PROTOCOL: Detection of Visual Faults in Wireless Mesh Networks through CNN Approach

Authors: parutagouda khanagoudar
Date added: 12th August 2021, 11:19:22

Linked Problems

No linked problems available.


Reviews

No linked reviews available.



Full text

Detection of Visual Faults in Wireless Mesh Networks through CNN Approach

       Authors: Parutagouda S Khanagoudar

                    Dr G M Patil

                    Dr Dinesha H A

                                               

 


Abstract

There are many challenges faced by network administrators. This paper examines the challenges with respect to detecting and identifying faulty network components like nodes or cables in a computer network environment. Here we reviewed several researches works on techniques used in building fault detection and identification mechanisms for LAN (Local Area Network). An active probing method is applied to deal with the matrix dependency building. And for fault identification and detections an internal method is designed. This proposed research method is tested in a non- production computer network environment. The results produced in this research work is able to identify and detect the nodes with faults within 0.22 seconds given sufficient computing resources. Therefore, the proposed mechanism can be best suited for any WMN.

Keywords — Fault Detection Identification, LAN, Dependency Matrix

 

I.   Introduction

 

To monitor the dedicated network structure indicates the malfunctioning or failure of the network devices and some components. This also indicates the fault in the network behaviors with fault detection and identification mechanism. An active and passive monitoring can be done for identified network structure. In order to manage or monitor effectively and successfully a targeted computer network, a huge number of data needs to be continuously obtained and processed about network devices.

 

In general fault nodes can be detected by two methods in WMN. First, active system which is based on the probe mechanism and second, is the alarm correlation mechanism based on passive system [10]. These two mechanisms involve in a certain challenging network and also comes with an alternative solution to identify the faulty behavior in the network structure.

Classification of Network Faults

The decision of the failure of component is rely solely on network alarms. Network faults are classified based on their time duration in the network into three categories [11].

(i)           Permanent faults: the faults that cannot be rectified and resolved can be classified as Permanent fault.

Like – Malfunctioned Cables, Network Interface Card, Switches, Routers

(ii)         Intermittent faults: the faults that can cause halt in the system for a period of time.

(iii)       Transient faults: the faultswhich can degrade the service at minor level and cab be masked with managementutilities.


II.   LiteratureReview

Numerous procedures have been examined in various literary works to distinguish deficiencies in computer networks. Coming up next are some audited writings on various strategies for distinguishing and recognizing issues computer networks.
Techniques for Fault Management in Network Structure

Fault management in the network structure can be analyzed and will be done to collect information about a given network structure. by using different techniques to identify the deficiency in the network component. This segment talks about some notable existing strategies utilizing four key areas.

     Techniques based on Artificial Intelligence

Some of the existing system works on AI based fault management techniques. Expert system utilizes a standard based technique to impersonate the human information or manner of thinking of a specialist.

According to the survey presented in [12] the main components which are loosely coupled are:

1.       Amonitor

2.       A problem clearingadvisor

3.       A trouble ticket creation system

4.       Networkdatabase repository

 

In [7] a Kohonen, Self-Organizing Map (SOM) fault detection system is proposed which works based on the neural network and clustering components. As the neural network uses a weight-based training method and can take too much time to train the models. Because of this no specific rules can be designed for selecting the number of neurons and layers using artificial neural networks.

 

Intelligent Probing-Based Techniques

A single node in a computer network will be installed with a probe a committed program or an application to detect the fault in the network. This node can be referred to as a major probing station to send, receive and analyze remaining nodes in the network most frequently.

In this a dedicated matrix called as dependency matrix is used to design and develop the station for probing. This matrix can assist to identify and rectifying the fault in the network [1,2,3]. There is an alternate way also for detecting the fault as given in [10] for the reduction of the total probing station and by making the detection of the fault effectively using constraint based fuzzy logic technique. The outcomes of the existing mechanism shows that the AI model is more efficient in order to identify and detect the fault in the computer networks.

Model-Based Technique

This is techniques based on the model which is an abstract layered approach and indicates the functional parameters of the physical network strcuture of the system segment.


 

 

 

III.  METHODOLOGY

The dependency matrix is specifically designed as an active probing station identifying the faulty nodes in a WMNs. Recently majority of devices connected through network where the cost of maintenance will be reduced. This can also reduce the cost of fault management in the WMN.  Assuring the high accessibility of the various network devices and its resources is the main goal of in the fault management system. An automatic monitoring process in implemented in the proposed research work which helps in identifying and detecting the fault in the network. This mechanism is implemented using the         probe-based mechanism.

An active-probe mechanism is obtained for aggregating the network object (NO) information. In this the probe mechanism software based diagnostic mechanism is carried out in order to ensure the NO is active or inactive. Hence, this active-probe (p) based a subset p Í NO. Here the presence of fault can affect some probes [2], while remaining probes will be unaffected. And there will be some components which may be in F also present in P. As fault F disturbs a probe P if F∩P ≠ Ø.

In the network the various physical entities like- nodes, switch, router and NIC are recognized as Objects. Probes will be received to the probe station from the faulty nodes along with the performance of the various components. There will be change of getting fault in any of the components of the network and cane deducted using the fault detection method proposed in this research work.

Proposed Fault deduction method:

SW = node*link

Where SW = Switch

Node=Computer

L = Link (Wired)

Here the set of processing network nodes is represented as N = {node1, node2, node3, node4, node5, …, node n }, while the set of processing links is denoted as L = { link1 , link2 , link3 , link4 , link5 , …, link n}

This finite set (NO) of objects can exist in many network states like -Node (N) and Link (L).



 

 

 

 

The proposed method is focused on adopting the Sequential predicting model and is implemented using Keras to build the network behavior prediction model. Using this sequential method, a layer-based model will be used

 

An ADD function implemented to merge the layers proposed on this work. The first layer will be on the Convolution neural network which deals with the various input images which are considered in 2 dimensional.

The convolution neural in this proposed method uses 64 layers in the first layer and 32 layers in the second layer with so many numbers of nodes in each defined layer. This number of layers along with the nodes can be varied according to the number of instances available in the dataset. In this case the dataset will work for 32 and 64 layers where the performance depends on the Kernel size base on the dimension of the available image.

 

The kernel size 3 indicates that the model is having 3*3 matrix filter. Each model working depends on the chosen activation function for the layer. The first two layers uses Rectified Linear Activation function whereas this activation function is proven effective over other neural network. The proposed model uses the greyscale input images with a pixel of 28,28,1 with a dense layer as a output layer where the dense layer is the standard and common Apart from the 2D convolution neural network and dense layer the proposed model adopts the flatten network layer. This layer establishes the effective connection between the 2D convolution network and the dense layer networks.

The proposed model is evaluated using a single output layer with 10 nodes with a single output node 0-9 where node id starts with 0 and ends with 9 for 10 considered nodes. And the activation function is SoftMax which is the output sum of at least one node is predicted as output. The proposed models are predicted based on the node having the maximum probability of occurrence.

 

The proposed model is compiled and evaluated using loss and metrics using the optimizer parameter which controls the learning rate. The optimizer used in our method is “ADAM” which is a very good general optimizer with an adjustable learning rate. And a loss function of the proposed model is based on the binary cross entropy with a binary classification. The prediction is done with the considered dataset with an array of 10 numbers with an input images representation from 0 to 9. The prediction model is the outcome of the highest array number and it is classified based binary outputs 0 and 1.

 


Graphs:

Test Accuracy vs Training Accuracy

 

 

 

 

 

 

 

 

 

Test Loss vs Testing Loss


1st graph is all about Test accuracy vs Training accuracy

 

 

•  Accuracy metrics: The number of actual classifications / the total number of predictedclassifications.

•   The training accuracy: The samples of a model which is good enough with high metrics of accuracy

•   The test accuracy of the model with the sample dataset.

•   Overfitting: Training the model to high accuracy for the given training dataset but this may result in low accuracy for the unobserved testing samples resulting in the performance degrade of the prediction system.

•   For more information:https://stats.stackexchange.com/questions/100194/real-world- challenge-large-difference-between-training-and-testing-set-accuracy

 

2nd graph is all about Test loss vs Training loss

 

 

•   Overfitting if: training loss << validationloss

•   Underfitting if: training loss >> validationloss

•   Just right if training loss ~ validationloss

•   Training loss: Training loss is the error on the training set ofdata

•   Test loss/validation loss: Validation loss is the error after running the validation set of data through the trainednetwork.

•   Unexpectedly, if, as the epochs increase both validation and training error drop. At a certain point though, while the training error continues to drop (the network learns the data better and better) the validation error begins to rise -- thisis

overfitting!


Outputs:

Validating Output - Fault image

 





 

Validating Output - noFault image






 

IV.S YSTEM TESTING

The result of the proposed probe-based mechanism of fault detection in shown in the table 1. This result were tested in the computer with the said specification mainly focusing on the RAM and Processing.

 

Table 1: System specifications

 

 

S/No

Processor

RAM

Quantity

1

Intel i7

4 GB

1

2

Intel i3

4 GB

1

3

Intel i3

2 GB

1

4

Intel Duo Core

2 GB

3

5

Pentium IV

2 GB

2

 

 

V.PERFORMANCE EVALUATION


The performance of the proposed methods is evaluated for identifying and detecting the fault by concentrating mainly on Detection Time (DT).  The main parameters like time are the key metrics used to measure the time taken to detect the fault in the system. The detected faulty nodes information is displayed with an inactive status and a status alert over the admin system. The proposed system resulted in 0.32 seconds of time taken to detect the fault in the system with Intel i3 processor having 4 GB capacity of RAM. And for the system with more high-end specification can show more good performance with increase in RAM and Computation Power like processing. Thereby resulting in less time to detect the fault by analyzing it and reporting it to the main probe station.

 

VI.              CONCLUSION

This research work mainly aimed in resolving the issues in the fault deduction system by addressing the major challenges. There were many researches works carried out in this area for the localization and fault detection based on the LAN environment and localization.  In the proposed research work the main concept focused was the Dependency matrix mechanism which were used to design and build the active probing station mechanism. This active probing mechanism found to be effective in detecting and resolving the faulty nodes with the local area network. And the results of the proposed methods shows that it is increased with performance as well over different system specification. The result also shows the detection in time to identify the faulty nodes is depended on the processor and RAM available for the particular system in the network with an active probing mechanism.

 

 

 

 

 




Related publications

View related publications