Study On Fault Diagnosis Method Of Mechanical Rotating Components Under Non-Ideal Data Conditions

With the improvement of information technology and the rapid development of modern industry,mechanical equipment is gradually moving towards to the direction of integration,electrification and automation,and its structure is becoming more and more refined and complex.Mechanical rotating components as the key components of mechanical equipment,once failure occurs will cause the entire equipment can not operate,the light increase in downtime,or cause a lot of economic losses and even casualties.Therefore,the research on condition monitoring and fault diagnosis technology can maintain the industrial safety of mechanical equipment,which has important application value and potential economic benefits.At present,most of the fault diagnosis of mechanical rotating components is carried out under ideal data conditions,ignoring the non-ideal data problems faced by the actual operation of mechanical equipment,such as samples with strong noise interference,partial samples without labels,missing fault samples,etc.,which makes it inadequate in responding to the fault diagnosis demands of mechanical rotating components under actual scenarios.Consequently,this project studies the fault diagnosis method under non-ideal data conditions for mechanical rotating components as follows.1.For the issue of low accuracy of fault diagnosis of mechanical rotating components caused by strong noise interference in practical application scenarios,an end-to-end deep learning fault diagnosis method based on interference suppression convolutional neural network(ISCNN)model is designed.A parallel convolutional flow with a dilated-width convolutional kernel is used to sparsely sample from noisepolluted signals to extract multi-scale short-time features,and noise interference is suppressed by a filtering operation.A multi-scale feature enhancement module is constructed to achieve adaptive adjustment of neuronal receptive fields using parallel selective convolutional kernel networks to exploit weak fault features hidden in noisy signals.A convolutional feature fusion method is adopted to integrate and input multidimensional fault features into the classifier,thus achieving accurate fault diagnosis of mechanical rotating components under strong noise interference.The experimental results show that the fault diagnosis performance of the ISCNN model is better than the existing methods in the field and can improve the identifiability of mechanical rotating component faults in noisy scenarios.2.A quasi-full supervised fault diagnosis method with pseudo label automatic learning(PLAL)is proposed to address the existing fully supervised learning methods that cannot utilize a large number of unlabeled samples and semi-supervised learning type methods that still have a large deficiency in fault diagnosis model accuracy.Acquiring deep representation feature sets of labeled and unlabeled samples using self-normalized convolutional adversarial autoencoder(SCAAE)in unsupervised learning mode.Introducing the constrained seed K-means(CSKM)algorithm into SCAAE to achieve optimization of depth representation features and improve the correct pseudolabel tagging of unlabeled samples.The original labeled samples and the labeled pseudo-labeled samples are employed to train the fault diagnosis model and derive the final classification results.The experimental results show that the PLAL fault diagnosis algorithm can make full use of the unlabeled samples to achieve the goal of improving the fault diagnosis accuracy of mechanical rotating components.3.Aiming at the existing fault diagnosis methods based on generative adversarial networks(GAN),which are difficult to generate high-quality fault samples of missing types and suffer from insufficient model feature extraction ability and unstable training,a fault diagnosis method based on upgraded generative adversarial networks(UGAN)for mechanical rotating components is proposed.The conditional variational autoencoder(CVAE)model is introduced as the generator of UGAN to achieve the goal of sampling from real samples rather than random noise distributions and to improve the quality of the generated samples under the fault sample missing condition.A discriminator based on the self-normalized convolutional autoencoder(SCAE)model is designed to enhance the model’s ability to identify faulty samples while maintaining network stability.By adding a classifier to the discriminator,the fault diagnosis process is simplified while improving the stability and convergence of the network.Experimental results show that UGAN achieves higher accuracy,F-measure and G-mean on the imbalanced dataset with missing fault samples.