M. P. B. Ekanayake

Residential appliance monitoring based on low frequency smart meter measurements

A Non-Intrusive Load Monitoring (NILM) method for residential appliances based on uncorrelated spectral components of an active power signal is presented. This method utilizes the Karhunen Loéve (KL) expansion to breakdown the active power signal into subspace components so as to construct a unique information rich appliance signature. Unlike existing NILM techniques that rely on multiple measurements at high sampling rates, this method works effectively with a single active power measurement taken at a low sampling rate. After constructing the signature data base, subspace component level power conditions were introduced to reduce the number of possible appliance combinations. Then, an algorithm was presented to identify the turned on appliance combination in a given time window. After identifying the turned on appliance combination, another algorithm was introduced to disaggregate the energy contribution of each individual appliance. The case study conducted using tracebase public data set demonstrate the ability of the proposed method to accurately identify and disaggregate individual energy contributions of turned on appliance combinations that contain single state, multi state and continuous varying appliances. Finally, the proposed method was modified to accommodate usage behavior patterns of each residence adaptively. The modification was validated using six US households in REDD public database. This significantly improves the convergence speed of the turned on appliances identification process.

A subspace signature based approach for residential appliances identification using less informative and low resolution smart meter data

This paper presents a subspace signature based approach for the identification of turned on appliances at a given observation time using one single-function smart meter. The novelty of the proposed approach compared to existing method is its capability for proper identification while relying on a significantly lower amount of measurement data. Unlike existing techniques which rely on multiple measurements at high sampling rate, the proposed method uses only the active power consumption of a grid sampled at second level sampling intervals. This is enabled by constructing signatures for each appliance based on a subspace analysis which incorporates amplitude and frequency information based on active power data. The identification is performed using a Maximum a Posteriori (MAP) criteria using the subspace database that was constructed. This accurately predicts which appliances are turned at any given observation time window. A performance validation is performed for various scenarios to highlight the methods superior accuracy while operating on minimal data at large sampling interval.

Human motion tracking under dynamic background conditions

This paper addresses the specific problem of human event detection from a video sequence in both indoor and outdoor environments. Foreground image pixels are identified through the principle of background subtraction by defining a reference background model using a mixture of time varying Gaussian distributions. Color filtering in the RGB space is then used to remove image distortions due to camera effects and shadowing. A novel approach to tackle the issue of sudden foreground bursts that appear as a result of impulsive environmental changes is also embedded in to the foreground segmentation algorithm. Objects are tracked throughout its presence in the video using an assignment problem based tracker which is capable of handling multiple object interactions such as merges, splits, re-appearances and disappearances. A feature space for each object is constructed and is refined using a Kaiman filter. A fusion of multiple features is used to obtain feature trajectories that closely represent real feature variations of objects. An important aspect of the proposed method is its ability to operate and produce satisfactory results in a scene where there are dynamic background changes and complex inter-human interactions.

Object identification, enhancement and tracking under dynamic background conditions

A real-time event tracking method is proposed that is immune to background variances. The proposed method models each pixel as a collection of Gaussian distributions to handle background variations and uses manipulations in the RGB space to mitigate the effects of foreground shadows. A two stepped connected component analysis method is also introduced in refining the estimated foreground and clustering pixels into silhouettes based on objects. Pixel clusters are formed by filling inter-cluster pixels on the basis of neighborhood solidity of individual pixels. Clustered pixels are defined as object fragments and objects are formed by combining object fragments considering their size and mutual distances. The proposed tracker employs an algorithm to boil down the multiple object interaction problems (objects merging, objects splitting, new object appearance and lost objects) into a simple matrix interpretation problem to construct a consistent feature space. Specifically, splitting of merged objects and temporary disappearances of objects due to occlusions with background objects are handled by the means of a feature correspondence matching. A novel object identification method is proposed for this purpose.

Material based acoustic signal classification - A subspace-based approach

In this paper, approaches to develop subspace-based classifiers for material based signal classification are presented. Material based signal classification is essential in many applications for the identification of the cause of the acoustic signal. Four approaches are presented in this paper for classification. They are namely, classification via eigen-filter characteristics, eigenvector projection method, classification through eigen-filter and eigen-filter bank outputs. For these methods, the optimum order of the eigen-filters is selected based on the maximum true positive counts. The optimal eigenvectors that are most suitable for classification are selected by minimizing the cross-correlation among different classes. Finally, the robustness and the superior performance of the proposed techniques are highlighted by comparing them with time and frequency domain techniques such as AR modeling, spectrogram based classification.

Individual power profile estimation of residential appliances using low frequency smart meter data

We propose a new Non-Intrusive Load Monitoring (NILM) approach for appliances power profile/signal estimation at low sampling rate (1 s or greater). The proposed method relay on two main phases: identification of turned on appliance combination in a given time period and estimation of the active power consumption signal of each individual appliances in that combination. Unlike most existing NILM method that rely on multiple measurement at high sampling rates, appliances identification of this paper rely on a Karhunen Loéve (KL) expansion based spectral signature approach which only need active power measurements at a low sampling rate. Then the estimation of the power signal of the identified appliances is newly presented in this paper based on modified version of mean-shift clustering algorithm and Bayesian classification. The proposed method was validated by using two public databases: tracebase and REDD. The presented results demonstrate the ability of the proposed method to accurately estimate individual active power signal of turned on appliance combinations in real households.

Simplified controller for three wheeled omni directional mobile robot

This paper presents the work on developing a controller for the three wheeled omni-directional mobile robot for the operation on a flat terrain. Central to a robot base are omni-directional wheels. In addition to providing traction normal to the rotor axis like an ordinary wheel, they are capable of sliding parallel to the rotor axis without much friction. We have designed and implemented a robot utilizing three omni-directional wheels mounted 120° apart. Trajectory of the robot is specified as the time variation of the position coordinates and the orientation of the robot with respect to a fixed frame. The trajectory is dictated by the speeds of the three omni-directional wheels. Been a holonomic robot, the designed robot can achieve any continuous trajectory. The trajectory is planned on the fixed world frame. Then, it is transformed to the time-varying robot body frame. The velocities of the wheels are computed with respect to the body frame. A PID controller is utilized for the control of each parameter. Consequently, several PID controllers are employed for trajectory control. The MATLAB® SIMULINK® platform is used to develop the kinetic model of the system. The designed robot has been implemented in hardware.

Vision based obstacle detection and map generation for reconnaissance

Obstacle detection and map generation is an essential tool for site reconnaissance applications. Further it enables optimal and efficient path planning for mobile agents to navigate in unknown environments. This paper proposes a solution to this problem through a stereo vision-based obstacle detection and depth measurement method for reconnaissance. The proposed approach employs a boundary tracing algorithm to identify the objects in the close vicinity to the vision system. Next, it generates a depth map that contains the obstacles and the distance to each obstacle with respect to the position of the mobile agent. An error model was developed to further improve the accuracy of the depth estimation. Depth maps from multiple angles for a given set of identified obstacles are superimposed in the same coordinate system to generate the aerial map. For objects with sharp edges, the aerial map was generated based on corners identified using a corner detection algorithm. A prototype of the system was implemented and tested in a controlled environment. The results show that percentage estimation error was significantly reduced after the depth calculation was refined using the proposed error model.

Visual event classification with human like perception

The primary objective of automated motion semantic classification would be to recognizing events in close similarity with human like perception. This work proposes novel modifications to the standard spectral clustering algorithm in enhancing its capacity to capture human like semantics for visual event classification. The proposed novel multi-feature aggregation strategy replicates human like decision making, incorporating the contextual information of features rather than attempting blind fusion of them. The structural alterations introduced in the Laplacian enabled the methodology to alter the scheme of an event to be detected as an anomalous activity similar to human interpretation. Results of the implemented methodology have been demonstrated for experiments conducted on video streams focusing on human motion patterns.

Adaptive free cylindrical mixture model for foreground estimation in rapidly fluctuating dynamic background conditions

A novel method of modelling pixel distributions for foreground detection in rapidly fluctuating dynamic background conditions is presented in this paper. A comprehensive study of the characteristics of pixel behaviour in videos of backgrounds in clear water under natural lighting conditions has been presented in this work. Videos from real world situations such as in swimming pools and ponds where foreground detection is important were analyzed and it was identified that the distributions of pixel intensity values in a single pixel appear to form cylindrical clusters in RGB space. Therefore, in order to model the highly dynamic rapidly fluctuating background scenes in aquatic conditions, a novel cylindrical model is proposed where the axis is freed to allow for the high dynamism. An adaptive free cylindrical mixture model (AFCMM), which learns the directions of orientation of the clusters using an eigenanalysis based approach, is proposed for foreground detection in aquatic conditions. The results from foreground estimation in a swimming pool using the adaptive Gaussian mixture model and the proposed AFCMM have been compared and it has been shown that the latter provides an improved estimate of the foreground while demonstrating its effectiveness as a better descriptor for the pixel dynamics under such conditions.