Abstract
In industrial automation, generating smooth and energy-efficient trajectories for servo drives is critical for reducing mechanical wear and energy consumption. This paper proposes a novel approach to piecewise polynomial (PP) approximation that integrates physical energy optimization into a machine learning framework. By utilizing TensorFlow's gradient descent optimizers to minimize a multi-objective loss function—comprising approximation error, continuity constraints, and energy expenditure—the authors produce trajectories that are both accurate and mechanically favorable. Experimental results on synthetic datasets demonstrate that significant energy savings can be achieved with negligible loss in approximation accuracy.
Introduction
Industrial machines frequently utilize "electronic cams" to define repetitive motion profiles for servo drives. These trajectories are typically approximated from input point clouds using Piecewise Polynomials (PPs). While traditional closed-form mathematical solutions are computationally efficient, they are often rigid and struggle to incorporate complex objectives such as energy minimization. Although parametric curves like B-Splines offer flexibility, industrial hardware often requires non-parametric PP functions. This research bridges this gap by framing trajectory generation as a machine learning optimization problem, allowing for the flexible integration of physical constraints and objectives.
Methodology
The authors formulate trajectory generation as an unconstrained multi-objective optimization problem within the TensorFlow framework, employing the AMSGrad optimizer. The model optimizes a parameterized PP function based on a composite loss function:
- Approximation Loss ($\ell_2$): Measures the deviation of the polynomial from the target data points.
- Continuity Loss ($\ell_{CK}$): Penalizes violations of $C^k$-continuity to ensure smooth transitions between segments.
- Energy Loss ($\ell_E$): A novel term calculated as the integral of the squared acceleration. This term physically corresponds to the elastic strain energy and is proportional to copper losses in motor coils.
To meet the strict continuity requirements of industrial applications, the output of the gradient descent process is refined using a post-processing algorithm ("CKMIN") that ensures the continuity loss is exactly zero.
Results and Discussion
The approach was evaluated using a synthetic dataset ($y_i = \sin(4\pi x_i^2) + n_i$) designed to simulate noisy industrial measurements. The inclusion of the energy loss term ($\ell_E$) serves as a powerful regularizer, preventing the model from overfitting to noise.
As shown in the visual comparison, optimizing solely for approximation error leads to oscillatory curves that track noise, which would cause excessive mechanical vibration. In contrast, the energy-optimized curve is significantly smoother.
Quantitatively, the framework allows for a fine-tuned trade-off between accuracy and efficiency. By adjusting the weights ($\alpha, \beta$) in the loss function, engineers can navigate a Pareto front of solutions.
A key finding is that a slight shift in optimization priority can yield massive energy benefits. For instance, reducing the weight $\beta$ from 0.90 to 0.45 resulted in a 4.4x reduction in energy loss (from 8.682 to 1.970), while the approximation error only increased from 0.005 to 0.014.
Conclusion
Integrating energy optimization into the PP approximation process provides a robust method for generating industrial trajectories that balance accuracy with mechanical longevity. The energy loss term effectively regularizes the fitting process, eliminating high-frequency oscillations caused by noise. Future research will explore the use of orthogonal polynomial bases, such as Chebyshev polynomials, to improve numerical stability and convergence when combined with energy-based loss terms.
Summary updated