Automatised quality assessment in additive layer manufacturing using layer-by-layer surface measurements and deep learning

---

CIRP ICME 2020 paper: Automatised quality assessment in additive layer manufacturing using layer-by-layer surface measurements and deep learning Additive manufacturing (AM) has gained high research interests in the past but comes with some drawbacks, such as the difficulty to do in-situ quality monitoring. In this paper, deep learning is used on electron-optical images taken during the Electron Beam Melting (EBM) process to classify the quality of AM layers to achieve automatized quality assessment. A comparative study of several mainstream Convolutional Neural Networks to classify the images has been conducted. The classification accuracy is up to 95 %, which demonstrates the great potential to support in-process layer quality control of EBM.And the error analysis has shown that some human misclassification were correctly classified by the Convolutional Neural Networks.

Getting started

---

Installation

• Clone this repo: git clone https://github.com/hy-son/Images-classification-extraction
• Create a new conda environement (Python 3.6) conda env create -f environment.yml

Some import may fail, to fix them please run:

• pip install --force-reinstall --no-cache-dir numpy
• pip install --force-reinstall --no-cache-dir --user torchvision

Configuration:

All configuration data are stored in data.py. model_name and trainned_dir must be list of the same size as for the n model_name the n trainned_dir file will be loaded.

Data

The ELO images data must be stored in the Data folder, split in 3 folders train, test, val and all images must be in a folder with of there class. Ex: Data\train\0\10-44-52_01.jpg The required images size is 224 by 224. You can download a small data sample here: https://drive.google.com/file/d/1yoarg6nhrNrUypNiRZ8vsZ9uN2Ji6gsQ/view?usp=sharing

Use

You can test the accuracy of the neural network by running main.py. The jupyter notebook Classification extraction execution time.ipynb will allow you to create the confusion matrix. To train launch 5CNN.py, (the trainning parameters are stored in his dataclass Project_data).

Training

The 5 networks can be trained as show in 5CNN.codeexaple.py

Citation

If you find this code useful, please consider citing my paper. DOI

 @article{le roux_liu_ji_kerfriden_gage_feyer_körner_bigot_2021,
  title={Automatised quality assessment in additive layer manufacturing using layer-by-layer surface measurements and deep learning},
  volume={99},
  DOI={10.1016/j.procir.2021.03.050},
  journal={Procedia CIRP}, 
  author={Le Roux, Léopold and Liu, Chao and Ji, Ze and Kerfriden, Pierre and Gage, Daniel and Feyer, Felix and Körner, Carolin and Bigot, Samuel},
  year={2021}, 
  pages={342–347}}