Sumeet Kumar

Urban Street Lighting Infrastructure Monitoring Using a Mobile Sensor Platform

We present a system for collecting and analyzing information on street lighting infrastructure. We develop a car-mounted sensor platform that enables collection and logging of data on street lights during night-time drive-bys. We address several signal processing problems that are key to mapping street illumination levels, identifying street lamps, estimating their heights, and geotagging them. Specifically, we highlight an image recognition algorithm to identify street lamps from the video data collected by the sensor platform and its subsequent use in estimating the heights of street lamps. We also outline a framework to improve vehicle location estimates by combining sensor observations in an extended Kalman filter framework. Our eventual goal is to develop a semi-live virtual 3-D street lighting model at urban scale that enables citizens and decision makers to assess and optimize performance of nighttime street lighting.

Infrastructure for data-driven agriculture: identifying management zones for cotton using statistical modeling and machine learning techniques

Advances in many areas of sensing technologies and the widespread use and greater accuracy of global positioning systems offer the prospect of improving agricultural productivity through the intensive use of data. By nature, agriculture is a spatial science characterized by significant variability in terms of yield and concentration of pests and plant diseases. Consequently, precision agriculture seeks to improve the effectiveness of various types of sensing information to give the grower more data and the ability to design the specific treatments for site-specific management of inputs and outputs. The intensive use of data in agriculture is at a relatively early stage and there remains much opportunity to refine modeling approaches and to build information infrastructure. With the overall goal of optimizing inputs to achieve the maximum output in terms of yield, this paper focuses on the application of a clustering algorithm to field data with the goal to identify management zones. We employ two sets of attributes, first yield and second field properties like slope and electrical conductivity to delineate the management zones. By definition, a management zone is a contiguous area defined by one or more features and may take on many different shapes. Building on the established machine learning approach of k-means clustering, we successfully identify a near optimal number of management zones for a cotton field.

Thermal Modeling for Design Optimization of a Microfluidic Device for Continuous Flow Polymerase Chain Reaction (PCR)

Polymerase Chain Reaction (PCR) is a molecular biological method for the in vitro amplification of nucleic acid molecules which has wide applications in the area of genetics, medicine and biochemistry. The typical three step PCR cycle consists of heating the sample to 90–94 °C to denature double-stranded DNA, cooling down to 50–54 °C to anneal the specific primers to the single stranded DNA and finally increasing the temperature to 70–75 °C for extension of the primers with thermostable DNA polymerase. The temperature sensitivity of the reaction requires precise temperature control and proper thermal isolation of these three zones. In this paper we present the design of a continuous flow PCR microfluidic device with the channels fabricated in (poly) dimethylsiloxane (PDMS) and thin film Platinum Resistance Temperature Detector (RTD) elements fabricated on glass substrate to define the three different temperature zones. The fluidic arrangement has a water jacket layer to minimize evaporation from the porous PDMS walls. A detailed thermo fluidic model of the device is presented to predict the performance and efficacy of the proposed design. Numerical simulations are carried out to find the temperature distribution and temperature gradients in the device and a parametric study is done by varying flow rate, heat flux and channel dimensions in order to optimize the design for achieving temperature isolation and sharp temperature gradients between different zones.© 2008 ASME

Air filter particulate loading detection using smartphone audio and optimized ensemble classification

Abstract Automotive engine intake filters ensure clean air delivery to the engine, though over time these filters load with contaminants hindering free airflow. Today’s open-loop approach to air filter maintenance has drivers replace elements at predetermined service intervals, causing costly and potentially harmful over- and under-replacement. The result is that many vehicles consistently operate with reduced power, increased fuel consumption, or excessive particulate-related wear which may harm the catalyst or damage machined engine surfaces. We present a method of detecting filter contaminant loading from audio data collected by a smartphone and a stand microphone. Our machine learning approach to filter supervision uses Mel-Cepstrum, Fourier and Wavelet features as input into a classification model and applies feature ranking to select the best-differentiating features. We demonstrate the robustness of our technique by showing its efficacy for two vehicle types and different microphones, finding a best result of 79.7% accuracy when classifying a filter into three loading states. Refinements to this technique will help drivers supervise their filters and aid in optimally timing their replacement. This will result in an improvement in vehicle performance, efficiency, and reliability, while reducing the cost of maintenance to vehicle owners.

Engine Misfire Detection with Pervasive Mobile Audio

We address the problem of detecting whether an engine is misfiring by using machine learning techniques on transformed audio data collected from a smartphone. We recorded audio samples in an uncontrolled environment and extracted Fourier, Wavelet and Mel-frequency Cepstrum features from normal and abnormal engines. We then implemented Fisher Score and Relief Score based variable ranking to obtain an informative reduced feature set for training and testing classification algorithms. Using this feature set, we were able to obtain a model accuracy of over 99 % using a linear SVM applied to outsample data. This application of machine learning to vehicle subsystem monitoring simplifies traditional engine diagnostics, aiding vehicle owners in the maintenance process and opening up new avenues for pervasive mobile sensing and automotive diagnostics.

Efficient Parametric Signal Estimation From Samples With Location Errors

We introduce an iterative linear estimator (ILE) for estimating a signal from samples having location errors and additive noise. We assume that the signals lie in the span of a finite basis and the location errors and noise are mutually independent and normally distributed. The parameter estimation problem is formulated as obtaining a maximum likelihood (ML) estimate given the observations and an observation model. Using a linearized observation model we derive an approximation to the likelihood function. We then adopt an iterative strategy to develop a computationally efficient estimator, which captures the first order effect of sample location errors on signal estimation. Through numerical simulations we establish the efficacy of the proposed estimator for one-dimensional and two-dimensional parametric signals, comparing the mean squared estimation error against a basic linear estimator. We develop a numerical approximation of the Cramér-Rao lower bound (CRB) and the Expectation-Maximization (EM) algorithm, and for a one-dimensional signal compare our algorithm against them. We show that for high location error variance and small noise variance the mean squared error (MSE) with ILE is significantly lower when compared to the baseline linear estimator. When compared to EM, our algorithm provides comparable MSE with a significant reduction in computational time.

Size matters: Why cell size is vital for minimizing cost of air-cooling in battery packs

There is wide disagreement in the automotive industry over what size cells are ideal for battery packs. Because of the inherent complexity of packs and the diversity of applications, current lines of thought are loosely opinion based. In order to start quantitatively examining the problem, this study measures the performance of a range of cylindrical cell sizes in a generic battery module (modeled in COMSOL) and estimates the cost of a fan for air-cooling. The results reveal that cell size has a large impact on the cost of cooling, and a clear minimum is reached.

Stable arrangements of mobile sensors for sampling physical fields

Todays wireless sensor nodes can be easily attached to mobile platforms such as robots, cars and cell phones enabling pervasive sensing of physical fields (say of temperature, vibrations, air quality and chemicals). We address the sensor arrangement problem, i.e. when and where sensors should take samples to obtain a good estimate of a field using mobile sensors. In particular, we focus on incidentally mobile sensors that move passively under the influence of the environment (e.g. sensors attached to floating buoys, cars and smartphones carried by humans). We model the field as a linear combination of known basis functions. Given the samples, we use a linear estimator to find unknown coefficients of the basis functions. We formulate the sensor arrangement problem as one of finding suitably characterized classes of sensor arrangements that lead to a stable reconstruction of the field. We consider a family of multidimensional δ-dense sensor arrangements, where any square disc of size δ contains at least one sample, and derive sufficiency conditions for the arrangement to be stable. δ-dense sensor arrangements are geometrically intuitive and are easily compatible with the incidental mobility of sensors in many situations. We present simulation results on the stability of such arrangements for two-dimensional basis functions. We also present an example for constructing basis functions through proper orthogonal decompositions for a one-dimensional chemical diffusion field in a heterogeneous medium, which are later used for field estimation through δ-dense sampling.

The Future Internet of Things: Secure, Efficient, and Model-Based

The Internet of Things’ (IoT’s) rapid growth is constrained by resource use and fears about privacy and security. A solution jointly addressing security, efficiency, privacy, and scalability is needed to support continued expansion. We propose a solution modeled on human use of context and cognition, leveraging cloud resources to facilitate IoT on constrained devices. We present an architecture applying process knowledge to provide security through abstraction and privacy through remote data fusion. We outline five architectural elements and consider the key concepts of the “data proxy” and the “cognitive layer.” The data proxy uses system models to digitally mirror objects with minimal input data, while the cognitive layer applies these models to monitor the system’s evolution and to simulate the impact of commands prior to execution. The data proxy allows a system’s sensors to be sampled to meet a specified quality of data target with minimal resource use. The efficiency improvement of this architecture is shown with an example vehicle tracking application. Finally, we consider future opportunities for this architecture to reduce technical, economic, and sentiment barriers to the adoption of the IoT.

Towards low-cost resolution optimized passive UHF RFID light sensing

This paper presents a low-cost, fully passive UHF RFID light sensor. We demonstrate how our design makes use of off-the-shelf photoresistors and LC components to establish a power frequency dependence which is used for improved sensor precision. Preliminary results show we can successfully identify 4 light intensity states between 0 to 1000 lux.