New Algorithm Promises to Detect and Fix 3D Printing Errors as They Happen

The primary advantage of 3D printing is that one can go from a CAD model to a physical prototype in just hours without any active labor. But various issues that pop up during the printing process can ruin parts and that forces operators to keep an eye on their printers, eliminating some of the benefit. And even when an operator does spot an error, it can be difficult to recognize the cause and address it in time to salvage the part. To solve both problems, a team of engineers from the University of Cambridge developed a machine learning algorithm that can detect and fix 3D printing errors as they happen.

This isn’t the first attempt the industry has made at print error detection. Three years ago, we wrote about The Spaghetti Detective (now Obico), which sounds an alarm when it detects major print failures. But this new algorithm takes that concept a couple of steps further. First, it detects print issues much earlier so it knows that there is a problem before total print failure occurs. Second, it can attempt to adjust the 3D printing parameters on the fly to save the part. For example, if it detects under-extrusion then it can increase the extrusion multiplier to compensate. Because most relevant parameters are adjustable during a print job, it can handle a wide range of potential issues.

Like all machine learning algorithms, the “secret” here is the data set used to train the model. To build that training data, the engineers captured 950,000 images over the course of 192 print jobs. Each image also carried data on the actual print settings, along with information on how far those settings deviated from known good settings. With that training data set, the machine learning model can recognize the early signs of a failure caused by specific parameter set to an improper value. The training images and the real-time input images are taken with a camera close to the hot end nozzle, so the algorithm can see issues that would be difficult for human operators to spot.

That training data set is large, but it still isn’t enough to cover all printer models, materials, and configurations. For that reason, the engineers envision this as a collaborative system that learns from a network of 3D printers in the real world. Each job from a printer in the network, whether it fails or succeeds, improves the training data and makes the algorithm smarter. To make that happen, the team formed a company called Matta to commercialize this algorithm.