Research On Fault Diagnosis Methods Of Key Components Of Rotating Machinery Based On Adversarial Transfer

As the key components of rotating machinery,bearings and gears usually operate under complex working conditions,are prone to failure,and fault diagnosis on them is critical to ensure industrial production.The development of deep learning has boosted the application of data-driven models in the field of rotating machinery fault diagnosis.However,it is difficult to acquire fault operation data of machine and labeling them.Available diagnostic models cannot be effectively generalized to real scenarios.Transfer learning can accomplish target machine diagnosis with the help of auxiliary domains which have abundant labeled data.As the key branch of transfer learning,adversarial transfer avoids the selection of feature metric functions during the phase of alignment domain distribution,so it is more generalizable.However,there are still a few key issues that need to be addressed,Therefore,research on the fault diagnosis method based on adversarial transfer is carried out in the paper,and the advantages of our proposed method will be illustrated by experimental analysis and methodology comparison.Main research contents of thesis are as follows:(1)A deep dynamic joint adaptive adversarial transfer network is proposed to address the problem that current diagnostic models cannot align domain distribution differences effectively under unsupervised transfer scenarios due to the difficulty in obtaining a prior knowledge of data distribution.Based on the residual module,a feature extractor is constructed to capture common discriminative features between domains.Multilinear conditioning is integrated into the domain discriminator to align conditional probability distribution.A weighting factor is introduced to dynamically reduce the differences between the two types of distributions during the phase of model training to maximize the common information transfer to the target domain.Three cases of bearings with diagnosis tasks of cross-condition and cross-domain and methodology comparison jointly demonstrate the advantages of our method.(2)A zero-fault sample diagnosis method fused with adversarial transfer and sample synthetic is proposed for the scenario where the actual equipment contains abundant health operation data but no fault operation data.Combined with generative adversarial networks,generative neurons with stronger nonlinear mapping capability in self-organizing operational neural networks are used to replace the convolutional neurons to construct a network framework for fault sample synthesis.After transferring the network parameters,the fault classifier is built using synthetic fault samples to replace the real fault samples in the target domain.The diagnostic model is established when there are no fault samples in tested equipment.Two cases of bearings with transfer diagnostic tasks are used to show validity of our method.(3)To deal with the problem that it is difficult to carry out homogeneous domain transfer diagnosis caused by the difficulty of acquiring data of actual equipment and missing labeled data of same components,a transitive adversarial transfer model is proposed for diagnosis in cross-machine component scene.Based on transitive transfer learning,Fuzzy Entropy is introduced to calculate the domain complexity,and Wasserstein distance is used to measure the inter-domain distribution discrepancy,while an intermediate domain selection strategy is proposed to quantitatively obtain intermediate domain from alternative datasets.Based on the non-negative matrix triple factorization with cosine similarity as the measurement,a deep data enhancement module is constructed to expand the target domain samples.Further,a generalized transitive transfer module is proposed for heterogeneous domain transfer diagnostic tasks.Both few-shot diagnostic tasks and method comparison illustrate the robustness of our method.