Finding Representative Sampling Subsets in Sensor Graphs Using Time-series Similarities

One of the greatest challenges in many machine learning applications is to build robust classification models that reduce calibration efforts, ensure model transferability to unseen data distributions, and handle high semantic resolution datasets with a relatively large number of target classes. We present an operational domain adaptation framework for crop type verification from remote sensing data for integrated administration and control and farm management systems. It is part of the ag|knowledge crop monitoring platform and demonstrates a more robust spatiotemporal model transferability than conventional supervised crop type classification approaches, which use multitemporal and multispectral features to classify the specific characteristics of crop phenology patterns. The proposed method is based on a machine learning model that has been trained on particular similarity metrics between all target crop types. The metrics are derived by quantifying the similarity of vegetation index time series between the observed parcel and the aggregated time series of reference objects for each parcel and crop type. This method achieves higher classification results than a model trained on the pure remote sensing time series data, as the algorithm does not learn the temporal vegetation index patterns, but instead, the adversarial characteristics and the differences between each crop type. Moreover, it also reduces dimensionality as the time series are summarized through the respective similarity metrics. We compare both approaches through different classification algorithms. Overall, the achieved classification accuracy for more than 67 crop types scores more than 80%.

Due to technology advances and complexity of designs, thermal issue is a bottleneck in electronics designs. Various dynamic thermal management techniques have been proposed to address this issue. To effectively apply thermal management techniques, providing an accurate thermal map of chips is highly required. For this goal, a network of temperature sensors ought to be provided. There are various implementations for temperature sensors and network of sensors on Field Programmable Gate Arrays (FPGAs). This work defines and formulates four metrics and criteria, in terms of area, thermal, and power overheads and thermal map accuracy for exploring and evaluating efficiency of different implementations of Ring Oscillator-based Temperature Sensor (ROTS) networks on FPGAs and reports the comparison results for 12 networks with various sensor configurations. According to our metrics and experiments, the sensor that it is composed of NOT gates with open latches and RNS ring counter has lower thermal and power overheads compared to other configurations. Moreover, in this work, a new ROTS is presented that occupies 25% less resources than the most compact temperature sensor. Also, it provides 1.72 times higher sensitivity than the best sensitive ROTS design.

Virtual sensing is a technique in which sensor values in a sensor network are estimated some of the time instead of taking direct measurements. Virtual sensing conserves energy on the sensor nodes, and it can be used for recovering missing data due to node or network failure, and for estimating values where permanent sensor installation is in-feasible. This paper considers the novel and challenging problem of designing an adaptive long range virtual sensing (LRVS) framework that is able to estimate values for days or months. The framework does not require any a priori modeling of generating functions for the sensor time series. A new correlation metric, called nearest hour of the day neighbor, finds similarity in multiple time series that is able to arbitrate the hidden contexts within the data. LRVS is evaluated on data from a building-monitoring sensor network with over 100 heterogeneous, noisy sensors from different suppliers.

Energy efficiency is an important design criterion for the development of sensor networking protocols involving data dissemination and gathering. In-network processing of sensor data, aggregation, transmission power control in radios, and periodic cycling of node wake-up schedules are some techniques that have been roposed in the sensor networking literature for achieving energy efficiency. Owing to the broadcast nature of the wireless channel many nodes in the vicinity of a sender node may overhear its packet transmissions even if they are not the intended reci ients of these transmissions. Reception of these transmissions can result in unnecessary expenditure of battery energy of the recipients. In this paper,we investigate the impact of overhearing transmissions on total energy costs during data gathering and dissemination and attempt to minimize them systematically.We model the minimum energy data gathering problem as a directed minimum energy spanning tree problem where the energy cost of each edge in the wireless connectivity graph is augmented by the overhearing cost of the corresponding transmission. We observe that in dense sensor networks, overhearing costs constitute a significant fraction of the total energy cost and that computing the minimum spanning tree on the augmented cost metric results in energy savings, especially in networks with non-uniform s atial node distribution. We also study the impact of this new metric on the well known energy-efficient dissemination (also called broadcasting) algorithms for multihop wireless networks.We show via simulation that through this augmented cost metric, gains in energy efficiency of 10%or more are ossible without additional hardware and minimal additional complexity.

Few-shot palmprint recognition based on similarity metric hashing network

In a wireless sensor network, a large number of small-size cheap sensor nodes are scattered, and have limited resources such as battery power, data processing capability and memory capacity. They should show fault-tolerance, energy efficiency, and topological adaptation to the change of external environment, therefore, an effective routing algorithm is essential. In general, if a network is large in scale, multi-hop transmission is more energy efficient than direct transmission. In case a large number of nodes are deployed densely within a fixed 1-hop communication range, however, many nodes participate in data transmission, multi-hop transmission consumes much more energy in such a case. Thus, this study proposes a routing algorithm that solves above problems at the same time in a dense and large scale sensor network. The result of experiment showed that routing algorithm proposed in this study has reasonable fault-tolerance, energy efficiency, and offers its adaptability to topological changes in a dense and large scale sensor network.

Recent advances in low-power sensing devices coupled with the widespread availability of wireless ad-hoc networks have fueled the development of sensor networks. These are typically deployed over wide areas to gather data in the environment and monitor events of interest. The ability to run spatial queries is extremely useful for sensor networks. Spatial query execution has been extensively studied in the context of centralized spatial databases; however because of the energy and bandwidth limitation of sensor nodes these solutions are not directly applicable to the sensor network. In this paper we propose a scalable and distributed way of spatial query execution in sensor networks. We develop a distributed spatial index over the sensor nodes that is used in processing spatial queries in a distributed fashion. We evaluate the behavior of our approach and show that our mechanism provides an efficient and scalable way to run spatial queries over sparse and dense sensor networks.

Wireless sensor networks have found many applications because of their facility for remote information monitoring and processing. One of the main applications of sensor networks is detection of an event occurrence. In these applications, a fusion center decides about an event occurrence based on the data of sensors. Here, the goal is to detect event occurrence correctly while avoiding its false detections. To reach this, either a sparse network of expensive precise sensors or a dense network of cheap sensors may be used. This paper attempts to examine these cases. It is shown that using a relatively dense network of cheap sensors may satisfy design parameters. In addition, it is shown that how a suitable decision fusion rule can reduce design costs.

With the recent development of low-cost sensors and the Internet of Things (IoT), air quality monitoring is facing new opportunities and challenges. This study examines the effect of sensor quality, quantity, and wind climatology on the accuracy of characterizing airborne emission sources for sensor network data. We consider a simulation domain with a point-source emission and virtual sensors evenly placed around the facility fence-line. A source characterization framework is introduced that consists of three parts: plume dispersion, sensor response, and Bayesian source inference. Two types of real-time sensors are considered as candidate sensors: point-scale sensors (PS) and line-integrated real-time sensors (LS). A parsimonious set of parameters, including sensor bias, noise, and detection limit, are used to describe sensor quality. Wind data are obtained from multiple geographic settings to represent typical wind conditions from inland and coastal regions. Overall, we find that LS offer superior performance of source characterization than PS, mainly due to their extended spatial coverage that leads to a large observational footprint. However, PS have some advantages over the LS for field deployment and cost. It was also found that the persistence and uncertainty of wind direction are critical for source characterization. Persistence of wind direction is shown to be negatively correlated with the performance of source characterization, such that places with more broadly varying wind direction can better utilize sensors on the fence-line, thus improving source characterization. However, uncertainties in wind direction data can severely affect source characterization, especially for the estimation of emission strength. Finally, it is shown that a trade-off between sensor quality and quantity exists under certain circumstances, such that a dense sensor network with relatively low-quality sensors (large sensor noise or low sensitivity) could outperform a sparse sensor network with high-quality sensors (small sensor noise or high sensitivity) for source characterization. This finding suggests that a relatively low-quality sensor network has the potential of being a cost-effective method for long-term air quality monitoring.

Frost protection is a significant concern among tree-fruit growers. In order to protect orchards from frost, temperature sensor networks have gained popularity to judge when and where to turn on wind generators and sprinklers. Currently these sensor networks only give growers a real-time alert, however such sensor networks would be more effective if they were integrated in a computation loop that uses past and current temperature data for predicting every overnight transition of orchard air temperature. Although a couple of algorithms using artificial neural network and empirically-formulated polynomials are available to the public, they need to be parallelized for useful on-the-fly prediction. To utilize a cluster of multi-core computing nodes (available through cloud services), we are developing MASS, a library for multi-agent spatial simulation, and parallelizing temperature prediction programs with MASS. This paper demonstrates the MASS library's suitability to parallelization of temperature prediction programs for on-the-fly sensor-data analysis by (1) porting the programs to MASS, (2) running them in a multi-core system, (3) feeding real-time sensor data to them, and (4) measuring their analyzing performance.

Accurate detection of time and location of damage in its early stages has motivated the researchers to develop several damage diagnosis methods in recent decades. With current improvements in monitoring hardware, many of the previously developed Structural Health Monitoring (SHM) methods encounter the BIG DATA problem in processing the data collected through dense sensor networks. Therefore, it is vital to improve the efficiency and scalability of today’s SHM procedures parallel to data acquisition enhancements. Toward this end, this paper presents a data-driven damage detection methodology based on compressed sensing techniques. The objective of the paper is accurate damage localization in a structural component instrumented with a dense sensor network, by processing data only from a subset of sensors. In this method, first a set of sensors from the network are randomly sampled. Measurements from these sampled sensors are processed to extract damage sensitive features. These features undergo statistical change point analysis to establish a new boundary for a local search of damage location. As the local search proceeds, probability of the damage location is estimated through a Bayesian procedure with a bivariate Gaussian likelihood model. The decision boundary and the posterior probability of the damage location are updated as new sensors are added to processing subset and more information about location of damage becomes available. This procedure is continued until enough evidence is collected to infer about damage location. Performance of this method is evaluated using a finite element model of a cracked gusset plate connection. Pre- and post-damage strain distributions in the plate are used for damage diagnosis. doi: 10.12783/SHM2015/279

Network connectivity and link quality information are the fundamental requirements of wireless sensor network protocols to perform their desired functionality. Most of the existing discovery protocols have only focused on the neighbor discovery problem, while a few number of them provide an integrated neighbor search and link estimation. As these protocols require a careful parameter adjustment before network deployment, they cannot provide scalable and accurate network initialization in large-scale dense wireless sensor networks with random topology. Furthermore, performance of these protocols has not entirely been evaluated yet. In this paper, we perform a comprehensive simulation study on the efficiency of employing adaptive protocols compared to the existing nonadaptive protocols for initializing sensor networks with random topology. In this regard, we propose adaptive network initialization protocols which integrate the initial neighbor discovery with link quality estimation process to initialize large-scale dense wireless sensor networks without requiring any parameter adjustment before network deployment. To the best of our knowledge, this work is the first attempt to provide a detailed simulation study on the performance of integrated neighbor discovery and link quality estimation protocols for initializing sensor networks. This study can help system designers to determine the most appropriate approach for different applications.

In this paper, we introduce the concept of building electronic sensate skins as extremely dense, multimodal, systolic sensor networks. In this fashion, the copious signals produced by the skin's receptors are reduced by the network itself, and only high-level features are routed out peer-to-peer, avoiding complex wiring requirements while promising to enable scalability across large areas. Our architectures and algorithms have been inspired by biological skin, where signals from receptors are enhanced or suppressed by processing in the receptor cells and nervous system before arriving at the brain. We illustrate this concept with work in progress on two test beds, the Pushpin Computer, an easily configurable, planar array of over 100 nodes set up as 'smart wallpaper', and the Tribble, a sphere coated with over 500 diverse tactile and noncontact sensor channels processed in 32 interconnected sensor 'cells'. We also introduce the Z-Tiles, a collaborative project that has resulted in a pressure-imaging floor realised through a collection of pixilated floor tiles configured as a mesh sensor network. We conclude with a technology roadmap for scaling to higher densities.

<p>The Starzach site in the upper Neckar valley, Germany, is known for its natural degassing of carbon dioxide (CO₂). During the last century, the gas has been mined industrially until yields stagnated and the extraction wells were sealed. Interestingly, CO₂ exhalation spots have reappeared across the site over the last years. Neither the total emission rate across the site, nor the diurnal or seasonal variations of the CO₂ degassing on the site have been quantified scientifically yet.</p><p>In this project, an autarkic wireless sensor network is developed to monitor the CO₂ concentrations in the Neckar valley continuously and investigate the spatial and temporal variability. In the project's second phase, gas transport simulations with a numerical dispersion model will be used to assess the actual CO₂ emissions into the lower atmosphere. Ultimately, the developed methods may be exported to other regions with similar gas emission phenomena.</p><p>To account for the spatial heterogeneity of the CO₂ outgassing, a dense sensor network is needed. Deploying several dozen stations requires each station to be of reasonable cost. Currently, commercially available, deployable and self-sustaining measurement systems for CO₂ are very expensive. So to facilitate a targeted network with a 10-meter scale mesh size, we developed an infrastructure that meets our requirements. The network’s modular setup permits flexible sensor extension or spatial expansion. A Campbell Scientific IRGASON eddy-covariance station supplements the network as a punctual high-precision reference. Live gathered data are offered to the science community via the publicly accessible OpenSenseMap.org measurement data platform.</p><p>In this talk, the sensor network infrastructure is introduced and the low-cost CO₂ sensor performance is assessed. First long-term measurements at the site reveal a clear diurnal cycle of meteorological parameters and especially for the atmospheric CO₂ concentration and stratification. The site's valley location with surrounding hills is shown to create complex flow dynamics.</p>

A major challenge in 1D 1H NMR-based metabolomics studies is the occurrence of slight shifts of the resonances of mixture compounds compared to the reference spectra in the metabolomics spectral databases due to variations in buffer conditions, temperature, and matrix effects. This hampers both the automated spectral deconvolution and metabolite quantification of crowded regions in 1D 1H NMR spectra of complex mixtures whose analysis is particularly susceptible to such effects. 2D NMR-based metabolomics, on the other hand, is substantially more robust but also much more demanding in terms of NMR spectrometer time. Here we introduce an approach, termed COLMAR1d2d, which uses selected 2D 1H-13C HSQC and 1H-1H TOCSY NMR spectra of a subset of samples along with 1D 1H NMR spectra of all samples to overcome this bottleneck. It relies on our 2D NMR-based platform COLMARm using 2D 1H-13C HSQC and 2D 1H-1H TOCSY spectra measured for a representative subset of samples to unambiguously and comprehensively determine the metabolite composition and the exact peak positions of the identified compounds under the sample conditions present. This information is then used to update the spectral database for the automated analysis of a potentially large cohort of 1D 1H NMR spectra using the COLMAR1d platform. It is demonstrated how this synergistic combination of 1D with selected 2D NMR spectra allows the analysis of a significantly larger number of metabolites than would be possible with 1D NMR alone. Moreover, COLMAR1d2d also improves quantitation, as is demonstrated for samples from mouse urine and Pseudomonas aeruginosa biofilm.