Open Access

We introduce a compact model for the prediction of performance figures of CMUTs prospectively used in airborne applications such as flow meters. The model comprises both a mechanical and an electrical approach resulting in a set of four parameters: resonance frequency, Q-factor, static capacitance and pull-in voltage. A small set of exemplary CMUTs has been intensively characterized by means of LDV and LCR measurements, eventually confirming the precision of the proposed prediction model.

This work provides insight into the potential of camera-based surface plasmon resonance sensors when utilising the spatial sensor by modulating the sensor response through lateral modification of surface parameters or input illumination. These modifications have to be coupled with adequate signal processing. Modelling the sensor response using Fresnel formulae allows for accurate representation of the sensor surface state, increasing confidence in measured values. Alternatively, data-based modelling approaches can be utilised which eliminate the need for approximation of a physical model to observed data but are limited by the data they are trained on. It has been explored how, especially, deep learning models that are specialised in image processing (CNNs) can be utilised for qualitative or quantitative assessment of analytes with the help of a receptor array on an SPR surface. Notably, this approach demonstrates strong generalization capabilities, performing effectively on unseen sensors used for the same task. Preliminary experiments indicate that performance enhancements are feasible when data from different points in time during binding or debinding are used during processing. In conclusion, deep learning-assisted spatial SPR sensors hold immense potential for diverse applications, including online quality control, anomaly detection, and biofouling quantification. However, realizing this potential hinges on carefully designed experiments that yield high-quality data. Furthermore, incorporating orthogonal data sources is crucial for more precise determination of the surface state during calibration. Suitable surface-sensitive techniques include Raman spectroscopy, ellipsometry, and impedance spectroscopy. Coupling these methods can improve knowledge on the observed state during training. This information together with the presented methods can simplify sensor development and improve performance. Diese Arbeit gibt einen Einblick in das Potenzial kamerabasierter Oberflächenplasmonenresonanzsensoren, wenn der räumliche Sensor durch die Modulation von Sensorreaktionen durch laterale Modifikation der Oberflächenparameter oder der Eingangsbeleuchtung genutzt wird. Diese Modifikationen erzeugen eine komplexe Sensorantwort und müssen daher mit einer geeigneten Signalverarbeitung gekoppelt werden. Die Modellierung der Sensorantwort mit Hilfe von Fresnel-Formeln ermöglicht eine genaue Bestimmung des Sensoroberflächenzustands und erhöht so die Konfidenz in gemessene Werte. Alternativ können datenbasierte Modellierungsansätze angewandt werden, die eine Annäherung eines physikalischen Modells an aufgenommene Daten ersetzen können, aber durch die Daten, auf denen sie trainiert werden, begrenzt sind. Es wurde untersucht, wie insbesondere tiefe neuronale Netze, die auf die Bildverarbeitung spezialisiert sind (CNNs), für die qualitative oder quantitative Bewertung von Analyten mit Hilfe eines Rezeptor-Arrays auf einer SPR-Oberfläche eingesetzt werden können. Insbesondere zeigt dieser Ansatz starke Generalisierungsfähigkeiten, die auch bei unbekannten (unkalibrierten) Sensoren, die für dieselbe Aufgabe verwendet werden, wirksam sind. Vorläufige Experimente deuten darauf hin, dass Genauigkeitssteigerungen möglich sind, wenn bei der Verarbeitung Daten von verschiedenen Zeitpunkten während der Bindung verwendet werden. Zusammenfassend lässt sich sagen, dass räumliche SPR-Sensoren mit Hilfe von tiefen neuronalen Netzen immenses Potenzial für verschiedene Anwendungen haben, darunter Online-Qualitätskontrolle, Erkennung von Anomalien und Quantifizierung von Biofouling. Die Realisierung dieses Potenzials hängt jedoch von sorgfältig geplanten Experimenten ab, die qualitativ hochwertige Daten liefern. Darüber hinaus ist die Einbeziehung orthogonaler Datenquellen entscheidend für eine genauere Bestimmung des Oberflächenzustands während der Kalibrierung. Zu geeigneten oberflächensensitiven Techniken gehören Raman-Spektroskopie, Ellipsometrie und Impedanzspektroskopie. Die Kopplung dieser Methoden kann das Wissen über den beobachteten Zustand während des Trainings verbessern. Diese Informationen können zusammen mit den vorgestellten Methoden die Sensorentwicklung vereinfachen und die Leistung des Sensors in der Anwedung verbessern.

Surface plasmon resonance (SPR) is limited by small-signal detectability and drift when subtraction occurs in software after digitization. We introduce an SPR detector that performs on-detector amplification and analog differential readout, eliminating moving parts and software-heavy correction. The hardware-native subtraction boosts the usable ADC range and suppresses illumination and environmental noise. In fixed-angle refractive-index steps (NaCl), the platform resolves Δn_min ≈ 1.8 × 10⁻⁷ RIU compared to 4.6–7.2 × 10⁻⁶ RIU on a commercial comparator and improves small-signal SNR by up to ∼5,000-fold, while remaining competitive at high signal levels. In a model IgG–BSA assay, the detector’s low noise floor clarifies early binding and equilibrium transitions. By generating inherently clean raw signals, this hardware-native approach dramatically enhances sensitivity and long-term stability for label-free biosensing and inline process analytics while rendering AI-based or complex post-processing entirely unnecessary. The concept generalizes across platforms and opens a compact route to robust, high-fidelity SPR in complex environments, with a clear path toward multi-wavelength and arrayed detectors for high-throughput chemical monitoring.

Embedded gesture recognition using radar sensors enables intuitive and robust human-machine interfaces, which is appealing for automotive applications such as trunk opening via foot gesture. Designing a radar-based classifier that runs ondevice (edge) under strict resource constraints poses several challenges. The system must achieve real-time inference (e.g. under 200 ms) on a microcontroller unit (MCU), while maintaining high accuracy, and minimize false detections. This paper focuses on a binary classification task, while distinguishing a valid trunkopening “kick” gesture from other motions. Firstly, we evaluate classical machine learning (cML) models (Random Forests or support vector machines (SVMs)) for baseline performance, however they face difficulties with higher false positive rates and do not meet real time criteria. We then employ Convolutional Neural Networks (CNNs) and apply neural architecture search (NAS) to discover a compact CNN tailored for a 60 GHz Doppler radar dataset of 50 subjects, each with 10 leg gestures. By constraining network depth, filter sizes, and hyperparameters, NAS yields a small but accurate “edge” CNN that fits within the limited memory of a STM32 Nucleo F446RE microcontroller. The final model achieves 95.1 % accuracy on the binary classification task with 96.6 % precision and 2.17 % false positive rate (FPR), and requires only 167.5 ms to run inference on the MCU.

Object recognition systems based on visual sensors often struggle in degraded environmental conditions such as fog, rain, darkness, or occlusion. Radar, in contrast, offers robust detection capabilities under such conditions but lacks the spatial resolution of optical sensors. This paper presents an ongoing research effort towards a multimodal sensor fusion framework that integrates camera data with radar signals, specifically exploiting micro-Doppler (μD) signatures, with the aim of improving object recognition robustness. The proposed setup uses an Intel RealSense Depth Camera D455 and an Infineon BGT60TR13C radar sensor, deployed on a low-power embedded platform with STM32 microcontrollers. The signal processing pipeline, currently under development, combines time–frequency radar analysis with convolutional neural network-based visual feature extraction for real-time edge AI inference. We describe the system design, preliminary data acquisition setup, and planned evaluation strategy.