Automated Diagnosis and Correction of Defects in 3D Printing With Machine Learning
Machine learning has revolutionized 3D printing by reducing manufacturing waste and improving dimensional accuracy. In addition, smarter printers incorporate closed-loop learnings that speed up part production and reduce labor costs.
Engineers have developed an algorithm that can train a neural network controller to watch the manufacturing process and correct errors on the fly. They used a simulator to teach the model to adjust printing parameters, and it performed better than existing controllers.
Real-time Fault Detection
Despite the availability of many machine learning techniques, few are suitable for the automated diagnosis and correction of defects in additive manufacturing. This paper develops a system that uses low-cost equipment to remotely monitor FDM printing, identify faults and correct them in real-time without manual intervention.
The approach is based on a single multi-head deep convolutional neural network (CAXTON) that augments any extrusion-based printer with error detection, correction and discovery of optimal parameters for new materials. Unlike previous methods that require human labelling of errors, CAXTON automatically labels errors in terms of their distance from optimal printing parameters.
This allows the system to identify and correct different types of errors, including stringing, with a high rate of accuracy. It is also capable of identifying multiple defects, even in highly-complicated geometries, and offers insights into the decisions the network makes by creating visualisations. This enables humans to understand why it is selecting certain features, thus building trust or providing traceability.
Researchers are predicting that machine learning can eliminate many errors during the printing process. This will allow companies to save on production costs by reducing the need for reprinting parts.
The algorithm they developed compares each layer that a 3D printer creates to its idealized representation in the CAD file. It works by treating each layer as a narrow cross-section of some mathematical function and comparing the shape of the real-world model to that of the CAD file.
Unlike traditional computer vision approaches, the feature extraction algorithm used by PrintFixer can be applied to a wide range of printers, part geometries, materials, and printing conditions. To test the accuracy of the algorithm, a random set of 3D models, slicing settings, and parameter values were downloaded from an online repository. Each resulting image was then uploaded to PrintFixer for analysis. The model achieved a training and validation accuracy of 98.1% and 96.6%, respectively. This means that the algorithm is able to correct for an erroneous parameter combination before it causes a failure.
Density-Wise Object Identification
ML can help identify various geometrical anomalies that occur in layers while printing. These anomalies can cause weak infills, inconsistent extrusion, and lack of support. Hence, it is essential to monitor the quality of the prints in real time. However, it can be challenging to detect these defects in the scanned data. ML can help automate the process of finding these defects and make corrections in real-time.
For this, ML models are trained on a full dataset that contains labelled and unlabelled data. The model’s output is compared with the outcomes of ultimate tensile strength tests to identify faulty 3D printed parts.
The experimental process involves setting up the printer to capture layer-wise images of the printed component. Then, the images are processed for noise reduction and segmentation. After that, a model algorithm is selected for identifying the best solution to solve the problem of faulty components. The resulting model can also be used to optimize the AM process and predict mechanical performance.
Determining the ideal parameters for a 3D print is often a long, expensive trial-and-error process. And even when technicians find a set of printing parameters that work well, they may have little data on how those settings will perform with different materials or printer hardware.
Using computer vision to capture images of an object as it’s being printed can help spot errors as they happen. But the amount of data needed to train a machine-learning system to understand this kind of image would be prohibitively expensive, requiring millions of real prints.
To avoid this, researchers developed a method to train a neural network-based controller without the need for actual physical prints. They showed the network 950,000 images captured automatically during the production of 192 objects, and labelled each one with the printer’s parameters and how far those parameters deviated from good values. Then they used the model to predict new parameters for the print, and found that it performed better than other printing controllers.