DSFD

Drill Sound Failure Detection

Software requirements:

• keras, librosa, matplotlib, numpy, pandas, scipy, sklearn, soundfile, tensorflow, tqdm. To download the dependencies:

pip3 install -r requirements.txt

• MATLAB 2020b.

Dataset:

Valmet dataset can find using this link: https://zenodo.org/record/6478574#.Ymj7ONpByUk

UrbanSound8K dataset can find using this link: https://urbansounddataset.weebly.com/urbansound8k.html

• Original features: Download

• Augmented features: Download

Note: All above datasets need to download and extract to the [data/features] folder.

Usage:

• First, clone the repository locally:

git clone https://github.com/thanhtran1965/DrillFailureDetection_SciRep2022.git

or

git clone https://github.com/nhattruongpham/DFSD.git

• Structure of repository:

|-data
|----features
|--------audio_test_varup2
|--------audio_train_varup2
|----Metadata
|--------Drill_Dataset_Test.csv
|--------Drill_Dataset_Train.csv
|----test
|----train
|-logs
|-weights
|-dataLoader.py
|-featureExtractor.py
|-Generate_Log-Mel_Spectrograms.ipynb
|-LICENSE
|-model.py
|-params.yaml
|-README.md
|-requirements.txt
|-ResultAnalysis.ipynb
|-Sound_Augmentation20Aug2021.m
|-test.py
|-train.py
|-utils.py

• Run Sound_Augmentation20Aug2021.m to generate the augmented dataset (if any).

• Run Generate_Log-Mel_Spectrograms.ipynb to extract and visualize log-Mel sepctrogram features (if any).

• Run python train.py -p params.yaml > logs/dumy.out to reproduce the several experiments and corresponding results using the proposed method. The experiments are configured by modifying lines 244 and 280 in train.py as follows:

  • CNN - Leaky ReLU:

  model = CNN_LeakyReLU(params_learn=params_learn, params_extract=params_extract)
  

  • CNN - LSTM - Leaky ReLU:

  model = CNN_LSTM_LeakyReLU(params_learn=params_learn, params_extract=params_extract)
  

  • CNN - LSTM - Att - Leaky ReLU:

  model = CNN_LSTM_Att_LeakyReLU(params_learn=params_learn, params_extract=params_extract)
  

  • CNN - LSTM - Att - ReLU:

  model = CNN_LSTM_Att_ReLU(params_learn=params_learn, params_extract=params_extract)
  

• Run ResultAnalysis.ipynb to show confusion matrix and classification report (if any).

Note: To reproduce the experiment and result for the original dataset, please change the dataset directory! You can also clone the same code from my co-author's Github https://github.com/nhattruongpham/DSFD

Citation

If you use this code or part of it, please cite the following paper:

Preprint version

@article{tran2021detecting,
  title={Detecting Drill Failure in the Small Short-sound Drill Dataset},
  author={Tran, Thanh and Pham, Nhat Truong and Lundgren, Jan},
  journal={arXiv preprint arXiv:2108.11089},
  year={2021}
}

Published version

@article{tran2022deep,
  title={A deep learning approach for detecting drill bit failures from a small sound dataset},
  author={Tran, Thanh and Pham, Nhat Truong and Lundgren, Jan},
  journal={Scientific Reports},
  volume={12},
  number={1},
  pages={1--13},
  year={2022},
  publisher={Nature Publishing Group}
}