Research On Rolling Bearing Fault Diagnosis Based On Open World Hypothesis

In the field of modern industrial machinery,mechanical components play a crucial role,and rolling bearings are an indispensable part of the mechanical operation process.However,the existing intelligent fault diagnosis methods for bearings have mostly been studied based on the closed-set assumption scenario,where the training and testing sets share the same fault information.This assumption is often not valid in practical industrial machinery scenarios.On one hand,data collection in industrial applications is usually limited,and the collected rolling bearing fault information cannot cover all fault categories,which means that models proposed under the closed-set assumption may fail.On the other hand,the working conditions of rotating machinery are diverse,and the distribution of training and testing data is relatively independent,which means that models trained under a single operating condition cannot be applied to the actual situation under multiple operating conditions.Therefore,the rolling bearing fault diagnosis based on the open-set assumption has significant application value and research significance.In order to better solve the problems existing in the intelligent fault diagnosis of rolling bearings in the field of real industrial machinery and form a complete industrial machinery application framework,this thesis mainly focuses on the intelligent fault diagnosis method of rolling bearings based on the open-world hypothesis and studies the bearing fault information under different working conditions.The following fault diagnosis model and cloud edge collaborative interaction framework are proposed to solve the problem of rolling bearing fault diagnosis under the above conditions.This thesis is supported by the national key research and development program: key technologies and application demonstration of high torque hub drive for engineering machinery,project number: 2019YFB20066400.The main research content of this thesis is as follows:(1)Aiming at intelligent fault diagnosis under a single working condition based on open world hypothesis,an open set convolutional neural network method is proposed.Compared with the last layer of the traditional model using Softmax multi-classifier method,this thesis uses 1-vs-rest classifier in the last layer,and through visualization to verify this method can effectively tighten the intra-class distance of known fault classes and expand the inter-class distance,so the overall performance is significantly improved.Then,the model is applied to the extreme value theory algorithm based on distance modeling as a deep network to learn the characteristics of training fault information to accurately detect unknown fault types in the test set.The overall effectiveness of the proposed method is verified based on experimental research.This method can not only realize the correct classification of known fault categories,but also accurately detect unknown fault categories.(2)Aiming at the intelligent fault diagnosis under multiple working conditions based on the open world hypothesis,an adaptive adversarial network framework based on open set domain is proposed.Firstly,in order to better learn the cross-domain bearing signal category features,this thesis uses the deep convolution autoencoder model for training.Secondly,in order to ignore the influence of unknown fault samples in the target domain,the unknown boundary is established in a weighted way by calculating the similarity between the known class and the unknown class.Under the premise of class alignment between the two domains,the known class is classified across domains to further identify unknown fault samples.(3)Based on the cloud-edge collaboration framework,the cloud-edge interaction function is systematically completed and the real-time intelligent fault diagnosis at the edge is realized.Firstly,the model is trained based on the labeled data set in the cloud.Then the edge end completes the task-based setting,pulls the model trained in the cloud to perform the intelligent fault diagnosis task based on the open set,outputs the diagnosis result,and forms the unknown fault sample contained in it.The unknown fault set is uploaded to the cloud database.Then,the clustering annotation work is performed on the unknown class set uploaded by the edge end to update the data in the cloud,so as to train a new version of the model in the cloud and continuously update the model in the cloud model library.The model is trained in the resource-rich cloud through clustering and labeling,and stored in the cloud model library.The sinking of the model is realized according to the edge fault diagnosis task,which meets the requirements of real-time fault diagnosis.This thesis has 46 figures,19 tables,138 references.