Research On Fault Prognosis Method And Application Of Equipment Based On Extreme Learning Machine

In order to improve the use efficiency, reduce the maintain cost, avoid the major accident and realize the fault prognosis of equipment, extreme learning machine and the methods related with artificial neural network was introduced and researched in this paper. The rolling bearings and gear was the object of study, and the time-domain feature parameters and morphology fractal dimension (MFD) of vibration signal was regarded as the prediction features. The gray neural network, Elman feedback neural network and extreme learning machine (ELM) was applied to realize the fault prognosis. Comparing with the traditional BP neural network, it proved that the ELM model was efficient and accurate in fault prognosis. This paper was divided into three following sections.Firstly, the method of gray neural network was studied. The time-domain feature parameters were extracted from the original vibration signal of fault rolling bearing. According to the time series trend, the equal number of data was cut from the different period of fault development and the mean value and variable value was calculated. The root mean square (RMS) value and kurtosis was chose as the feature characters with comparing the sensibility and stability of the early stage fault. The BP network and gray network was utilized in the different data set and the comparing application of fault gear was studied as well.Secondly, the method of Elman neural network was studied. As a dynamic feedback neural network, the difference and character of ELM network was introduced briefly. Through selecting rolling bearing and gear as the studying object, the prediction result was obtained from the feature parameters of RMS and kurtosis based on the ELM model. Meanwhile, the influence of different number of neuron to the prediction result was researched in this section.Thirdly, the method of ELM based on MFD was studied. The structure and algorithm advantages of ELM was described briefly and the prediction result of it was acquired in the faulty bearing and gear. Meanwhile, the accuracy of variable neuron to the prognosis was analyzed as well. In order to improve the accuracy of result, the fractal dimension and mathematics morphology was introduce to the fault prognosis. It revealed the application probity by calculating the MFD value of original time-domain signal. The prediction result was compared with ELM model after the MFD and time-domain process. The performance of all prediction models was summarized in the end.The result showed that ELM model had the best accuracy and stability that was not influenced by the number of neuron based on time-domain feature parameters. The accuracy of gray neural network come to the second place and it could realize more accurate result under the condition of less data. The BP and Elman neural network had a worse prediction result that was not influenced by different objects. The prediction result of ELM model was more accurate after the MFD process comparing with the time-domain process, and the combination of ELM and MFD was a better method in fault prognosis.