Applocation Of Deep Learning In Rolling Bearing Fault Diagnosis

With the tremendous opportunities and challenges brought about by the rapid development of modern machinery and equipment,mechanical fault diagnosis technology is also burgeoning toward intelligence.Rolling bearings are vulnerable parts of rotating machinery.The complexity,variability and uncertainty of their faults make their fault diagnosis requirements particularly prominent.The enthusiasm of artificial intelligence triggered by deep learning technology has engulfed many research fields including fault diagnosis.Deep-learning-based fault diagnosis methods have achieved better results than traditional methods in many aspects.The ability to handle big data and multi-source heterogeneous data is uniquely advantageous.Therefore,the rolling bearing fault diagnosis technology focuses on the knowledge-driven approach of signal processing,and is gradually transitioning to a data-driven approach that focuses on the study of data intelligence learning.However,the deep learning in the field of rolling bearing fault diagnosis research and development is still insufficient,and it lags behind other application fields.A deep learning model based on vibration acceleration signal data samples of multiple failure positions and different damage states of rolling bearings is build for training learning and intelligent classification.The data sources include two types,which are the bench test data of the electrical engineering laboratory of Case Western Reserve University and the measured service data of locomotive lines.There are seven kinds of failures,including normal operation and failures of the outer ring,inner ring,and rollers.Two kinds of deep learning models were designed and compared,and the accuracy of different methods was analyzed through experiments.First,a model based on LSTM is proposed.LSTM is a variant of a recurrent neural network and is the most natural way to deal with time series problems such as vibration signals.The DNN is mixed in the network to enhance the nonlinear feature mapping capability of the network.A batch normalization layer is introduced to solve the network "gradient dispersion".And a random search algorithm is designed to automatically optimize the hyperparameter.The original data is directly used for training and testing to avoid the loss of original information due to the extraction of feature values.Under various experimental conditions,more than 99.8% accuracy can be achieved.Subsequently,a model based on the concept of FDCNN is designed and implemented.The invariance of the convolutional network can well deal with the vibration signals with diversity and heterogeneity.Using the Fourier transform spectrum image of the vibration signal as input,accurate diagnosis of the above-mentioned 7 types of faults with different rotation speeds,different sources,and different dimension data is performed.For existing diverse,multi-source,heterogeneous data,the diagnostic accuracy is as high as 100%.Compared with other machine learning diagnostic methods,the deep learning model presented in this paper has higher recognition accuracy and resolution.The effective diagnosis of diverse,multi-source,and heterogeneous fault data also illustrates its generalization ability.Finally,the improved CDCGAN model is designed and implemented to implement a single fault and some composite fault samples from known data.The training network is used to generate the specified category compound fault samples of unknown data.It is proved that the linear superposition of a single fault in hidden space can generate compound faults..Analysis of composite fault sample generation conditions,design of data orthogonalization filter to meet generation conditions,design of conditional image generation method for compound faults,concrete realization of generator and discriminator model structure construction,and adoption of an improved FDCNN model for data Classification,proving the advantages of generating samples.