Skip to content

Latest commit

 

History

History
36 lines (26 loc) · 2.16 KB

README.md

File metadata and controls

36 lines (26 loc) · 2.16 KB
Raw

GAOSD

One-shot learning with Generative adversarial network

Code accompanying the paper "Generative Adversarial One-Shot Diagnosis of Transmission Faults for Industrial Robots" by Authors (Ready to be submitted for publication).

  • Tensorflow 1.15.0 implementation
  • Inspired by Jeff Donahue $et$ $al$. [Adversarial feature learning] (https://arxiv.org/pdf/1605.09782.pdf)(Bi-GAN)
  • This repository contains several experiments mentioned in the paper
  • The proposed GAOSD was verified with the local six-degree-of-freedom industrial robot dataset.
  • One of the implementations for Bi-GAN using the MNIST dataset is shown at (https://github.com/jeffdonahue/bigan)

Requirements

  • python 3.7.13
  • Tensorflow == 1.15.0
  • Numpy == 1.19.5
  • Keras == 2.3.1

Note: All experiment were excecuted in Google colab with Tesla T4 GPU alt text

Main file discription

  • --main: The GAOSD model we build for runing some experiments. It is a class and based on tensorflow 1.15.0.
  • --main_saprseae: Main Functions about Sparse auto-encoder.
  • --main_dcae: Main Functions about deep convolutional auto-encoder.
  • --encoder_bigan: To project the dataset into the features space with a trained encoder from Bi-GAN (main.py).
  • --encoder_dcae: To project the dataset into the features space with a trained encoder from the deep convolutional auto-encoder (main_dcae.py).
  • --encoder_sae: To project the dataset into the features space with a trained encoder from the sparse auto-encoder (main_saprseae.py).
  • --encoder_wpt: To project the dataset into the featurets space of Wavelet packet transform (wpt).
  • --model: Model architectures

Implementation details

  • The overall experiments include GAOSD,OSD-SAE,OSD-DCAE and OSD-WFE are included in src. Note that users should change the directory to successfully run this code.
  • Hyperparameter settings: Adam optimizer is used with learning rate of 2e-4 in both the generator and the discriminator;The batch size is 64, total iteration for Bi-GAN is 1000. For the random forest, 100 nodes were chosen with their default setting for running 100 times to get an optimal result.