Research On Cross-domain Intelligent Fault Diagnosis Methods Based On Transfer Learning For Rotating Machinery

Rotating machinery is a kind of widely used mechanical devices,and is playing an important role in industrial production.Facing its development direction of becoming more complicated and intelligent,it is a significant initiative to guarantee the reliability,safety,and economy of industrial production processes that developing the health management technology which can timely monitor the health condition and its trend of devices.Data-driven intelligent fault diagnosis is one of the supporting contents of the health management technology,and its ultimate purpose is to learn fault characteristics and diagnosis rules from massive historical data using machine learning methods.However,for conventional data-driven fault diagnosis methods,a prerequisite of ensuring their performance is that the data for training diagnosis models and the data to be tested should be followed the same distribution.It is very difficult to satisfy this prerequisite on actual engineering diagnosis tasks,which is a principal reason that prevents data-driven diagnosis methods from being applied to actual fault diagnosis.To break this limitation,this research work focuses on the cross-domain fault diagnosis problem that utilizes non-identically distributed data from multiple sources to build diagnosis models.Inspired by the idea of transfer learning,this dissertation mainly studies several transfer strategies of diagnosis knowledge,which can mine and transfer diagnosis knowledge from one or multiple source domains and further promote the generalization performance of the models with respect to the target task.Specifically,the research contents of this dissertation are summarized as follows:(1)As the research motivation,the necessity and transferability on developing cross-domain diagnosis methods are discussed.The necessity of our research is discussed from two aspects,engineering demand and method demand.In this part,the cross-domain diagnosis performance of conventional data-driven methods is focused on.The transferability of implementing cross-domain diagnosis is preliminary explained from the point of view of vibration generation model.(2)The cross-domain diagnosis problem under single source domain.For the task scenario with inadequate information of target domain during the training stage,a relevance assumption that defines the a priori distribution structure information of training data is proposed.Meanwhile,a feature transform method,called transfer locality preserving projection,which can reduce the distribution discrepancy between domains in the new space,is proposed under the guidance of the relevance assumption.By this means,the cross-domain diagnosis performance of the model is promoted.The proposed method achieved superior diagnosis results than several data-driven diagnosis methods both on gear and bearing cross-domain diagnosis tasks.The effectiveness of the proposed method is verified.(3)The cross-domain diagnosis problem under multiple source domains.The essential idea of this part is promoting the model’s generalization performance on the target task by learning the common discriminant knowledge across multiple source domains.Specifically,the discriminant structure of each domain is firstly described as a point of Grassmann manifold by the local Fisher discriminant analysis.Then,the common knowledge across all domains is learned by solving the Karcher mean on the Grassmann manifold.On the diagnosis tasks of bearing,the proposed method achieved superior performance than several conventional supervised methods,single-source transfer learning methods,and multisource transfer learning methods.(4)The cross-domain diagnosis problem based on deep neural networks under the scenario of multiple sources.The essential idea of this part is learning discriminative and domain-invariant fault features from the vibration signals of multiple sources by leveraging the representation learning ability of deep neural networks.Specifically,the network inputs with consistent meaning across domains are firstly given by a signal preprocessing stage guided with the a priori diagnosis knowledge.On this basis,a deep domain generalization network is elaborated to learn domain-invariant features,and then generalize the learned knowledge to identify unseen target samples.The proposed method achieved superior diagnosis performance than several deep-learning based diagnosis methods and deep transfer based diagnosis methods on both gear and bearing cross-domain diagnosis tasks.Our research work provides the new options to handle the diagnosis demands in the actual engineering,to break the data dilemma of conventional data-driven diagnosis methods,and to promote the level of intelligence of fault diagnosis.