Quantitative Detection Method for Wheel Flats Driven by Time-Frequency Information Fusion
In the analysis of Envelope Spectrum(ES),both the envelope time-domain signal and the envelope frequency spectrum can effectively extract the features of wheel flats.To combine the advantages of these two data forms,we proposed a quantitative detection method for wheel flats driven by time-frequency information fusion.First,the gearbox vibration response under different scale conditions of wheel flats is obtained through constructing a metro vehicle-track rigid-flexible coupling dynamics model.Second,the gearbox vibration response is processed with overlapping sampling and ES analysis to generate two different types of data samples.Then,the accuracy of the Multi-input Convolutional Neural Network(MCNN)under two different model architectures is compared.The study focuses on analyzing the generalization of the MCNN model under the optimal model architecture and examining the differences in recognition performance when each part of the sample set is input separately.The results show that the model with optimized analysis has the best comprehensive recognition performance,with an average absolute percentage error of 3.812%and a coefficient of determination(R2)of 0.990.When dealing with random unknown data,the relative error does not exceed 7.2%.Although there is a decrease in timeliness,it still meets the requirements for online monitoring.
NETL
NSTL
万方数据
