Research On Key Technologies Of Machine Fault Diagnosis Based On Deep Learning

With the industrial machines more and more complicated,it’s important to ensure stable operation in the current industry.To detect faults of key elements in machines such as bearings and gearboxes,many scholars apply data-driven approaches to build straightforward connections between collected vibration signal data and operating conditions.In recent years,various deep learning based fault diagnosis methods have been studied.However,the labeled fault data obtained from some machines and some operation conditions in practice are small and data-imbalanced,which seriously affects the fault diagnosis performance.In order to solve the above problems,the main work of this thesis can be summarized as follows:（1）A fault data generation model SSIGAN based on semi-supervised learning is designed and implemented for the problem of data imbalance.In order to solve the mode collapse problem caused by the imbalanced labeled training data,SSIGAN unifies the domain classification task with the fault classification task.Thus,the generator could focus on generating new fault samples with high quality.In order to solve the hard training problem of conditional GAN under the small training data,the model use a mass of unlabeled fault data and small labeled fault data to train the unsupervised cluster model Va DE firstly.And then the model initializes generator and discriminator with the parameters of trained Va DE’s encoder and decoder.Besides,this thesis designs a conditonal latent variable generator based on the unsupervised cluster result.It could provide conditional latent variables rather than random noise to the GAN’s generator and increase the model’s ability of generating high quality samples.The results of several experiments show that SSIGAN has good ability of generating fault samples with high quality.（2）A few-shot fault diagnosis model DFSMMN based on unsupervised transfer learning and metric-based meta learning is designed and implemented.Condidering that the vibration signals contains multi-scale and complicated spatial related information,this thesis designs the MSA Block based on multi-scale architecture and attention mechanism firstly.And then this thesis embeds several MSA Blocks into the model’s feature extractor,to make the model focus on more signigicant features of different scales and different spatials.Considering that the Proto Nets classify samples according to the distance between samples and prototype centers,this thesis designs a penalty term for Proto Nets’ s loss function.It can make the samples’ features in the metric space be distinguished more easily.In order to solve the domain shift problem between the samples contained in metatraining tasks and meta-test tasks respectively,this thesis pretrains the model’s feature extractor through unsupervised transfer learning firstly.Thus the model can extract domain-invariant features in the meta-learning process and obtain a better performance for few-shot fault diagnosis.The results of experiments based on several few-shot fault diagnosis scenarios and different datasets prove that DFSMMN model can performance well on fault diagnosis tasks when fault samples are scarce.