Information Fusion Fault Prediction Method Based on Multi-Class Support Vector Machines

In this paper, a novel fault prediction method is proposed for complicated equipments, which extracts features by energy comparing method and realizes pattern recognition by multi-class support vector machines (SVM). Firstly, signal features are extracted by energy comparing method, which combines multi-resolution analysis of wavelet packets and energy spectrum. Through three scales wavelet packet decomposition, it takes energy as the feature vector of element, and establishes corresponding relations between feature and fault state. Then, using the experimental data and Riemannian geometry analysis of kernel function, an improved RBF is obtained as a new kernel function for the multi-class SVM. Conformal function of the kernel function is expressed by Euclidean distance in improved RBF, which can decrease the number of support vectors and reduces the workload, and it can fit the actual problems nicely. Finally, the multi-class SVM is adopted to realize fault information character-level fusion and pattern recognition for complicated equipments. Samples of some kinds of engine fault prediction are verified, and the result proves this method is effective and commendable.