Research On Fault State Identification Method Of Key Parts Of Gearbox Based On Deep Learning Model

Gearbox is the core component of mechanical equipment,and its operation condition determines the operation state of mechanical equipment.Failure of gear box will not only bring serious economic loss,but also bring great hidden danger to people’s life safety and smooth and normal operation of equipment.In order to make the gear box run smoothly,it is of great importance to develop intelligent fault diagnosis technology for gear box fault feature extraction and classification and to carry out health monitoring and fault diagnosis.With the increase of the number of device detection points and sampling frequency,the field of mechanical fault diagnosis has entered the era of "big data".Traditional intelligent fault diagnosis method based on feature extraction and classifier of signal processing requires high experts’ experience,and has the problems of low efficiency,low accuracy and universality.For this problem,in view of the powerful data representation and analysis ability of deep learning technology,this thesis builds an adaptive gear box fault diagnosis model through deep learning theory,and focuses on the key problems of gear box in the field of fault diagnosis,such as signal noise reduction,fault feature extraction and fault status recognition.Firstly,according to the problems existing in gear box fault diagnosis at home and abroad,the content of this thesis is determined.The structure of gear box used in the experiment is introduced,the common fault types and their causes are summarized,and the vibration mechanism of gear box parts with the largest proportion of fault is studied.The data used in the experiment are collected and divided into data sets by the test bench.Then,using the powerful feature extraction performance of convolution neural network,a one-dimensional Inception convolution neural network(1D-ICNN)is built and applied to gear box fault diagnosis.The model can automatically complete feature extraction and fault identification.It is found that after training the model on the collected data sets,the fault identification rate of the gear box can reach more than 99%.Then,in order to solve the problem that the vibration signals collected in the harsh working environment of gearbox contain a lot of noise and it is difficult to reduce the noise of this non-linear and non-stationary signal by conventional noise reduction methods,this thesis integrates the complete adaptive noise into an empirical mode decomposition(CEEMDAN).Combining Permutation Entropy(PE)and Adaptive Wavelet Threshold Noise Reduction(AWTNR),a combined noise reduction method based on CEEMDAN-PE-AWTNR is used to reduce the noise of vibration signal.This method uses the permutation Entropy as the criterion to determine whether each intrinsic mode function(IMF)needs noise reduction and the threshold value is 0.60.Firstly,the noisy signal is decomposed by CEEMDAN and the arrangement Entropy of each eigen mode function is obtained after the decomposition.Then,the eigen mode function components with the arrangement Entropy greater than 0.60 are denoised by adaptive wavelet threshold.Finally,the signal is recombined.The experimental research shows that the signal-to-noise ratio of the signal after noise reduction is significantly improved,and 1D-ICNN can accurately extract fault features from the signal after noise reduction to achieve fault classification.Finally,in order to compensate for the weakness of 1D-ICNN model in extracting characteristic information of vibration signal in time dimension,this thesis adds GRU(Gate Recurrent Unit)channel based on 1D-ICNN model to form 1D-ICNN-GRU dual-channel fault diagnosis model.GRU network is good at extracting information features in time from vibration signal sequence and 1D-ICNN is good at extracting information features in space.The model combines the advantages of both and uses parallel structure to extract fault characteristics of vibration signals in space and time.Finally,the fault features extracted by the two channels are fused into a feature vector and input into the Softmax layer for fault classification.The experimental results show that the 1D-ICNN-GRU dual-channel fault diagnosis model has higher fault identification accuracy than the 1D-ICNN single-channel model and is more suitable for gear box fault detection.