Machine Learning Optimizes 3D Concrete Printing Accuracy

By Nidhi DhullReviewed by Susha Cheriyedath, M.Sc.Jul 23 2024

A recent article published in Inventions introduced a computer vision setup for real-time precision in three-dimensional (3D) concrete printing, employing machine learning (ML) to control layer morphology.

Background

Rising environmental concerns and the push for sustainability are driving innovations in the construction industry, such as 3D printing technologies. 3D concrete printing (3DCP) offers design flexibility and reduces material waste and costs, but it requires precise parameter adjustments due to concrete's variable properties. Manual interventions often lead to errors and inconsistencies.

Machine learning vision models can address these challenges by analyzing images and videos during printing to detect defects and deviations. They can also enable automatic adjustments to print parameters, improving quality and reducing the need for human intervention.

This study integrated a camera-based system with a machine learning framework to autonomously control the motion speed of a 3DCP robot, aiming to maintain the desired layer width throughout the printing process.

Methods

This study focused on four key aspects:

1. Generating a Comprehensive Dataset for 3DCP: A synthetic dataset was created using numerical simulations tailored to the properties of the concrete mix employed. These simulations helped derive various printing parameters, such as extruder speed and aimed to achieve specific printed layer shapes with measurable attributes like width and height. Calibration and fine-tuning ensured that the digital material accurately represented the physical concrete.

   The numerical simulations utilized the “overInterDyMFoam” solver from OpenFOAM v2112, an open-source library for engineering and scientific computations. This dataset, consisting of diverse printing scenarios, was used to train an Extreme Gradient Boosting (XGB) model.

2. Real-Time Monitoring of Printed Layer Width: To ensure precise layer thickness during the printing process, a real-time monitoring system was implemented. This system captured data continuously to maintain the desired layer width and adapt to variations.

3. Assessing the Accuracy of XGB in Predicting Printing Parameters: The XGB model was trained on the dataset obtained from simulations and used in real-time alongside a computer vision system. The model’s accuracy in predicting printing parameters was evaluated by comparing predicted values with actual measurements during the printing process.

4. Evaluating the XGB-Controlled Design Process: A data-capturing channel was employed to refine the XGB model using experimental data. Feature extraction was carried out using the Grasshopper optimization algorithm, which helped identify relevant combinations of locations and features. The final dataset, consisting of 2,245 entries and 9 features from various simulations, was used to train the machine learning (ML) model. This ML framework was then examined for its accuracy in estimating print time and material requirements, taking into account unpredictable changes in print speed and material flow.

Overall, the study aimed to enhance the precision and efficiency of the 3DCP process by integrating ML techniques with real-time monitoring to control printing parameters dynamically.

Results and Discussion

The researchers evaluated various model architectures to identify the one with the best performance using the ScikitLearn 1.2.2 ML package on a synthesized dataset.

The linear regression model showed very low accuracy (0.62). In contrast, the deep neural network model yielded promising results on the training data but struggled with unseen data, indicating overfitting. Similarly, a gated recurrent unit (GRU) model also experienced overfitting. The XGB architecture, however, demonstrated the highest performance, achieving an accuracy of 0.92 in predicting motion speed.

The XGB model was trained in two variations: variation 1 predicted layer width given a state with motion speed, while variation 2 predicted motion speed given a state with layer width. The layer width values extracted from the simulation and those predicted by variation 1 of the XGB model were consistent throughout the print path.

Nonetheless, careful analysis revealed that predictions for the curved portions of the print path were poor. Expanding the dataset to better represent the print path geometry could address this issue.

The proposed computer vision system for monitoring layer width and motion control involved a 360-degree rotating camera aligned with the printing direction, controlled by an external Arduino unit through a stepper motor. Image-processing operations were conducted on a Raspberry Pi system. The Grasshopper environment served as the central hub for real-time communication between computational processes and fabrication equipment.

The depth camera’s video stream was processed using Python v3.9 to extract layer width information. This data was then fed into the ML model, which used the movement data and printing parameters to determine the optimal extruder speed for maintaining or achieving a target layer width.

Conclusion

Overall, the researchers successfully demonstrated the automation of the 3DCP process using an ML algorithm. The XGB model showed promising results, with accuracies exceeding 80 % in both design (predicting speed for a given width) and monitoring (predicting speed corrections based on camera measurements and target width) applications.

However, these results need to be validated in real-world printing scenarios. The researchers recommend expanding the dataset to include various concrete mixes to improve the system’s versatility and broaden its potential applications.

Journal Reference

Silva, J. M. et al. (2024). Real-Time Precision in 3D Concrete Printing: Controlling Layer Morphology via Machine Vision and Learning Algorithms. Inventions, 9(4), 80. DOI: 10.3390/inventions9040080, https://www.mdpi.com/2411-5134/9/4/80

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.