Martin von Mohrenschildt

Virtual reality (VR)-based training provides a safe and cost-effective solution for truck-driver-training tasks. However, the effectiveness of VR training in replicating real-world driving conditions remains underexplored. This study examines how vestibular motion cues and field of view (FOV) affect driving control, VR sickness, and user experience in truck-driving simulations under crosswind disturbances. Fifty-four participants (35 undergraduates and 19 truck drivers) performed simulated highway driving tasks under Motion (vestibular feedback) and No Motion conditions, using either a head-mounted display (HMD) with a larger vertical FOV or a three-screen setup. Steering control was assessed via wind-steering correlation (the alignment between steering adjustments and wind disturbances), entropy, and reversal rate (corrective steering adjustments), while VR sickness and user experience were evaluated using standardized questionnaires. The absence of vestibular motion cues in the HMD condition led to poorer steering control, reflected in lower wind-steering correlation, higher entropy, and increased corrective adjustments. It also received the lowest ratings for perspicuity and efficiency, particularly among experienced drivers. Simulator sickness symptoms (nausea, oculomotor strain, disorientation) increased in the absence of motion cues in the HMD condition. Overall, motion conditions improved steering performance and were rated higher in efficiency, perspicuity, and novelty. These findings underscore the importance of integrating vestibular motion cues and realistic disturbances in VR truck training to support skill transferability. Optimizing FOV and motion feedback can enhance training fidelity, reduce simulator sickness, and improve driver adaptation. This research informs VR simulator design, improving safety, efficiency, and cost-effectiveness in driver education programs.

Damping identification from transient vibration data can be posed as an envelope-tracking problem for narrowband modal components. We develop an envelope-level state-space formulation for impact-type experimental modal analysis (EMA), in which each mode is represented by a pair of complex rotators at a pre-identified damped frequency, and the modal envelope is inferred via Kalman filtering with Rauch-Tung-Striebel (RTS) smoothing. Two variants are examined: a linear complex-state RTS-Kalman filter (RTS-KF) with fixed rotators, and an RTS-extended Kalman filter (RTS-EKF) that estimates envelope and rotators jointly. Synthetic three-mode tests with linear time-invariant and amplitude-dependent (nonlinear) damping, and with both well-separated and closely spaced modes, are used to compare these trackers against window- and filter-based envelope estimators and standard frequency-domain baselines. Across all scenarios, the simpler RTS-KF yields more stable and lower-variance envelopes than RTS-EKF. It is particularly advantageous for closely spaced modes, and extends naturally to nonlinear decay laws where frequency-domain methods do not apply. Because the per-mode state dimension is small and the transition is block-diagonal, the algorithm scales linearly and is suitable for real-time and embedded implementations.

Data association and filter robustness are two key problems for general multiple target tracking. Classic data association methods have high computational complexity, and the gold standard Kalman Filter (KF) suffers when the motion model does not match target behavior. Random Finite Set (RFS) based filters such as the Probability Hypothesis Density (PHD) filter solves the computational complexity issue of data association by using one multiple-target filter instead of multiple single-target filters in parallel, and the Smooth Variable Structure Filter (SVSF) improves filter robustness by bounding the estimation error using sliding mode control theory. However, there has not been an attempt to merge these two filtering strategies in literature. This paper presents 3 contributions to this problem: a methodology to combine the SVSF filtering strategy into the PHD filtering framework, results that show improvement in performance over traditional PHD filtering, and modifications to improve computational complexity. Three nonlinear filters are created and tested against the two common nonlinear PHD filters in a complex simulation with time varying targets, model mismatch, and high process noise. The variants show stability and robustness in situations where the original filters diverges. The filters are evaluated using the Optimal Subpattern Assignment (OSPA) metric.

Electromyography (EMG)-based biofeedback can facilitate high volumes of exercise in individuals with severe motor impairments. Biofeedback systems based on EMG classification can integrate muscle activities from multiple sensors in a way that is flexible and scalable, allowing the choice and number of trained muscles to be tailored. The classification record can also provide insights about biofeedback quality and the exercise dose. We previously developed a class-based system and tested it within one session. We now test the feasibility of using the same system to deliver a tailored six-week intervention in three participants with advanced MS symptoms. Participants attended up to three sessions every week for 30 minutes of training. For two participants, training was divided into two 15-minute blocks with separate tailored configurations. Single-session healthy reference datasets were acquired for each tailored intervention and sliding window analyses were performed on the classification records from all sessions. We thus attempted five tailored interventions. The primary success criteria were: 1) creation of accurate classification models, and 2) participant ability to successfully interact with the biofeedback during the 15- or 30-minute session. Four of the interventions were successful with typical classification accuracy above 95% and stable feedback control during all sessions. One intervention was not successful, with median accuracy of 86% and negligible feedback control in three of seven sessions. This work suggests that class-based EMG biofeedback is feasible in multiple sclerosis patients with severe impairment, and that conducting larger studies of the technology’s impact on health outcomes is a practical next step.

The Bayes Error Rate (BER) is the fundamental limit on the achievable generalizable classification accuracy of any machine learning model due to inherent uncertainty within the data. BER estimators offer insight into the difficulty of any classification problem and set expectations for optimal classification performance. In order to be useful, the estimators must also be accurate with a limited number of samples on multivariate problems with unknown class distributions. To determine which estimators meet the minimum requirements for ‘‘usefulness’’, an in-depth examination of their accuracy is conducted using Monte Carlo simulations with synthetic data in order to obtain their confidence bounds for binary classification. To examine the usability of the estimators for real-world applications, new non-linear multi-modal test scenarios are introduced. In each scenario, 2500 Monte Carlo simulations per scenario are run over a wide range of BER values. In a comparison of k-Nearest Neighbor (kNN), Generalized Henze-Penrose (GHP) divergence and Kernel Density Estimation (KDE) techniques, results show that kNN is overwhelmingly the more accurate non-parametric estimator. In order to reach the target of an under 5% range for the 95% confidence bounds, the minimum number of required samples per class is 1000. As more features are added, more samples are needed, so that 2500 samples per class are required at only 4 features. Other estimators do become more accurate than kNN as more features are added, but continuously fail to meet the target range.

Accurate damping estimation is crucial for structural health monitoring and machinery diagnostics. This article introduces a novel wavelet-based framework for extracting the damping ratio from multiple impulse responses of vibrating systems. Extracting damping ratios is a numerically sensitive task, further complicated by the common assumption in the literature that impacts are perfectly aligned—a condition rarely met in practice. To address the challenge of non-synchronized recordings, we propose two wavelet-based algorithms that leverage wavelet energy for improved alignment and averaging in the wavelet domain to reduce noise, enhancing the robustness of damping estimation. Our approach provides a fresh perspective on the application of wavelets in damping estimation. We conduct a comprehensive evaluation, comparing the proposed methods with four traditional algorithms. The assessment is strengthened by incorporating both numerical simulations and experimental analysis. Additionally, we apply the analysis of variance (ANOVA) test to assess the significance of algorithm performance across varying numbers of recordings. The results highlight the sensitivity of damping estimation to time shifts, noise levels, and the number of recordings. The proposed wavelet-based approaches demonstrate outstanding adaptability and reliability, offering a promising solution for real-world applications.

Bearing fault diagnosis is a well-developed field and an active area of research in which the combination of model-free machine learning techniques with vibration data has become a popular approach. However, vibration data from rotating machines has the potential to contain domain shifts beyond the accepted causes in this research area (different part models, operating conditions and sensor locations) which can enable data leakage between training and test datasets. To demonstrate the impact of data leakage, six common bearing diagnosis methods are applied to two datasets using three data splitting methods to compare classification performance. Diagnosis is preformed using Principal Component Analysis (PCA), Supervised Principal Component Analysis (SPCA) and Linear Discriminant Analysis (LDA) in combination with frequency analysis and envelope analysis feature extraction methods. Datasets from McMaster University and Paderborn University are used as experimental data sources, and produce vastly differing results (over a 40% drop in accuracy) depending on the selected dataset splitting method, revealing a previously unknown domain shift. Despite great results for diagnosis methods using frequency response analysis on the data from McMaster, these results are not expected to generalize due to possible data leakage. Out of fifty-five previous works using the Paderborn dataset, ten are identified as likely to be affected and only six properly address the problem. Recommendations are given for future experiment design, model creation and model evaluation.

Classification, the task of discerning the class of an unlabeled data point using information from a set of labeled data points, is a well-studied area of machine learning with a variety of approaches. Many of these approaches are closely linked to the selection of metrics or the generalizing of similarities defined by kernels. These metrics or similarity measures often require their parameters to be tuned in order to achieve the highest accuracy for each dataset. For example, an extensive search is required to determine the value of K or the choice of distance metric in K-NN classification. This paper explores a method of kernel construction that when used in classification performs consistently over a variety of datasets and does not require the parameters to be tuned. Inspired by dimensionality reduction techniques (DRT), we construct a kernel-based similarity measure that captures the topological structure of the data. This work compares the accuracy of K-NN classifiers, computed with specific operating parameters that obtain the highest accuracy per dataset, to a single trial of the here-proposed kernel classifier with no specialized parameters on standard benchmark sets. The here-proposed kernel used with simple classifiers has comparable accuracy to the ‘best-case’ K-NN classifiers without requiring the tuning of operating parameters.

Electromyography (EMG)-based biofeedback has considerable potential as a mode of therapeutic exercise for individuals with motor impairments. Advances in technology now allow the delivery of biofeedback with cheap, compact computers and sensors. Reliable machine learning classification of EMG signals in real-time, which is a core component of many biofeedback systems, is also now easily accessible through various open-source software libraries. Despite this progress and the attention garnered by EMG biofeedback among researchers, broad clinical acceptance remains elusive. We aim to open this technology to a broader audience by proposing an accessible standard approach to the design and implementation of EMG-based biofeedback systems. We highlight important considerations when designing a system to deliver potent biofeedback, including maximizing motivation, minimizing constraints on sensor number or configuration, and maximizing replicability by other researchers. Based on relevant neuroscientific and technical literature, we recommend methods and procedures by which these goals can be achieved. Finally, we create and test a biofeedback system in a sample of both healthy and motor impaired volunteers. We found that the EMG biofeedback system supported accurate and stable control by healthy and impaired users and could be implemented with minimal access to coding expertise and an off-the-shelf EMG device. This work expands awareness of effective design principles for EMG biofeedback systems and will advance the state-of-the-art in this field.