Fault Diagnosis Method Of Rolling Bearing Based On Improved Convolutional Neural Network

In the context of industrial big data,the traditional rolling bearing health status identification method requires rich a priori knowledge of inspectors and is inefficient,and the existing convolutional neural network-based bearing fault diagnosis method has the problems of complex network,many training parameters and weak model generalization.The models were validated on variable load environment,high noise environment and model generalizability in the experimental dataset.The main research work is as follows.(1)Rolling bearing datasets for training and testing models were built.Starting with the basic structure of bearings and their vibration mechanisms,several typical types of failures of rolling bearings,the characteristic frequencies generated and the causes leading to the occurrence of failures are analyzed.Wavelet analysis with better localization characteristics is selected from various signal processing methods,and time-frequency maps are generated by continuous wavelet transform for Case Western Reserve University bearing dataset and Xi’an Jiaotong University bearing dataset,so as to establish the dataset for the training model.(2)A bearing fault diagnosis method combining the improved Incetpion V2 module and CBAM is proposed and its accuracy is verified.The model is improved and optimized by adding average pooling branches to the Inception V2 module,introducing CBAM attention mechanism,adding BN layer and other operations,and verifying its performance with the help of open-source bearing datasets from two universities and obtaining 100% accuracy with an average accuracy of94.1% and 94.92% under variable load as well as high noise conditions,respectively.(3)A second rolling bearing fault diagnosis method based on depth-separable convolution is proposed and its accuracy is verified.Combining the features of Google Efficient Net V2 model and Inception structure,a fault diagnosis model with better performance is obtained by virtue of introducing operations such as depth-separable convolution,residual structure,and Swish activation function.By validating the model under the same dataset,the model not only converges fast but also achieves 100% accuracy and 96.72% accuracy under variable load conditions.In this thesis,we propose two bearing fault diagnosis methods incorporating multi-scale convolutional features,which are validated in two types of rolling bearing experimental datasets,showing that they have good recognition effect under variable load and high noise environment,and have good generalization and stability.