Research On Fault Diagnosis And Health Assessment Of Rolling Bearing Based On Auto-encoder

Rolling bearing is one of the most widely used parts in rotating machinery equipment,and it is also the most easily damaged parts.It is of great significance to study its fault diagnosis and health assessment.As a kind of deep learning model,auto-encoder is widely used in many fields.It can extract depth features from monitoring data adaptively without supervision,which brings a solution for fault diagnosis and health assessment of rolling bearings.Aiming at the problem of fault diagnosis and health assessment of rolling bearing,the improvement of auto-encoder structure and target output is proposed.The vibration signal denoising,fault feature extraction,fault classification,health indicator construction and early fault diagnosis are studied.Experiments are carried out on simulation signals and two public rolling bearing data sets to verify the effectiveness and practical value.The main research contents and conclusions are as follows:(1)Research on rolling bearing vibration signal denoising based on stacked denoising auto-encoder.An improved stacked denoising auto-encoder model is designed and implemented.Experiments are carried out on sinusoidal signals,amplitude modulation signals and simulated bearing fault signals,and good denoising effects are achieved.It is verified that the stacked denoising auto-encoder has strong capability of feature extraction and signal reconstruction,which lays a foundation for rolling fault feature extraction,fault diagnosis and health assessment based on auto-encoder.(2)Research on rolling bearing fault feature extraction and diagnosis based on frequency domain auto-encoder.A frequency domain auto-encoder model is proposed which can extract frequency domain features directly from time domain signals.Different from traditional auto-encoder,the target of frequency domain auto-encoder reconstruction is spectrum.Experiments are carried out on the open rolling bearing data set,and the experimental results show that the frequency domain auto-encoder has better feature extraction ability,stronger generalization ability and less label dependence.(3)Health assessment of rolling bearings based on depthwise separable convolutional auto-encoder.A construction method of rolling bearing health indicator based on depthwise separable convolution auto-encoder is proposed.Based on the health indicator,an early fault detection method based on savitzky-Golay filter smoothing optimization is designed.Experimental results show that the proposed health indicator can reflect the degradation trend of rolling bearings,is sensitive to the occurrence of early faults,and has good generalization performance.The proposed early fault detection method can obtain the threshold adaptively,and can detect early faults effectively,with low false alarm rate and missing alarm rate.The proposed model is lightweight and the detection speed can meet the demand of real-time.