Research On Intelligent Diagnosis Method Of Rotor System Based On Residual Network

The rotor system is a key component of rotating equipment.In case of failure,it is easy to cause unplanned shutdown of the equipment,which will bring heavy economic losses and even casualties to enterprises.Therefore,the research on fault diagnosis of rotor system is of great significance.With the advent of the era of "industrial big data" and the rapid development of deep learning theory,the research on rotor system has gradually changed from the traditional diagnosis method with "feature engineering" as the tool to the intelligent diagnosis method with "deep learning" as the tool.However,there are still some problems to be solved in the intelligent diagnosis of rotor system:(a)the input of conventional deep learning model is required to be two-dimensional signals,while the vibration data monitored on rotor system is one-dimensional signals.(b)The traditional feature extraction and fusion methods of multi-source heterogeneous data of rotor system are not effective.(c)The performance of the diagnostic model decreases when the data are not uniformly distributed due to the change of working conditions.In view of the above problems,this thesis studies the intelligent diagnosis method of rotor system,and the main contents are as follows:(1)Fault feature extraction method of rotor system based on multi-channel one-dimensional residual network.Aiming at the problem that the conventional deep learning model can not directly process one-dimensional signals,this thesis constructs one-dimensional residual network to automatically mine fault sensitive features directly from one-dimensional monitoring signals.At the same time,"multi-channel input" is introduced,that is,the monitoring information of multiple positions of the equipment is collected at the same time and input into the one-dimensional residual network in the way of multiple channels in parallel,so as to solve the problem that single channel input can not fully characterize the health status of the equipment.The fault diagnosis experiments of rotor system and rolling bearing show that: The accuracy rates of the proposed one-dimensional residual network method are 24.31% and 23.42% higher than that of the traditional diagnosis method based on "feature engineering".Compared with single channel input,the accuracy rates of multi-channel input in the two experiments are improved by 15.6 percentage points and 3.7 percentage points respectively.(2)Coupling fault diagnosis method of rotor system based on multi-source heterogeneous data fusion.In view of the problems that the existing single structure diagnosis models can not deal with multi-source heterogeneous data and the poor effect of traditional feature extraction and fusion methods,a multi-mode residual network is constructed to extract features from one-dimensional vibration signals and two-dimensional infrared images,and discriminant correlation analysis is used to realize feature fusion.The fault diagnosis experiment of rotor system shows that after multi-source heterogeneous data fusion,the diagnosis accuracy rate reaches 99.79%,and can accurately identify rotor coupling faults of rub-impact and misalignment,which is 7.26 percentage points and 6.46 percentage points higher than that only using vibration signals or infrared images.(3)Transfer diagnosis method of rotor system with variable operating conditions based on residual adversarial neural network.In view of the problem that the performance of the diagnosis model decreases when the data is not uniformly distributed due to the change of working conditions,a residual adversarial neural network is constructed for domain adaptive training to extract domain invariant features and overcome the interference of working conditions on the diagnosis effect.The two transfer diagnostic experiments of rotor and rolling bearing show that the diagnostic accuracy rates are improved by 23.73% and 14.17% respectively after domain adaptation training.Compared with shallow transfer learning methods such as TCA,the diagnosis accuracy rates of the proposed deep transfer method based on residual adversarial neural network in the two experiments are improved by 35.42% and 24.23%respectively.