Study On The Fault Diagnosis Method Of Rotating Machinery Based On Machine Learning

Rotating machinery,especially bearings and gears,is an important part of mechanical equipment,which has been widely used in manufacturing,transportation,chemical industry,metallurgy,aerospace and other fields.Rotating machinery,as a basic component,once it breaks down,it will do great harm,and even cause catastrophic accidents such as machine damage and human death.Therefore,it has a high practical significance for rotating machinery fault diagnosis.Once these basic parts fail,they will cause catastrophic accidents such as aircraft damage and human death.Therefore,it has a high practical significance for rotating machinery fault diagnosis.Based on the analysis and experimental research of typical fault mechanism and fault feature expression of rotating machinery,it is found that the fault diagnosis of rotating machinery based on the calculation of fault characteristic frequency is fuzzy.The possible reason is that the modulation of the noise signal to the effective signal and the single sensor detection and monitoring are not comprehensive and the perception is insufficient.Aiming at the above two possible reasons,machine learning is applied to the fault diagnosis of rotating machinery under different working conditions and different fault types.1.Aiming at the fuzziness of diagnosis caused by noise signal modulation,the bearing fault diagnosis method in high noise environment is studied.A diagnostic method based on wavelet packet and self-organizing feature mapping(SOFM)is proposed.Firstly,the noise bearing signal is denoised by wavelet packet with five layers of adaptive threshold,and six kinds of time-domain features of the denoised signal are extracted for feature extraction and dimension reduction of the data.Then,after unsupervised training with SOFM network,the unknown bearing fault can be diagnosed independently.Through the experiment on the rolling bearing signal set of Case Western Reserve University which is artificially superimposed with Gaussian white noise and actual factory noise in high noise environment,the accuracy of diagnosis is more than 96%.This method realizes the effective diagnosis of bearing fault in high noise environment.2.Aiming at the fuzziness of diagnosis result caused by incomplete monitoring and insufficient perception of single sensor,a diagnosis method of rolling bearing based on data-driven and stochastic intuitionistic model decision is proposed.This method does not need to calculate the fault characteristic frequency,but only needs to construct and match the fuzzy expert system and the samples to be tested,and takes the maximum value of the cross coordinate,that is,the likelihood measure value,as the membership degree of the supporting proposition.Then,the essential meaning of uncertain parameters is analyzed,and the membership degree of fuzzy set in the frame of random set is transformed into the membership function of intuitionistic fuzzy set.Finally,single sensor multi feature fusion and multi-sensor information fusion is transformed into intuitionistic fuzzy set multi-attribute decision fusion.The experimental results show that the accuracy of the test set reaches 99% when the training set data and the test set data are set as different sources,and the model has strong reliability and generalization ability.3.Based on the advantages of multi-sensor information fusion,a fault diagnosis method of GAFs/MTF-Res Net based on multi-sensor information fusion is proposed.Firstly,we use Gramian angular summation/difference fields and Markov transition field(GAFs/MTF)algorithm to transmit the multi-channel time series signals of different types and lengths collected by the experimental platform through multi-dimensional time Sseries(MTSI).Then,the deep residual network(Res Net)is used to train and classify the image set,and finally different fault types are diagnosed.The experimental results show that: the GAFs/MTF-Res Net model realizes the accurate classification of 520 types of fault tags,the accuracy of training set can reach 100%,and the accuracy of test set can reach 79.15%.This model bypasses the gradient vanishing and gradient exploding effects which can not be overcome by classical deep neural network in processing long time series data,so it has high research value and application value.