Application Of Transfer Learning In Bearing Fault Diagnosis

Bearings are important components of machines.Different from the other components of the machine,bearings always run with the operation of the machine.As the connection components between the mechanical parts,bearings always work in harsh environments while they are often affected by other components.The component is very susceptible to wear and tear,as the wear become severer,the bearing fault which may damage the bearing would appear soon.Once the bearing is damaged,it will inevitably cause some machine problems which would make the machine fail to operate.The situation may reduce productivity and cause the factory to suffer significant economic losses,and even lead to a tragedy.If we can find bearing problems early,then many disasters can be avoided.Therefore,it is very important to do research on bearing fault diagnosis.Due to the wide variety of bearings and complex working conditions(such as changes in temperature,speed,and other working environments),the difference between the data is huge for which it is difficult to train a proper classifier.In order to solve this problem,we apply transfer learning to bearing fault diagnosis.With transfer learning,we can classify the data in the target domain with a classifier trained on the source domain data set,so that the data in the target domain can be classified across domains,and this task only need a really small amount of data.In the field of bearings,due to the similarities between various types of bearings,their vibration signal data could be transferable.In industrial production,although most of the bearing data is unlabeled,we can still get some labeled bearing data.A large number of experiments have proved that for this bearing data and some unlabeled bearings,the classifier obtained by the transfer learning training of the data also has a good diagnostic effect.This dissertation mainly includes two key parts: feature extraction and transfer learning.In the process of transfer learning,we analyze various features and choose the features suitable for transfer to train a classifier.For training,the main research on these two parts will be shown below.This article mainly proposes two bearing fault diagnosis methods for variable working conditions.The first is a bearing fault diagnosis method in variable conditions based on transform component analysis and bag of words.For the labeled data to be used for training(called source domain data)and unlabeled data for test diagnosis(called target domain data),first it used short-time Fourier transform to convert the two types of data into frequency domain data,and secondly mapped the spectrum energy of the two to the same distribution through transfer component analysis in order to make the bag of words for it as a feature of the data.Finally,it trained a suitable classifier on the model of the source domain data and diagnosed the target domain data with that.The experimental results under the Siemens SQI-MFS platform experimental data set,Case Western Reserve University public data set and Machinery Failure Prevention Technology data set show that the algorithm is valuable.The second is a variable conditions bearing fault diagnosis algorithm based on spectrum and joint distribution adaptive.It converts the one-dimensional vibration signal into a two-dimensional spectrum for the data under various working conditions,which is more intuitive.And then we take out the features that can describe the signal characteristics with image segmentation and dimensionality reduction.Especially,we make the gray-scale value of pixels of the vibration signal spectrogram the feature of the bearings.After that,we use the joint distribution adaptive method to reduce the distance between the source domain and the target domain.And then we train a corresponding classifier to test the data.The results of the algorithm on many data sets show the effectiveness of the algorithm.