Cut Quality Estimation in Industrial Laser Cutting Machines: A Machine Learning Approach.

Giorgio Santolini, Paolo Rota, Davide Gandolfi, Paolo BosettiWorkshops @ CVPR 2019 · OpenReview

Abstract: The use of machine learning models to improve industrial production quality is becoming more popular year after year. The main reason is the huge data availability and the impressive boost of performance of such methods achieved in the last decade. In this work we propose an adaptation of three well known machine learning algorithms to estimate the quality of cut in industrial laser cutting machines. The challenge here is to use a pool of multimodal parameters coming from different sensors and fuse them in order to detect the cutting status of the machine in a near-online modality. We analyze then generative and discriminative approaches based on Gaussian Mixture Models, Recurrent Neural Networks, and Convolutional Neural Networks in a supervised setting. Results are computed on a brand-new dataset that is freely available for reference.

Introduction: The Challenge of Real-Time Quality Control

In high-precision manufacturing, laser cutting is a cornerstone technology. However, maintaining consistent cut quality remains a significant challenge due to material inconsistencies such as rust spots, surface reflections, or uneven welding seams. Traditionally, these issues required exhaustive manual calibration and post-production inspection.

Visualizing cut quality defects: (a) sample geometries from the ALCIDE dataset, (b) cross-section highlighting a bad cut, and (c) surface view showing missed and bad cut anomalies.

When the laser loses continuity, it results in either a "Missed Cut" (incomplete detachment) or a "Bad Cut" (unstable, low-quality edges). Detecting these failures in real-time is essential to "close the loop," allowing machines to sense their own performance and intervene immediately, thereby reducing material waste and downtime.

Methodology

Researchers from the University of Trento and Adige S.p.A. evaluated three machine learning architectures to classify cut status into three categories: Good Cut, Bad Cut, and Missed Cut.

  1. Gaussian Mixture Models (GMM): A generative baseline that models class probability distributions. While theoretically sound, it often struggles with the high-frequency noise typical of laser processes.
  2. Long Short-Term Memory (LSTM): A recurrent neural network optimized for sequential data, utilizing sequences of 100 readings to capture temporal dependencies.
  3. Fully Convolutional Network (FCNN): A discriminative approach designed to exploit mutual information between sensors. By using kernels that convolve multiple parameters (e.g., laser power vs. light reflectance), the FCNN effectively captures complex inter-sensor relationships.

The ALCIDE Dataset

The study introduces the ALCIDE (Adige Laser Cut Interruption DatasEt), a comprehensive benchmark for industrial process monitoring.

The industrial Adige LT8.10 3D laser cutting machine used to generate the ALCIDE dataset.

The dataset was collected using an Adige LT8.10 3D laser cutting machine and includes data from over 400 sample cuts across 20 tube pieces with varied geometries. Features include multimodal data such as light intensity (captured via optical fiber), laser power, feed rate, and gas pressure set points.

Results and Performance

The FCNN emerged as the superior model, achieving an average accuracy of 76.65%, followed by the LSTM at 76.15% and the GMM at 72.10%.

Beyond accuracy, industrial viability depends on inference speed. The FCNN achieved a sub-millisecond inference time of 0.91ms on standard hardware, significantly faster than the machine's 20ms sampling rate. Qualitatively, the models demonstrated robustness by identifying anomalies even in "unseen" geometries where ground-truth labels were noisy.

Key Insights and Future Work

A critical finding was the importance of the photodiode signal (backscattered light), which proved to be the most informative feature, though multi-sensor fusion remained necessary for distinguishing between "Good" and "Bad" cuts. Additionally, the FCNN provided more stable predictions compared to the "spiky" outputs of the GMM.

Future research will address the "moving target" problem, where machine software updates shift data distributions, requiring robust domain adaptation. The ultimate goal is to transition from quality estimation to active control, enabling real-time parameter adjustments to prevent defects before they occur.

Summary updated

Abstract

The use of machine learning models to improve industrial production quality is becoming more popular year after year. The main reason is the huge data availability and the impressive boost of performance of such methods achieved in the last decade. In this work we propose an adaptation of three well known machine learning algorithms to estimate the quality of cut in industrial laser cutting machines. The challenge here is to use a pool of multimodal parameters coming from different sensors and fuse them in order to detect the cutting status of the machine in a near-online modality. We analyze then generative and discriminative approaches based on Gaussian Mixture Models, Recurrent Neural Networks, and Convolutional Neural Networks in a supervised setting. Results are computed on a brand-new dataset that is freely available for reference.
BibTeX
@inproceedings{santolini2019cut,
  author = {Giorgio Santolini and Paolo Rota and Davide Gandolfi and Paolo Bosetti},
  title = {{Cut Quality Estimation in Industrial Laser Cutting Machines: A Machine Learning Approach.}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
  url = {https://openreview.net/forum?id=Soi-eAzlOTH},
}
Figure