Yuxun Zhou

Abnormal event detection with high resolution micro-PMU data

Power system has been incorporating increasing amount of unconventional generations and loads such as renewable resources, electric vehicles, and controllable loads. The induced short term and stochastic power flow requires high resolution monitoring technology and agile decision support techniques for system diagnosis and control. In this paper, we discuss the application of micro-phasor measurement unit (μPMU) for power distribution network monitoring, and study learning based data-driven methods for abnormal event detection. We first resolve the challenging problem of information representation for the multiple streams of high resolution μPMU data, by proposing a pooling-picking scheme. With that, a kernel Principle Component Analysis (kPCA) is adopted to build statistical models for nominal state and detect possible anomalies. To distinguish event types, we propose a novel discriminative method that only requires partial expert knowledge for training. Finally, our methods are tested on an actual distribution network with μPMUs, and the results justifies the effectiveness of the data driven event detection framework, as well as its potentials to serve as one of the core algorithms to ensure power system security and reliability.

Adaptive Localization in Dynamic Indoor Environments by Transfer Kernel Learning

Accurate Location Based Service (LBS) is one of the fundamental but crucial services in the era of Internet of Things (IoT). WiFi fingerprinting-based Indoor Positioning System (IPS) has become the most promising solution for indoor LBS. However, the offline calibrated received signal strength (RSS) radio map is unable to provide consistent LBS with high localization accuracy under various environmental dynamics. To address this issue, we propose TKL-WinSMS as a systematic strategy, which is able to realize robust and adaptive indoor localization in dynamic indoor environments. We developed a WiFi-based Non-intrusive Sensing and Monitoring System (WinSMS) that enables WiFi routers as online reference points by extracting real-time RSS readings among them. With these online data and labeled source data from the offline calibrated radio map, we further combine the RSS readings from target mobile devices as unlabeled target data, to design a robust localization model using an emerging transfer learning algorithm, namely transfer kernel learning (TKL). It is able to learn a domain-invariant kernel by directly matching the source and target distributions in the reproducing kernel Hilbert space instead of the raw noisy signal space. The resultant kernel can be used as input for the SVR training procedure. In this manner, the trained localization model can inherit the information from online phase to adaptively enhance the offline calibrated radio map. Extensive experiments were conducted and demonstrated that the proposed TKL- WinSMS is able to improve the localization accuracy by at least 26\% compared with existing solutions under various environmental interferences.

SugarMate: Non-intrusive Blood Glucose Monitoring with Smartphones

Inferring abnormal glucose events such as hyperglycemia and hypoglycemia is crucial for the health of both diabetic patients and non-diabetic people. However, regular blood glucose monitoring can be invasive and inconvenient in everyday life. We present SugarMate, a first smartphone-based blood glucose inference system as a temporary alternative to continuous blood glucose monitors (CGM) when they are uncomfortable or inconvenient to wear. In addition to the records of food, drug and insulin intake, it leverages smartphone sensors to measure physical activities and sleep quality automatically. Provided with the imbalanced and often limited measurements, a challenge of SugarMate is the inference of blood glucose levels at a fine-grained time resolution. We propose Md3RNN, an efficient learning paradigm to make full use of the available blood glucose information. Specifically, the newly designed grouped input layers, together with the adoption of a deep RNN model, offer an opportunity to build blood glucose models for the general public based on limited personal measurements from single-user and grouped-users perspectives. Evaluations on 112 users demonstrate that Md3RNN yields an average accuracy of 82.14%, significantly outperforming previous learning methods those are either shallow, generically structured, or oblivious to grouped behaviors. Also, a user study with the 112 participants shows that SugarMate is acceptable for practical usage.

Data-driven event detection with partial knowledge: A Hidden Structure Semi-Supervised learning method

Enabled by the advancement of data acquisition and data analysis technologies such as sensor networks and machine learning, recently data-driven event detection has shown its advantage in dealing with complex systems especially those with significant stochastic and dynamic behavior. However, existing methods usually adopt supervised learning framework and depend on explicit expert labeling in the learning phase, which is expensive even impractical in many situations. In this work, we propose a new data-driven event detection method, namely Hidden Structure Semi-Supervised Machine (HS3M), that only requires partial expert knowledge. The key idea is to combine unlabeled data and partly labeled data in a large margin learning objective to bridge the gap between supervised, semi-supervised learning and learning with hidden structures. Difficulties do arise as the incorporation of extra learning terms makes the problem non-convex. To optimize the learning objective we establish a novel global optimization algorithm, namely Parametric Dual Optimization Procedure (PDOP), by showing that the parametrized dual problem has local explicit solutions and the corresponding optimality is convex in hidden variables. The proposed approach is applied to power distribution network event detection, and the result justifies the effectiveness of both HS3M and the new global optimization algorithm.

FreeDetector: Device-Free Occupancy Detection with Commodity WiFi

Occupancy detection is playing a critical role to improve the efficiency of building management system and optimize personalized thermal comfort, among many other emerging applications. Conventional occupancy detection methods, such as Passive Infra-Red (PIR) and camera, have several drawbacks including low accuracy, high intrusiveness and extra infrastructure. In this work, we propose FreeDetector, a device-free occupancy detection scheme that is able to detect human presence accurately just using existing commodity WiFi routers. We upgrade the firmware of the routers so that the channel state information (CSI) data in PHY layer can be obtained directly from them. With only two routers, FreeDetector is able to reveal the variations in CSI data caused by human presence. We leverage signal tendency index (STI) to analyze the shape similarity of adjacent time series CSI curves. The most representative subset of subcarriers is selected by greedy algorithm and we utilize machine learning algorithm to construct a detection classifier. Extensive experiments are conducted and the results demonstrate that FreeDetector is able to provide outstanding occupancy detection service in terms of both accuracy and efficiency.

Virtual power sensing based on a multiple-hypothesis sequential test

Virtual-Sensing, which is achieved through the disaggregation of composite power metering signals, is a solution towards achieving fine-grained smart power monitoring. In this work we discuss the challenging issues in Virtual-Sensing, introduce and ultimately combine the Hidden Markov Model and the Edge-based methods. The resulting solution, based on a Multiple-hypothesis Sequential Probability Ratio Test, combines the advantages of the two methods and delivers significant improvement in disaggregation performance. A robust version of the test is also proposed to filter the impulse noise common in real-time monitoring of the plug-in loads power consumption.

Causal analysis for non-stationary time series in sensor-rich smart buildings

Advances in low cost sensor and networking technologies in smart buildings have given researchers access to a multitude of time series data, including temperature, humidity, real and reactive power consumption of specific nodes or devices, occupant presence and activities, etc. Time series generated by sensor networks reflect various phenomena in buildings and are naturally related to each other. Hence quantitative techniques are required to exploit dependence among different types of sequences in order to allow smart applications such as non-intrusive activity detection, energy usage prediction, demand side management and control. Past research of relational analysis has focused on symmetric correlative statistics. On the other hand, asymmetric causal relations can capture more dynamic and complex relationships and is able to reveal directed influence among series. However, most traditional causal analysis relies on stationarity, while the statistics of real sensor measurement in smart buildings is rarely time invariant. In this paper, a statistical time series analysis framework is proposed to examine causal relationships among time series that are highly non-stationary. The Granger causality identification is extended to sensor data in buildings and the issue of non-stationarity is initially addressed by using modified Hodrick-Prescott (HP) filter which is able to extract simpler trend components. Subsequently, Autoregressive Integrated Moving Average model with exogenous variables (ARIMAX) model is trained for different components of two series. Finally, Granger causality is tested for both directions by F-statistics. The above procedure is performed on actual energy-consumption time series to exploit potential causal relations.

Optimal Sensor Configuration and Feature Selection for AHU Fault Detection and Diagnosis

Experiments show that operation efficiency and reliability of buildings can greatly benefit from rich and relevant datasets. More specifically, data can be analyzed to detect and diagnose system and component failures that undermine energy efficiency. Among the huge quantity of information, some features are more correlated with the failures than others. However, there has been little research to date focusing on determining the types of data that can optimally support fault detection and diagnosis (FDD). This paper presents a novel optimal feature selection method, named information greedy feature filter (IGFF), to select essential features that benefit building FDD. On one hand, the selection results can serve as reference for configuring sensors in the data collection stage, especially when the measurement resource is limited. On the other hand, with the most informative features selected by the IGFF, the performance of building FDD could be improved and theoretically justified. A case study on air-handling unit (AHU) is conducted based on the dataset of the ASHRAE Research Project 1312. Numerical results show that, compared with several baselines, the FDD performances of conventional classification methods are greatly enhanced by the IGFF.

Optimal Training and Efficient Model Selection for Parameterized Large Margin Learning

Recently diverse variations of large margin learning formalism have been proposed to improve the flexibility and the performance of classic discriminative models such as SVM. However, extra difficulties do arise in optimizing non-convex learning objectives and selecting multiple hyperparameters. Observing that many variations of large margin learning could be reformulated as jointly minimizing a parameterized quadratic objective, in this paper we propose a novel optimization framework, namely Parametric Dual sub-Gradient Descent Procedure (PDGDP), that produces a globally optimal training algorithm and an efficient model selection algorithm for two classes of large margin learning variations. The theoretical bases are a series of new results for parametric program, which characterize the unique local and global structure of the dual optimum. The proposed algorithms are evaluated on two representative applications, i.e., the training of latent SVM and the model selection of cost sensitive feature re-scaling SVM. The results show that PDGDP based training and model selection achieves significant improvement over the state-of-the-art approaches.

Online learning of Contextual Hidden Markov Models for temporal-spatial data analysis

The problem of mining a network of time series data naturally arises in many research areas including energy system, quantitative finance, bioinformatics, environmental monitoring, traffic monitoring, etc. Among others, two emerging challenges shared by manifold applications are (1) the modeling of temporal-spatial dependence with contextual information and (2) the design of efficient learning algorithms for big data (exceedingly long sequence) analytics. In this paper, we study a Contextual Hidden Markov Model (CHMM) that describes infinite temporal dependence and contextual spatial relations in an unified framework. More importantly, to make model training feasible for growing number of data samples, we develop an Online Expectation-Maximization (OEM) algorithm that avoids the usual forward-backward pass of the entire time sequence. Two typical applications, missing value recovery and novelty detection, are considered to test CHMM and the online algorithm. Experiments are conducted on real world data collected from power distribution network monitoring. We compare CHMM with other popular methods and the results not only justify the benefit of incorporating temporal-spatial and contextual information, but also demonstrate the effectiveness of the proposed OEM algorithm.