A Study Of Sub-healthy Recognition Algorithm Based On Particle Swarm Neural Network And D-S Theory

With the development of science and technology, modern industry is graduallyto large-scale production equipment, complex, high-speed and automation direction.The development of the theory of fault diagnosis will promote the rapid developmentand wide application of the fault monitoring and surveillance system, which canfurther improve the system reliability and security, and the resulting huge economicand social benefits.This paper is made on the background of the rolling bearing fault diagnosisresearch. Wavelet denoising is selected to reduce the noise of the collected data forits own characteristics, and a wide set of features which can better reflect thecharacteristics of the failure are selected. Two diagnostic models are establishedbased on the strong nonlinear mapping ability of neural network, particle swarmalgorithm for global optimization and D-S evidence theory for dealing withuncertainty information. Based on the two models, we investigate their application inball bearing fault diagnosis.In order to overcome shortcomings of traditional BP neural network such as lowstudy efficiency, slow convergence speed, easily trapped into local optimal solution,in this paper, this paper proposes a new improved particle swarm optimization (IPSO)algorithm. This algorithm adjusts the inertia weight coefficients and learning factorsadaptively and therefore can be used to optimize the weights in the BP network.After establishing the improved PSO-BP (IPSO-BP) model, it is applied to solvefault diagnosis of rolling bearing. At the same time, the sub-health is defined forfault types in this paper. To avoid the misclassification caused by the BP neuralnetwork, the DS evidence theory is used to fuse the output of the BP neural networkand the probability given to the type of fault so that we can get the final failureclassification.Simulation experiment is made under the MATLAB platform. The proposedIPSO-BP algorithm is effective by comparing with the traditional BP, PSO-BP and linear PSO-BP (LPSO-BP) algorithms. And IPSO-BP network outperforms aboveother algorithms with faster convergence speed, lower errors, higher diagnosticaccuracy and learning ability. The D-S evidence theory is used to fuse the output ofthe BP network, which can overcome misclassification of traditional BP network anduncertainty classification to some extent. It is proved that the method has betterclassification ability than a separate BP network, and the diagnosis result is quitegood.