Siemens (Belgium)

Abstract In this paper, a Machine Learning surrogate model is developed for Noise, Vibration, and Harshness (NVH) analysis of a 48-slot, 8-pole Permanent Magnet Synchronous Motor (PMSM) stator. A parametric stator model is used to generate a Finite Element based dataset. The resultant Frequency Response Functions (FRFs) due to dedicated force waves are the output to be predicted. To this end, several pre-processing data reduction techniques combined with Machine Learning (ML) regressions are investigated. The findings confirm the workflow ability to generalize effectively, offering a robust and efficient method for NVH analysis of PMSM without the need for an FE model.

<p>The estimation of thermal parameters varying during the lifecycle of Electronic Components and Systems (ECS) is enabled by Digital Twins (DTs). The estimation performance depends on the number and utility of the chosen sensing locations. Moreover, large thermal gradients in ECS may cause misleading measurements resulting in wrong estimates. The presented Optimal Sensor Placement (OSP) algorithm considers parameter uncertainties thanks to the utilized DT framework composed of an Augmented Kalman Filter employing a parameter dependent thermal Reduced Order Model. Robustness in the sensor selection process is enforced by providing the OSP algorithm with the thermal gradients in the candidate sensing locations. A simulation study highlights the validity of the proposed methodology and opens to experimental validation.</p>

In this paper, we detect spike-like measurement errors in accelerometer sensor signals using continuous wavelet transform based machine learning methods. The wavelet coefficients are computed in automatic feature extraction layers (CWT layers), producing sparse representations of the input signals. The proposed methods ensure low model complexity, which allows real-time application. We complement a previously proposed variable projection based method to estimate wavelet coefficients with a numerical quadrature based approach. We present a qualitative and quantitative comparison of the CWT layers. To demonstrate the generality of the method, we introduce support vector machines supplemented with CWT layers in addition to previously used neural networks. Sensor fault detection experiments are conducted on real measurements using a low-cost accelerometer. The results of the experiments show that the proposed method achieves perfect accuracy on our dataset while outperforms previously used approaches in terms of interpretability and online usage. In addition to convincing classification accuracy, our results illustrate the interpretability of the proposed model driven machine learning framework.

This paper presents a novel approach to imitation learning from observations, where an autoregressive mixture of experts model is deployed to fit the underlying policy. The parameters of the model are learned via a two-stage framework. By leveraging the existing dynamics knowledge, the first stage estimates the control input sequences and hence reduces the problem complexity. At the second stage, the policy is learned by solving a regularized maximum-likelihood estimation problem using the estimated control input sequences. We further extend the learning procedure by incorporating a Lyapunov stability constraint to ensure asymptotic stability of the identified model, for accurate multi-step predictions. The effectiveness of the proposed framework is validated using two autonomous driving datasets collected from human demonstrations, showcasing its practical capability in modelling complex nonlinear dynamics.

The design optimization of acoustic metamaterials represents a critical challenge in modern noise control applications. Global sensitivity analysis (GSA) has emerged as a powerful tool for quantifying the correlation between metamaterials' design variables and corresponding sound quality metrics (loudness, sharpness, and speech intelligibility index). However, GSA's high computational intensity limits the practical application of design optimization, especially for sound quality metrics which vary with sound stimulus. This study introduces a computational framework using radial basis function (RBF) interpolation to accelerate the GSA of acoustic metamaterials, with a periodic arrangement of micro-perforated panels (MPPs) serving as an exemplary case study. The methodology employs RBF interpolation to create a surrogate model of metamaterial acoustic response, demonstrating the potential for computational cost reduction while maintaining accuracy. It is validated across three acoustic scenarios that represent common challenges in noise control: stationary signals (broad/narrowband noise), environmental sounds (mechanical noise, traffic), and dynamic signals (speech, music). The study systematically compares sensitivity indices derived from direct numerical simulation and interpolated predictions, assessing the RBF's impact on GSA accuracy. [The European Commission is gratefully acknowledged for supporting the Marie Sklodowska Curie program through the Horizon Europe DN METAVISION project (GA 101072415)]

In rotating machinery, noise source identification plays a vital role in early-stage acoustic refinement. When machines operate at varying speeds, traditional beamforming techniques like frequency-based beamforming is no longer suitable, as the generated vibroacoustic phenomena are nonstationary. Order-tracking is a popular way to deal with such nonstationary regimes and can extract orders, which are time-varying phasors rotating proportionally to shaft speed. This paper introduces an Order-Based Beamforming (OBBF) technique, which performs beamforming on extracted orders by mapping complex envelopes at the array's microphones to the source complex envelope to localize and quantify sources causing tonal noise at variable speeds. The wealth of information provided by OBBF facilitates the construction of tonal acoustic radiation movies along the course of an order with each frame being a beamforming map, thereby enabling source characterization in space, time, frequency, angle and angular speed. The proposed technique is comprehensively detailed and verified both in numerical simulations and experimentally on an induction motor test bench.

Abstract The numerical prediction of spiral bevel gear contact dynamics is often hampered by the complexity of the gear tooth geometry and the three-dimensional (3D) nature of the motion transfer, which forms a major challenge for efficient contact detection and accurate contact force calculation. This works proposes a novel spiral bevel gear contact force element that enables the accurate prediction of 3D bevel gear contact force distributions in flexible multibody system dynamics simulations (MBDS). An efficient contact detection strategy, based on the surface of action, is used, while the accurate contact force calculation is obtained by inclusion of a new semi-analytical bevel gear contact stiffness model. This stiffness model relies on model order reduction (MOR) of finite element method (FEM)-based models to describe the bevel gear tooth and blank deformation, and analytical modeling to predict the tooth contact deformation. The solution can be used to model and analyze lightweight bevel gear dynamics. The method is validated against high-fidelity nonlinear finite element analysis-based contact simulations to evaluate the predicted transmission error, contact pattern, and contact pressure distribution. Finally, a use case is presented to demonstrate how the methodology enables detailed insights in spiral bevel gearbox dynamics, and noise and vibration (NVH) performance.

In semiconductor industry, CD-SEMs (Critical Dimension Scanning Electron Microscopes) play a pivotal role in metrology and process control. Furthermore, additional applications harnessing the potential of CD-SEM equipment have been developed, including contour extraction algorithms, to address the increasing metrological demands. The task of detecting deformity in semiconductor processes has become a formidable challenge due to the continuous reduction in device feature sizes as we progress from one node to another. Deformity in this case can be described as simulation extra prints or detected manufacturing defects. In this paper, we have utilized vector calculations to detect deformity and calculate roughness. Our product, Calibre ContourCal, facilitates the computation of differences between lithographical simulations and measurements. By evaluating the Root Mean Square Error (RMSE) value, we can identify defects in the SEM image or unexpected extra prints in simulation. Additionally, we have proficiently positioned the sites in critical areas based on contour curvature. To evaluate the quality of the model’s simulated contours, we have introduced a novel indicator based on the difference between the simulated model's Power Spectral Density (PSD) and the fitted PSD. The computation of RMSE for a large FoV (Field of View) containing high site count, exceeding one million sites for example, is a demanding task. For that we parallelized our calculation method using GPUs. With the utilization of GPUs, we have achieved significant time savings, simplifying the processing of high site count. Consequently, we can efficiently detect deformity and calculate roughness for large FoV.