Valve Fault Diagnosis Research Based On SVM

Valve is one of the key equipment of the reciprocating compressor, also because it operates at high frequency frequently by vibration and shock, the failure rate is much higher than other components, so the research on valve fault diagnosis is full of significance. Terrible working environment and complicated fault type of valve greatly increase the difficulty of fault diagnosis. The traditional fault diagnosis method obviously can not meet the requirements of the enterprise to the valve fault diagnosis, while the fault diagnosis method based on artificial intelligence will be the future trend.In this paper, fault information of the original signal is not obvious because of the noise disturbance. Empirical mode decomposition(EMD) method is adopt to decomposed original signal into several smooth intrinsic mode functions(IMF), according to the change of the signal complex degree, sample entropy is introduced for describing this kind of fault information, then use the Hilbert transform to draw the Hilbert spectrum for IMF, energy is proposed to represent the fault information by analysing the change of vibration energy. Due to the limited valve failure data, this paper sets support vector machine(SVM) as s classifier which has a good performance on the recognition of small sample. The sample entropy and energy of IMF are used as inputs of SVM, then compare the ability of the cross validation, genetic algorithm and particle swarm optimization algorithm in parameters optimization, finally, choose the cross validation as the parameter selection method for training SVM in this paper.Feature selection is presented to improve the test accuracy, firstly, Relief F weights and cross validation accuracy is putting forward to measure feature performance, then eliminate redundancy between features by calculating their Pearson correlation coefficient, finally, Sequence backward selection(SBS) method is applied to search the best subset, experiments prove that feature selection has the good effect in determining the fault type. Meanwhile, according to the weakness of cross validation accuracy in the evaluation of feature performance, putting forward an improved feature selection method based on SVM-SBS, verified effectively by experiment.