The benefits of an early-stage defect detection for critical Additive Manufacturing applications
by Maud Chidiac, Program Manager AI, AMEXCI
A couple of months ago we published an article about the open innovation hackathon that we initiated with PhD students and researchers in the field of metal Laser Powder Bed Fusion (L-PBF) data. Read more here. Now that the labelling is done -both in-house and through all the experts, who contributed and who we thank here again – it felt natural to step back a little bit and take some perspective on the ongoing Artificial Intelligence related developments at AMEXCI.
The point of labelling our metal L-PBF defect data was to obtain a structured and labelled database that we could work with existing AI techniques. In other words, we transferred the defect detection knowledge from the domain experts to the model. Once labelled, the working data was used to train a classification model that could recognize learnt defect patterns on a new and unlabelled set of data of the same type.
Overall, our goal is to use AI techniques to classify L-PBF data, to detect early-stage defects happening in the process, and understand what corrective action to take during, and after the print. This will enable AM users to save material, machine capability, testing time, and effort by avoiding failure, crashes, and unnecessary testing on non-defective parts. Automatically recognizing defaults within a component using layer-by-layer monitoring, can be a truly transformational technique, that can support industries to speed up their adoption of AM.
We would like to exemplify those benefits through a case that we have built together with Siemens Energy, to answer the question – how can AM users benefit from this new intelligence?
1. The case of Siemens
Today, Siemens Energy prints 8 gas turbines per build plate, a total of 9500 layers build job corresponding to a 1 week printing time. To certify that the component is compliant to quality standards, Siemens Energy needs to do conventional flow testing to detect geometrical deformations. Those testings are costly and induce long lead times. To identify how an AI based defect recognition can improve those quality control loops, Siemens Energy has kindly agreed to share build data with us so that we could process it and evaluate the outcomes from the AI analysis.
The original scope of the project for Siemens Energy was to perform the classification, to analyse the geometrical deformation of inner channels of a burner, and more generally, to detect all process deviations and potential defects. We have labelled a few examples of obstruction of internal channels within our training dataset, but the amount of training data for this class of defect turned out not to be sufficient for the classifier to recognize them on new datasets. Therefore, no area was recognized as internal obstruction in any of the test jobs.
However, the analysis has recognized different process deviations areas, classified into several types of defects. We have trained our classifier with the seven classes of defects, which compose our annotated database: lack of fusion, spatters, recording problems, recoating irregularities, process dependant overheating, surface deformation, false positive.
2. Trying out the machine learning
Within the scope of the project, Siemens Energy has provided AMEXCI with an STL file for the body and lattice structure of the gas turbines. To reproduce the situation, and showcase how the AI powered classification can bring tangible value to Siemens Energy, as an AM user, we have printed 8 dummy burners made from AlSI10Mg – we are printing the experiment in aluminium instead of 316L since we have today a labelled training dataset for aluminium – with the same set-up on the build plate. The printing time was 50 hours for a 5146 layers build job, and has generated the following data: 5146 Powder Bed pictures after melting in .jpeg, 5146 Powder Bed pictures after recoating in .jpeg, 5146 Optical Tomography .raw pictures and .eosot log files, historical sensor data, a list of 80 detected process deviations showing abnormal hot or cold spot in the melt pool area and general statistics about the parts.
After processing Optical Tomography pictures and Powder Bed pictures into our AI classification model, the algorithm has classifed and recognized defaults. The outcome from the analysis is a .csv file with recognized classes of defects, their coordinate and prediction confidence of the class recognition. The chart below shows the proportion of defects recognized by the model.
Since the outcome from our model is a .csv file, we have worked with the company Interspectral to match the different data inputs (OT,PB, STL file) and to integrate the outcome from the AI on the 3D object, to visualize the area where a defect was found. The video below shows a visualization, in AM Explorer, of the classification results on the turbines, enabling to locate the defaults and understand where they are in the part.
3. Working with those results
When a class of defect, or of process deviation is recognized, the goal is to understand if it requires action, or if it is to be ignored (false positive). We want to correlate the image classification for process deviation and defect recognition with other build process conditions (e.g. build chamber related sensors such as oxygen, temperature, gas flow, as well as process parameters). This correlation would enable to interpret the cause for a defect and to generate recommendations on what action to take.
This outcome needs to be verified. This analysis combines all labels in one model, yet all labels are showing different buckets of abnormalities. While some might recognize process deviations and send warnings to the machine operators to evaluate the right live corrective action to take during the print (for example with the recognition of recoating irregularities, surface deformation, overheating, etc), some other labels are focused on formed defects within the part (lack of fusion for example). Two levels of analysis are therefore considered. While the first type of labels (process deviations) is quite straightforward to confirm, the actual defects that are recognized by the model should be verified with specific targeted testing. At a later stage we will separate the different “labels buckets” into distinct models trained for each use case, and perform strong verification testing.
Conclusion
We are happy that we came to the end with this prototype, having the ability to prove the technical feasibility of extracting and working with AM datasets, to apply AI machine learning and deep learning techniques like clustering and classification. We have proved the technical feasibility of recognizing process deviations on commercially available monitoring pictures. Now the approach can only be improved from quality and quantity of label to verification of the model analysis with testing, as well as diversification to other monitoring systems and materials. We will continue the work in this field, and we look forward to letting you know about the next development.
Since we have used a data-centric approach for the AI based development, the main groundwork for us was to work around the different challenges related to data collection, data clustering, data labelling and data classification. Iterations on the data collection strategy to find the right scope, lack of data quantity, lack of ground truth and experience with AM metal L-PBF process data, were for us the main problematic tasks. One could say that dealing with the data is almost more challenging than applying the AI techniques per see. Before being able to do the latter, the data must be structured, consistent, and must capture the right content. It is only when the groundwork on the data is complete, that one can apply generic Artificial Intelligence techniques onto the dataset.
We would like to thank Siemens Energy for contributing to this model evaluation and for giving requirements as potential users. Feel free to reach out if you would like to know more about it!