Frictional Signal Analysis And State Recognition Under Non-stationary Conditions

The friction state recognition has great significance for monitoring the operation state of friction pairs of mechanical equipment, but the common reciprocating frictional force signal is often non-stationary signal. To reflect the friction state feature information more reasonably and fully, it is necessary to extract multiple characteristic parameters of the frictional force signal by using the appropriate signal processing methods.Based on the method of empirical mode decomposition, the core of Hilbert-Huang transform, which is based on its own characteristics of signal decomposition, it could be more adaptive to non-stationary signal analysis. In the reciprocating motion experimental platform, the frictional forces of initial friction state, steady friction state and starved lubrication state were measured as the research object, and multiple characteristic parameters were extracted in order to reflect the characteristics of different friction state more effectively. Then the friction state recognition system was established for identification based on probabilistic neural network. The main conclusions of the whole thesis are summarized as following:(1)The thesis puts forward the method of signal de-noising based on empirical mode decomposition combined with wavelet soft threshold de-noising, which played a good wavelet threshold de-noising ability and the empirical mode decomposition’s adaptive advantage. Through simulating signals and testing experimental signals of the reciprocating frictional force, the results are obtained, which show that compared with the traditional de-noising method, the method of signal de-noising based on empirical mode decomposition and wavelet soft threshold de-noising has good adaptability and stability, and better de-noising effect, improves the signal-to-noise ratio, better keeps the details of the signal.(2)The friction kurtosis value, the largest singular value, the singular value entropy, IMF energy entropy, the time-frequency entropy, the energy standard deviation and the low, medium and high frequency band energy were extracted respectively as the characteristic parameters by different methods. Through observation, the size or distribution of characteristic parameters in different friction states was different, and these characteristic parameters can effectively reflect the characteristics of different friction states. At the same time, the different characteristic parameters complement each other, to create the conditions for the friction state recognition.(3)The friction state recognition system was established based on probabilistic neural network which has faster training speed, fault tolerance, good scalability and no need to repeat the training. According to the sample’s verification, the friction state recognition system based on probabilistic neural network has better classification precision which the recognition rate is up to 96.86%.