Developing A Fault Diagnosis Approach Based On Artificial Neural Network And Self Organization Map For Occurred ADSL Faults
الموضوعات :Vahid Golmah 1 , Mina Tashakori 2
1 - Department of Computer Engineering, Neyshabur Branch, Islamic Azad University, Neyshabur,Iran
2 - Computer Engineering Department
Ferdowsi University of Mashhad
Mashhad, Iran
الکلمات المفتاحية: data mining, Self Organization Map (SOM), Fault Detection and Diagnosis (FDD), multilayer perceptron Artificial Neural Network (MLP-ANN),
ملخص المقالة :
Telecommunication companies have received a great deal of research attention, which have many advantages such as low cost, higher qualification, simple installation and maintenance, and high reliability. However, the using of technical maintenance approaches in Telecommunication companies could improve system reliability and users' satisfaction from Asymmetric digital subscriber line (ADSL) services. In ADSL systems, there are many variables giving some noise for classification and there are many fault patterns with overlapping data. Therefore, this paper proposes a multilayer perceptron (MLP) classifier integrated with Self Organization Map (SOM) models for fault detection and diagnosis (FDD) of occurred ADSL systems. The interest of this paper is to improve the performance of single MLP by dividing the fault pattern space into a few smaller sub-spaces using SOM clustering technique and triggering the right local classifier by designing a supervisor agent. The performances of this method are evaluated on the fault data of Iranian Telecommunication Company which develop ADSL services and then the proposed algorithm is also compared against single MLP. Finally, the results obtained by this algorithm are analyzed to increase user's satisfaction with reducing occurred faults for them with predicting before they face it.
1 McKay, H.: ‘Development of the contractual relationship between an ISP and its customers–Is a fairer deal in sight?’, Computer Law & Security Review, 2005, 21, (3), pp. 216-225
2 Carrera, أ., Iglesias, C.A., Garcأa-Algarra, J., and Kolaإ™أk, D.a.: ‘A real-life application of multi-agent systems for fault diagnosis in the provision of an Internet business service’, Journal of Network and Computer Applications, 2014, 37, pp. 146-154
3 Hanafizadeh, P., and Mirzazadeh, M.: ‘Visualizing market segmentation using self-organizing maps and Fuzzy Delphi method–ADSL market of a telecommunication company’, Expert Systems with Applications, 2011, 38, (1), pp. 198-205
4 Gao, Y., Liu, S., Li, F., and Liu, Z.: ‘Fault detection and diagnosis method for cooling dehumidifier based on LS-SVM NARX model’, International Journal of Refrigeration, 2015
5 Muralidharan, V., Sugumaran, V., and Indira, V.: ‘Fault diagnosis of monoblock centrifugal pump using SVM’, Engineering Science and Technology, an International Journal, 2014, 17, (3), pp. 152-157
6 Hakim, S.J.S., Razak, H.A., and Ravanfar, S.A.: ‘Fault diagnosis on beam-like structures from modal parameters using artificial neural networks’, Measurement, 2015, 76, pp. 45-61
7 Lashkari, N., Poshtan, J., and Azgomi, H.F.: ‘Simulative and experimental investigation on stator winding turn and unbalanced supply voltage fault diagnosis in induction motors using Artificial Neural Networks’, ISA Transactions, 2015
8 Lau, C.K., Ghosh, K., Hussain, M.A., and Hassan, C.R.C.: ‘Fault diagnosis of Tennessee Eastman process with multi-scale PCA and ANFIS’, Chemometrics and Intelligent Laboratory Systems, 2013, 120, pp. 1-14
9 Mnassri, B., and Ouladsine, M.: ‘Reconstruction-based contribution approaches for improved fault diagnosis using principal component analysis’, Journal of Process Control, 2015, 33, pp. 60-76
10 Huang, J., and Yan, X.: ‘Dynamic process fault detection and diagnosis based on dynamic principal component analysis, dynamic independent component analysis and Bayesian inference’, Chemometrics and Intelligent Laboratory Systems, 2015
11 Zhao, Y., Wen, J., and Wang, S.: ‘Diagnostic Bayesian networks for diagnosing air handling units faults–Part II: Faults in coils and sensors’, Applied Thermal Engineering, 2015, 90, pp. 145-157
12 Zhang, Z., and Dong, F.: ‘Fault detection and diagnosis for missing data systems with a three time-slice dynamic Bayesian network approach’, Chemometrics and Intelligent Laboratory Systems, 2014, 138, pp. 30-40
13 Eslamloueyan, R.: ‘Designing a hierarchical neural network based on fuzzy clustering for fault diagnosis of the Tennessee–Eastman process’, Applied soft computing, 2011, 11, (1), pp. 1407-1415
14 Rad, M.A.A., and Yazdanpanah, M.J.: ‘Designing supervised local neural network classifiers based on EM clustering for fault diagnosis of Tennessee Eastman process’, Chemometrics and Intelligent Laboratory Systems, 2015
15 Mendoza, O., Melأn, P., and Castillo, O.: ‘Interval type-2 fuzzy logic and modular neural networks for face recognition applications’, Applied soft computing, 2009, 9, (4), pp. 1377-1387
16 Sanchez, D., Melin, P., and Castillo, O.: ‘Optimization of modular granular neural networks using a hierarchical genetic algorithm based on the database complexity applied to human recognition’, Information Sciences, 2015, 309, pp. 73-101
17 Sanchez, D., and Melin, P.: ‘Optimization of modular granular neural networks using hierarchical genetic algorithms for human recognition using the ear biometric measure’, Engineering Applications of Artificial Intelligence, 2014, 27, pp. 41-56
18 Chicco, G.: ‘Overview and performance assessment of the clustering methods for electrical load pattern grouping’, Energy, 2012, 42, (1), pp. 68-80
19 Lopez, J.J., Aguado, J.A., Martin, F., Munoz, F., Rodriguez, A., and Ruiz, J.E.: ‘Hopfield K-Means clustering algorithm: A proposal for the segmentation of electricity customers’, Electric Power Systems Research, 2011, 81, (2), pp. 716-724
20 Guerrero, J.D.M., Marcelli, D., Soria-Olivas, E., Mari, F., Martinez-Martأnez, J.M., Bech, I.S., Martinez-Sober, M., Scatizzi, L., Gimez-Sanchis, J., and Stopper, A.: ‘Self-Organising Maps: A new way to screen the level of satisfaction of dialysis patients’, Expert systems with applications, 2012, 39, (10), pp. 8793-8798
21 Segev, A., and Kantola, J.: ‘Identification of trends from patents using self-organizing maps’, Expert systems with applications, 2012, 39, (18), pp. 13235-13242
22 Davies, D.L., and Bouldin, D.W.: ‘A cluster separation measure’, Pattern Analysis and Machine Intelligence, IEEE Transactions on, 1979, (2), pp. 224-227
23 Szymczyk, P., and Szymczyk, M.: ‘Classification of geological structure using ground penetrating radar and Laplace transform artificial neural networks’, Neurocomputing, 2015, 148, pp. 354-362
24 Prakash, M., Pradhan, S., and Roy, S.: ‘Soft computing techniques for fault detection in power distribution systems: A review’, in Editor (Ed.)^(Eds.): ‘Book Soft computing techniques for fault detection in power distribution systems: A review’ (IEEE, 2014, edn.), pp. 1-6
25 Chojaczyk, A.A., Teixeira, A.P., Neves, L.C., Cardoso, J.B., and Soares, C.G.: ‘Review and application of Artificial Neural Networks models in reliability analysis of steel structures’, Structural Safety, 2015, 52, pp. 78-89
26 Palanichamy, A., Jayas, D.S., and Holley, R.A.: ‘Predicting survival of Escherichia coli O157: H7 in dry fermented sausage using artificial neural networks’, Journal of Food Protectionآ®, 2008, 71, (1), pp. 6-12
27 Gosukonda, R., Mahapatra, A.K., Liu, X., and Kannan, G.: ‘Application of artificial neural network to predict< i> Escherichia coli</i> O157: H7 inactivation on beef surfaces’, Food Control, 2014, 47, pp. 606-614
28 Li, Y., Xia, J., Zhang, S., Yan, J., Ai, X., and Dai, K.: ‘An efficient intrusion detection system based on support vector machines and gradually feature removal method’, Expert Systems with Applications, 2012, 39, pp. 424–430
29 Esa Alhoniemi, J.H., Jukka Parviainen and Juha Vesanto: ‘Som toolbox for matlab’, in Editor (Ed.)^(Eds.): ‘Book Som toolbox for matlab’ (1999, edn.), pp.
30 Vesanto, J., Himberg, J., Alhoniemi, E., and Parhankangas, J.: ‘Self-organizing map in Matlab: the SOM Toolbox’, in Editor (Ed.)^(Eds.): ‘Book Self-organizing map in Matlab: the SOM Toolbox’ (1999, edn.), pp. 16-17
31 Kuzmanovskia., I., Trpkovskaa., M., and Soptrajanov, B.: ‘Optimization of supervised self-organizing maps with genetic algorithms for classification of urinary calculi’, Journal of Molecular Structure, 2005, 744–747, pp. 833–838
Developing A Fault Diagnosis Approach Based On Artificial Neural Network And Self Organization Map For Occurred ADSL Faults
Abstract— Telecommunication companies have received a great deal of research attention, which have many advantages such as low cost, higher qualification, simple installation and maintenance, and high reliability. However, the using of technical maintenance approaches in Telecommunication companies could improve system reliability and users' satisfaction from Asymmetric digital subscriber line (ADSL) services. In ADSL systems, there are many variables giving some noise for classification and there are many fault patterns with overlapping data. Therefore, this paper proposes a multilayer perceptron (MLP) classifier integrated with Self Organization Map (SOM) models for fault detection and diagnosis (FDD) of occurred ADSL systems. The interest of this paper is to improve the performance of single MLP by dividing the fault pattern space into a few smaller sub-spaces using SOM clustering technique and triggering the right local classifier by designing a supervisor agent. The performances of this method are evaluated on the fault data of Iranian Telecommunication Company which develop ADSL services and then the proposed algorithm is also compared against single MLP. Finally, the results obtained by this algorithm are analyzed to increase user's satisfaction with reducing occurred faults for them with predicting before they face it.
Keywords— Fault Detection and Diagnosis (FDD), Data mining, Self Organization Map (SOM), multilayer perceptron Artificial Neural Network (MLP-ANN)
1- Introduction
The Internet has revolutionized the world of people life over recent years. Anyone who wishes to place a web page on the Internet, or undertake e-mail communications will invariably have to enter into a contract for Internet access [1]. Such Internet services could be provided by different telecommunications companies. Whilst there is a significant array of companies which can provide Internet services, it would appear that technical advancement may well reduce competition for the provision of such services. Should this trend persist, then it is likely that the remaining companies will face far more stringent regulation [1].
Fault Detection and Diagnosis (FDD) is one of the most important technical approaches in network management tasks for telecommunications companies. Traditionally, this process has been carried out by humans and software systems that work in a cooperative way. The constant increase in the size and complexity of the network makes fault diagnosis a critical task that should be handled quickly and in a reliable way. Highly skilled engineers are required to carry out this task, although even these individuals are not always able to deal with the increasing heterogeneity and complexity of the networks. Although automated fault diagnosis processes have been developed, such as surveillance systems for symptom detection in the core or backbone networks, fault diagnosis is mainly a manual process managed by human operators [2].
Telecommunication industry is a growing industry in Iran which has an enormous potential market for different telecommunications services [3]. Comparing to all the existing researches, there aren’t many researches on analyzing the Iran satisfaction improvement of the ADSL users’ with reducing faults. Moreover, many Iranian telecommunication companies do not fallow any clear strategies for fault diagnosis and solve it before users face them. Our main contribution is based on the real register ADSL faults to diagnosis them for improving users’ satisfaction. Fig. 1 shows the picture of the ADSL equipment and how ADSL’s work is illustrated in under of figures.
Fault Detection and Diagnosis can be seen as a supervised classification task whose objective is to assign new observations to one of the existing classes. Various supervised learning models such as support vector machines (SVM)[4, 5], Artificial Neural Networks (ANN) [6, 7], adaptive network-based fuzzy inference system (ANFIS)[8], Principal Component Analysis (PCA) [9, 10] and the Bayesian network classifier[11, 12] have been applied for fault classification of industrial processes. Among the mentioned methods, the artificial neural network of multilayer perceptron (MLP) type in Fault Detection and Diagnosis has received considerable attention over the last three decades [13] and high efficiency as non-linear classifier. However, its application in large scale industries is not without difficulties[13, 14].Despite simplicity and versatility of MLP networks, their application to Fault Detection and Diagnosis in wide scale systems is not without difficulties. One of the most important issues of employing MLP networks in Fault Detection and Diagnosis is related to its training. As the size and complexity of a system increases, the number of measured variables and fault patterns rise. This results in a bigger and more complex MLP network to diagnose the system faults properly. Sometimes the structure of the network becomes so large that it could not learn all fault patterns with an acceptable performance goal. To overcome this problem in artificial neural network, hierarchical and modular artificial neural networks (HANN) have been proposed, developed and applied by researchers [15-17] in different research areas. According to these research works, diagnosis of single faults and simultaneous double and triple faults are hierarchically divided among neural networks specifically trained for that purpose. The concept behind this technique is that, instead of using a single neural network for diagnosing all faults, whole pattern space of faults is divided in a hierarchical way into a number of smaller sub-spaces. For each set of faults, in a particular sub-space, a special diagnostic agent is trained. Clustering methods may be employed to divide the whole space of fault patterns into alike fault pattern sets based on a similarity criterion [13].
FIG. 1. (a) Main Distribution Frame (MDF), (b) ADSL Panel Blocks and (c) ADSL Line Cards. From the MDF panel, telephone line will go to an ADSL line card, each line card ties up and labeled according to the MDF panel which your line is attached to.
In ADSL systems, there are many variables giving some noise for classification and there are many fault patterns with overlapping data. As a result, the MLP classifier cannot learn all the fault patterns with an acceptable performance goal. Therefore, the aim of this article is to present a new method for improving the performance of MLP classifier for fault diagnosis of ADSL systems. This method is based on dividing the huge amount of fault pattern dataset into smaller fault subset using the Self Organization Map (SOM) clustering algorithm. At each specific fault pattern subset, a local MLP classifier is trained. For a new test observation, the SOM clustering algorithm acts as a supervisor agent by assigning the data to the cluster corresponding to the highest mean similarity.
2- Methodology
The idea behind this approach is to provide mechanisms for improving customer relationship management for an Iranian telecommunication company which develop ADSL services through increasing user's satisfaction with reducing occurred faults for them. There are various methods related to solve Fault Detection and Diagnosis problem in various ways. Among the mentioned methods, the MLP is more popular due to its simplicity and high efficiency as non-linear classifier. However, it hasn’t suitable performance in large scale systems. In regard to high variables, numerous noise data and overlapping data in occurred ADSL fault, a single MLP cannot learn all the fault patterns with an acceptable performance goal. Therefore, proposed approach to this problem is based on dividing the occurred faults for ADSL's user dataset into smaller fault subset using the SOM clustering algorithm and training MLP for patterns of faults any cluster. FIG. 2 shows our model based on collecting, preparing, clustering and training process schematically for Fault Detection and Diagnosis in ADSL systems. This Fig. provides a general overview of our model, specifying the components and characteristics.
FIG. 2 Developed methodology diagram for detecting and diagnosing of occurred faults for ADSL’
2-1- Data gathering and preparing
The Collecting and preparing stage starts with an initial data collection and proceeds with activities in order to get familiar with the data, to identify data quality problems, to discover first insights into the data or to detect interesting subsets to form hypotheses for hidden information. This phase covers all activities to construct the final dataset (data that will be fed into the modeling tool(s)) from the initial raw data. Data preparation tasks are likely to be performed multiple times and not in any prescribed order.
2-2- Clustering phase
The clustering methodologies are generally used to perform load pattern grouping with grouping them in such a way that patterns in the same group are more similar (in some sense or another) to each other than to those in other groups. In particular, it is possible to identify unsupervised learning based techniques, such as supervised learning-based techniques (such as the ones adopting multilayer perceptron or Elman neural networks), the Kohonen’s self organising map (SOM), or vector quantization, fuzzy logic-based techniques, statistical techniques such as k-means (KM) and multivariate analysis, and hybrid techniques such as probability neural networks (PNN) and fuzzy k-means (FKM)[18, 19].
All these techniques have advantages include high speed, the possibility of re-allocations in successive iterations and the fact that they do not require previous training but their obtained solutions depend on the initial centroids and number of predefined clusters (k-means (KM) and fuzzy k-means (FKM)), the internal distance threshold (follow the leader (FDL) algorithm), the initial neuron weight (The Adaptive Vector Quantization (AVQ) method) and they suffer from non-reallocation of data in successive iterations (hierarchical algorithm) and lack of objective function (hierarchical algorithm) [18]. Among these methods, SOM has other important advantages with less disadvantages. The SOM modifies the search space to represent the results on a bi-dimensional map, but does not generate the final clusters directly. Hence, a post-processing stage is needed to form the clusters, with arbitrary assumptions, so that different numbers of clusters can be formed starting from the same SOM outcomes, by using a specific technique to identify the final clusters. Likewise, support vector clustering (SVC) requires a first stage in which the support vectors are formed, followed by a second stage in which the groups are formed for the desired number of clusters. SOM has able to learn both the distribution and the topology of the input vector training set, and it also allows to identify new patterns and the interpretation and visualization of the results with the unified distance matrix method or U-matrix. The most appealing characteristic of SOM is that the underlying mathematics ensure that the map is a faithful representation of the original data, e.g. two data points (two faults in our case) are represented close to each other in the resulting map when they have similar features (e.g. Row, Terminal and etc.). Therefore, the U-matrix can be used to discover otherwise invisible relationships in a high-dimensional data space and classify data sets into clusters of similar values.[20, 21].
In order to dividing the huge amount of fault pattern dataset into smaller fault subset we use SOM clustering algorithm and to confront the randomness of the solution obtained drawback in the SOM algorithms we can run the algorithm more. We used the Davies–Bouldin index (DBI) [22] to measure the performance because of its popularity for cluster evaluation. Low DBI values indicate high clustering performance.
2-3- Training phase
Classification is a problem of identifying to which set of categories a new observation belongs, on the basis of a training set of data. Containing observations whose category membership is known. An algorithm that implements classification is known as a classifier [23].
Each classifier has its own strengths and weaknesses. The most widely used classifiers are the neural networks. Neural networks have been selected for further consideration because they are simple, easy to build, efficient and reliable classifier and they are also give good results [23]. An artificial neural network is a system based on the operation of biological neural networks, in other words, is an emulation of the central nervous system in animals. Neural networks comprise of many artificial fundamental units, called neurons. An artificial neuron is simply a computational model of a biological neuron [24].
In the ANN, the neuron is a processing element with several inputs and one output. Each neuron m receives an input signal vector X =x1 + x2 + … + xn from n input channels. Next, the weighted sum of x is calculated by multiplying each element xk by a coefficient wmk demonstrating adequate importance of the input channel k. The m-neuron activation a m is given by [25]:
(1) |
| (2) |
| (3) |
| (4) |
ِDescription | Attribute |
User's name | A1 |
User's telephone number. | A2 |
Service name | A3 |
Center name | A4 |
Row | A5 |
Terminal | A6 |
Port | A7 |
Port state | A8 |
User address | A9 |
User mobile number. | A10 |
Register date | A11 |
Registrar | A12 |
Fault title | A13 |
Fault description | A14 |
The data sources are first located, accessed, and integrated. Next, selected data is put into a tabular format in which instances and variables take place in rows and columns, respectively. A labeling is used for a better understanding of results. By means of this labeling, a number is assigned to each registered fault.
Since the input to the data mining model affects the choice of a data mining algorithm and the results, we attempted to remove polluted data such as incorrectly coded input (e.g. typos) or inconsistent input (e.g. outliers or anomalous answers) from the database by filtering out the Excel file. Effective (2192) faults were amassed other than 47 invalid pieces containing omissions and incomplete registered faults.
3-2- SOM Training and Evaluation
One batch SOM algorithm is used for designing SOM and training it. Because it is more stable than the online version, as its results do not depend on the order in which the input patterns are presented to it. Therefore, the integration of the batch SOM in an industrial environment is more feasible, since it is a deterministic and reproducible algorithm. In addition, the batch version is faster and can be parallelized to reduce computational time.
All the simulations were programmed using Matlab 2014a and the specific Matlab SOM Toolbox [29]. See [30] for further references on the design of SOM networks.
Before the training process begins, data normalization is often performed. The linear transformation formula to [0, 1] is used:
Method | DBI |
k-means | 0.93 |
SOM | 0.90 |
AVQ | 1.23 |
SVC | 1.08 |
The DBI value for these four methods indicate SOM has the lowest DBI and as it has highest cluster quality. In general, however, the overall performance of SOM is better from others in this dataset. Therefore, we divide the whole dataset of faults into alike fault pattern sets based on a similarity criterion by using of SOM method and feed them to MLP for more precise classification.
The designed MLP is trained based on specified parameters in previous section. By using of Eq. (3) and average results of 10-fold validation, MSE calculated as 0.04425. These low amounts for MSE can demonstrate the appropriateness and precise of our modeling and forecasting. Whereas, alone MLP results 0.09641 as MSE measurement. Therefore, the proposed approach based on SOM and MLP has better classification precision.
Second experiment is for detecting new types of fault during the monitoring process. 438 new observations of ADSL faults (1-6) have been monitored and classified as new type of fault. According to TABLE 3, 414 of 438 new observations are correctly classified as new faults, i.e., the classification error rate of the classifier is 5.48%. Therefore, the proposed hybrid approach based on SOM and MLP can detect new types of ADSL faults with high performance (correct classification rate of 94.52%) without reducing the performance of discriminating between known fault classes (misclassification rate has increased from 22.58% to 23.21%).
TABLE 3. Confusion matrix of proposed hybrid approach based on SOM and MLP for ADSL
67 | 2 | 1 | 2 | 0 | 1 |
1 | 69 | 0 | 1 | 1 | 1 |
1 | 0 | 70 | 0 | 0 | 2 |
2 | 1 | 0 | 68 | 0 | 0 |
0 | 1 | 2 | 2 | 71 | 0 |
3 | 0 | 0 | 0 | 1 | 69 |
5- Conclusion
We developed expert system based on Artificial Neural Network (ANN) use a MLP with clustered data by using of SOM for fault diagnosis in ADSL services. The 10-fold cross validation for DBI value and MSE value is used for evaluating of SOM and MLP respectively. Low amounts for resulted DBI value for SOM compared with other used clustering methods (KM, SOM, AVQ and SVC) demonstrate the appropriateness of SOM to divide ADSL faults dataset faults into alike fault pattern sets. Moreover, the lower MSE value for MLP on clustered data compared with whole data demonstrate better forecasting of our proposed hybrid approach.
In this study, we regard 6 ADSL’ faults, whereas variety of ADSL faults is high. Supposing more ADSL faults can be subject of future research. Moreover, we developed a new proposed approach based on MLP and SOM to data classify, but it is has evaluated by using it to diagnosis faults of ADSL. Whereas it is not sufficient to evaluate performance of a new proposed approach. Therefore, performance evaluation of proposed approach based on MLP and SOM by using it in other fields can regard as future researches.
6- References