Bearing Fault Diagnosis Based On Sparse Decomposition And Deep Belief Network

Rotary machinery plays an important role in industry,and rolling bearing is one of the core parts of rotary machinery.Rolling bearings are widely used in rotating machinery and play an important role in energy conversion.They are the key link of the system and the most malfunctioning parts of the entire machinery.The working environment of the rolling bearing in the rotating machine is harsh,and it is in a high-load and high-speed working environment all year round.If the fault signal can be detected earlier,subsequent major accidents can be prevented.However,the mechanical equipment is working in a strong noise environment all year round.The fault signal of rolling bearing is often obscured by noise.It is difficult to find fault signal frequency with the traditional diagnosis method.Ensemble Empirical Modal Decomposition(EEMD)is a modified method based on Empirical Modal Decomposition(EMD).The Hilbert spectrum based on EEMD has the characteristics of high resolution and anti-modal aliasing.This paper proposes a selection method based on energy ratio and correlation coefficient for the problem of contains false components.This method can remove the false components of IMF by EEMD,and the validity of the scheme is verified by bearing fault data.In this paper,the singular value difference spectrum is used to reduce noise.The Hankel matrix is obtained from the IMF function,the singular value before the maximum difference spectrum is preserved and reconstructed according to the fault signal.The experiment shows that the process of noise reducing can keep useful information and remove noise signal.In order to ensure the richness of the features,and time-frequency-domain mixed features are extracted for bearing fault signals.At the same time,the energy distribution and correlation coefficient obtained by EEMD decomposition are extracted for the input of the classifier.The sparse representation classifier based on sparse decomposition can effectively classify faults.Sparse representation includes sparse coding and dictionary learning.Compare with analyzing dictionaries,dictionary learning can learn the internal strong self-adaptability structure of signals without prior knowledge.Each atom in the dictionary is a typical feature.Greed algorithm and base tracking method can find sparse coefficients well.Through the repeated iteration of sparse coefficient solving and dictionary learning,the original features can be sparse expression.Each type of signal will produce a dictionary suitable for this type of signal..Based on dictionary learning,the sample is classified by judging the size of the reconstruction error using sparse classifiers,and the diagnosis of the fault mode is realized.This paper also adopts deep confidence network to realize fault classification.The deep belief neural network is made up of RBM.The training method uses a layer by layer greedy algorithm and reverse fine-tuning.The parameters of the depth confidence network are obtained by experimental verification and empirical formulas.The extracted features are used as the intelligent fault diagnosis method for the input of DBN to realize the accurate identification of various types of faults of rolling bearing inside the bearing.Finally,this paper will illustrate the advantages of deep learning through model comparison.