Shortening time to market in 3D printing with vison-based industrial AI

Additive manufacturing is a technology, which allows completely new design and geometry of 3D parts which can’t be produced with classic subtractive manufacturing technologies. These new possibilities and freedom in design come with new challenges on the way from part design to industrialization. Naturally, printing 3-dimensional objects is not as easy as printing 2D text on paper.

**Picture 1** Additive Manufacturing allows a flexible parts design which is difficult to produce with classic manufacturing technologies, e.g., the air flow channels in the blade

At Siemens we are aiming to significantly improve quality assurance in Additive Manufacturing (AM) with industrial artificial intelligence and machine-learning to accelerate the time from prototype to industrialization as well as the efficiency in large-scale serial production.

In this blog post we will highlight how we were able to detect and evaluate quality issues in ongoing and finished 3D metal printing jobs for the turbine production of Siemens Energy in Finspång/Sweden.

Additive Manufacturing

Additive manufacturing is a specific 3D printing process. This process builds parts layer by layer by melting metal powder according to digital 3D part design. A thin layer of the powder material is added to a build platform. There, a powerful laser beam melts the powder precisely according to the part design. Next, the build platform is lowered, and the next layer of powder is added. The material is melted once again, which connects it to the layer below at the specified points.

In this powderbed recoating process, we are typically exposed to the following two quality issues:

Missing powder and
Powder dropping

Missing Powder

Picture 1 shows missing powder on several printed parts. This typically occurs when the powder dose factor for recoating is too low. Missing powder must be detected so that a quality engineer can decide if the issue is critical. The criticality depends on the type of printed part and affected layers.

Powder Dropping

Powder droppings as shown in the Picture 2 occur when under certain circumstances powder accumulates on the recoater and stochastically drops down during the recoating. When the powder falls on a printed part, it will be melted by the laser, creating a super-elevation potentially causing a recoater crash or at least severe recoater bumping that can also affect quality.

**Picture 2-left** Missing powder on printed parts
**Picture 2-right** Dropping powder from the recoater into the powder bed

The application has been developed for selective laser melting with the focus on detecting recoating issues in the powder bed. An optical camera, which is already installed in the EOS printers is used to capture the image of every single layer after recoating.

In the process of bringing a 3D model from design to serial production (internal product and process qualification process), our application supports both prototype development and serial production.

Workflow and Technical Realisation

The workflow typically starts with a prototyping phase where the goal is to optimize the 3D model design and/or the print parameters (e.g., scan strategy, laser power). For example, overheating often leads to super-elevations, which can jeopardize the print job if the recoater blade collides with the elevated melted material.

Using a powder bed camera of industrial printers (currently, already available for many models), an image is taken after the recoating of each layer. The quality of the recoating is automatically evaluated and expressed as a numerical value, called severity score, per layer.

High severity scores indicate major recoating issues like missing powder or severe scratching of the recoater blade on the melted material. Visualizing the severity scores over all layers gives a good indication of the print (recoating) quality and provides an important aid for the AM experts to shorten the number of iterations needed to improve the printing behavior in the prototyping phase.

Data of all print jobs are collected in a virtual private cloud (encrypted and secured by two-factor authentication), which facilitates the analysis and comparison across multiple print jobs and factory locations.

A profile of the severity scores of the final prototype can be used to define upper control limits for the serial production, which are then the basis for an automatic monitoring of the printing quality in the industrial phase. This could include, for example, the automatic creation of non-conformance reports (NCR).

The application calculates a severity score per printed part on the layer and additionally a severity score for the whole build plate. The severity score per part is calculated on the area of the bounding box of every single part, which helps to focus on those issues in the powder bed that can negatively impact the part’s quality. It allows a detailed monitoring of every part during the print process and is used by technical experts to evaluate if further Non-Destructive-Evaluation (NDE) of the finished part is required.

**Picture 3** AM process monitoring shortens the prototyping phase and continuously tracks the print quality in the industrial phase

**Picture 4** Severity score per part supports evaluation of every single part

The Scalability Challenge

Every printer is unique in terms of small differences in lighting conditions, optical distortions of the camera lens, and camera settings. To eliminate these differences, all images are transformed in normalized, distortion corrected images of the powder bed. A basic principle of the severity score calculation is to measure deviations from an ideal recoating in the sense of an anomaly detection, that is, the system is not explicitly trained on specific defect types but measures any deviation from the expected ideal recoating. This allows detecting even unexpected defect types like the powder dropping shown in Picture 2 (right) – an unknown defect type when the system was trained.

The deviations are mapped to a severity score value using a machine-learning based regression model in the sense that major deviations yield a high score and minor deviations a small score. The training of the machine-learning algorithm is done on a training set of a few hundred example images whose severity scores were defined by AM experts. Due to significant differences between the powder bed images of the two considered printer types.

Due to the different camera resolution and lighting between different printers, we developed a machine-learning algorithm able to learn printer type-specific recognition models. This means, our machine learning algorithm only distinguishes between the printer type, but not between individual printers and can thus be scaled across large fleets of printers.

We at Siemens AI have spent over 30 years leveraging artificial intelligence in the arts of engineering to create safe and sustainable solutions for real-world applications

The patents we were able to develop along these developments, allow us in this case to provide a state-of-the-art possibility to monitor real time quality deviations and reduce cycle time for part industrialization.

We did not only apply this technology for selective laser melting but we also successfully transferred it to other additive manufacturing techniques such as laser jet binding. We were able to prove, that in different industrial context of additive manufacturing optical surface inspection improved combined by Industrial AI is a highly beneficial technology.

Thus, the main benefit for AI supported monitoring application for powderbed printing are:

We can reduce the numbers of prototypes, which reduces costs for prototypes, and shortens the prototype cycle. This results in a shorter time to market phase.
It makes optical inspections much easier, as only critical layers need to be evaluated by a human expert.
The critical layers can be both, online and offline, analyzed.
Different print jobs can be easily comparted by a quick overview looking on the severity scores
Failed print jobs can be stopped before they end, and print time is saved

If you would like to find out more about our research or our AI services please be invited to get in touch, directly or via the Siemens AI Lab.