Study On Support Vector Machine-based Faults Diagnosis For Water Distribution Systems

Support Vector Machine (SVM) is a Statistical Learning Theory(SLT) in themid-nineties evolved on the basis of machine learning techniques. Statistical LearningTheory focuses on the small sample statistical regularity and learning methods, machinelearning problem it created a good theoretical framework. Support Vector Machine inthe small sample case, the use of structural risk minimization principle, thus has a verygood learning performance.Support Vector Machine applied to fault diagnosis has the advantage that it issuitable for small sample decition, its essence is a pattern classification problems, itsnature of learning method is the ability to feature information in limited circumstancesas to maximize classification knowledge which is implicits in the data. From apromotional point of view, it is more suitable for fault diagnosis on such practicalengineering problems.This paper first describes support vector machine fault diagnosis in the field ofapplication and research status. Then introduced the statistical learning theory relatedknowledge, support vector machine principle and two major multi-class support vectormachine classification algorithm, namely "one versus the rest" and "one versus one".SVM’s performance is greatly influenced by its parameters, so there were a lot ofrelevant parameters optimized SVM algorithm. In this paper, genetic algorithm (GA)and particle swarm optimization (PSO) of the basic theory, the study was to investigatethe ability of SVM parameters optimization of its advantages and disadvantages. Studieshave found that genetic algorithm is an efficient global search capability, but there arealso disadvantages of slow convergence; while PSO algorithm has fast convergenceproperties, but also easy to fall into local optima.SVM in fault detection for vulnerable diagnostic accuracy is not high, parameteroptimization, training, slow, easy to fall into local optimal value and the shortcomingsof poor generalization ability, this paper proposes the establishment based on geneticalgorithm and particle swarm optimization algorithm for SVM parameters foroptimization algorithms.The improved algorithm to achieve a balanced global and local search; than theGA-SVM model has higher diagnostic accuracy in small samples, nonlinear case, anadvantage to solve the problem of over-learning and less learning; while its fast convergence the ability of the algorithm to converge rapidly and the optimal solutionthan GA-SVM model calculation time is short; This algorithm also has good robustnessand generalization ability.Finally, in the experimental simulation environment for small urban waterdistribution network model for fault diagnosis. Through experiments we can see theadvantages of the new diagnostic model, in the actual production practice, based onsupport vector machine fault diagnosis technology has practical significance forindustrial applications to broaden the application of the new direction.