This paper introduces a non-intrusive privacy-preserving method for detecting people's presence and movement patterns. This approach tracks WiFi-enabled personal devices carried by individuals, leveraging network management messages to associate those devices with available networks. Randomization protocols are implemented in network management messages, a necessary measure to protect privacy. This prevents identification based on elements like device addresses, message sequence numbers, the data fields, and the total data content. Consequently, a novel de-randomization approach was presented, identifying individual devices by clustering comparable network management messages and their correlated radio channel attributes using a novel matching and grouping algorithm. The proposed approach began with calibrating it using a publicly available labeled dataset, confirming its accuracy through controlled rural and semi-controlled indoor measurements, and finally assessing its scalability and accuracy in an uncontrolled, densely populated urban setting. The proposed de-randomization method demonstrates over 96% accuracy in identifying devices from both the rural and indoor datasets, with each device type validated individually. Accuracy of the method diminishes when devices are grouped, though it surpasses 70% in rural areas and 80% indoors. The urban environment's people movement and presence analysis, using a non-intrusive, low-cost solution, confirmed its accuracy, scalability, and robustness via a final verification, including the generation of clustered data useful for analyzing individual movements. this website In spite of its strengths, the process revealed inherent limitations regarding exponential computational complexity and precise parameter determination and fine-tuning, requiring significant efforts toward optimization and automation.
Using open-source AutoML tools and statistical methods, this paper presents a novel approach to robustly predict tomato yield. Five selected vegetation indices (VIs) were acquired from Sentinel-2 satellite imagery over the 2021 growing season (April-September), with data points taken every five days. Evaluating Vis's performance across different temporal dimensions, 108 fields, covering a total of 41,010 hectares of processing tomatoes in central Greece, had their actual yields recorded. Beside this, the crop's visual indexes were associated with crop phenology to define the yearly progression of the crop. Vegetation indices (VIs) exhibited a powerful relationship with yield, as demonstrated by the peak Pearson correlation coefficients (r) within the 80-90 day period. The growing season's correlation analysis shows the strongest results for RVI, attaining values of 0.72 at 80 days and 0.75 at 90 days, with NDVI achieving a comparable result of 0.72 at 85 days. The AutoML method affirmed this output and concurrently identified the greatest performance by VIs within the same timeframe. Adjusted R-squared values were observed to fluctuate between 0.60 and 0.72. Employing the synergistic combination of ARD regression and SVR led to the most precise results, showcasing its superiority for ensemble construction. R-squared, a measure of goodness of fit, equated to 0.067002.
The state-of-health (SOH) of a battery evaluates its capacity relative to its specified rated capacity. Despite the creation of numerous algorithms using data to estimate battery state of health (SOH), they often encounter difficulties with time series data, as they fail to fully capitalize on the valuable information within the sequence. In addition, algorithms fueled by data frequently fail to develop a health index, a metric assessing battery condition, thereby neglecting capacity deterioration and enhancement. To tackle these problems, we introduce a model optimized to compute a battery's health index, meticulously portraying the battery's degradation trend and improving the accuracy of predicting its State of Health. In addition, a deep learning algorithm employing attention mechanisms is introduced. This algorithm constructs an attention matrix that reflects the relative significance of data points within a time series. This empowers the predictive model to prioritize the most important segments of the time series when estimating SOH. Our numerical evaluation of the algorithm confirms its effectiveness in establishing a reliable health index, and its ability to precisely predict battery state of health.
Hexagonal grid layouts, while advantageous in microarray technology, appear in various fields, particularly with the ongoing development of novel nanostructures and metamaterials, making image analysis of these patterns an indispensable aspect of research. This research presents a shock-filter-based method, leveraging mathematical morphology, for the segmentation of image objects within a hexagonal grid arrangement. The original image is disassembled into a pair of rectangular grids; their superposition results in the original image's formation. For each image object's foreground information within each rectangular grid, the shock-filters serve to focus it into a particular area of interest. The proposed methodology's successful application to microarray spot segmentation is highlighted, underscored by its general applicability in two additional hexagonal grid layouts. The proposed approach for microarray image analysis demonstrated high reliability, as indicated by strong correlations between computed spot intensity features and annotated reference values, evaluated using quality measures including mean absolute error and coefficient of variation in segmentation accuracy. The computational complexity of determining the grid is minimized by applying the shock-filter PDE formalism to the one-dimensional luminance profile function. Compared to leading-edge microarray segmentation methods, from traditional to machine learning-based ones, the computational complexity of our approach demonstrates a growth rate that is at least one order of magnitude smaller.
Industrial applications frequently utilize induction motors, owing to their robustness and affordability. Industrial operations can halt, unfortunately, due to the nature of induction motors and their potential for failure. this website Accordingly, further research is essential for achieving swift and precise fault detection in induction motors. This research involved the creation of an induction motor simulator, which could be used to simulate both normal and faulty operations, encompassing rotor and bearing failures. A total of 1240 vibration datasets, each containing 1024 data samples, were ascertained for each state using this simulator. Subsequently, support vector machine, multilayer neural network, convolutional neural network, gradient boosting machine, and XGBoost machine learning models were applied to diagnose failures from the gathered data. Cross-validation, using a stratified K-fold approach, confirmed the diagnostic precision and calculation rapidity of these models. Along with the fault diagnosis technique, a user-friendly graphical interface was developed and incorporated. The experimental evaluation demonstrates that the proposed approach is fit for diagnosing faults within the induction motor system.
Considering the influence of bee activity on the health of the hive and the increasing presence of electromagnetic radiation in the urban landscape, we analyze ambient electromagnetic radiation as a possible predictor of bee traffic near hives in a city environment. At a private apiary in Logan, Utah, two multi-sensor stations were deployed for 4.5 months to meticulously document ambient weather conditions and electromagnetic radiation levels. Omnidirectional bee motion counts were extracted from video recordings taken by two non-invasive video loggers, which were placed on two hives located at the apiary. Employing time-aligned datasets, 200 linear and 3703,200 non-linear regressors (random forest and support vector machine) were assessed to forecast bee motion counts based on time, weather, and electromagnetic radiation. Across all regression analyses, electromagnetic radiation demonstrated predictive ability for traffic volume equivalent to that of weather patterns. this website Electromagnetic radiation and weather patterns, in contrast to mere time, were more accurate predictors. Analyzing the 13412 time-stamped weather data, electromagnetic radiation readings, and bee activity logs, random forest regression models demonstrated superior maximum R-squared values and more energy-efficient optimized grid searches. Both types of regressors were reliable numerically.
Passive Human Sensing (PHS) allows for unobtrusive monitoring of human presence, movement, and activities without demanding any equipment from the monitored individuals. PHS is frequently documented in the literature as a method which capitalizes on variations in channel state information of a dedicated WiFi network, where human bodies affect the trajectory of the signal's propagation. The implementation of WiFi in PHS networks unfortunately encounters drawbacks related to power consumption, the substantial costs associated with extensive deployments, and the possibility of interference with other networks operating in close proximity. Bluetooth Low Energy (BLE), a refinement of Bluetooth, provides a compelling solution to WiFi's drawbacks, its Adaptive Frequency Hopping (AFH) method being particularly effective. Employing a Deep Convolutional Neural Network (DNN) to enhance the analysis and classification of BLE signal distortions in PHS using standard commercial BLE devices is the subject of this work. The suggested approach was implemented to ascertain the presence of human inhabitants in a large, complex space with minimal transmitters and receivers, under the stipulated condition that occupants did not interrupt the direct line of sight between devices. The experimental findings confirm that the proposed approach yields a significantly superior outcome compared to the most accurate technique identified in the literature, when tested on the same data.