Bearing Fault Diagnosis Of Asynchronous Motor Research Based On Deep Learning

Asynchronous motors are widely used in national defence,transport,production and daily life.If the motor malfunction,it will affect the operation of the entire system,even causeing economic losses and casualties.Among many faults of the motors,bearing faults account for the highest proportion.Timely detection and replacement of faulty bearing can avoid cascading faults effectively.Therefore,it is important to diagnose the fault diagnosis of asynchronous motor bearing is significant.Thesis proposes a new feature extraction method which focuses on bearing geometry,operating speed and operating frequency.Then,a fault diagnosis scheme based on convolutional neural networks is designed to achieve cross-conditions classification of bearing faults.Besides,a data feature-based transfer learning is designed to achieve cross-bearing state classification without changing the model parameters.The main research content of thesis are as follows:(1)Firstly,the theoretical characteristic frequencies of the different components of rolling bearings in failure are analyzed conjunction with their structure and the mechanism of fault generation.Secondly,the empirical modal decomposition is used to decompose the vibration signal of bearing into a series of intrinsic modal functions.Then energy contribution of each function is computed,and spurious components and noise in the signal are screened out according to their contribution.Finally,the vibration signal is reconstructed to complete the feature information enhancement.(2)A sliding window segmentation of the reconstructed signal is performed.Then energy spectrum is obtained using a variational modal decomposition and Hilbert transform.According to the geometrical parameters of the bearing,the characteristic frequency in the corresponding state is calculated,and the fault characteristic band is constructed with the theoretical value as the centre.The components of the band are then filtered according to their energy,and the frequency and energy values of the retained components are extracted to construct a feature matrix in chronological order.Afterwards,to address the problem of large order of magnitude differences between the feature frequencies and the energy of the corresponding vibration signals,a linear amplification strategy is used to increase the energy values and narrow the gap between frequency and energy to achieve local energy value enhancement within the feature matrix.Finally,a convolutional neural network is built to learn the feature set,and the model structure and parameters are modified according to the classification performance.The accuracy of 99.7% can be achieved for cross-bearing state classification.(3)Different classification models and different feature extraction methods are used as variables for the ablation,experimentation to validate the effectiveness of the diagnostic scheme,respectively.Based on this,the feature reinforcement coefficients are obtained by combining the migration learning theory based on sample features.Then,two classification methods,namely overall feature reinforcement and single-class feature reinforcement,are used for bearing state voting.Finally,the accuracy of state classification for another types of bearings without changing the model is able to be87.6%.The feasibility and effectiveness of the state classification scheme has been verified experimentally,and state classification of cross-conditions and cross-bearings can be achieved without adjusting the model.