Deep Convolutional Neural Network Using Transfer Learning for Fault Diagnosis

Fault diagnosis is critical in industrial systems since early detection of problems can not only save valuable time but also reduce maintenance costs. The feature extraction process traditional fault time-consuming and laborious work. Recently, with the rapid development deep learning (DL) method, it has shown its superiority an end-to-end been applied to classification other fields. To a certain extent, solves disadvantages manual method. However, available training data often limited, will degrade performance DL methods. A new method that combines convolutional neural network (DCNN) transfers (TL) for proposed this paper handle different types. signal processing converts one-dimensional time-series signals into grayscale images firstly applied, eliminate effect handcrafted features. Secondly, optimal DCNN designed trained ImageNet datasets, which extract high-level features massive images. Finally, TL further developed apply knowledge learned source distribution target distribution, greatly reduces dependence on improves generalization DCNN. Three well-known including bearing vibration dataset from Case Western Reserve University (CWRU), self-priming centrifugal pump (SPCP), force Paderborn, are utilized Some popular methods added comparison. Results show precisely identify types have highest accuracy among