TLDR: A new deep learning framework for surface metrology accurately predicts surface texture parameters (Ra, Rz, RONt) and their uncertainties. It uses single-target models for better performance and integrates quantile, heteroscedastic, and conformal methods for calibrated uncertainty intervals. The system achieves high accuracy in classifying measurement instruments (92.85%) and predicting parameters (R2 up to 0.99), offering trustworthy decision support for metrological workflows.
This research introduces a new deep learning approach for surface metrology, a field focused on measuring and analyzing surface characteristics. The goal is to accurately predict surface texture parameters, like Ra, Rz, and RONt, along with their associated standard uncertainties. This is crucial for making informed decisions in manufacturing and quality control.
Traditionally, many methods for predicting surface parameters focus on single point estimates and often overlook the important aspect of measurement uncertainty. This new framework addresses this by simultaneously modeling both the primary surface parameters and their reported standard uncertainties. It also tackles the challenge of using data from multiple types of instruments, including tactile and optical systems.
The core of this framework involves deep learning models that can perform two main tasks: classifying the type of measurement system used and regressing (predicting) continuous surface parameters and their uncertainties. The researchers found that using separate, specialized models for each target (single-target regressors) generally performed better than a single model trying to predict multiple outputs at once. This is because different parameters have varying scales and noise characteristics, and a single model can struggle with these differences.
To ensure reliable predictions, the framework incorporates several techniques for uncertainty quantification. These include quantile regression, which provides asymmetric prediction intervals, and heteroscedastic Gaussian regression, which accounts for input-dependent variations in uncertainty. Additionally, a post-hoc conformal prediction method is used to adjust these intervals, ensuring that the predicted uncertainty ranges are empirically calibrated and trustworthy.
The models were trained on a large dataset of approximately 40,000 instances, collected from various tactile and optical instruments. This dataset included measurements from reference roughness standards and machined specimens made of different materials. Data augmentation techniques, such as bootstrap resampling and adding controlled noise, were used to expand the training data and simulate real-world variability.
In terms of performance, the classification model achieved a high accuracy of 92.85% in identifying the measurement system type. For the regression tasks, the single-target models showed excellent fidelity. For instance, the R-squared values for predicting Ra, Rz, and RONt were 0.9824, 0.9847, and 0.9918, respectively. The models also successfully predicted the uncertainties for Ra and Rz (R-squared of 0.9899 and 0.9955). However, predicting the uncertainty for RONt proved more challenging, with a lower R-squared of 0.4934. This difficulty is attributed to the greater variability and less precise uncertainty reporting associated with RONt measurements across different instruments.
The researchers highlight that these calibrated predictions can be highly valuable in practical metrological workflows. They can help in selecting the most appropriate instrument for a given measurement and in making acceptance decisions based on quantified confidence levels. The framework is designed to be reproducible, with all code and scripts publicly available for verification and further development.
Also Read:
- Enhancing AI Reliability: A Dual Approach to Mitigate Overconfident Errors
- CARLE: A New AI Framework for Predicting Bearing Lifespan with Clarity and Robustness
This work represents a significant step towards integrating advanced AI into surface metrology, providing not just predictions but also a clear understanding of their reliability. For more in-depth technical details, you can refer to the full research paper available at this link.
