| Mechanical equipment is an important basic support for promoting the development of the national economy,and is widely used in important national fields such as aerospace,ships,construction machinery,energy,petrochemicals,and robots.Bearing,gearbox and other rotating mechanical components as a typical key core parts of mechanical equipment,in order to avoid heavy losses to the national economy and safety of people’s lives and property due to the occurance of fault,the fault diagnosis technolgy of rotating mechnical components has become an important technical means to ensure their safe and reliable operation.However,rotating mechanical components are usually in a complex and changeable operating environment,and their monitoring data distribution varies greatly,which severely limits the diagnostic ability and generalization ability of machines and deep learning intelligent fault diagnosis methods subject to the same distribution constraint of training data and test data.Transfer learning,as a promising mechine learning method,can eliminate the distribution differences in similar fields and improve the generalization performance of the model,which provides a new diagnostic idea for the fault diagnosis of rotating mechanical components in complex and variable service environments.Based on this,the advantage of transfer learning is used to narrow the difference fault feature distribution of rotating mechanical components under cross-working conditions,learn common fault information representing variable operating conditions,explore a fault diagnosis method of rotating mechanical components with adaptive changing working conditions to improve the selfsensing,self-adaptive and self-learning ability of mechanical equipment with common fault knowledge.And accelerate the application of intelligent fault diagnosis technology in practical engineering.Therefore,this paper takes deep learning and transfer learning as the theoretical basis to conduct research on fault diagnosis methods of typical rotating mechanical components based on deep transfer learning,which mainly solves three key issues: how to improve the model’s domain generalization ability under cross-working conditions,how to improve the learning ability of common fault knowledge under multicondition incomplete fault information and how to improve the self-learning ability of the model for continuously adding condition and fault mode scenarios,and construct four kinds of deep transfer fault diagnosis models for different transfer diagnosis scenarios.The main contents of this paper are as follows:Aiming at the issue that domain generalization ability of the fault diagnosis model is insufficient under cross-working condition environment of single-source domain,a dynamic domain adaptation fault diagnosis model based on deep multiple auto-encoder network with attention mechanism is established.This model integrates various autoencoders with different activation functions to extract rich fault features.It also uses a multi-layer attention mechanism to automatically assign weights to deep feature of different layers,and improve the diagnostic accuracy of the model.In addition,dynamic domain adaptation strategy is proposed to adaptively adjust the weight of the marginal and conditional distribution for learning domain invariant fault features to enhance the domain adaptation ability of the model,and learn domain-invariant fault feature,and imporove generalization ability of model under single source domain transfer diagnosis scenario.Compared to other transfer learning methodss,the proposed diagnostic model has better diagnosis accuracy,generalization performance and bobustness under single source domain cross-working condition fault diagnosis tasks of bearing and gearboxTo address the need of improving domain generalization ability of fault diagnosis model in cross-working condition scenarios of multi-source domain,a multi-source domain transfer fault diagnosis model based on ensemble of anchor adapters is established.The model learns more comprehensive common fault information by constructing multiple source domaintarget domain anchor adapter matrices,and integrating multiple source domain fault information through anchor adapter integration strategies.Furthermore,the multi-source domain adaptation transfer network considers the mutual difference between feature distribution in multisource domain to learn domain-invariant common fault diganosis knowledge and enchance the domain generalization ability of the model.The transfer diagnostic tasks for bearings and gearboxes are used to verify the effectiveness of the proposed diagnostic model,which shows that multi-souce domain transfer diagnostic model has better diagnosis performance and generalization ability than single-source domain.In order to further perfect the application scenario of multi-source domain transfer fault diagnosis,improve the ability of model to learn common fault knowledge from multiconditions with incomplete fault information,this paper establishes a transfer fault diagnosis model based on cycle generative adversarial network and multi-level domain adversarial network.This model constructs multi-source domain dataset with complete fault information by cycle generative adversarial network,and then uses deep multi-level domain adversarial network to learn fine-grained discriminant and domain-invariant common fault features from two perspectives of global domain alignment and local class alignment,which improves the diagnosis accuracy and stability of the model.The effectiveness of the proposed method is verified by two experimental cases of bearing.The results show that the proposed diagnosis model can effectively solve the fault diagnosis problem of coexistence of domain and fault label shift,learn common fault knowledge from multiple working conditions dataset with incomplete fault categories,and further improve the applicability of the model.For the rotating mechanical components will maybe gradually produce new fault modes with continuously incremental new working conditions,the traditional transfer diagnosis method with known fault types is difficult to continuously learn new fault knowledge.An incremental transfer fault diagnosis model based on prototype continuous domain adaptation is established to further improve the diagnostic ability of model to learn new fault knowledge.The model can effectivelyi dentify new fault types though prototype alignment and decoupling domain adaptation modules,and enhance model’s ability to autonomously learn new fault knowledge.Additionally,the model can also ensure memory retention for old fault knowledge by utilizing prototype sample set replay and model parameter transfer.The effectiveness of the proposed method is verified by the transfer diagnosis task of continuously new fault modes in bearings and gearboxes.The results show that the proposed incremental transfer fault diagnosis based on prototype continuous domain adaptation can have the ability to learn new fault knowledge under new working conditions,and further provide a new way to realize online transfer fault diagnosis modeling. |
