A Novel Binary Chaotic Cloud Flower Pollination Algorithm for Analog Fault Diagnosis

In analog fault diagnosis, support vector machine (SVM) classifier is a popular diagnosis technique. Before training the SVM classifier, feature extraction is an essential preprocessing step to enhance the identifiability of the fault samples. However, in the feature set, there exist the redundant and uncorrelated features, which have the inferior influence on the learning performance of the SVM classifier. Feature selection is to select the most representative feature subset to enhance the diagnosis accuracy. In addition, the penalty coefficient and kernel width parameter in SVM also impact the classification performance. Obviously, the selection of feature subset and SVM parameters can be considered as an optimization problem. Flower pollination algorithm (FPA) as a swarm intelligence optimization is derived from the pollination behavior of flowers. In this paper, we propose a novel binary version of FPA and then incorporate a cross search strategy based on chaos theory and cloud model into the novel BFPA to further enhance the solution quality, named BCCFPA. Moreover, different kinds of features are extracted by various wavelet basis functions based wavelet packet energy spectrum (wpes) and wavelet packet energy entropy (wpee) methods. The original feature set is built by combining these feature vectors. BCCFPA is used to simultaneously search the optimal feature subset and SVM parameters. The simulation experimental results prove that BCCFPA can effectively balance the diagnosis accuracy and the dimension of feature subset, and achieves the better diagnosis performance when compared with other methods.