In this paper we propose and present a straightforward generalization of the Hamburger moment problem for a single random variable to higher-order stochastic processes via generalized tensor approach. The problem is considered as the Nevanlinna-Pick type tensor moment interpolation (often termed also the Schur interpolation) vs. covariance interpolation problem for 2M-th order stochastic processes. Its orthogonal solution via the generalized tensor Schur Transform and the Lossless Inverse Scattering (LIS) theory is presented in our second paper submitted to this conference.
This paper presents a multi-layered augmented image encryption algorithm that combines the Knight’s Tour scrambling algorithm, DNA encoding, a 4D hyperchaotic system-based key and S-box, and a Mersenne Twister-generated key. The process begins with the generation of a pseudo-random key using the 4D hyperchaotic system, which is XORed with the original image data. The scrambled image is then further processed using Knight’s Tour paths generated by a Hamiltonian path algorithm. DNA encoding is applied next, preparing the data for logical operations and the application of the hyperchaotic S-box. Finally, a second key generated by the Mersenne Twister is used in an additional XOR operation to produce the fully encrypted image. The proposed method is evaluated using metrics such as Peak Signal-to-Noise Ratio (PSNR), histogram analysis, and information entropy, demonstrating excellent encryption performance and strong resistance to various attacks. Moreover, the scheme features an extensive key space of 22604, ensuring robust security against brute-force attacks.
The paper presents the design and implementation of a modular perception and control system for an autonomous surface vehicle (ASV), developed on a System-on-Module (SoM) embedded platform. The system was created with the goal of participating in the international RoboBoat competition. The high-level architecture integrates data from a stereo camera, LiDAR sensor, and a GNSS unit with an Inertial Measurement Unit (IMU), enabling object detection, simultaneous localization and mapping (SLAM), and autonomous navigation. We evaluated the use of a quantized convolutional neural network (YOLOv11) for object classification, obstacle detection based on LiDAR data, and control module operation using NAV2 within the ROS 2 environment under limited computational resources. Accelerating visual detection algorithms with CUDA and TensorRT libraries on the onboard NVIDIA Jetson Orin Nano platform enabled effective real-time and low-energy classification. Tests carried out in simulation demonstrated the correct operation of the perception and control system. The average runtime of the system using the quantised YOLO model with GPU was 26.01 ms. This equates to an approximate speed increase of 6.7 times compared to inference performed on an ARM processor, and an approximate speed increase of 5.5 times compared to a standard CPU.
The article concerns the evaluation of the impact of microwave radiation with a power of 700 W on the neutral and acidic food safety heated in currently popular plant-based materials intended for contact with food (so called food contact materials, FCMs). Low molecular weight carbonyl compounds (especially formaldehyde, acetaldehyde and acetone) were treated as markers of undesirable dynamic FCM-food interactions. In the migration studies, appropriate food simulants were used, imitating neutral food (distilled water, 10% ethanol) and acidic food (3% acetic acid). The preliminary preparation of samples for testing, including derivatization, extraction and purification, enabled the use of a sensitive method: gas chromatography with an electron capture detector (GC-ECD) to identify and quantify migrating organic contaminants from plant-based FCMs to neutral and acidic food. It enabled to assess the safety of ecological materials for rapid heating of food in microwave ovens.
Event-based cameras are gaining popularity as the sensor of choice for mobile robotics, due to their high performance in dynamic environments. However, these applications require efficient real-time data processing with low latency and power consumption. One strategy to meet these stringent requirements is hardware acceleration of efficient algorithms that preserve the temporal sparsity of event data. In this work, we propose an optimization strategy for Graph Convolutional Neural Networks models aimed at adapting their architecture to the limited resources of embedded heterogeneous FPGA platforms. Our method incorporates hardware-aware pruning and quantization, taking into account the trade-off between on-chip memory savings and inference accuracy. Strategic exploration of the design space with Fine Grid Search and Greedy layer-wise Iterative Deepening Search methods enables flexible adaptation of the model architecture to the target platform. Our approach was evaluated across various network configurations and multiple datasets, resulting in BRAM memory reductions of 28.8
This paper introduces a deep learning-based approach for classifying welding quality in the MAG welding process using features extracted from sensor signals via discrete wavelet transform (DWT). Data was collected from sensors such as microphones, accelerometers and current sensors during the welding operation. Welding quality is classified into classes such as good, bad, and partial welds using a deep neural network (DNN). To address class imbalance, balancing techniques such as SMOTE, SMOTEENN, ENN, Tomek Links, SMOTE-Tomek Links, and ADASYN were explored. DWT is used to acquire high-frequency detail coefficients with transient features and local signal variations, which are crucial for assessing the quality of the welded joint. These features help the DNN to learn and distinguish among several welding results depending on complex underlying patterns. Results from experiments conducted emphasized the performance of all balancing techniques, where SMOTEENN was the outperformer. Test accuracy achieved by SMOTEENN in the DNN was 98.80%, greatly exceeding the results from other approaches such as SMOTE, ENN, Tomek Links, SMOTE-Tomek Links, and ADASYN. These findings demonstrate that SMOTEENN is much more efficient at solving class imbalance problems and improving the generalization ability of the model, validating that the combination of wavelet feature extraction and deep learning is effective and reliable for weld quality classification.
In this paper, a novel multiple image encryption algorithm is proposed based on the iterative application of a memristive coupled neural network (MCNN) system exhibiting hyperchaotic behavior. The system generates dynamic pseudo-random sequences and substitution-boxes (S-boxes) to perform multiple rounds of confusion and diffusion on medical and general-purpose images. Designed for both security and performance, the algorithm ensures that small changes in the input or key yield significant alterations in the ciphertext, satisfying Shannon’s principles of confusion and diffusion. The encryption process is implemented in hardware using an FPGA platform, demonstrating suitability for real-time applications. Experimental results confirm the algorithm’s robustness through high entropy, low structural similarity, and strong resistance to statistical and differential attacks. Comparative analysis with existing methods highlights the proposed system’s competitive performance in terms of security metrics, key space expansion, and hardware efficiency.
Nodules found in the lungs are classified by physicians as benign or malignant based on lung images. If malignant nodules are detected, methods such as surgical intervention, radiation, or chemotherapy can be used during the treatment process. Early diagnosis is of great importance in the treatment of lung cancer because the extensive vascular network can cause the disease to spread rapidly. With the help of artificial intelligence (AI), fast and precise diagnosis of nodules and monitoring of their size over time using lung imaging data, makes the job of specialist doctors easier and enables early intervention. In this study, we proposed a novel framework using deep learning-based AI techniques supported by YOLOv8 for the detection of lung nodules and multi-class classification according to their size. Our proposed model based on YOLOv8x and YOLOv8n achieved lung nodule classification tasks with average accuracy rate of 90% and 91%, respectively.
Autism Spectrum Disorder (ASD) is a neurodevelopmental disease characterized by a range of symptoms including repetitive behaviors, fascination with rotating objects, and communication deficits. However, one of the most prominent symptoms has been observed to be avoidance of eye contact. When diagnosed early, appropriate treatment and education methods significantly improve patients’ quality of life. In this study, artificial intelligence (AI)-based deep learning models were developed using 547 eye-tracking image data from a total of 59 children, including children with typical development (TD) and those diagnosed with ASD. With the Convolutional Neural Network (CNN) model trained on the data augmented by preprocessing steps such as cropping and shifting, 74% test accuracy was achieved. In addition, with the ResNet50, VGG16 and InceptionV3 models trained with transfer learning techniques, 72%, 70% and 73% accuracy values were reached, respectively. When the metrics such as accuracy, precision, recall and F1 score of the developed models were examined, it was seen that they could be used in the early stage detection of the disease in ASD patients.
In this paper, we present a compact four-element circularly polarized microstrip patch antenna array engineered for dual-band operation and optimized for the 2.4 GHz ISM band. The proposed design exhibits two primary resonances at 2.20 GHz and 2.45 GHz, each providing return losses better than –10 dB and a continuous impedance bandwidth spanning 2.15–2.50 GHz. An axial ratio below 1.2 dB confirms high circular polarization purity, while a realized gain of 3.6 dBi at 2.45 GHz demonstrates effective radiation performance that remains stable through 5 GHz. Moreover, the array supports a secondary resonance between 4.156 GHz and 4.664 GHz, enabling true dual-band functionality within a single aperture and compatibility with 5G NR n79 channels, fixed wireless access, WiMAX near 4.5 GHz, and C-band satellite downlinks. These features collectively underscore the array’s potential as an element for multi-band MIMO for radar, mobile, backhaul, and IoT applications.
This paper proposes a method for early detection of inter-turn short-circuit faults in squirrel-cage induction motors using vibration signal analysis. The approach combines wavelet packet decomposition and statistical correlation analysis to extract features sensitive to low-level faults under steady-state conditions. Correlation surfaces are generated to evaluate fault-related changes under varying load conditions. Additionally, machine learning algorithms are applied to enhance diagnostic accuracy, with regression models providing interpretable assessment of fault severity. The method shows improved sensitivity compared to conventional techniques and offers a non-invasive diagnostic framework suitable for industrial applications.
The Hilbert transformer (HT) is a digital filter that changes only the phase by 90 degrees without changing the amplitude. Depending on the symmetry of the filter coefficient matrix, four types of 2-D HT can be defined, each with a square filter-coefficient matrix of odd-order. This paper introduces the four types of symmetries. Furthermore, we show the relationship between the frequency response of the HT designed based on each symmetry and the symmetry of the filter coefficients.
Thanks to the advancement in computing power and the processing of sensor data collected from machines, great efforts have been made in recent years to predict machine tool breakage and wear that may cause production to stop. Long-term data collected from sensors constitutes big data. A certain part of this large amount of data is labeled by experts and successful predictions can be made for future. In this study, the number of features was reduced from 134 to 10 with three different feature selection methods using big data collected in the industry. Inferences can be made from the data of cutting tools in production using artificial intelligence algorithms. The main motivation for this study is the tools used for machining. During the production of the tools, data is received from the machine with the help of sensors. Estimation of the locations where the location data will go on the three axes, which are among the selected discriminative features, was made using LSTM, RNN and ARIMA algorithms. The LSTM algorithm has the lowest error rate. According to the results, preventing production disruption by predicting tool tip wear is possible with the hybrid approach of discriminative feature selection and deep learning-based prediction.
This paper presents the differences in the time-frequency properties of bonafide and spoofed recordings. A comparative analysis was conducted on the ASVspoof 2019 dataset, which contains real and fake English utterances synthesized by generative models. In our study, fake recordings generated by 3 models were used. At the initial stage, three time-frequency representations: spectrogram, mel spectrogram, and constant-Q transform (CQT) spectrogram, were compared. Based on the conducted comparisons, the representation that exhibited the highest degree of differentiation between real and fake pairs was selected. Subsequently, a series of tests was conducted to examine the differences in spoofed audio. Using histogram analysis of energy distributions in selected frequency bands of the analyzed signals, distinct features that characterize fake utterances were identified. Besides comparing the energy distribution of each utterance, the differences in speech prosody were evaluated based on the dynamics of changes in the fundamental frequency of the speech signal. As a result of the study, increased energy in fake audio was observed. Moreover, it did not cover the full range of fundamental frequency values computed from the real subset, and it displayed a tendency to produce longer voiced segments.
In this article, the authors explore how well-known computer vision methods can support path planning in static environments. Such methods can play an important role in real-time industrial applications, providing accurate and efficient results while ensuring a high level of safety. The proposed solution uses a computer vision algorithm for boundary detection, supported by a convolutional neural network (CNN), to identify obstacles in an input image. The grid map is then updated with obstacle edge information obtained from the computer vision algorithm before the path is recalculated using heuristic planning algorithms. This method was tested in a simulated environment with various numbers and locations of obstacles. The performance of each case was evaluated by calculating the path length, planning time, and the number of nodes explored. The results suggest that using a computer vision-based obstacle detection algorithm to prepare the map improves the overall efficiency and accuracy of the planning process. Although the environment remains static, using visual data makes the setup more realistic and closer to real-world conditions.
Blood testing in cattle is essential for feed adjustment and the early detection of diseases, thereby improving reproductive performance and growth. Metabolic Profile Tests (MPTs), however, require laboratory analyzers and trained personnel, limiting their on-farm applicability. We therefore propose a non-invasive method for estimating nine blood component concentrations by combining handheld near-infrared spectroscopy with a ResNet-18-based deep learning model. After denoising each spectrum by a k-point moving average and formulating the loss with label distribution learning, the model achieved estimation accuracies exceeding 80% for all components.
Human color perception is a complex process involving the interaction of light, the eye, and the brain. It is not just about the eye detecting wavelengths and the brain registering them; rather, it is a multifaceted sensory experience shaped by both physical and psychological factors.Therefore, research in this field is still being intensively developed. However, many aspects and phenomena related to it still remain unexplored and require a correct theoretical approach and appropriate mathematical representation. This latter issue is addressed in this article, which presents mathematical descriptions of the most important color spaces used to represent colors in images, namely: RGB, RGBW, CMY, CMYK, HSB and HSL using linear algebra, geometric algebra, and abstract algebra of split quaternions.Some of the described ideas, resulting from the performed algebraic considerations, namely RGBW (red, green, blue, white) family of color spaces and the RGBW pixel configuration, have been proposed earlier, but are not widely known. However, the HSD (hue, saturation, darkness) cone color space concept is new.It has been shown that the description of the double HSL cone and the simultaneous description of both single cones: HSB and HSD is possible using a subalgebra of pure imaginary split quaternions with positive quadratic form. This is the idea of the authors, who are still working on assessing its usefulness.
Alzheimer’s disease is a common neurodegenerative disorder that significantly impairs individuals’ cognitive and communicative abilities. Subtle changes in speech signals have been identified as potential indicators for the early detection of this condition. In this study, we propose a novel deep learning-based approach for Alzheimer’s diagnosis by leveraging acoustic features extracted from the ADDReSS 2020 and ADDReSSo 2021 speech corpora. While previous studies have predominantly focused on extended speech segments or transcript-based features, this study introduces a novel segmentation approach that departs from conventional methods by targeting specific spoken words. Mel-spectrograms and chromagrams derived from the recordings were utilized as input features to train Convolutional Neural Network (CNN)-based models. The two proposed models, trained separately on spectrograms and chromagrams, achieved accuracy rates of 99% and 98%, respectively. These findings, evaluated using standard classification metrics, underscore the considerable potential of CNN-based methods utilizing raw speech data for the diagnosis of Alzheimer’s disease.
Recent advances in neural networks have enabled the development of numerous tools that facilitate daily life. A notable example is an application designed to assist visually impaired individuals in navigating urban environments. Although the use of segmentation, depth estimation, and object detection models offers significant potential in this area, the necessity of making reliable decisions about the passability of the predicted path in a dynamically changing environment presents a substantial challenge for neural networks, ranging from selecting an appropriate model, through dataset preparation, to the training process itself. Existing publicly available tools did not meet our stringent requirements in terms of both quality and computational efficiency. Consequently, we opted for a significantly simpler yet markedly more effective solution for this specific context. By comparing and verifying selected key points, we developed the so-called "trapezoid method", which consistently outperformed the neural network-based models evaluated in our study.
This paper presents a wireless biosignal acquisition system for human-computer interaction (HCI) based on surface electromyography (EMG) and electrooculography (EOG). The device captures facial and ocular gestures using four differential channels and processes them in real time on a low-cost ESP32 platform. Five gestures were tested: jaw clenching, cheek puffing, cheek sucking in, eye blinking, and eyebrow raising. The system achieved high classification accuracy (up to 97%) and was also validated in case of limited facial muscle control. User evaluations confirmed high gesture comfort. The results support the feasibility of this approach for intuitive, gesture-based control in assistive and immersive HCI environments.
