Research On Small Sample Deep Learning In Bearing Fault Diagnosis System

Rolling bearings as one of the key parts of rotating machinery,its running state greatly determines whether the rotating machinery can work normally.The fault diagnosis technology of mechanical equipment can accurately identify the running state of the bearing.To a certain extent,it can ensure the safe and stable operation of bearings.The fault diagnosis technology based on deep learning surpasses the traditional fault diagnosis technology in many aspects,but there are still deficiencies in some aspects.Therefore,three innovative bearing fault diagnosis methods based on deep learning are proposed:(1)A bearing fault diagnosis model based on LSTM-SAE combined network is proposed.This model superimposes LSTM and SAE sequentially,and has the functions of feature extraction of LSTM network and noise reduction filtering of SAE network.The model directly takes the original vibration signal as input,which meets the needs of end-to-end intelligent diagnosis in practical applications.It is verified on the data sets from different sources that the diagnostic accuracy of the model is above 99.00%.The premise of this model is that the target data is sufficient and the data distribution of the training set and the test set is the same.However,in actual production,the working conditions of rolling bearings are complex and changeable,and cannot meet the above conditions.Therefore,the fault diagnosis method for bearings under a small sample is proposed.(2)A fault diagnosis model based on WKTCA-SAE is proposed,which realizes intelligent diagnosis of bearing faults under a small sample.The model first uses the FFT transform to obtain the frequency domain features of the source and target domain data,then uses the WKTCA algorithm to map the frequency domain features of two domains to the same feature space,and finally the SAE network with classification function is used to realize feature self-learning and fault diagnosis.The diagnostic accuracy of this model is higher than 99.50%under different transfer data sets.The model needs the help of source data close to the target data distribution.(3)A fault diagnosis model based on IBNN is proposed to realize intelligent diagnosis of bearing faults under small samples.IBNN converts the fixed value of the network weight into a gaussian distribution,and uses reparametrization tricks to convert the global uncertainty into a local uncertainty,thereby speeding up the training.At the same time,the variance of the gaussian distribution is used as a parameter to participate in model training,which effectively prevents the network from falling into a local optimum.The model first uses the FFT transform to obtain the frequency domain features of the training data,and then inputs the frequency domain features into the IBNN to obtain the distribution of the predicted values.The model can learn from a small amount of data without the aid of source data.It is verified on the data sets from different sources that the test diagnostic accuracy of the model is above 99.20%.