Efficient sub-pixel fully connected neural network: an intelligent fault diagnosis method for signal resolution enhancement

Using deep learning to augment the dataset has become a hot topic in
various fields. That is, to generate a more simulated dataset from the
limited dataset. However, the premise of the study is to use high
resolution sampling equipment to collect experimental data. In this paper,
we propose a simple and effective algorithm -- efficient sub-pixel fully
connected neural network (ESPFCN). Different from other data enhance
algorithms, ESPFCN does not explicitly generate more simulation data. On
the contrary, it performs the fully-connected operation on the original
input data and outputs the results of four-channel multi-feature maps.
Through the sub-pixel fully connected layer, the data resolution is
changed from low to high and increased to 4 times of the original.
Finally, two set of bearing and gearbox experiments are set up to evaluate
the performance of the generated model. The experimental results verify
the effectiveness of the ESPFCN model, and the feature learning process of
ESPFCN is visualized.