Research On Non-liner Vibration Prediction Method

Vibration is a common phenomenon in nature and engineering. Some equipment and instruments that are produced by vibration property are widely used in engineering. Although the vibration applied to actual engineering work has brought great convenience to industrial production in some cases, in most cases, vibration is harm to the safe operation of machinery and equipment. With the development of economic and industrial technology, various industrial equipment have increasingly become more and more complex, large-scale, heavy and precise, which make the adverse effects of vibration on equipment more prominent and is a big challenge to vibration control and condition monitoring. The most traditional vibration system model is precise mathematical model based on certain assumptions, which presents some limitations. Based on these considerations above, vibration prediction models based on BP network, support vector machine and echo state network are built from the perspective of vibration sequence. The sample data was obtained from a space camera.Firstly, BP network, support vector machine and echo state network are introduced to build multi-step prediction model of nonlinear vibration signal, and the simulation results are compared. According to the characteristics of the three methods, a combined model was proposed from BP networks and support vector machine, named BP-SVR model. To assess the prediction performance, prediction accuracy of each prediction model was computed, which show that the prediction performance of echo state network is the best.Then, phase space reconstruction is introduced to reconstruct the vibration sequence and mine the implicit information of the vibration signal. According to the appropriate delay time and optimal embedding dimension, phase points are grouped to predict vibration sequence. Echo state network is used to build multi-step prediction model of each phase point groups. The results of simulation show that the modeling method extends the prediction steps of the vibration prediction model effectively.In order to lengthen the prediction time of the model, wavelet analysis is introduced to optimize the prediction model. Through wavelet decomposition and reconstruction of original vibration signal, the trend of the vibration is separated. Then, the vibration trend is reconstructed by phase space reconstruction method. According to the appropriate delay time and optimal embedding dimension, phase points are grouped and prediction models for long time are built based on echo state network. Keeping certain precision, this modeling method can greatly lengthen the prediction time.