Research On Fault Diagnosis Of Aeroengine Bearings Based On Imbalanced Datasets

Aero-engine spindle bearing is an important component of aero-engine transmission system,whose reliability directly affects the running state and performance of the whole engine.In addition,bearing failure is also one of the weak links of aero-engine reliability.Once occurs,it will lead to the failure of the whole transmission system,hence serious economic losses and catastrophic accidents.Therefore,fault diagnosis of rolling bearings can provide effective support for engine health management and maintenance decision supporting,which is of great significance.However,due to the accidental nature of bearing faults,the number of fault samples in the data collected during the actual operation of the bearing is often much smaller than the number of healthy samples.The training dataset composed of those is called imbalanced dataset.Using such dataset for training will bring difficulties to bearing fault diagnosis.To solve this problem,this study focuses on the following aspects :Firstly,this study analyzed the fault mechanism of rolling bearings.By explaining the corresponding relationship between fault types and vibration characteristics,this paper proves that vibration characteristic is an important basis for fault diagnosis,which can also guide the sample augmentation network to synthesize high-quality pseudo-fault samples.Secondly,considering the imbalance problem of fault types,this paper compared SVM,CNN and DBN to explore the influence mechanism of imbalanced data on the diagnosis accuracy and stability.The experimental results show that with the increase of imbalance ratio,the mistake diagnostic rate and false negative rate of majority class also increase.Under the same imbalance ratio,DBN has the best diagnosis accuracy and stability,so it is selected as the pattern recognition network of fault diagnosis model.Then,aiming at the problem of bearing fault signal feature coupling and insufficient learning ability of GAN for disentangled representations,this study proposed an Info GANbased model with regards to learning the vibration features.By adding the latent codes to the input,the vibration features in the original data can be extracted and utilized into the sample enhancement.To testify the effectiveness of the proposed method,a series of comparative experiments were carried out on a bearing simulation dataset.Using feature parameter estimation and similarity analysis to test the quality of synthesis,the results show that the average synthesis error for vibration characteristics is only 2.159 %.Compared to SMOTE,GAN,and WGAN,the proposed method has better sample enhancement effects.A rolling bearing fault diagnosis model for imbalanced datasets is further proposed.The model combines Info GAN with DBN regarding to the vibration features.To testify the effectiveness of the proposed method,a series of comparative experiments were carried out on CWRU bearing dataset.The results show that compared with the original imbalanced dataset,the mistake diagnostic rate of the dataset processed by this model decreases from 4.25 % to0.6 %,and the false negative rate decreases from 13.8 % to 0.25 %.Finally,compared with other diagnosis models,the results show that the proposed model always performs better in terms of accuracy,precision,recall and F1-score.The model can make good use of synthetic samples and extract rich and representative features for final diagnosis.