The current identification of AM components at the end of the overall process chain represents a non-scalable and cost-intensive manual, labor intensive process. The variety of geometries in prototyping leads to complex challenges where existing automation solutions cannot be implemented. AI-based image recognition can be an improvement in this regard, as demonstrated by a BMW-led study focusing on AM Flow‘s AM VISION automated solution.
With time-to-market in the automotive industry steadily decreasing, demand for prototyping components is higher than before and the vision of large-scale production, delivering just-in-time to assembly lines, is emerging. This is not only a question of increasing output quantity and production speed but also of economic viability. The process chain of current available AM technologies still includes a high amount of labor intensive work and process steps, which lead to a high proportion of personnel costs and decreased product throughput. Also, these operations lead to bottlenecks and downtimes in the overall process chain.
For this reason, the study’s authors, who have built extensive experience within the BMW Additive Manufacturing Campus, highlight the ongoing development towards automation and industrialization of the entire AM industry, which is shown by new solutions, applied patents and announcements of collaborations and government-funded projects. In order to effectively industrialize an AM process chain, it is necessary to understand and analyze the overall process chain of various technologies and to identify automation potential in form of current production bottlenecks caused by manual operations.
Current available AM process chains reach a limit of productivity for large production volumes. In addition, process steps after printing include even higher amounts of manual process steps which lead to bottlenecks, when scaling up the production. One of these manual operations is the identification and assignment to the acceptor or customer of the components by labeling the components for further logistical transportation.
Even though part identification is a small part of the overall process chain, it is still a process step that does not scale and requires a high amount of manual work (in comparison, for example, to cooling for example, which does not require personnel capacities). To achieve the best economics (i.e. lowest cost per part), the components are nested tightly with the support of specialized software. This leads to batches with a large number of different parts in one build job. However, this also means, that the trackability of single components gets lost and the assignment to customer orders needs to be done after the production. Usually, it is a manual process step that takes time and space since the operator has to identify and compare every single physical part with 2D images from a list. Increasing production throughput at this level increases the cost of additional employees, slows down component delivery and requires more space. Therefore, automation of these manual operation steps should be implemented.
However, in contrast to conventional production processes, automation of AM component identification is associated with additional challenges. Other than the basic idea of automation as a replacement for recurring process sequences in order to increase productivity and to reduce labor costs, a solution for this task needs to be highly flexible. Flexibility in AM enables the economic production of different batch sizes, different workpieces in any sequence and is a transition to the basic idea of Industry 4.0.
Innovation in the field of Artificial Intelligence (AI) has made the here necessary combination of flexibility and automation feasible. The enormous development of especially deep learning algorithms within the last decade enables systems to mimic human cognitive abilities that require strategic thinking. It enables machines to take over flexible tasks depending on data input that previously could only be performed by humans. The input data can be for example parameters, sensor measurement data or image material. In addition, AI has the ability to constantly improve and differentiate from image to image.
Scaling with the AM Flow
Therefore, camera-based imaging recognition with image recognition in combination with AI provides the most promising approach. AI has already arrived in the AM sector. It is used for screening suitable components, generating complex designs and for monitoring quality control. Until recently, there was no automated solution on the market capable of solving these complexity factors. BMW’s internal market analyses have shown that AM-Flow’s AM-VISION is designed for the AM market and does not require high training efforts for objects that might only appear once. The system uses an AI-based algorithm that enables the recognition of AM components based on their unique geometric fingerprint at a high processing speed. Rendered STL files and scans of different component batches are used for cyclewise retraining for continuous improvement.
The algorithm processed in two computers uses pattern recognition in order to match an optically scanned part with the digital CAD model. The CAD model needs to be uploaded and analyzed in advance by the system, which can be done automatically via a manufacturing execution system (MES). The processing of the CAD files includes rendering from different angles and a gravity analysis to determine how the object is likely to lie on the conveyor belt. Therefore, physical recognition does not need a full 360° view of the AM component. The system can be calibrated to specific materials and different colors in order to increase recognition rates. Since the recognition takes place digitally, next machining step can be determined by communicating with a MES via an Application Programming Interface (API).
AI image classification model of the AM-VISION system was trained with a large number of digital models. The models were expanded with simulated variations generated with a CAD engine. Therefore, it is not necessary to train the system with a high quantity of physical parts beforehand. If the recognition score (0-100%) is below a certain score, the top three matches detected by the algorithm are shown and an operator has to select and approve the correct one. This guided learning is also data to train the system.
This validation is synthetic training and ensures further accuracy improvements. The machine learning model becomes more accurate the larger the set of objects processed with AM-VISION. Test studies of build jobs containing a high geometry mix already proof that the pilot setup saves time for identification and labeling. Components can be processed up to 50% faster compared to manual operation. Final machine setups can be integrated in the production line with an automated one-way conveyor belt.
AM-Flow estimates that this will result in 6 to 10 times faster processing time per part. The recognition rate is between 80% and 95 % if the build job contains a high diversity of geometries. The more similar the geometries are, the more the recognition rate can decrease. During testing, calibration to the gray HP MJF component appearance leads to an increase in the detection rate. In the ideal case, the technology used and therefore the appearance color of the physical component is also included in the name or code of the CAD file visible for the system.
The deployment of AI is effective here. BMW engineers found that the investigated AM-VISION System from AM-Flow was able to perform reliable object detection of high mixed AM components based on a partial representation of the geometry. The implementation results in a reduction in throughput time that leads to cost savings due to reduced identification time and reduced failure costs, even if an operator is still required in the shown setup. It is crucial to choose the right field of application for automated identification. Identical parts, parts that only differ on specific component areas or too little volume of manufactured components do not lead to an economic benefit today. It is this particular field that should be optimized in the future since similar components are also difficult to distinguish for humans and lead to bottlenecks.
Current research efforts are focused on modular and flexible process steps. This facilitates flexibility in terms of the process steps used and enables loose linking of the manufacturing steps. Thus, not every component goes through all process steps, such as coloring and finishing. In this case, it will be necessary to identify components frequently in the overall chain. Occurring fluctuations of the recognition rate are solved through the continuous improvement of the deep learning algorithm. Updates during testing have already helped to distinguish mirror-inverted automotive components. In addition to the AI image classification model, it also has an AI decision model, which can handle multiple simultaneous inputs.
Further input could help enable automated comprehensive quality control by using point clouds for example to measure dimensional accuracy. The study’s authors also pointed out that the algorithms could additionally be applied in grippers for component handling which is important in order to fill remaining bottlenecks for the fully automated process chain. Nevertheless, the currently available automated identification of a high mix at high volume with the AM-VISION is already another step towards a large-scale AM production.