This paper presents a method for data-driven identification of passive systems, with a focus on systems with nonlinear friction models. This is achieved by approximating the passive dynamics of a given system in a reproducing kernel Hilbert space (RKHS). The main novelty of this approach is the use of a quadratic kernel model to ensure that the estimated friction dynamics is passive. The regularized regression is done by optimizing a convex Fenchel dual of the optimization problem. The method was validated with a simulation study, where it was first applied to the estimation of a Karnopp friction model with Stribeck effects. It was also applied to the identification of systems dynamics of a mass-spring-damper (MSD) system with the same friction model. The identification was performed using two different forms of input data: sampled input and output signals and the Legendre coefficients corresponding to these signals. Results demonstrated accurate identification of training data, as well as examples of generalization to test data. Passivity and causality of the approximated continuous system do not guarantee passivity and causality of its corresponding discretized or spectrally approximated system. Although simulation results demonstrate this, increasing the sample size could increase representation accuracy and thereby the preservation of these properties.
While physics-informed neural networks (PINNs) are commonly employed for time-domain problems, this work extends their application to the frequency domain for parameter estimation. The framework integrates the governing equation into the neural network (NN) training loss function by using a composite complex domain formulation to enforce data consistency, minimize physics residuals, and respect parameter bounds. The methodology is presented as a transfer function-based formulation, whose parameters can be estimated from available data. The effectiveness of the methodology is validated through both simulation and experimental datasets on axial active magnetic bearing (AMB) systems, which are open-loop unstable and thus make frequency-domain parameter estimation particularly relevant. The results show excellent agreement between measured and predicted responses. Furthermore, the proposed approach is compared with an existing frequency-domain identification approach. PINN-based estimation of physical parameters is closer to reference values due to physics residual and parameter bounds. The findings validate the proposed methodology, which indicates that the PINN-based frequency-domain framework can accurately estimate system parameters, demonstrating its accuracy and robustness for physics-constrained learning in the frequency domain.
Determination of the pose of workpieces is important for robotic applications in manufacturing, including handling, assembly, machining and welding. Established methods based on 3D sensors may fail for workpieces with highly reflective materials. In this paper, we take advantage of recent development in machine learning to determine the pose of reflective workpieces without the use of depth data. Our proposed method is based on deep iterative matching of image data of the workpiece with a computer-aided design model. Starting with an initial estimate of the workpiece pose, the method iteratively aligns the computer-aided design model projections with an image of the actual workpiece, adjusting the pose until computer-aided design model matches the image of the workpiece. The deep learning-based approach optimizes this alignment by updating the pose estimate at each iteration, achieving high precision even for geometrically complex or reflective surfaces. This refinement process enhances accuracy in robotic applications where precise workpiece positioning is critical, such as in automated welding and assembly tasks. We use photorealistic rendering to create two datasets for pretraining the network, which reduces both training time and the need for real labeled data. After the network is trained on synthetic data, it is fine-tuned and tested on real images of reflective aluminium workpieces. We show that the proposed deep iterative matching method outperforms established methods based on iterative closest point with two 3D scanners due to large errors in the scans caused by reflections.
This paper presents a data-driven framework utilising neural networks to approximate the Koopman operator for a discrete-time representation of an electrically actuated two-link mechanical system, enabling the application of linear control techniques. A validated dynamic model is used to generate training data. The Physics Informed Fully Embedded Koopman (PIFEK) framework embeds the original state space directly into the Koopman space, leading to sensible tuning of the controllers. The tuning was done using bandwidth analysis of the Koopman dynamics, yielding estimates of the true dynamics. This resulted in satisfactory control performance, including accurate tracking of a trajectory. Comparisons with a conventional PI controller show that while performance is similar, the data-driven PIFEK approach requires significantly less system-specific knowledge, underscoring its potential for efficient control design.
Species that have persisted over millions of years have done so because they have been able to track peaks in an adaptive landscape well enough to survive and reproduce. Such optima are defined by the mean phenotypic values that maximize mean fitness, and they are predominantly functions of the environment, for example the sea temperature. The mean phenotypic values over time will thus predominantly be determined by the environment over time, and the trait history may be found in the fossil record. Here I use fossil data from four cases found in the literature, and show that adaptive peak tracking models give better results than alternative weighted least squares and directional random walk models. The model performances are compared by use of weighted mean squared errors and Akaike information criterion results.
