Research On Small Sample Intelligent Diagnosis Of Mechanical Faults

At present,the diagnostic method based on deep learning has achieved good diagnosis effects,but such method requires magnanimous samples.Nevertheless,in industry,most of them are sudden faults,and the collection and marking of fault samples are time-consuming and laborious,which often leads to insufficient fault samples,namely,small samples.In cross condition of fault diagnosis,the recognition accuracy decreases,the network model often needs to be redesigned,and the generalization performance is weak.Therefore,this paper carries out relevant research on how to build a highprecision diagnostic model in small sample scenarios and enhance the generalization performance of the model.The main research work is as follows:(1)In view of small samples and unbalanced problems,A fault diagnosis model based on GAN and CNN is constructed.From the point of view of data enhancement,GAN is designed to generate samples to expand and balance data sets.GAN is used to generate one-dimensional original time domain signal collected to retain effective fault information.In view of the long training time of GAN and the difficulty of reaching Nash equilibrium,the signal is truncated in this paper,make several separate samples,and take each sample as input,which shorens the time to reach Nash equilibrium and improves the quality of sample generation.This article designed a quantitative and qualitative sample quality evaluation method.Qualified analysis is performed through the waveform diagram,spectrum diagram,and packaging spectrum,and quantitative analysis is performed through time domain indicators,fault characteristics frequency,which realized the sample quality analysis and screening in time domain and frequency domain.The algorithm verification was carried out on two kinds of bearing data sets,and the deep CNN was constructed to realize fault classification on the expanded and balanced data sets,which proved the effectiveness of the proposed model in small sample fault diagnosis.(2)Aiming at small samples and zero sample problems,a small sample diagnosis method based on mechanism simulation and data-driven was built.The simulation model of fault vibration signal was constructed according to the fault mechanism.Parameters were selected from a quantitative perspective through impact amplitude,natural frequency,fault characteristic frequency,correlation evaluation,and parameters and signals were fine-tuned from a qualitative perspective through time domain waveform diagram,spectrum diagram and envelope spectrum diagram.The GAN model is designed to further generate the simulation signals to increase the richness of fault knowledge in the simulation signals,which solves the problem of insufficient accuracy of diagnosis results due to the error between the simulation signals and the actual signals.In view of the lack of diagnostic ability of a single simulation model,the lack of mechanism of a single data-driven model and the requirement for generous samples,the simulation signal and generated signal are randomly fused as the training set,and the real signal is taken as the test set,which solves the problem of small samples or even no samples and improves the diagnostic accuracy.A Deep CORAL domain adaptive classification network model was constructed to further reduce the impact of differences between simulated data,generated data and real data on test accuracy.Multiple convolution blocks were designed for feature extraction,and CORAL Loss function of nonlinear transformation was used to align the features of training set and test set,reducing the gap between training set and test set,and improving the generalization performance.The bearing data set is used to test the availability of the method in this chapter.(3)To solve the problems of small sample and weak generalization ability under cross-working and cross-device scenarios,a small-sample fault diagnosis model based on DML and GCN is constructed.The metric metalearning is integrated with GCN,the meta-learning idea is used to construct tasks and set extraction methods,and tasks and samples rather than simple samples are taken as the input of the whole network.An automatic measurement module is constructed,GCN is designed to replace the traditional distance measurement function,and the connection weight between nodes in GCN is used as a measure of similarity to realize the function of automatic measurement and avoid the manual selection of measurement function.A multi-convolution block feature extraction module is constructed,and a BN layer is added to speed up network convergence.A layer of Dropout is added after the nonlinear activation layer of the last two convolution blocks to suppress the occurrence of overfitting during element training.Besides,a kind of graph sample fused with label information was constructed.The feature extracted by the feature extraction module and its label vector were used as the nodes of the graph sample.MLP was designed to replace the distance function to automatically calculate the similarity between nodes,which enables the automatic construction of edges.Through the cross-working condition and cross-equipment fault classification of small sample experiment verifies the significance of the method.