Hamid Krim

Nuclear power plants predominantly operate at or near full power. As such, ML-based anomaly detection methods will be strongly biased towards full power and when the power drops (often called transients) may have difficulty detecting anomalies or may provide false alarms due to the lack of associated data. The current approach to handling this difficulty is to turn detectors off during transients when the power falls below some threshold. This means that the significant benefits from automated anomaly detection will be unavailable when plant conditions are changing most rapidly (and consequently could incur above averages stresses on process components). The objective of this effort is to extend anomaly detection capabilities to transient conditions. This effort defines two different types of transients that might be encountered in plants. First, there are fast transients, where the power may drop quickly in reaction to some condition, but the reduced power will only last for a few days before returning to full power. Second, there are slow transients, where the power is more likely to drop slowly (although could drop quickly), but the reduced power may last for weeks or months. An example of this is a slow ramping down of power before a plant shutdown. The current scope will explore methods of extending anomaly detection methods to transient conditions where there is significantly less operating data compared with full power. This means these methods must learn from full power conditions and extend that learning to transient conditions to be able to differentiate between normal transient conditions and anomalous transient conditions.

An interest in a position at ARO in the capacity of an IPA contract employee is expressed. The interest is particularly focussed in exploiting the PI's knowledge and familiarity with DOD challenges to provide the vision, leadership and help shape the research agenda priorities to ensure that the current as well future sensing and information and machine learning challenges be met.

Our goal in this effort is to apply recently developed multi-modality fusion algorithms to vehicle data collected at ORNL. The testing of additional sensors would ensure the applicability of our approaches to the ORNL target goal of potentially deploying these algorithms in their system which is an on-going project. This testing project is to hence harness a number of sensing modalities including, acoustic, magnetic, video, possibly laser, and jointly exploit them to carry out target inference. We will focus on a vast database of vehicles most of which have been characterized by the previously mentioned modalities. In coordination with the ORNL Project Lead, we will work to identify the potential stress factors for the developed algorithms in the field, and will define the proper strategies to preserve a robust or at least gracefully degrading performance of the algorithms.

This work is to thoroughly evaluate and test algorithms that have been developed within VISSTA Laboratory. These pursued remote sensing data sets with associated applications will strengthen our collaborative work with Lawrence Livermore by a number of internship periods of the student at LLNL.

With a focus on Diabetes in Phase 1, we propose the development of a comprehensive tool which will systematically and seamlessly navigate across the various hybrid data accessible through UNC Medical Records, with a health assessment enabling capability as well as various possible trends. More specifically, we plan on fully exploiting the tools of machine learning and bring them to bear on each step of the data analysis and exploitation. In close consultation with the health specialists, we will develop a mapping mechanism of qualitative data to the quantitative space, thus homogenizing the data. In addition, we plan to design a “Decision Tree” (DT) adapted to our homogenized data. We will exploit the characteristic computational efficiency of tree structures to comb through each patient’s data to yield a quantifiable assessment of interest (e.g. patient is cured as a result of treatment, follow up visits, prescription follow up, and cross validated with State Vital Records). Note that the proposed DT-based analyses not only play a key role in the analysis of trends across populations of patients, but also can also be used to conduct other global statistical analyses (e.g. probabilistically determine a permanent cure conditioned on a life style). Our plan is to have a Personal Computer-based menu-driven tool with a comprehensive set of options, including visualizations, to carry through a thorough exploration of data; the planned interactive operation will necessitate some attention to identifying and solving all computational bottlenecks in this process.

An interest in a position at ARO in the capacity of an IPA contract employee is expressed. The interest is particularly focussed in exploiting the PI's knowledge and familiarity with DOD challenges to provide the vision, leadership and help shape the research agenda priorities to ensure that the current as well future sensing and information and machine learning challenges be met.

NC State University, in partnership with University of Michigan, Purdue University, University of Illinois at Urbana Champaign, Kansas State University, Georgia Institute of Technology, NC A&T State University, Los Alamos National Lab, Oak Ridge National Lab, and Pacific Northwest National lab, proposes to establish a Consortium for Nonproliferation Enabling Capabilities (CNEC). The vision of CNEC is to be a pre-eminent research and education hub dedicated to the development of enabling technologies and technical talent for meeting the grand challenges of nuclear nonproliferation in the next decade. CNEC research activities are divided into four thrust areas: 1) Signatures and Observables (S&O); 2) Simulation, Analysis, and Modeling (SAM); 3) Multi-source Data Fusion and Analytic Techniques (DFAT); and 4) Replacements for Potentially Dangerous Industrial and Medical Radiological Sources (RDRS). The goals are: 1) Identify and directly exploit signatures and observables (S&O) associated with special nuclear material (SNM) production, storage, and movement; 2) Develop simulation, analysis, and modeling (SAM) methods to identify and characterize SNM and facilities processing SNM; 3) Apply multi-source data fusion and analytic techniques to detect nuclear proliferation activities; and 4) Develop viable replacements for potentially dangerous existing industrial and medical radiological sources. In addition to research and development activities, CNEC will implement educational activities with the goal to develop a pool of future nuclear non-proliferation and other nuclear security professionals and researchers.

This work addresses a problem of space debris detection, and target parameters estimation from both optical and radar data. It aims at: - A network-based experimental design to make measurement of existing debris - Exploiting two or more sensing modalities, and fusing information of at least Optical and Radar measurements potentially made at geographically distinct locations, and enhancing the data for analysis - Developing a Bayesian inference framework to overcome the diversity of tracks and targets - The optical data will be captured by a 60 cm telescope which is at disposal to CTU Prague team. The radar measurements will be obtained from a Czech amateur radio astronomy network. The optical and radar measurement will be synchronized, i.e. the same orbiting object will be seen simultaneously in both sensor modalities.

The topic of Union of Subspaces has recently emerged as a promising alternative to PCA and Robust PCA. In addition to one’s ability to retrieve a noise-free component satisfying such a model as we have recently shown, we propose to approach the problem as a subspace pursuit problem, much akin to basis pursuit, using the formalism of a Grassman Manifold. We also propose to investigate more challenging spaces where singularities appear and the underlying spaces are not necessarily flat. This so-called stratified space promises to provide additional flexibility to capture sudden changes of scenes in imagery data for instance.

In this proposed effort, we plan on building a powerful and flexible GPU computational infrastructure platform which consists of (i) a GPU cluster with more than 79,000 GPU cores and around 80 CPU cores including two cutting-edge GPU supercomputers and ten existing desktops with new high-end GPU graphics cards upgraded, and (ii) a mobile base with on-board computer for robots. The platform will be used by both the Vision, Information, and Statistical Signal Theories and Applications (VISSTA) lab directed by the co-PI Krim, and the Interpretable Visual Modeling and Computing Lab (iVMCL) currently created by the PI Wu in the department of Electrical and Computer Engineering at NC State University (NCSU). The proposed computational platform will enable the two to address ongoing research projects on heterogeneous Big Data analysis and deep understanding, as well as to smoothly prepare for future ones. This new platform will complement and match the input modalities already present at the two labs, and expand current capabilities in aggregating, parsing, fusing and ultimately analyzing and understanding heterogeneous Big Data in the following applications relevant to the Department of Defense (DoD): (i) Sensor Networks of various modalities (static or mobile) such as deep understanding of scene and events of a camera network; (ii) Social Networks with information stored locally, and (iii) Bioinformatics data, specifically related to the brain connectome, both based on topological data analysis theory; (iv) Robot autonomy by learning from situated dialogue (i.e., verbal instructions grounded on visual demonstration) and physiological sensing. Not only will the proposed GPU cluster enable PI Wu and co-PI Krim to investigate and pursue a genuine parallel implementation of many already successful centralized models and algorithms, but also be beneficial to students (undergraduates and graduates) in the classes regularly taught by PI Wu and co-PI Krim, as well as other research groups at NCSU since the proposed cluster will be seamlessly integrated, and bring new feature, into the university's computing platform.