Publications & talks

Assessing whether datasets follow a normal distribution is essential for valid statistical inference, yet GNSS positioning errors often deviate from Gaussian assumptions under real conditions. Despite the widespread use of normality tests, their performance on GNSS data remains insufficiently characterized, especially under heavy-tailed regimes. This study investigates long-duration GNSS latitude residuals as a one-dimensional case using stationary observations collected over a continuous 48-h period. Normality is evaluated using graphical diagnostics, descriptive statistics, and hypothesis testing, including histograms, box plots, Q–Q plots, skewness, kurtosis, and seven parametric tests. To assess robustness and sensitivity, a Monte Carlo bootstrapping procedure is applied to more than 160,000 latitude samples. For multiple sample sizes and significance levels, 10,000 replicates are used to estimate empirical power and median p-values. The results indicate clear departures from normality, with Shapiro–Wilk and D’Agostino showing the highest sensitivity, particularly for small and moderate samples. GNSS latitude errors are additionally modeled using Student’s t distribution and a Gaussian–Student’s t mixture. These models better represent the empirical distribution, especially in the upper tail, than the Gaussian model. The findings confirm that Gaussian assumptions may underestimate uncertainty in GNSS analysis. They also show that combining normality diagnostics with flexible statistical models improves error characterization under non-Gaussian conditions.

Excessive speed remains one of the main factors contributing to road‑traffic fatalities, while current Intelligent Speed Assistance (ISA) and map‑ or vision‑based systems often fail because of environmental limitations, outdated data, or high infrastructure costs. This paper proposes a cloud‑based Unified Driver Speed‑Assistance System (UDSAS), inspired by the European Train Control System (ETCS). The aim of the concept is to create a centralised and authoritative digital source of speed‑limit information and to enable real‑time computation of vehicle‑specific braking curves. The system integrates existing in‑vehicle components, including navigation, GNSS positioning, and speed‑limiter interfaces, to ensure low implementation cost and high scalability. The driver remains in the loop, while the system provides continuously updated speed guidance along the planned route. Optional extensions include offline operation, integration with vision‑based recognition, digital signage, Vehicle‑to‑vehicle (V2V) communication, and support for autonomous vehicles. The proposed solution addresses the current lack of a unified and trustworthy speed‑limit database and provides a simple, cost‑effective, and evolvable framework that can enhance road‑traffic safety, environmental performance, and driver comfort.

High-impedance ion-selective electrodes are susceptible to electromagnetic interference (EMI), which degrades potentiometric sodium and pH monitoring in automation. This study presented a disturbance-aware reconstruction pipeline for dual-channel recordings sampled at 1 kHz under laboratory conditions. A windowed sinusoidal regression was estimated and the 60 Hz interference component was subtracted using overlapping 0.5 s windows. Disturbances were detected per channel using robust amplitude and derivative scores with hysteresis, minimum duration rules, and mask dilation. Masked samples were replaced by an interpolated quiet-state baseline estimated from locally weighted scatterplot smoothing (LOWESS) using unmasked data only. Disturbed datasets contained events that masked up to 33% of the sodium record, which were reconstructed without changing the time axis. Two outputs were generated, including a drift-preserving trace and a drift-removed trace anchored to an initial reference level. Six smoothing kernels and a zero-phase second-order sections (SOS) Butterworth low-pass filter were benchmarked using quiet-only metrics. The SOS low-pass filter outperformed six alternative kernels in 22 of 24 experimental test cases, including one tie. Under the best quiet segment, robust variability decreased by approximately 50 times after reconstruction and low-pass filtering overall. Power spectral density values near -200 dB reflected a numerical reporting floor set by the safeguard, not floating-point quantisation.

Reliable ion-selective electrode (ISE) monitoring in industrial matrices is often compromised by electromagnetic interference (EMI). Traditional uncertainty models assume independent Gaussian noise; however, ISE data sampled at 1 kHz reveals temporal instabilities and cross-channel dependencies that violate these frameworks. This study presents a two-phase metrological evaluation of ISE responses (EH-01 and ENa-01). First, statistical characterization reveals extreme non-Gaussian behavior, with excess kurtosis reaching κ=20, negative skewness (γ = -3.40), and a Pearson correlation shift from r ≈ 0.26 to r ≈ 0.69 during impulsive activity. In the second phase, we evaluate uncertainty propagation using the Guide to the Expression of Uncertainty in Measurement (GUM) and validate results via Monte Carlo simulation (MCs). Our findings demonstrate that GUM underestimates measurement risk for EH-01 by 15.0% due to heavy-tailed noise, while overestimating ENa-01 uncertainty by 22.0% due to extreme central peak density. These results invalidate the second-moment-only GUM approximation in high-interference settings, necessitating multivariate, non-Gaussian uncertainty evaluation. These findings motivate the development of ML-based filtering approaches to mitigate EMI-induced transients and potentially restore GUM validity in real-time industrial process control.

Air-quality measurements performed using unmanned aerial vehicles (UAVs) enable observations that are difficult or impossible to obtain with stationary monitoring systems. Although low-cost CO2 sensors are widely applied in such work, their accuracy is restricted by environmental influences. This study assesses the metrological performance of inexpensive NDIR CO2 sensors using a controlled test chamber. The TESTO probe results show strong temperature sensitivity, with CO2 indications varying by approximately 17 ppm per 1 °C. Measurements at −2.6 °C produced implausibly low concentrations of 275–280 ppm, despite the global baseline being about 430 ppm. Electromagnetic interference and humidity produced negligible effects on the indications. No differences appeared between measurements taken during UAV flight, after landing, or under laboratory conditions. Comparison with the manufacturer-calibrated Figaro CDM7160 sensor revealed a substantial shift in the characteristic at the lowest CO2 concentration level and a marked reduction in sensitivity, which shows that the sensor needs recalibration. The findings confirm that investigated low-cost CO2 sensors provide reasonably accurate absolute measurements only when environmental conditions are correctly compensated. However, their relatively high measurement uncertainty prevents reliable detection of small concentration changes and therefore limits their suitability for precise UAV-based air-quality studies.

Modeling the dynamics of the flow rate system is challenged by the nonlinear behavior and the noisy measurement data. Accurate models require a comprehensive understanding of fluid mechanics, as well as knowledge of all instruments in the measurement chain. This study presents a black-box optimization approach to develop a nominal Fractional-Order (FO) model of a laboratory-scale flow system. The model was constructed by repeatedly solving an optimization problem using preprocessed experimental data and averaging the resulting optimal parameters. The nominal FO model was then validated against unseen, unprocessed measurement data to assess its robustness. The parameter sensitivity of the proposed model was analyzed by introducing +10% and +20% perturbations in each parameter individually. Error analysis evidences that root mean squared, mean absolute, and mean absolute percentage errors with the proposed model have reduced to 9.3%, 5.1%, and 5.3%, respectively, compared to those integer-order models. Furthermore, residual-based distribution analysis confirms the robustness of the approach, with residuals tightly concentrated around the lowest values. Although the FO model incurs a higher computational cost during optimization, it was significantly reduced using an online optimizer. The proposed model demonstrates superior robustness and accuracy, making it a compelling choice for precise modeling.

As part of the project, a system was developed to measure and control the conditions within a terrarium. The system measures air temperature and humidity, as well as soil moisture, and uses this information to control lighting, heating, and misting. Information about the status of the devices can be accessed via a custom-built mobile application (web-based). The application also allows users to set selected parameters, such as lighting periods and temperature. The device is equipped with a microprocessor system based on a Raspberry Pi. The entire system is aesthetically designed. The students equipped the terrarium with decorations, sensors, and actuators, while the electronics were enclosed in a housing produced using 3D printing technology.

As part of the project, a system was developed to maintain optimal moisture levels in two flower pots. Each pot is controlled independently. The input data includes soil moisture and temperature. Depending on the implemented algorithm, the system dispenses water from a reservoir. The entire setup is controlled by a microprocessor-based system built on a Raspberry Pi. The current operating status, temperature, and humidity are displayed on an e-paper display. The whole system is aesthetically designed using 3D printing.

Accurately characterising datasets is crucial for effective statistical modelling, particularly when analysing Global Navigation Satellite System (GNSS) data. While traditional approaches often assume a Gaussian distribution, real-world GNSS datasets frequently exhibit heavy-tailed and skewed properties, prompting the need to explore alternative statistical models. The study examines the suitability of non-Gaussian distributions, specifically the Laplace, skew-normal, skew-t, and generalised hyperbolic (GH) distributions, for modelling GNSS data obtained from a stationary receiver. Using empirical GNSS datasets, we estimate parameters within confidence intervals (CIs) through weighted maximum likelihood estimation (WMLE). Model performance is assessed using log-likelihood analysis, Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and root mean squared error (RMSE). Our comparative analysis shows that heavy-tailed and skewed distributions, particularly those offering greater flexibility in capturing extreme deviations, consistently outperform the conventional normal distribution. Among the non-Gaussian models considered, the GH distribution provides the best overall performance. These results emphasise the importance of selecting appropriate statistical models to improve uncertainty quantification in GNSS-based measurements.

Air quality measurements conducted using unmanned aerial vehicles offer researchers opportunities that were previously difficult to achieve. An increasing number of publications on this topic are appearing in the literature. However, the popularity of low-cost sensors is often associated with their limited accuracy and sensitivity to environmental changes. This work demonstrates that CO2 sensors can accurately indicate the actual concentration of this gas in the air, provided that the influence of temperature is considered.

Regulated flow systems exhibit a variable dynamic behavior when subjected to distinct inputs. The significant non-linear behavior of the system at low inputs and the difference in the output pattern for the increase and decrease in flow pose challenges in modeling. One way is to identify separate sets of parameters for each input-output data set and develop distinct models. However, this approach is computationally expensive when designing a single controller covering the whole input range. To overcome the said limitation, we used a Principal Component Analysis (PCA) based estimation technique. The developed Linear Parameter Varying (LPV) dynamic model describes real measurement values of a laboratory flow system. The data treatment consists of (1) an automated input-output data acquisition, (2) 3rd-order model estimation from measured data, (3) the reduction of the dynamic parameters using PCA, and (4) the development of a reduced LPV model for the entire input range. The LPV model presents its output response comparable to the experimentally taken flow rates. The proposed modeling technique can help design a single controller sufficient to achieve the desired output applicable in the whole measuring range. The effectiveness of our approach suggests its use for synthesizing an LPV controller for flow systems.

Accurate positioning obtained from the Global Navigation Satellite System (GNSS) is crucial for many applications; however, environmental and instrumental factors, such as satellite geometry, signal multipath, and atmospheric delays, often degrade its accuracy. Traditional uncertainty quantification and filtering methods, such as the Kalman Filter (KF), assume that the measurement noise follows a Gaussian distribution. In practice, real-world GNSS data often violate this assumption, showing skewness, heavy tails, and outliers that distort the estimates and reduce the accuracy. This study systematically evaluates the impact of non-Gaussian noise on GNSS positioning accuracy and introduces a unified framework that integrates robust statistical methods and nonlinear filtering. Specifically, we compared the performance of the KF, Median Absolute Deviation (MAD), and Particle Filter (PF) on stationary GNSS data. The evaluation used comprehensive accuracy and uncertainty diagnostics, including the root mean square error (RMSE), probable circular error (CEP), estimated position error at 95% (R95), covariance analysis, and nonparametric bootstrapping. The results demonstrate that Gaussian-based models fail under heavy-tailed noise, whereas MAD and PF achieve superior reliability by suppressing outlier influence and capturing complex error distributions. The proposed framework advances GNSS uncertainty quantification by bridging robust statistics and nonlinear filtering. It offers a statistically sound and practical approach for real-world environments in which non-Gaussian noise is prevalent.

In statistical analysis, evaluating the normality of large datasets is crucial for validating parametric tests, particularly in areas such as Global Navigation Satellite System (GNSS) measurements, where data often exhibit non-normal characteristics resulting from their variability and errors. This research aims to transform the measured GNSS data and to assess the effectiveness of transformation methods in achieving normality. Techniques like logarithmic, quantile and rank-based Inverse Normal Transformation (INT) were evaluated using visual methods (histograms, Q-Q plots), descriptive statistics (skewness, kurtosis) and statistical tests, including Kolmogorov-Smirnov (KS), Anderson-Darling (AD), Lilliefors (LF), D’Agostino K-squared (DA), Shapiro-Wilk (SW), Jarque-Bera (JB), Cramérvon Mises (CM), and Pearson Chi-square (Chi2) tests. The sensitivity of these tests to deviations from normality was assessed through the Receiver Operating Characteristic (ROC) analysis and the Area Under the Curve (AUC) values at a significance level of 0.1, using Monte Carlo (MC) simulations across the varying sample sizes. The results showed that untransformed latitude data consistently failed normality tests, while transformed data displayed normal characteristics. The rank-based INT showed superior effectiveness, influenced by the original distribution and characteristics of the dataset. The findings underscore the importance of tailored transformations in large-scale data applications, enhancing the accuracy and applicability of parametric statistical methods in geospatial and other industrial domains.

Due to additional parameters and similar structure with PID, fraction-order (FO) PID controllers offer tuning flexibility for better performance. More degree of freedom poses challenges in applying simple tuning rules for FO-PID controllers. The dilemma of practical realization of the FO controllers further limits their benefits. This work proposes simple tuning rules for the parameters of the FO-PI controllers by following the frequency-domain methods. We apply three different approximation approaches to realize the controller implementation and validate its performance on a simulation example. Comparison with a FO-PID controller evidences the effectiveness of the proposed approach

Electromyography (EMG) signals, crucial for neuromuscular assessment, are frequently corrupted by noise, impairing signal fidelity and subsequent analysis across diverse applications. Conventional filters often inadequately address non-stationary noise or introduce signal distortion. This paper introduces an advanced deep learning framework for EMG denoising, centred on a U-Net-inspired convolutional autoencoder with integrated residual blocks and skip connections. Training utilised synthetic EMG data, closely emulating physiological frequency bands and burst dynamics, subsequently corrupted by a comprehensive noise model encompassing electrode, crosstalk, electronic, drift, and contact artefacts. Training was guided by a custom loss function that combined weighted mean squared error (MSE) with signal-to-noise ratio (SNR). The proposed autoencoder achieved substantial improvements, SNR increased from -0.95 dB (noisy) to 14.64 dB (denoised), and MSE was drastically reduced from 0.001493 V2 to 0.000041 V2 on the test dataset. Qualitative analysis confirmed effective noise suppression while retaining crucial EMG burst characteristics. This advanced framework offers a promising solution for robust restoration of EMG signals in practical settings.

Recent review articles highlight an exponential rise in publications on Digital Twin (DT) technology. Despite its recognised potential, DTs have yet to achieve widespread practical use. Following a presentation of the state of the art and an original conceptual diagram illustrating the technology, this work presents the key factors contributing to the gap between conception and implementation. Using experimental data from a laboratory water system and a reliability-based perspective, this analysis examines the practical limitations of DT application. The findings indicate that broader use of DTs depends on the technological maturity and reliability of all system components, which still require further development.