efficiently even in complex situations. and its application to node failure detection. The inputs to the
Therefore, to contribute in the area of fault detection and neural network are taken from the nodes that are modeled and
diagnosis and overcome the limitations in existing approaches from neighboring nodes. The authors have used received signal
and in WMNs, we propose an Artificial Neural Network strength based confidence factors to model communication
(ANN) based fault detection and diagnosis approach. We use uncertainties. The approach is compared with Kalman filtering
the back-propagation learning algorithm to train our ANN to technique based approach. The simulation results show the
detect and classify faults in WMN. We use the event data and proposed scheme is more efficient.
network statistics that are produced during network operation Jabbari et al. [4] use a modified version of Probabilistic
and due to changes in the network as dataset for training the Neural Networks (PNN) for fault detection in sensor net-
ANN. Once learned, the ANN is able to detect, classify faults works. Their proposed approach is called Fault Detection
in the network. To the best of our knowledge there is no and Isolation (FDI). They used two types of neural network
existing fault detection and diagnosis approach in the research architectures to generate and compare measurement residuals.
domain of WMNs that is based on ANN. Hence, our proposed The authors have first employed an ANN to generate residuals
approach is novel in this essence. by comparing measurement with temperature values and sec-
The rest of the paper is organized as follows. Section II ondly, they have used a PNN to detect faults and classify them
presents the related work. In section III a brief description of accordingly. Their approach was able to differentiate between
normal and faulty transducers by combining current tempera- IV. T HE P ROPOSED A PPROACH
ture measurements and additional environment information. To contribute in the area of fault detection and diagnosis in
III. BACKGROUND WMNs we propose a fault detection and diagnosis approach
based on Artificial Neural Networks (ANN). To the best of our
The notion of Artificial Neural Network is based on neural knowledge none of the existing fault detection approaches in
network in human body. ANN is made of a number of the literature have employed ANN in the area of fault detection
interconnected neurons that process information to provide the and diagnosis in WMNs. Furthermore, the existing approaches
expected solution of a problem. There are two types of ANN; deal with single type of faults whereas our proposed approach
feed-forward and recurrent neural networks. The recurrent is able to detect individual and composite faults as well as
neural network’s structure is similar to a cyclic graph in which classifying them. In this approach we consider three faults;
the signals can loop in the network. On the contrary feed- 1) node failure, 2) link failure and 3) traffic congestion. The
forward neural networks do not form any loops. Feed-forward ANN designed in this approach detects and classifies the faults
neural networks are also called multilayer perceptrons because by learning the normal and faulty status of the WMN. In the
they are consisted of an input layer, one or multiple hidden training phase we pass a vector of inputs and desired output
layers and an output layer [10]. values to the ANN and in the testing phase the ANN delivers
The ANN takes n number of inputs Xi with associated learned results. We use the back-propagation algorithm to
weights Wi and computes the sum nof the product of given train the ANN. We use the sigmoid function as the activation
inputs and their weights as z = i=1 Xi · Wi The function function in the back-propagation algorithm.
will return 1 if the value fed to it is greater than or equal to
a threshold and otherwise it will return 0 [9]. A. Network and Fault Model
The back-propagation algorithm in ANN is used to solve We consider a backbone WMN which is decentralized and
complex and non-linear problems. The back-propagation al- infrastructure-based. The components of the network include
gorithm works by gradually computing the errors from the Gateways, Mesh Routers where some of the routers have
input layer to the output layer and propagate the error back to moderate level of mobility.
the previous layer to update the weights. [9]. The algorithm We have distinguished WMNs faults in the following cate-
back-propagates the value using gradient descent G which gories: Node failure, communication failure, traffic overload,
is the derivative of the activation function f with respect to routing protocol failure, network scalability issues, network
the weights. With every iteration of the process the error is topological changes, base station and backbone failures [6].
minimized and the output gets closer to the desired output The target faults in this paper include; node failure, link failure
[10]. and traffic congestion.
n
1 1) Node Failure: Node failure is a situation when a node is
E= (Desiredoutput − Obtainedoutput ) (1)
2 i=1 unresponsive to messages from other nodes. Node failure can
happen in a WMN in different ways. Software failure, hard-
δE
G= (2) ware defects and batter power drainage can be probable causes
δwij of node failure [6]. Node failure degrades the performance of
The most commonly used activation function in feed-forward the network and results in dropped or corrupt messages and
multilayer perceptron networks is the logistic/sigmoid function routing protocol overhead.
[10] displayed as follows: 2) Link Failure: Link failure happens when wireless links
1 face problems such as interference, limited bandwidth, unpre-
f (z) = (3)
1 − e−z dictable circumstances, multi-path fading and weak signal [6].
The following formula shows the updated weights: Link failures degrade the performance of the network.
Δwij = −η · E · G + α · Δwij (n − 1) (4) 3) Traffic Congestion: Traffic congestion is caused by large
amount of data transmitted over a link simultaneously, high
where, η is the momentum and α is the learning rate and n−1 throughput demand, and limited bandwidth in links. Traffic
is the previous layer. The values of momentum and learning congestion causes degradation in link quality and degrade the
rate are given by trial and error approaches. If momentum is performance of the WMN. Furthermore, traffic congestion can
too high the network moves fast and cannot learn correctly and happen as a side effect of other network faults and network
if it is too low causes the network to back-propagate slower. topological changes [6].
Similarly, an α value too high makes the network to over-
learn and produce false results. Relatively a too low learning B. Artificial Neural Network Architecture
rate makes the network lazy to learn. An ANN works in two Our proposed fault detection and diagnosis approach for
phases; training and testing. A vector of input values and its WMNs is based on multilayer feed-forward ANN shown in
desired output value is passed to the neural network in the Figure 1. The architecture we use for the ANN for the purpose
training phase. The network learns normal and anomalous of fault detection and diagnosis of WMN is made of three
states of the system and in the testing phase the network layers; one input layer, one hidden and one output layer. We
delivers results based on the learned information. have 5 neurons in the input layer. Each neuron in the input