A New Bearing Fault Diagnosis Framework With Deep Adaptation Networks For Industrial Application

In the past decades, a host of fault diagnosis methodologies have been designed and successfully used for bearings. However, most of them still have two deficiencies. (1) Traditional methods extract and select features manually according to a specific issue, but these features may be not appropriate for other tasks, leading to performance degradation of fault diagnosis. (2) Many studies assume that the dataset for model learning obey the uniform distribution as the testing dataset do, which seldom accords with the practice. To remedy these problems, we devise a novel framework for bearing fault diagnosis. First, the raw condition monitoring data are converted to 2D images with continuous wavelet transform. Then the classification model is learned with these 2D images, during which the transfer learning scheme, deep adaptation networks, is introduced for adapting the deep model trained with source data for use in new but related target domain. The presented approach is demonstrated with bearing condition monitoring information, and the results indicate it can identify bearing faults effectively under different operational conditions and has a higher accuracy than conventional approaches.