Research On Fault Diagnosis Method Of Machinery Based On Deep Learning

This dissertation was supported by the National Natural Science Foundation of China(No.51675258,51261024,51075372),the Open Found of The State Key Laboratory of Mechanical Transmissions(SKLMT-KFKT-201514).Based on the Deep learning as the center,a series of new algorithms are proposed for the fault diagnosis of Machinery.Combining with the new compressive sampling theory,some innovative achievements are obtained.The main contents are summarized as follows:The first chapter discusses the significance of the research and expansion of the research.At the same time,the research status of deep learning at home and abroad,and the research status of deep learning in the field of mechanical fault diagnosis are introduced.Finally,the main contents and innovations of this paper are listed.In the second chapter,a new mechanical fault diagnosis method based on deep belief network is proposed,according to deep belief network can gradually learn from the lower level to a more abstract and complex high-level representation.The experimental results show that the proposed method can achieve self-adaptive extraction of fault features and intelligent identification of bearing health conditions,and overcomes the inadequacies of the traditional methods in feature extraction and fault recognition that rely on a large amount of signal processing knowledge and engineering experience,reducing the labor Factors involved in troubleshooting.The comparison experiments show that the proposed method has higher failure recognition rate than traditional wavelet-based methods,multivariate features and feature kurtosis methods.The third chapter proposes a BN-WCNN convolutional neural network aiming at the disadvantage that traditional convolutional neural networks can only obtain small receptive fields with multiple weights and the intranet-network covariates are easy to transfer during training.In the proposed network,the original convolutional neural network was redesigned,in which the first convolution kernel was designed as a large convolution kernel,and the rest of the convolution layer remained unchanged while maintaining the network extraction failure.Reduction of convolution kernel parameters in the case of feature capabilities.Secondly,BN processing was added after the convolution operation to prevent network internal covariates from being transferred.In order to verify the effectiveness of the network,the BN-WCNN network was applied to bearing fault diagnosis.The experimental results show that the proposed method can directly extract the fault features from the time-domain vibration signals and realize the high-precision knowledge of bearing faults.In the fourth chapter,the deep neural network based on noise Denoising autoencoders is applied to the fault diagnosis of rotating machinery with noise interference.First,the frequency domain signal of the rotating machine is input into a noise reduction self-encoder,and a plurality of noise reduction self-encoders are used to stack layers to obtain a deep neural network.After the deep neural network is obtained,the network is fine-tuned using dropout technology to complete the training of the network and perform fault diagnosis.Experimental results show that the proposed method can overcome the dependence on signal processing technology and engineering diagnostic experience,and interfere with the noise signal,and adaptively extract fault features directly from the frequency domain signal.The fifth chapter discusses the shortcomings of the traditional method of using isolated information to monitor and diagnose faults in mechanical equipments.It also introduces the Vector-bispectum technique based on multi-channel information fusion and the two-channel full-spectrum digital method for orthogonal installation of two sensors.Combining the advantages of the whole vector spectrum technique in multi-information fusion and the powerful modeling and characterization capabilities of deep neural networks,a Vector-bispectum-DNN algorithm is proposed.Experimental results show that the proposed method can fuse multi-channel information very well,and has good adaptability and high prediction accuracy.Chapter sixth summarizes the chapters,points out the directions for improvement and further research in the future.