With the increasing adoption of Additive Manufacturing in the industries, quality assurance has become a critical and increasingly challenging process. Especially in an industrial setup, in-situ monitoring is crucial to facilitate faster prototyping and industrialisation. Towards this objective, there have been increasing efforts towards real-time detection of recoating anomalies that occur in Laser Powder Bed Fusion. These defects are observed to have a detrimental effect on the quality of the final product. Machine learning methods have seen great success in addressing this problem; however, it usually involves crafting features specific to each printer type. This thesis explores how deep learning, particularly Convolutional Neural Networks, could be leveraged for further performance. A deep regression model is formulated to predict a score defining the quality of recoating at each layer from the imaging data made available through inbuilt monitoring solutions. The presented methodology is built by employing transfer learning with recent state-of-the-art architectures and demonstrated robust performance with reduced computation effort. These models also showed generalizability to changes in illumination conditions observed with different machines.
Additive Manufacturing is an emerging technology with the potential to disrupt conventional manufacturing processes by bringing digital flexibility and efficiency to manufacturing operations. However, along with the benefits it brings, it also comes with a share of challenges like processing defects and inconsistencies. The process reliability is crucial, primarily for the production of safety-related parts like turbine blades. This makes process monitoring and quality control play a critical role in addressing these challenges. There have been increasing efforts to learn high-dimensional dependencies in various Additive Manufacturing (AM) processes like Selective Laser Melting. The current quality control processes in industries are mostly open-loop with limited real-time monitoring capabilities. Hence, there has been a rise in demand to bring in more in-situ process monitoring to assist in the early detection of process instabilities preventing product failure. The rapid rise of data science and artificial intelligence has provided an opportunity to understand these process anomalies and ways to address them. Efforts have been made to review various factors that directly impact these manufacturing processes, like analysing various process parameters, correlation with material properties, and discovering different anomaly types. Different sensors and process monitoring solutions provided by AM printer manufacturers help performing a closer analysis during the prototyping phase of printing. This analysis led to a realisation that most processing defects and anomalies occur during the powder spreading phase. This perception has called for developing various methods for successfully detecting these anomalies, and this thesis is also targeted at solving this problem.
Several approaches have endeavoured to leverage machine learning for the autonomous detection of these anomalies during this phase. The classical computer vision approach is explored initially through feature extraction from the imaging data collected from the Powder Bed Monitoring tools. However, this involves tedious manual tuning of parameters and thresholds. Then there was a shift towards incorporating feature-based machine learning to address this problem. This gives more flexibility in the selection of features with enhanced interpretability. This approach further requires merely a moderate amount of data to generate predictions. However, this is a very time-consuming process due to the manual feature engineering involved and might also require readapting the features to different AM printer types and manufacturers. Deep Learning could be a potential solution for this predicament because of its recent progress and the evolution of various architectures. This approach could aid in showing a path in tackling anomaly detection problems much efficiently. It should also be noted that this would require a significant amount of labelled data to develop such models. This thesis is a step in this direction and examines the advantages and efficacy of this approach.
In this work, the following research questions are addressed:
- Can the deep learning approach yield similar results to machine learning models built with a feature-based approach?
- How does the absence of human-in-the-loop (HIL) labelling impact the training and accuracy of CNN models?
- How robust are CNNs in detecting the recoating anomalies concerning varying illumination conditions?
- A deep regression approach is formulated to facilitate feature extraction and produce an indicative score of the severity of anomalies present without any necessity for manual feature engineering.
- The deep learning models are shown to be robust to label noise which is typical in machine learning methodologies.
- The CNNs are shown that they detect these anomalies despite the dynamic lighting configuration without many adjustments to the model or requirement for more data.
The models are developed using Python 3.7, Tensorflow 2.3.1, Keras 2.4.0 and relevant libraries such as Numpy, Pandas, PIL. The training is performed on Nvidia Tesla P40 GPU with 24 GB RAM.
The model is constructed in cnn_networks.py and pretrained ImageNet weights are used for training.
The code for training and generating predictions is found in cnn_regressor.py and and hyperparameters can be changed there. Use python cnn_regressor.py csv_file image_patches_path model_network no_of_epochs batch_size learning_rate to try the code.
In the following table, the performance comparision of different CNN networks is reported below
| Model | Test MSE | Rsquare |
|---|---|---|
| ResNet50 | 0.0067 | 0.916 |
| MobileNetV2 | 0.0085 | 0.903 |
This thesis proposed a deep regression model that can successfully detect the surface-visible anomalies observed during the Powder Bed monitoring. It has improved upon the previous machine learning methodologies by employing more contemporary architectures with less complexity and further used transfer learning to increase computational efficiency. The deep learning approach also shows more potential in learning representations than machine learning methods obviating any effort for feature engineering and contextual heuristics. The novelty of this approach is using a regression model for powder bed anomaly detection as previous work usually followed a Classification approach. This procedure has also reduced the manual labelling effort involved and prevented the issue of class imbalances typically observed. Furthermore, the usage of Convolutional Neural Networks for regression is also slightly distinctive.
The results reinforce the hypothesis that predominantly a general-purpose network (ResNet-50 or MobileNetV2) adequately tuned can yield results close to the state-of-the-art without resorting to more complex and ad-hoc models. These models successfully extracted features from grayscale images, although these networks are usually optimised to learn from images with all colour channels. The efficacy of the previous feature-based machine learning model was also implicitly assessed. An impressive development is discovered that deep learning models can make the correct predictions even with small scale inconsistencies in the ground truth. This is validated by a training procedure that could achieve decent performance with deep learning models even in the presence of label noise.
The presented methodology can also be tuned further to enable its implementation in a real-time environment. The robustness of anomaly detection models to illumination changes, which is relatively underexplored, is also investigated in this study. This result shows promise that models trained with image data of one AM printer type could be used for another and reduce the need for machine-specific calibration or lighting configuration adjustments. It could also save a lot of training and data collection effort.