ANC-LSTM-fault-detection

The source code, pretrained models and dataset are released here for our IROS 2021 paper of "Soft Manipulator Fault Detection and Identification Using ANC-based LSTM" by Haoyuan Gu^†, Hanjiang Hu^†, Hesheng Wang* and Weidong Chen.

Get started

Install pytorch and tensorboardX first.

Clone this repo:

git clone https://github.com/HanjiangHu/ANC-LSTM-fault-detection.git

Prepare the dataset

This repo has been well organized with dataset in dataset folder and pretrained models in outputs folder, where the experimental results could be easily reproduced or extended for further research.

The training set and validation set are randomly spilt and each sequential sample is formatted in json with the inputs collected from the real-time system and the corresponding labels.

Train the model

To train the ANC-LSTM model for the first time use the following command under the root path of the repo.

python train.py --name ANC_LSTM

For the vanilla-LSTM model without ANC module for the comparison experiment,

python train.py --name vanilla_LSTM --att_dim 0

To ine-tune the pretrained model at XXX iteration,

python train.py --name ANC_LSTM --continue_train --checkpoint_epoch XXX

For more details about the settings of training,

python train.py -h

Validation and the real-time implementation

To validate the pretrained ANC-LSTM or vanilla LSTM model at XXX iteration on the validation set,

python validate.py --name ANC_LSTM --checkpoint_epoch XXX

python validate.py --name vanilla_LSTM --checkpoint_epoch XXX --att_dim 0

For the real-time implementation in C/C++, get the input vector from the system at the end of each control period first. Then use python.h to use the functions in the validate.py given the real-time input to infer the real-time sequential classification results with confidence.

More

Our paper will be available soon and welcome to our lab if you are interested in conducting more research with soft manipulator.