Gas Chromatography - Ion Mobility Spectrometry (GC-IMS) serves as a valuable analytical tool for enhancing non-targeted screening of volatile organic compounds in a wide range of different applications. Nowadays, authenticating various food samples such as honey and olive oil, as well as identifying microorganism cultures, is essential for practical applications and can be effectively achieved using GC-IMS alongside various classification methods. However, limited resources and the time-intensive data collection process have led to a scarcity of datasets with a relatively low amount of data. As a result, training of generalized classification models becomes more challenging and complex. To address issues, such as effectively learning essential spectral features, compressing the data into a more compact, representative latent space of less than 1% of the input parameters, and classifying unseen spectra with limited training data available, we propose utilizing various configurations of CNN-based Deep Auto-Encoders. Moreover, we propose a detailed pre-processing pipeline along with three different denoising protocols for GC-IMS spectra to achieve a better generalization of our models. Through the evaluation of the proposed models utilizing three publicly available datasets, the experimental results demonstrate that these models outperform state-of-the-art methods reported in the literature across various tasks, including the classification of complex food samples and the identification of microorganism cultures.
Accurate tool wear measurement is crucial for optimizing tool life and product quality. Therefore, our research aim is to realize an optomechatronic sensor solution for in-situ tool wear measurements, using a chromatic confocal sensor and a robotic arm integrated into a turning machine. To enable tool wear measurements with maximal repeatability, i. e. to enable precise sensor positioning despite the robotic arm's position uncertainty of 50 mu m, a tool coordinate system (TCS) is defined based on the tool's unworn edge. Validation experiments on a reference sphere show enhanced repositioning capabilities of the optomechatronic sensor system with the proposed TCS-based registration method in the order of 6 mu m.
Oxygen saturation (SpO2) is a critical indicator of peripheral circulatory health, particularly in conditions such as peripheral artery disease. This study focuses on developing an innovative approach to quantify SpO2 at varying skin depths using multispectral illumination. By employing light sources emitting multiple wavelengths (e.g., 660 nm, 850 nm, 890 nm, and 940 nm) and capturing reflected skin images, the method analyzes SpO2 at different depths and constructs three-dimensional SpO2 distribution maps, enabling visualization and quantification of oxygen saturation across tissue layers. The study further examines the enhancement and retention effects of near-infrared (NIR) light exposure (600-1700 nm) on SpO2 at various depths. Results demonstrate that after 10 minutes of NIR illumination, SpO2 in superficial vascular regions increases by 7%. For retention analysis, the most effective layers for vascular and tissue SpO2 were identified as 940 nm and 890 nm, respectively. Notably, following 15 minutes of NIR exposure, SpO2 in vascular regions decreased by less than 2% within 15 minutes after the light source was turned off. This innovative, non-invasive, and quantitative approach offers insights into oxygen dynamics in tissues and blood vessels. It provides a promising tool for clinical diagnostics and therapeutic monitoring, showcasing the efficacy of NIR in enhancing blood oxygenation.
The temperature field of falling film is crucial of the heat transfer characteristics. However, falling film interface topology and temperature distribution are constantly changing, which is challenging for the analysis of heat transfer characteristics. The Planar Laser-Induced Fluorescence (PLIF) measurement method enables highly accurate and non-intrusive. The PLIF40 method effectively minimizes the influence of total reflection, which ensures the true film thickness more accurately and reliably from complex interfacial waves. In this paper, the super resolution reconstruction algorithm is designed for the falling film by PLIF40. It is utilized to improve the resolution of the falling film images. Moreover, noise is eliminated and the edge of the liquid film is smoothed. The temperature information is accurately extracted, and the heat transfer coefficient and Nusselt number of the temperature field are quantified. The effectiveness of the method is verified by numerical calculations and experiments. Finally, the heat transfer characteristics of the falling film are further investigated.
Inversion of sea surface wind direction using nautical radar images has significant practical value in engineering applications. While traditional single-curve fitting methods effectively address the issue of comprehensive data provision, the method is prone to substantial errors under unstable sea surface condition. This paper proposes a one-dimensional discrete wavelet transform-based algorithm for inverting wind direction from nautical radar images. The algorithm first extracts static low-frequency wind signals from radar images by applying the one-dimensional discrete wavelet transform (1D-DWT) to the one-dimensional radial echo intensity distance distribution data. This step effectively reduces the interference of high-frequency noise and wave signals on echo intensity, while preserving the relationship between echo intensity and azimuth. Secondly, prior to curve fitting, 1D-DWT is applied to extract the smoothed mean value of the echo intensity along each azimuthal line, thereby reducing the influence of echo intensity anomalies caused by breaking waves on the sea surface. Finally, the wind direction is extracted by fitting a cosine squared function using the least squares method. In this paper, real X-band nautical radar image data are utilized, compared to the traditional single-curve fitting algorithm, and the results demonstrate that the correlation coefficient between the inverted wind direction and the sensor wind direction improves by 0.07, the mean bias is reduced by 1.45 degrees, and the root mean square error decreases by 1.64 degrees.
Cerebrovascular diseases rank as the third leading cause of death, following cancer and heart disease, according to the World Health Organization (WHO) [1]. Stroke-related mortality remains difficult to control, with one person dying from a stroke every three minutes on average [2]. This challenge arises because strokes occur suddenly, depriving the brain of oxygen. In acute ischemic stroke, the condition is often unstable, and brain tissue deterioration continues even after hospitalization. Therefore, early diagnosis and timely, appropriate treatment during the critical window before deterioration are essential to reducing stroke-related mortality [3-6]. Clinical data indicate that a BUN/Creatinine ratio greater than 15 is a key indicator for stroke risk assessment [7]. However, current hospital and home care practices fail to implement effective real-time monitoring. This is primarily due to hospitals relying on large central laboratory analyzers to measure BUN and Creatinine concentrations, making immediate BUN/Creatinine ratio assessments unavailable. To address this issue, a point-of-care creatinine detection device was designed to assess acute stroke deterioration. The developed system measures BUN/Creatinine levels in blood or bodily fluids and provides a critical golden window for timely intervention when acute dehydration occurs (BUN/Creatinine > 15). The device was developed as a feasible diagnostic system using commercially available test strips that meet clinical and regulatory standards. A portable prototype was created to detect blood creatinine concentrations within 10 minutes of sampling, allowing healthcare workers or caregivers to make timely evaluations and treatment decisions. By enabling continuous creatinine monitoring, the device helps reduce the risk of stroke progression.
Ancient coins are valuable historical artefacts for revealing insights into ancient civilizations and cultural practices. Typically, numismatists grade the coins relying on their own experience rather than on science and objective measurement tools. For ensuring reliability of a coin-grade system it has to be reproducible and this can only be achieved by building a system that does not fully rely on manual object inspection. In this work, we propose a transfer learning approach to grade ancient Roman coins. In particular, we use coin data comprising three fine grades, such as Fine (F), Very-Fine (VF), and Extremely-Fine (EF) coins. The transfer learning implemented yielded a reliable result in recognizing the grade of these ancient Roman coins. By building a robust data processing pipeline and leveraging transfer learning, an accuracy of 0.84 was achieved. This result shows promise for automated coin grading or assisting manual grading processes.
A double D-shaped eddy current probe was proposed for quantitative non-destructive evaluation. Simulation results demonstrate that the uniform induced eddy currents in the middle gap area have significant potential for defect characterization. A Mn-Zn ferrite core is added in the semi-cylindrical bobbin to enhance the capability of subsurface detection. Experiments were conducted to validate the reliability and efficiency of the proposed system. The results confirm that the probe can effectively detect both surface and subsurface defects, with a maximum detection depth of 6 mm below the surface of an aluminum plate specimen. Calibration equations for both cases are provided, showing a high coefficient of determination. Under consistent testing conditions, this sensor can be effectively utilized for the quantitative characterization of defect depth.
In this paper, a reference-free damage imaging method is proposed based on path matching for quasi-isotropic carbon fiber reinforced plastic (CFRP) structures. The patch matching concept in this technique considers paths with the same propagation distances and similar directions. A structural health monitoring (SHM) system was developed using a sensing array with twelve circular piezoelectric transducers (PZTs) for the generation and reception of ultrasonic guided waves. Among different sensor combinations, there are 66 sensing paths belonging to 12 groups according to the principle of path matching. The excitation frequencies to perform experimental studies are first determined based on the dominance of a single non-overlapping fundamental guided wave mode. Then the group velocities of A0 and S0 modes are measured to determine the time windows applied to the receiving signals to extract the maximum envelopes. And the mode that has a higher sensitivity to defects is selected for damage imaging. The proposed method uses the maximum envelope of the first arrival wave packet as the damage features, ensuring independence from baseline signals. Finally, the effectiveness of the proposed method is validated through four distinct defect cases at various positions, demonstrating both the feasibility and accuracy of the path-matching approach for inspecting anisotropic material.
This work presents a voltametric potentiostat designed for electrochemical sensors based on an STM 32 microcontroller and realizing several measurement techniques such as cyclic voltammetry (CV), square wave voltammetry (SWV) and differential pulse voltammetry (DPV) based on software-configured timing control. The entire measurement system consists of seven hardware elements: a microcontroller, signal generation module, potentiostat, I/V converter, signal conditioning circuit, signal measurement module, and DC-DC power supply module. For precise signal generation, a 16-bit external DAC has been implemented. For an accurate signal acquisition, a 24-bit external ADC has been implemented. To enhance precision, several noise-reduction techniques have been implemented. The system performance has been validated by a dummy cell testing and demonstrates a voltage control accuracy within +/- 1 mV in the central operational range. Electrochemical measurements in ferri-ferrocyanide show comparable results to a commercial PalmSens device, achieving a limit of detection of 24.65 mu M and sensitivity of 3.286 mu A mu M(-1)cm(-2). This work presents a viable approach for developing accessible, high-performance electrochemical instrumentation suitable for research and analytical applications.
In applications where safety and reliability are of utmost importance, such as in the automotive and aerospace industries, the standards for defect coverage are exceptionally stringent. These high standards are necessary to ensure that systems function correctly under all conditions and to prevent failures that could lead to significant losses. Such rigorous requirements are further reinforced by various industry regulations, including ISO 26262 for automotive safety and AEC-Q100 for the qualification of automotive components. A particularly pressing challenge is that nearly 80% of automotive integrated circuit (IC) failures are due to analog defects, highlighting the need for specialized and effective defect detection techniques for analog circuits. This paper introduces a robust and efficient defect detection methodology for Phase-Locked Loops (PLLs), which are key components in analog and mixed-signal (AMS) systems. By leveraging the capabilities of digital control and observer circuits, the proposed method offers key advantages, including enhanced robustness and operational simplicity. Overall, using transistor-level simulations in conjunction with traditional defect models, the proposed methods achieves a high defect coverage of over 97.5% with minimal (< 0.1%) area overhead.
Gas Chromatography - Ion Mobility Spectrometry (GC-IMS) is a powerful analytical technique for separating and identifying chemical compounds. The resulting data is represented as a matrix image, where peak locations correspond to specific compounds that must be accurately detected, isolated, and correlated with the concentrations present in the sample. This work presents a tailored approach for peak detection in GC-IMS data, with a focus on tracking concentration variations. The methodology is designed for relative short columns and applied across diverse solution matrices. Short-column GC-IMS is particularly valuable for rapid detection applications, such as the identification of chemical warfare agents (CWAs). However, its reduced separation efficiency leads to broader peaks and increased signal overlap, making peak differentiation more challenging, especially in complex matrices. To address these limitations, we employ signal processing techniques to preprocess data, define regions of interest (ROIs), and apply localized peak detection, enhancing both sensitivity and accuracy. As a practical application, we detect tributylphosphate (TBP), a toxic industrial chemical (TIC) commonly used as a simulant for nerve agents, across various matrices, including diesel, oil, solvent, sand, and mixtures containing other hazardous chemicals. A short GC column is used to reduce acquisition time and minimize exposure to hazardous substances, albeit at the cost of resolution. The results demonstrate the effectiveness of the proposed method, successfully detecting, distinguishing, and quantifying peaks in the acquired data.
The oil-gas-water three-phase flow state is an important factor affecting the safety and stability of the oil and gas industry production process. However, due to the complexity, variability and strong coupling of three-phase flow, there is no universal definition and no common description for the three-phase flow states. In order to realize on-line monitoring of the flow states, an attribute causality-guided state monitoring model is proposed in this work. Firstly, the state attributes are designed for the flow process description. Then, the linear discriminant analysis and transfer entropy (LDA-TE) are combined to excavate the causal relationship between flow state attributes, resulting in a hierarchical attribute framework that describes the different flow states in terms of causal connections. Subsequently, the convolutional neural network (CNN) backbone is used to extract attribute features with an attribute guiding mechanism, which can integrate cause attribute information into effect attribute recognition and transfer knowledge based on attribute causality. Finally, the attribute heat map gives the monitoring results of the flow state evolution process and the identification results of typical flow states. The experiments show the feasibility and superiority of the proposed method.
Industries collect and analyze sensor data from multiple sources to enhance operational processes in Industry 4.0. Evaluating data quality examines sensor data for various quality categories to gain extended insights into the internal and external environments being sensed. Sensor data consistency in industrial systems ensures that various types of sensors monitoring the same event provide mutually consistent readings. Inconsistencies in sensor data can indicate sensor malfunctions or anomalies within monitored processes. We investigated unsupervised learning techniques for event detection and clustering algorithms to assess process events. This article evaluates sensor data consistency in the context of multi-sensor diaphragm valve event monitoring in the pharmaceutical industry. Multiple heterogeneous sensor systems for pressure, acceleration, and an event camera are installed to monitor a single valve. Process events are detected using autoencoder and K-means clustering models, identifying sustained and transient events in sensor readings. A consistency score is calculated for each sensor by comparing event-state signals across all sensors monitoring the same valve cycle of operation. This score quantifies how closely a sensor’s data aligns with that of others during the same events. Experimental results demonstrate that our methods effectively identify sensor discrepancies and offer a practical solution for detecting multiple sensor data consistencies.
Electrical Impedance Tomography (EIT) is a non-invasive imaging technique for volumetric visualization of objects in an electrically conductive environment. This paper presents a multifunctional, DSP-based impedance measurement device designed primarily for physiological applications, including electrical bioimpedance (EBI), with additional support for electrocardiography (ECG) and photoplethysmography (PPG) inputs. The device features one excitation channel and two parallel differential sensing channels, supporting up to 32 multiplexed electrodes. Operating across a frequency range of 1 kHz to 200 kHz, it provides both magnitude and phase impedance measurements with a refresh rate of 2 frames per second. The DSP-based implementation allows for high-resolution measurements using a built-in 16-bit ADC and supports implementing of multifrequency measurement modes. This capability is particularly advantageous for addressing challenges in EIT imaging, such as frequency difference imaging. The device was tested using Python-based PyEIT software and effective imaging was demonstrated in various experimental setups. Results show the potential of the proposed device to produce high-contrast EIT images while minimizing background interference. Overall, the developed system offers a portable, versatile, and efficient solution for EIT applications in biomedical and industrial domains.
In this work, we aim at realizing an eye tracker functionality on an off-the-shelf low cost web-camera. In particular, we address the problem of appearance gaze estimation employing deep learning methods for solving the regression task of predicting the point of gaze on a display. This solution is validated using a commercial infrared eye tracker: we achieve 73.843 Root Mean Square Error (RMSE) in pixels between the predictions. The proposed solution outperforms a relevant web-cam solution in terms of pixel-normalized RMSE 0.022 against 0.116 within the carried out comparative study, respectively.
In modern intelligent manufacturing systems, fault diagnosis is of great significance to ensure the safe operation of equipment. As the complexity of equipment increases, various components are tightly coupled, leading to occurrence of compound faults either between components or within the same component. This study proposes a compound fault diagnosis method based on health state feature decoupling. Firstly, the input signals are decomposed into normal and single fault-related components in the feature space. Meanwhile, the distance between different health state components of the signal is maximized by orthogonality constraints to enhance the independence of each other. Then, the feature decoupling is achieved by minimizing the distribution discrepancy between the same health state features of signals from different categories. During the decoupling process, signal reconstruction based on the sum of feature components is utilized to ensure the decoupled features are meaningful. Finally, a classifier is used to achieve accurate compound fault diagnosis. Experimental results verify the superiority of the proposed method on compound fault diagnosis.
Detection of physical effort through the analysis of physiological parameters is a challenging endeavor, due to the intrinsically complex nature of the physiological condition of fatigue status. Indeed, the possibility to categorize physical effort is critical to reduce the associated risks, particularly among people exposed to heavy workload. On the other hand, as the effects of physical effort on physiological signals are not univocal, data labeling commonly applied in machine learning is not always possible, and unsupervised methods must be investigated. This paper presents a method for unsupervised learning of physical effort. The method is based on a metric exploiting two physiological signals that can be easily acquired by wearable devices, namely, heart rate and skin conductance. Such metric has the same unit of measurement of heart rate, i.e., the beats per minute, but it is characterized by greater information content related to effort levels. Then, the experimental analysis is carried out on signals acquired before, during and after the execution of physical activity. The Davies-Bouldin index, chosen as figure of merit for performance analysis, shows that the proposed method is able to identify the proper number of clusters, corresponding to the different levels of physical effort.
This paper reports the electrical characterization of the sensing properties of nanostructured Nb2O5 films deposited by Pulsed Laser Deposition ( PLD) towards ethanol gas. In particular, a sensing thin film was deposited by 2000 pulses of Nd:YAG UV laser (operating at lambda = 266 nm and energy of 112 mJ) obtaining a porous layer composed by vertical nano-pillars with a thickness of about 500 nm. The sensing film was characterized in a temperature operating range from 340 degrees C to 450 degrees C, identifying the optimal working temperature at 420 degrees C. Sensing response toward ethanol was determined in a concentration range from 50 ppm to 100 ppm. Response repeatability, stability, reversibility and recovery times were tested as well. The preliminary characterization highlighted the very promising sensing performance of PLD niobium oxide thin films. Therefore, the investigation of ethanol gas sensors based on PLD sensing layers grown in different conditions to tune nanostructure and thickness, is in progress in order to further increase the sensor sensitivity and selectivity.
Deep learning (DL) models are increasingly used in industrial applications for precise object measurement, particularly in resource-constrained environments where monocular cameras are favored for their cost-effectiveness. However, the traditional evaluation metrics, such as mean Average Precision (mAP) and validation loss, cannot accurately capture measurement task’s accuracy and reliability. This study covers the use of Mean Absolute Error (MAE) and Standard Deviation (Std Dev) as specialized metrics to evaluate the trueness and consistency of DL models in measurement applications. We trained two DL based frameworks with different sets of hyperparameter configurations and assessed their performance using a dataset of images captured under identical conditions. The results show no significant correlation between mAP and the proposed metrics of MAE and Std Dev, further indicating the deficiency of the conventional metrics for measurement quality assessment. Instead, a positive linear relationship between MAE and Std Dev was recorded. Our analysis shows a strong correlation between MAE and Std Dev (0.92 for Detectron2 and 0.93 for YOLO). The random forest algorithm confirmed MAE and Std Dev as the most important feature (0.89 for Detectron2, 0.78 for YOLO), while validation loss and mAP had lower significance in feature importance and correlation. This work highlights the importance of incorporating statistical metrics into evaluating DL models to ensure the selection of configurations that deliver reliable, accurate, and efficient camera-based measurements.
