| Rolling bearings as core components of rotating machinery,which are prone to failure.Therefore,maintaining the monitoring,fault identification and localization of rotating machinery is crucial for the safe operation of rotating machinery.With the advent of the era of big data,the diagnosis of bearing faults gradually turns to the direction of intelligence.In this thesis,we will propose three intelligent fault diagnosis models about the viewpoints of variable load,variable speed and variable working conditions based on the characteristics and processing methods of signals and structural improvement of convolutional neural networks.The respective methods as follows.(1)Traditional fault diagnosis algorithms always be dependent on manual extraction of features,which are affected by noise and have low accuracy,a GAF combined with residual network is proposed for bearing fault diagnosis.The method first uses wavelet noise reduction on the original one-dimensional signal to reduce the noise,then uses GAF to convert the one-dimensional time-series signal into a twodimensional time-frequency image,further divides the data set into a training set and a test set,and uses data normalization to make the training set and the test set learn the same distribution to reduce the model training time.Batch Normalization is added to the model to control the input distribution,coordination distribution and weights of each layer,and dropout and regularization are used to suppress overfitting to achieve accurate classification of bearing faults.(2)The default contribution of all feature maps to the current task is the same for the neural network.In fact,each feature map has a different contribution to the current task.The contribution of the feature maps to the model is not consistent,and the model converges slowly,which weakens the feature extraction ability of the model.An attention mechanism combined with a deep residual network is proposed for the bearing fault diagnosis model.The method reconstructs the one-dimensional timeseries signals into grayscale maps as the input of the network;introduces a hybrid attention mechanism in the appropriate parts of the model to autonomously learn the sensitivity of each channel to faults,effectively suppressing the propagation of redundant features and enhancing the feature extraction ability of the model.Finally,the accuracy of the method is verified using the Case Western Reserve University bearing dataset and the high speed rail wheel pair faint fault bearing dataset.(3)In response to the challenges posed by the variable working conditions of rolling bearings in actual production to the traditional fault diagnosis methods and the feature migration problem of the fault time-frequency map,a bearing fault diagnosis method of multi-scale feature extraction combined with capsule network is proposed.The method first improves the Inception structure to carry out multi-scale adaptive extraction of faults to obtain the weak changes of frequency under variable working conditions.The extracted features are input into the Inception module to further extract the features in the fault information to obtain the local sensory field and parameter sharing,and finally the capsule network is introduced to refine the features for identification.Accurate identification of single fault of bearing under complex working conditions with different loads and different sizes is achieved. |
