Identifying Methods Of Rotating Machine Faults Based On Sparse Representation And Compressive Sensing

As important components of automobile gearboxes and wind turbines,gears and rolling bearings are inclined to be influenced by impact faults and stable faults under complex and changing running conditions.It has been a popular topic to identify,diagnose and monitor the faults in real time both quickly and precisely.Based on the sparsity of fault signal,a series of researches are implemented in fields of Compressive Sensing and Sparse Representation.A new SWD-KSVD algorithm based on K-SVD and sliding window denoising method is proposed to diagnose the rolling bearing impact fault.The algorithm uses only one small segment of time domain signal containing impacts to perform sliding window dictionary learning and select an optimal pattern with oscillating information of the rolling bearing fault according to a maximum variance principle.An inner product operation between the optimal pattern and the whole fault signal is performed to enhance the characteristic of the impacts' occurrence moments.Lastly,the signal is reconstructed at peak points of the inner product to realize the extraction of the rolling bearing fault features.Because of the short optimal“pattern” and the correlation analysis,the shape and phase of the impact are reconstructed relatively precisely.Meanwhile,the calculating speed is fast,and the method is robust to noises.Both simulation and experiments verify that the algorithm could diagnose the impact faults of automobile gearbox's rolling bearings.The adaptive penalty parameter method is embedded into ADMM algorithm to improve the original one,which is then utilized on the sub-sampling identification of gears' impact and compound faults.Combining the fault features,a double dictionary system including wavelet dictionary and stable composition dictionary is constructed to reconstruct the impact component and stable component respectively.On the one hand,the extrapolation method is applied to accelerate the algorithm,on the other hand,the adaptive penalty parameter updating method is added into the method to increase sparsity and convergence.By comparing the reconstruction error and success rate of different reconstructing methods,it shows that the wavelet dictionary is sensitive to impact fault,while the stable harmonic dictionary is satisfactory with the stable faults thanks to the specific physical meaning.As the double dictionaries couple with each other on extracting fault features,the reconstructing performance of the double dictionary method is worse than that of the single dictionary method.The simulation shows that the single dictionary algorithm could identify the fault feature when the compressing rate is 10% and the SNR is over-5dB,while the double dictionary algorithm could accomplish that when the compressing rate is 20% and the SNR isover 0dB;the experiment proves that the single and double dictionary algorithms could identify faults when the compressing rate is 10% and 20% respectively.