This repository contains the source code (written in python 3.8), trained CNN model and examples of the master thesis.
The Proceedings Paper, based on this work can be found here:
CNN based powder bed monitoring and anomaly detection for the selective laser melting process
https://github.com/Ay-De/SLM-CNN
The main difference between the proceedings paper and the complete master thesis is the real time prediction of the entire powderbed. This prediction takes not only the CNN output into account, but the prediction results of the previous printed layers. Rationale behind this approach is, that errors in the printing process are developing over time.
Example of classifications during printing of an object.
Note: Green patches are classified as powder, Blue patches are classified as printed objects, Red patches are classified to have an error.
This repository contains:
| Folder/file | Description |
|---|---|
| model | The trained CNN model, as described in the above mentioned paper |
| source | Source code for training and evaluating the model |
| CNN\preprocess_dataset.py | Preprocesses the raw powderbed images |
| CNN\train_classifier.py | File used to train the model |
| CNN\predict_powderbed_realtime.py | File used to classify entire powder bed image layer |
| CNN\predict_classifier.py | File used to classify small patches extracted from the powder bed |
| CNN\modules\* | Modules used in the source code |
| Label_Tool | Created tool to label the layer images and create the dataset of smaller patches |
| Test_Data\Layers\ | Sample full layer images from the test set |
| Test_Data\Patches\ | Sample patches, extracted from layers of the test set Note: Patches are sorted by class |
| requirements.txt | Containing the required python modules |
| images | Training and validation history Tensorflow/Keras graphical output of the trained model Classified Layer samples |
Classification of small patches and their heatmaps, showing the regions which were regarded as important for the neural network during classification. The Baseline is the starting point. Integrated Gradients are trying to find datapoints which are important for the classification. Attribution Mask is then applied to the original image as an overlay.
Patch wise classification with a patch size of 128x128 pixels and then applied to the entire powder bed layer.
Colours:
White: Powder | Blue: Objects | Red: Error
The model architecture, as seen in the previous section, was trained five times. Classification results were obtained by averaging the classification results of the test set.
| Class | Precision | Recall | F1-Score |
|---|---|---|---|
| Powder | 0.8418±0.0279 | 0.9804±0.0071 | 0.9057±0.0190 |
| Object | 0.9039±0.0150 | 0.7540±0.0291 | 0.8216±0.0127 |
| Error | 0.8343±0.0150 | 0.8607±0.0213 | 0.8471±0.0113 |
| Accuracy | 0.8574±0.0080 | ||
| Macro Average Accuracy | 0.8600±0.0070 | 0.8650±0.0076 | 0.8581±0.0084 |
| Weighted Average Accuracy | 0.8618±0.0064 | 0.8574±0.0080 | 0.8552±0.0082 |
If you want to run the model on the provided samples, please install the requirements.txt first:
pip install -r requirements.txt
After the installation, you can run the trained model by starting the python files.