Detecting Anomalies in Additive Manufacturing using signalAI
Altair EmployeeA customer success story on how Renishaw used signalAI to flag anomalous builds in additive manufacturing.
Picking up from where we have left last time, in today’s article let’s learn how Renishaw, one of the leading players in additive manufacturing, had used signalAI in their workflow to detect anomalies. The last blog already discussed what are some of the common techniques used for anomaly detection, today let’s get our hands dirty with the data coming from Renishaw’s printing machines and see how it works in the field.
As we all know, in the field of data analytics the data under consideration is often the most important part of the process. So, let’s start with data we are dealing with here. In this use case, we are provided with 3 types of data coming out of spectral sensors corresponding to 3 printed builds (Build number 1,2 & 3) and the goal is to detect if there is an anomalous build hiding in these 3 printed builds by analyzing these sensor readings. This is phase 1. Building on this, once we detect the anomalous build, in phase 2 we want to go a level deeper and also want to find out which sub block in that particular build is problematic. The sensors we are using here have a very high sampling rate of 100kHz and they record the data over a span of 4.6 seconds for each of the 333 layers each build is consist of. So, we are dealing with datasets which easily run into gigabytes and top of this we don’t have any corresponding “good” vs “bad” labels to learn and train from. Long story short: we have to play a bit like Sherlock Holmes here 
.
As we are dealing with a very high sampling rate, as step number one, we first performed a bit of a preprocessing/feature engineering to extract 5-time domain statistical features per layer to both reduce the noise in the raw data as well as to make the training process of the downstream anomaly detection models a lot faster. This also helps in reducing the sheer size of the raw data. In step 2, these extracted features are then fed into the signalAI’s hyperplane, density & tree-based algorithms to see if we can distinguish the outliers. After playing with all the three algorithms, we have observed that the tree-based method has performed the best here. We benchmarked it on two factors. First one being by analyzing each spectral parameter independently, it flagged the “Build 3” as the anomalous one. So, it’s very consistent in its prediction. On top of this the tree-based algorithm also has higher confidence probabilities in its predictions. Once we have detected the bad build, we have basically repeated the same process and also successfully detected the anomalous sub block in that particular build. From the implementation perspective within Altair’s product line, we have demonstrated the capability of performing the data analytics part here using signalAI both in Knowledge Studio as well as in Activate. Expanding on this we have also showcased the benefits of adding Panopticon dashboards to visualize and monitor the final results in real time using a MQTT connection if remote monitoring is required. This data driven anomaly detection workflow will help Renishaw make printers that stand out from the from the competition with faster and more accurate part qualification and simpler QA for production.
In conclusion, this use case shows that it is possible to successfully detect anomalies from a data driven approach without prior knowledge. This has exciting implications for wider use of machine learning in metal additive manufacturing and even the whole manufacturing domain, as well.
Thanks again for the read! Happy learning!