Counting down to zero
Capgemini ·
Here’s a new approach to additive manufacturing process control. The goals: zero defects in finished products, and manufacturing speeds increased by a massive ten times The post Counting down to zero appeared first on Capgemini .
Additive manufacturing has significant potential for the volume production of complex, high-value parts. However, for adoption to become widespread in industry, manufacturers must be confident: confident that parts meet quality standards; that defects are detected early; and that production time and material are not wasted.
Many quality control approaches still depend heavily on retrospective inspection, which only identifies problems after time, energy, and material have already been committed.
That’s why Capgemini Engineering is working with its strategic research partner, ETH Zurich, a world-class university, on a three-year project to monitor and respond to additive manufacturing production environment conditions in real time.
Much of the basic technology needed for real-time monitoring already exists. Optical, thermal, and acoustics sensors together with high-speed processing are already in place, and many modern approaches combine them with large-scale data crunching. The trouble with this method is that the data is very large-scale indeed – too large in many cases to process at the speeds required in those conditions.
The approach being taken by Capgemini Engineering and ETH Zurich relies, in part, on prior knowledge of the physics and chemistry of the manufacturing processes at work, in order to narrow the operational metrics of the data to more manageable levels.
But that’s just the start. When you’re monitoring and responding in real time to additive manufacturing conditions, your calculations need to be made on the spot, as each layer is put down. This means that algorithms and data must be moved as much as possible to the edge, where the action is, to reduce both data overheads and production time latency.
What’s more, this is happening with each successive additive layer. So, the data itself needs – in a sense – to be layered too, assessing what is happening on each pass, determining what is relevant, what is actionable, and making any necessary process adjustments live, right there on the manufacturing production floor.
In fact, while it might be the norm to refer to additive manufacturing taking place in layers, that’s almost a figure of speech in this case. The machine in question is rotary, which, strictly speaking, means there is no succession of passes: it just keeps going. This in turn means the data is not linear but continuous, measurable in terabytes per day, massively increasing the need for the lowest possible latency.
Potential anomalies gauged by the project include porosity, keyhole defects, and lack of fusion. If levels are too high and prints fail, they can be immediately aborted; if they are redeemable, laser parameters can be dynamically adjusted to heal flaws mid-build. This prevents the wasted expenditure of expensive metal powders, and of machine hours lost on parts destined for scrap.
In addition, algorithms need to be able to distinguish between these true structural defects and benign process variations in the harsh, high-temperature environment of the print chamber. For instance, no one wants parts to be wrongly rejected just because the ambient temperature today is higher than yesterday.
The idea of using AI for real time process control is not new. However, now that computing power and storage capacity have increased exponentially and we have cloud computing, it is a practical possibility.
To take advantage of AI in this application area, AI needs to act on data captured from different sensors, most of them video or photographic, and to be able to process that information in a very short cycle time between layers.
The challenge in achieving this is to make the neural networks on which AI depends run on the edge. The project team is currently working with an AI company that emerged from a world-leading US university on a way to significantly reduce the layers in the neural network that are needed. This, of course, would be highly appropriate for edge applications.
The word “transformative” is used liberally these days, but in this case it’s justifiable. The joint project team anticipates the approach will first reduce defects in the finished product to zero, considerably lowering production costs; and second, increase the speed of manufacturing ten times. That’s right: not by ten per cent – ten times .
For anyone manufacturing machine parts using circular symmetry, like in aerospace, maritime engineering, automotive, or space, the implications are huge.
The post Counting down to zero appeared first on Capgemini .