COMPRESSED SENSING FOR MATRIX TRAFFIC RECONSTRUCTION

• Main Author: Dyah Irawati, Indrarini
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/40484
• Institution: Institut Teknologi Bandung

Compressed sensing/sampling (CS) is a new paradigm in the field of signal processing that has been widely applied to various applications such as compression of video and audio signals, direction of arrival estimation on radar, weather radar detection, telecommunication traffic modeling, and others. This theorem utilizes signal sparsity in the transformation region to reduce the number of samples, which sampled below the Shannon-Nyquist sampling rate. In this study, the CS technique was applied to reconstruct of internet traffic matrix in the internet network. This is useful for monitoring network traffic, predicting links sensitive, and predicting anomalous events. The traffic matrix is a representation of traffic that flows between routers on the network at certain times of observation. Exploration of the sparsity in the traffic matrix is done by comparing the sparsity technique consisting of Principal Component Analysis (PCA), Singular Value Decomposition (SVD), and Singular Value Decomposition Mean (SVDM). In internet traffic data, the amount of energy concentration after transformation is expressed as rank. This study uses rank as a parameter to express information misery in the sparsity region. The test results on the use of rank determine that the SVD technique is best used to obtain sparsity in internet traffic data. In normal traffic conditions, the minimum rank for reconstruction with NMSE targets < 20% is 10%. Whereas in traffic that is randomly missed, the minimum limit of rank is 60%. The acquisition scheme was obtained from an experiment of eight measurement matrices generated randomly using Uniform, Normal, Binary, Half-normal, Log-normal, Binomial, Poisson, and Exponential distributions. In the simulation, the measurement matrix measuring m × r is used, by testing for a different number of measurements and m <r. Test parameters are expressed in a minimum compression ratio (CR), which is 1 with the smallest error. The simulation results show that the measurement matrix with the Binomial distribution produces the smallest error of the reconstruction results. The reconstruction algorithm in the CS scheme consists of two major schemes, namely Base Pursuit (BP) that meets the minimum l1-norm and greedy. The l1-norm based algorithm received considerable attention among researchers because it produced a good accuracy in the results of reconstruction, but this algorithm has disadvantages because computing is quite heavy. The greedy algorithm excels in terms of computational speed with deficiencies in reconstruction results. In this study, we used the algorithm for reconstructing SVD????1, IRLS, and Orthogonal Matching Pursuit (OMP) algorithms. The focus of this research is to compile CS modeling for the internet traffic matrix reconstruction, which is represented spatial-temporally. Traffic matrix sparsity is obtained from SVD decomposition. The use of a measurement matrix is generated randomly from the Normal distribution. Testing of measurement number is done to find out the minimum number of samples resulting in a reconstruction matrix with an error of <20%. Testing the success of reconstruction is done by removing elements in the traffic matrix randomly with the probability of missing values of 2% -98%. In addition, compressive sensing modeling was developed with the scheme of sparse vectors and the sparse matrix applied to the SVD????1, IRLS, and OMPO, EDMIN reconstruction algorithms. To improve the performance of the OMP reconstruction algorithm are carried out with the addition of optimized interpolations to overcome zero value problems in the reconstruction results. As well as the proposed new reconstruction algorithm, namely EDMIN. In addition, performance improvements were also made after SVD reconstruction, namely with the addition of Bilinier interpolation. Simulation results show that the proposed model can improve accuracy by reducing NMSE, overcoming missing six-type traffic problems, namely Missing Row Elements (MRE), Missing Column Elements (MCE), Missing Rows at Random (MRR), Missing Columns at Random ( MCR), Missing Elements at Random (MER), and Combine Missing Patterns (CMP), determine the location of sensitive links, and detect the sensitive time.