Incipient Fault Diagnosis Methods For Reciprocating Compressors Based On Stacked Denoising Autoencoder

Recently,the intelligent fault diagnosis in the prognostic and health management of reciprocating compressors has been a research hotspot for a long time.In the actual industrial production,the working environment of reciprocating compressor is complex,and the vibration signals generated during operation are often mixed with various and complex field noises,therefore the collected vibration signal usually has strong noise interference.The traditional signal processing methods have poor ability of feature extraction and fault diagnosis.Deep learning can achieve the modeling between complex data relationship through multi-layer non-linear transformation.Therefore,the research on fault diagnosis of reciprocating compressor has great significance based on the theory of deep learning.In this paper,the feature extraction and fault diagnosis model based on signal processing and deep learning for reciprocating compressor is studied.The main work is as follows:(1)Aiming at the problem that the system signal and noise signal are overlapped in frequency band,and the traditional vibration signal denoising methods,such as wavelet analysis and empirical mode decomposition,have some limitations in separating the overlapped signal,a signal denoising method based on local mean decomposition and independent component analysis is proposed.The virtual noise channel is constructed by decomposing and reconstructing the vibration signal,and the virtual noise channel and the vibration signal are used as the signal input of independent component analysis to effectively separate the noise signal and the effective signal.The results of signal denoising for reciprocating compressor valve data show that the signal to noise ratio is the largest,the root mean square error is the smallest,and the correlation coefficient is the largest after denoising by this method.This method reserves the internal information of the original data on the greatest degree,and the effect of noise reduction is obviously better than the other traditional methods,such as local mean decomposition combining with wavelet denoising and empirical mode decomposition combining with independent component analysis.(2)In view of the complex structure and large number of nodes of the stacked denoising autoencoder neural network,the network super parameters are mostly determined by experience gained from many experiments,and the lack of self-adaptability,a fault diagnosis optimization model based on the combination of genetic algorithm and stacked denoising autoencoder is proposed.The super parameters of stacked denoising autoencoder network are selected adaptively by genetic algorithm with good global search ability and fast convergence ability.In the example,the genetic algorithm is used to optimize the fault diagnosis model of reciprocating compressor.The diagnosis effect is obviously better than that of the non-optimized model and the traditional fault diagnosis optimization model.(3)Aiming at the problems of strong subjectivity in manual feature extraction of traditional signal processing methods,the traditional machine learning classification algorithms can only learn the low-dimensional shallow features of samples,low learning ability and lack of necessary generalization ability,a fault diagnosis model combining local mean decomposition and stacked denoising autoencoder is proposed.The fault diagnosis effects of the proposed method are validated by the reciprocating compressor valve datasets.The experimental results show that the classification accuracy can reach 92.72% when the signal-noise ratio is-10 d B.Compared with other fault diagnosis method,such as the local mean decomposition and depth confidence network,the proposed method improves by 5 percentage points,which shows the effectiveness and robustness of the proposed method.