Chebyshev fitting of complex surfaces for precision metrology

Xiangchao Zhang, Hao Zhang, Xiaoying He, Min Xu, Xiangqian JiangMeasurement 2013 · DOI

TL;DR

This paper introduces a novel Chebyshev fitting method for precisely measuring the form errors of complex surfaces, crucial for high-precision components. By transforming the challenging minimax optimization problem into a differentiable one using exponential penalty functions and an adaptive active-set strategy, the method achieves accurate and stable results efficiently, outperforming existing techniques.

What Problem Does This Paper Address?

In precision engineering, the "form error" of a component – how closely its actual shape matches its ideal design – is critical for its functionality. For optical components, the Peak-to-Valley (PV) value is a common metric for this form accuracy. However, the standard method for calculating PV, least squares fitting, often overestimates the error, leading to potentially good parts being rejected. The theoretically correct method, Chebyshev fitting (also known as minimum zone fitting), aims to find the narrowest possible tolerance band that encloses the measured data. This method is mathematically complex because it involves a minimax optimization problem that is not easily differentiable, making it computationally difficult to solve. Existing approaches often struggle with convergence, accuracy, or computational efficiency, especially for complex surface geometries.

Methods

The core contribution of this paper is a new approach to solving the Chebyshev fitting problem for complex surfaces. The authors tackle the non-differentiable minimax problem by converting it into a continuously differentiable minimization problem using exponential penalty functions (EPF). This transforms the objective function into a form that can be handled by standard optimization algorithms.

To further enhance the method, they employ an adaptive active-set strategy. This strategy intelligently selects a subset of data points to focus on during the optimization process, significantly improving computational efficiency without sacrificing accuracy. The smoothing parameter 'p' in the EPF and other configuration parameters are adjusted adaptively during the optimization to balance fitting accuracy and numerical stability. The paper also details how to calculate the gradient and Hessian matrices for this transformed problem, enabling the use of Newton's method for iterative updates. A modified Hessian matrix is used to ensure positive definiteness and avoid divergence.

Datasets

The paper evaluates its method using three benchmark datasets, which were previously used in other studies:

  • An ellipsoid surface.
  • A hyperbolic paraboloid surface.
  • An aspheric surface.

These datasets vary in complexity and the number of data points (ranging from 42 to 8100 points), allowing for a comprehensive validation of the proposed algorithm.

Key Insights and Ideas

  • Bridging the Gap: The paper successfully bridges the gap between the theoretically correct Chebyshev fitting and practical, efficient computation, especially for complex surfaces.
  • EPF as a Differentiable Proxy: Exponential penalty functions are shown to be an effective way to approximate the non-differentiable minimax problem with a differentiable one, making it amenable to powerful optimization techniques.
  • Adaptive Active-Set Efficiency: The adaptive active-set strategy is crucial for making the method computationally feasible for large datasets, as it avoids processing all points in every iteration.
  • Overcoming Least Squares Limitations: The results clearly demonstrate that Chebyshev fitting, as implemented by the authors, provides a more accurate assessment of form error compared to the commonly used least squares method, reducing over-estimation.

Key Results

The proposed Chebyshev fitting method using exponential penalty functions (EPF) demonstrated superior performance compared to the differential evolution (DE) algorithm on the tested datasets:

  • Accuracy: The EPF method consistently achieved results as good as or better than DE, and significantly more accurate than orthogonal distance least squares fitting. For example, the PV parameters obtained by EPF were 36.3%, 26.2%, and 18.0% lower than least squares for the ellipsoid, hyperbolic paraboloid, and asphere, respectively.
  • Efficiency: The EPF method was substantially faster than the DE algorithm, especially for larger datasets. For the aspheric surface with 8100 points, the EPF method took 2.34 seconds, while DE (even with point reduction) took 5.40 seconds. Without point reduction, DE would take over two minutes.
  • Reliability: The EPF method always obtained global optima, whereas DE produced different solutions on each run, indicating a lack of determinism and higher uncertainty.

Limitations and Future Work

  • Limitations: While the paper highlights the efficiency and accuracy of the EPF method, the computational effort is still concentrated on calculating the gradient and Hessian matrices. The paper notes that the convergence rate can be affected by the dependency between projection points and parameters, which is difficult to calculate precisely.
  • Future Work: The paper implicitly suggests that further optimization of the gradient and Hessian calculation, perhaps through more advanced approximations or specialized hardware, could lead to even greater efficiency. The adaptive nature of the algorithm also opens avenues for exploring more sophisticated parameter tuning strategies.

Summary updated

BibTeX
@inproceedings{zhang2013chebyshev,
  author = {Xiangchao Zhang and Hao Zhang and Xiaoying He and Min Xu and Xiangqian Jiang},
  title = {{Chebyshev fitting of complex surfaces for precision metrology}},
  booktitle = {Measurement},
  year = {2013},
  doi = {10.1016/j.measurement.2013.04.017},
}
Figure