LiConvFormer: A lightweight fault diagnosis framework using separable multiscale convolution and broadcast self-attention

• Core codes for the paper:
  LiConvFormer: A lightweight fault diagnosis framework using separable multiscale convolution and broadcast self-attention
• Created by Shen Yan, Haidong Shao, Jie Wang, Xinyu Zheng, Bin Liu.
• Journal: Expert Systems With Applications

Our operating environment

• Python 3.8
• pytorch 1.10.1
• numpy 1.22.0 (If you get an error when saving data, try lowering your numpy version!)
• and other necessary libs

Datasets

• Case1: XJTU gearbox
• Case2: XJTU spurgear
• Case3: OU bearing
• Save dataset

Guide

• This repository provides a lightweight fault diagnosis framework.
• It includes the pre-processing for the data and the model proposed in the paper.
• We have also integrated 7 baseline methods including 4 CNN methods and 3 fault diagnosis methods based on CNN-Transformer for comparison.
• train_val_test.py is the train&val&test process of all methods.
• You need to load the data in above Datasets link at first, and put them in the data folder. Then run in args_diagnosis.py
  Pay attention to that if you want to run the data pre-process, you need to load Case1, Case2 and Case3 in Datasets,
  and set --save_dataset (in args_diagnosis.py) to True; or you can just load the Save dataset, and set --save_dataset to False.
• You can also choose the modules or adjust the parameters of the model to suit your needs.

Initial learning rate

• Liconvformer: Case1--0.01; Case2--0.001; Case3--0.01
• CLFormer: Case1--0.01; Case2--0.001; Case3--0.01
• convoformer_v1_small: Case1--0.001; Case2--0.001; Case3--0.001
• mcswint: Case1--0.001; Case2--0.001; Case3--0.01
• MobileNet: Case1--0.01; Case2--0.001; Case3--0.001
• MobileNetV2: Case1--0.01; Case2--0.001; Case3--0.001
• ResNet18: Case1--0.001; Case2--0.001; Case3--0.001
• MSResNet: Case1--0.001; Case2--0.001; Case3--0.001

Pakages

• data needs loading the Datasets in above links
• datasets contians the pre-processing process for the data
• models contians 8 methods including the proposed method
• utils contians train&val&test processes

Citation

If our work is useful to you, please cite the following paper, it is the greatest encouragement to our open source work, thank you very much!

@paper{
  title = {LiConvFormer: A lightweight fault diagnosis framework using separable multiscale convolution and broadcast self-attention},
  author = {Shen Yan, Haidong Shao, Jie Wang, Xinyu Zheng, Bin Liu},
  journal = {Expert Systems With Applications},
  volume = {237, Part A},
  pages = {121338},
  year = {2023},
  doi = {doi.org/10.1016/j.eswa.2023.121338},
  url = {https://www.sciencedirect.com/science/article/pii/S0957417423018407},
}