Bearing Pattern Recognition Based On Image Classification With CNN

Traditional bearing fault diagnosis methods based on data-driven need to know relevant signal processing techniques and to construct algorithms manually for feature extraction and selection,and then classifiers are adopted for classification and recognition,of which the support vector machine and artificial neural networks are widely employed.However,for mechanical big data with unclear and variable modes and multi-fault information coupling,it is extremely difficult to artificially and manually design fault features that cover all information.In addition,features extracted and selected for a particular problem may not be applicable to other issues.Therefore,it is very necessary for models to adaptively extract characteristics rather than manually extracting and selecting features.In addition,mechanical faults tend to be uncertain,concurrent and coupled.However,the shallow models employed in traditional methods have weak self-learning capability,and feature extraction and model establishment are isolated,resulting in low fault recognition accuracy and weak generalization ability.It is imperative that the diagnosis models transform from shallow models to deep models.In order to address the two problems aforementioned,two bearing pattern recognition methods based on image classification with convolutional neural network(CNN)are presented in this paper,with the consideration of the wide and successful application of CNN in the area of image classification and recognition.The two bearing pattern recognition methodologies adopted in this paper can automatically extract characteristics from time-frequency diagrams and then classify and recognize those features,which avoids the problem of extracting and selecting features by artificial constructing algorithms in conventional bearing fault diagnosis methods based on data-driven,reduces the dependence on the labor and improves the intelligence level of the fault diagnosis process.In the two methodologies employed in this paper,the high classification results are obtained on the four datasets without any adjustment of model structures and parameters,which indicates that the models have better generalization capability.The main contents of this paper are described as follows:With the ensemble empirical mode decomposition(EEMD)method,bearing vibration signals were decomposed into a finite number of intrinsic mode functions(IMF).Some IMF components were selected according to the criterion of correlation coefficient.With the pseudo Wigner-Ville time-frequency distribution(PWVD)method,time-frequency diagrams were obtained from selected IMFs and preprocessing was carried on time-frequency diagrams.A CNN with three convolution layers and a fully connected layer was constructed,which then classifies and recognizes bearing time-frequency diagrams of sixteen states of four datasets with the average accuracy of over 94.60%.Based on transfer learning,the AlexNet is fine-tuned and then bearing time-frequency diagrams of sixteen states of four datasets are classified and recognized by the AlexNet with the average accuracy of more than 96.30%.