Salah Zidi

Least squares support kernel machines (LS-SKM) for identification

This paper presents a novel approach for non-linear systems identification called ‘least squares support kernel machines (LS-SKM)’. Instead of using a least squares support vector machines (LS-SVM) with a single kernel function, the proposed approach combines several kernels in order to take advantage of their performances and also reflects the fact that practical learning problems often involve multiple, heterogeneous data sources. The idea is to divide the regressor vector in several regressor vectors, and, for each vector a kernel function is used. The choice of kernel function and the corresponding parameters is an important task which is related to the non-linear system degrees. A constrained particle swarm optimisation (CPSO) is used to give solution for the determination of optimised kernel parameters. Two examples are presented for qualitative comparison with the classical LS-SVM. The results reveal the accuracy and the robustness of the obtained model based on our proposed hybrid method.

Feature Selection using an SVM learning machine

In this paper we suggest an approach to select features for the Support Vector Machines (SVM). Feature selection is efficient in searching the most descriptive features which would contribute in increasing the effectiveness of the classifier algorithm. The process described here consists in backward elimination strategy based on the criterion of the rate of misclassification. We used the tabu algorithm to guide the search of the optimal set of features; each set of features is assessed according to its goodness of fit. This procedure is exploited in the regulation of urban transport network systems. It was first applied in a binary case and then it was extended to the multiclass case thanks to the MSVM technique: Binary Tree.

Tennessee Eastman Process Diagnosis Based on Dynamic Classification With SVDD

The support vector domain description (SVDD) is an efficient kernel method inspired from the SV machine (SVM) by Vapnik. It is commonly used for one-classification problems or novelty detection. The training algorithm solves a constrained convex quadratic programming (QP) problem. This assumes prior dense sampling (offline training) and it requires large memory and enormous amounts of training time. In this paper, we propose a fast SVDD dedicated for multiclassification problems. The proposed classifier deals with stationary as well as nonstationary (NS) data. The principle is based on the dynamic removal/insertion of informations according to adequate rules. To ensure the rapidity of convergence, the algorithm considers in each run a limited frame of samples for the training process. These samples are selected according to some approximations based on Karush–Kuhn–Tucker (KKT) conditions. An additional merge mechanism is proposed to avoid local optima drawbacks and improve performances. The developed method is assessed on some synthetic data to prove its effectiveness. Afterward, it is employed to solve a diagnosis problem and faults detection. We considered for this purpose a real industrial plant consisting in Tennessee Eastman process (TEP).

A nonlinear MIMO system identification based on improved Multi-Kernel Least Squares Support Vector Machines (Improved Multi-Kernel LS-SVM)

In this paper, a new method for the identification of nonlinear Multiple Input-Multiple Output (MIMO) systems is proposed. An improved Multi-Kernel Least Squares Support Vector Machines (Improved Multi-Kernel LS-SVM) based on Constrained Particle Swarm Optimization (CPSO) is given. The basic LS-SVM idea is to map linear inseparable input data into a high dimensional linear separable feature space via a nonlinear mapping technique (kernel function) and to carry out linear classification or regression in feature space. The choice of kernel function and the corresponding parameters is an important task which is related to the system nonlinearity degrees. The suggested approach combines several kernels in order to take advantage of their performances. The CPSO technique is used to give solution for the determination of optimized kernel parameters and their evolved weights. Simulation results show that the CPSO can quickly obtain the optimal parameters and therefore satisfying the required precision.

New approach for systems monitoring based on semi-supervised classification

In this paper, we consider the problem of fault diagnosis for systems with many possible functioning modes. A new methodology has been proposed combining both supervised and unsupervised learning methods. Since supervised learning requires necessarily a broad labelled base that may not always available in a sufficient cardinality, we aim at first an unsupervised grouping of a critical faults set (classes) though a Self-Adaptive Clustering Algorithm (SACA). Within this framework, the presented algorithm is based on the evaluation of a metric distance between cluster centroids and samples. An integrated process for optimization allows the tuning of confidence threshold for decision. Next, an additional supervised classification step using Artificial Neural Network (ANN) provides practical information for decision-making. The network is trained according to the classification multi-levels dedicated for multi-class problems. The developed approach is assessed on a hydraulic system consisting of three connected tanks.

A New Feature Extraction Method Based on Clustering for Face Recognition

When solving a pattern classification problem, it is common to apply a feature extraction method as a pre-processing step, not only to reduce the computation complexity but also to obtain better classification performance by reducing the amount of irrelevant and redundant information in the data. In this study, we investigate a novel schema for linear feature extraction in classification problems. The method we have proposed is based on clustering technique to realize feature extraction. It focuses in identifying and transforming redundant information in the data. A new similarity measure-based trend analysis is devised to identify those features. The simulation results on face recognition show that the proposed method gives better or competitive results when compared to conventional unsupervised methods like PCA and ICA.

Feature extraction for atmospheric pollution detection

Atmospheric data sets are represented by an amount of heterogeneous and redundant data. As number of measurements grows, a strategy is needed to select and efficiently analyze the useful information from the whole data set. The aim of this work is to propose a feature extraction technique based on construction of clusters of similar features. The main objective of the proposed process is to attempt to reach a more accurate classification task and to achieve a more compact representation of the underlying structure of the data. The paper reports the results obtained using the above extraction and analysis procedure of a real data set on atmospheric pollution. It is shown that the proposed approach is able to detect underlying relationship between features and thus get to ameliorate classification accuracy rate.

An Efficient Density-Based Algorithm for Data Clustering

In this paper, we propose a novel density-based clustering method in which we deal with data appearing sequentially. In data mining, a cluster is a high-density region gathering a set of objects which are similar according to a prefixed criterion. For purposes of modelling, we restrict a cluster to be the contour of the region including these objects. The bounded contour function is obtained by applying a B-spline interpolation on the convex hull vertices enclosing the cluster. This procedure, named Cluster Domain Description (CDD), may give a realistic approximation of the cluster area. The clustering process is achieved afterwards with respect to the variation of the internal density of that area. In order to improve performances, a supplementary merge mechanism of evolving clusters is as well proposed. The method is assessed firstly on artificially generated data, and then on data extracted from a chemical system consisting of the Tennessee Eastman Process.

Support kernels regression for NARMA system identification (SKR NARMA)

This paper deals with the identification of nonlinear systems using multi-kernel approach. First, we have improved the support vector regression (SVR) method in order to identify nonlinear complex system. Our idea consists of dividing the regressor vector in several blocks, and, for each one a kernel function is used. This blockwise SVR approach is called support kernel regression (SKR). Furthermore, we have proposed two methods, SKR(rbf-lin) and SKR(rbf-rbf). Second, the SKR approach is improved to deal with the problem of NARMA system identification. Therefore, a new method called support kernel regression for NARMA model (SKRNARMA) is suggested. The basic idea is to consider the terms of auto-correlation and cross-correlation of the nonlinearity of input output discrete time processes, and for every term a kernel function is used. An example of MIMO system is presented for qualitative comparison with the classical SVR approach based on a single kernel function. The results reveal the accuracy and the robustness of the obtained model based on our proposed (SKRNARMA)-based approach.

Système d'aide à la régulation et à la reconfiguration des réseaux de transports: SVM et algorithme à colonie de fourmis

The public transport networks regulation is a real-time task and it becomes more and more difficult for a human operator, with a number of stations, vehicles and modes that does not stop increasing. Sometimes there are very complicated cases of disturbances and the regulator must propose a new planning of the network with a spatial reconfiguration and an hourly regulation. It is a task even more delicate than the regulation. This article proposes a regulation and reconfiguration decision support system. We use a first classification algorithm SVM (Support Vectors Machines) for the regulation and the second ant colony algorithm for the spatial and hourly reconfiguration. In this approach, we used a new idea of dynamic local search. The obtained results confirm the good performance of both approaches for the adaptation of a multimodal transport network to the real exploitation conditions.