Let me give a simple introduction to the artical. With all my codes.
Here's the pdf of the artical: PRN-20240829.pdf
In this artical, we proposed such a method for Beijing PHM 2024 conference
This model mainl contains 3 part:
The RSN is designed to extraction the high-quality features by itself from time-frequency images. Unlike ResNets, RSN adds a soft threshold as a nonlinear adjustment layer. This layer effectively removes noise-related information, thereby extracting meaningful features.
In order to describe the similarity of features between samples, we established the concept of sample pairs. In the feature concatenation module, features from different source samples are concatenated together as sample pairs.
In the KARN module, we use Kolmogorov-Arnold Networks (KAN) to serve as relation networks. KAN is an innovative neural network architecture that proposes an alternative to the traditional multilayer perceptron (MLP).
Model structure of PRN
An ASL is constructed to help extract high level features. Samples in this lib are generated by the soft Brownian offset method. To know more about this method, you can read this artical: Out-of-distribution Detection and Generation using Soft Brownian Offset Sampling and Autoencoders
Datasets are not included cause it's too large
You can download it on conference website: https://www.icphm.org
Official Website for PHM 2024