With everyday growth of Internet of Things (IoT), number of connected sensor devices increases as well, where each sensor consumes energy while being constantly online. During that time they collect large amounts of data in short intervals leading to the collection of redundant and perhaps irrelevant data. Moreover, being commonly battery powered, sensor batteries need to be frequently replaced or recharged. Former requires smarter and less frequent data collection, while latter being complementary to the former requires putting them to sleep while not being used in order to save energy. The focus of this research are low cost gas sensors as they need to preheat for several minutes to reliably collect gas concentration. However, instead of waiting for a sensor to heat up, a transient, i.e., a data trend that the sensor collects while heating up is analyzed. It is shown that Long Short-Term Memory (LSTM) Neural Network can be used to learn and later predict actual gas level from a part of the transient. This way, instead of being constantly online or fully preheating, the sensor needs to be turned on for only 20 seconds and then sleep for 120 seconds. With high accuracy our approach decreases energy consumption up to 85% compared to a system where sensors are constantly online, and more than 50% compared to a system where a sensor collects actual values instead of a part of the transient.
Predictive models of bathing water quality are a useful support to traditional monitoring and provide timely and adequate information for the protection of public health. When developing models, it is critical to select an appropriate model type and appropriate metrics to reduce errors so that the predicted outcome is reliable. It is usually necessary to conduct intensive sampling to collect a sufficient amount of data. This paper presents the process of developing a predictive model in Kaštela Bay (Adriatic Sea) using only data from regular (official) bathing water quality monitoring collected during five bathing seasons. The predictive modelling process, which included data preprocessing, model training, and model tuning, showed no silver bullet model and selected two model types that met the specified requirements: a neural network (ANN) for Escherichia coli and a random forest (RF) for intestinal enterococci. The different model types are probably the result of the different persistence of two indicator bacteria to the effects of marine environmental factors and consequently the different die-off rates. By combining these two models, the bathing water samples were classified with acceptable performances, an informedness of 71.7%, an F-score of 47.1%, and an overall accuracy of 80.6%.
Remote sensing is commonly performed via airborne platforms such as satellites, specialized aircraft, and unmanned aerial systems (UASs), which perform airborne photography using mounted cameras. However, they are limited by their coverage (UASs), irregular flyover frequency (aircraft), and/or low spatial resolution (satellites) due to their high altitude. In this paper, we examine the utilization of commercial flights as an airborne platform for remote sensing. Namely, we simulate a situation where all aircraft on commercial flights are equipped with a mounted camera used for airborne photography. The simulation is used to estimate coverage, the temporal and spatial resolution of aerial imagery acquired this way, as well as the storage capacity required for storing all imagery data. The results show that Europe is 83.28 percent covered with an average of one aerial photography every half an hour and a ground sampling distance of 0.96 meters per pixel. Capturing such imagery results in 20 million images or four petabytes of image data per day. More detailed results are given in the paper for separate countries/territories in Europe, individual commercial airlines and alliances, as well as three different cameras.
16 September 2022 in Journal of Computational Science
With the increase in number and size of Internet of Things systems, there is an ever-growing risk of (meta)data loss, as well as the maintenance overhead to mitigate such risks. The experts recognize three main challenges in this area that need to be tackled, namely (1) downsizing the manual work required for configuring sensor networks, (2) recovering metadata, such as sensor type, in case of connection issues, malfunctions or malicious actions in sensor networks, (3) rebuilding metadata lost due to unexpected problems within a data storage. Fortunately, all three challenges can be tackled with a uniform solution, namely the signal type classification approach, which is able to match raw signal to an appropriate data type. In this research, we evaluate and compare different approaches for signal type classification that can be used to recognize a signal type being read from an IoT sensor. This is done by using machine learning methods for modelling a signal represented as raw time series data. Three machine learning classification approaches are taken into a consideration, namely one class, two class and multi-class. According to the results of the evaluation, the most accurate multi-class random forest algorithm can correctly classify unknown signals in $$\sim {75}\%$$ of the cases based on only 20 consecutive sensor readings. Moreover, multi-class random forest can detect two most probable classes of monitored signal with the accuracy of 95%.
1 January 2023 in Journal of Communications Software and Systems
Various distance-based clustering algorithms have been reported, but the core component of all of them is a similarity or distance measure for classification of data. Rather than setting the priority to comparison of the performance of different clustering algorithms, it may be worthy to analyze the influence of different similarity measures on the results of clustering algorithms. The main contribution of this work is a comparative study of the impact of 9 similarity measures on similarity-based trajectory clustering using DBSCAN algorithm for commercial flight dataset. The novelty in this comparison is exploring the robustness of the clustering algorithm with respect to algorithm parameter. We evaluate the accuracy of clustering, accuracy of anomaly detection, algorithmic efficiency, and we determine the behavior profile for each measure. We show that DTW and Frechet distance lead to the best clustering results, while LCSS and Hausdorff Cosine should be avoided for this task.