Amulya K. Garga

Model-based predictive diagnostics for primary and secondary batteries

The model-based effort described here is focused on predictive diagnostics for primary and secondary batteries. However, this novel approach can also be applied to other electrochemical energy sources such as fuel cells. This method is based on accurate parametric modeling of the transport mechanisms within the battery. This system knowledge was used for the careful development of electrochemical and thermal models. These models have been used to extract new features to be used in conjunction with several traditional measured parameters to assess the condition of the battery. The resulting output and any usable information available about the battery is then evaluated using hybrid automated reasoning schemes consisting of neural network and decision theoretic methods. The focus of this paper is on the model identification and data fusion of the monitored and virtual sensor data. The methodology and analysis presented in this paper is applicable to mechanical systems where multiple sensor types are used for diagnostic assessment.

Neural network design using Voronoi diagrams

A novel approach is proposed which determines the number of layers, the number of neurons in each layer, and their connection weights for a particular implementation of a neural network, with the multilayer feedforward topology, designed to classify patterns in the multidimensional feature space. The approach is based on construction of a Voronoi diagram over the set of points representing patterns in feature space and this finds added usefulness in deriving alternate neural network structures for realizing the desired pattern classification.

Hybrid reasoning for prognostic learning in CBM systems

Reasoning systems that integrate explicit knowledge with implicit information are essential for high performance decision support in condition-based maintenance and prognostic health management applications. Such reasoning systems must be capable of learning the specific features of each machine during its life cycle. In this paper, a hybrid reasoning approach that is capable of integrating domain knowledge and test and operational data from the machine is described. This approach is illustrated with an industrial gearbox example. In this approach explicit domain knowledge is expressed as a rule-base and used to train a feedforward neural network. The training process results in a parsimonious representation of the explicit knowledge by combining redundant rules. A significant added practical benefit of this process is that it also is able to identify logical inconsistencies in the rule-base. Such inconsistencies are notorious in causing deadlock in large-scale expert systems. The neural network can be periodically updated with test and operational data to adapt the network to each specific machine. The flexibility and efficiency of this hybrid approach make it very suitable for practical health management systems designed to operate in a distributed environment.

Model-based predictive diagnostics for electrochemical energy sources

The model-based effort described here is focused on predictive diagnostics for primary and secondary batteries. However, this novel approach can also be applied to other electrochemical energy sources such as fuel cells. This method is based on accurate parametric modeling of the transport mechanisms within the battery. This system knowledge was used for the careful development of electrochemical and thermal models. These models have been used to extract new features to be used in conjunction with several traditional measured parameters to assess the condition of the battery. The resulting output and any usable information available about the battery is then evaluated using hybrid automated reasoning schemes consisting of neural network and decision theoretic methods. The focus of this paper is on the model identification and data fusion of the monitored and virtual sensor data. The methodology and analysis presented is applicable to mechanical systems where multiple sensor types are used for diagnostic assessment.

Architectural optimizations for a floating point multiply-accumulate unit in a graphics pipeline

Scientific visualization and virtual reality have pushed three-dimensional graphics engines to their limits for updating scenes in real-time. One bottleneck of graphic systems is the transformation of an objects vertices into normalized space based on an evaluated transformation stack. This operation as often done in floating point, requiring a fast floating point multiply-accumulate unit. This paper presents architectural optimizations to a graphics pipeline floating point multiply-accumulate unit by using block floating point and parallelism to bypass or merge trivial operations in the matrix multiplications.

MASM: a market architecture for sensor management in distributed sensor networks

Rapid developments in sensor technology and its applications have energized research efforts towards devising a firm theoretical foundation for sensor management. Ubiquitous sensing, wide bandwidth communications and distributed processing provide both opportunities and challenges for sensor and process control and optimization. Traditional optimization techniques do not have the ability to simultaneously consider the wildly non-commensurate measures involved in sensor management in a single optimization routine. Market-oriented programming provides a valuable and principled paradigm to designing systems to solve this dynamic and distributed resource allocation problem. We have modeled the sensor management scenario as a competitive market, wherein the sensor manager holds a combinatorial auction to sell the various items produced by the sensors and the communication channels. However, standard auction mechanisms have been found not to be directly applicable to the sensor management domain. For this purpose, we have developed a specialized market architecture MASM (Market architecture for Sensor Management). In MASM, the mission manager is responsible for deciding task allocations to the consumers and their corresponding budgets and the sensor manager is responsible for resource allocation to the various consumers. In addition to having a modified combinatorial winner determination algorithm, MASM has specialized sensor network modules that address commensurability issues between consumers and producers in the sensor network domain. A preliminary multi-sensor, multi-target simulation environment has been implemented to test the performance of the proposed system. MASM outperformed the information theoretic sensor manager in meeting the mission objectives in the simulation experiments.

New perspectives on Level four processing in data fusion systems

One of the top-level processing functions defined in the Joint Directors of Laboratories data fusion processing mode is Level 4 processing. This is a meta-process that monitors the on-going data fusion process to optimize the fusion of data. To date, two basic approaches have been used for Level 4 processing. The first approach uses classic constrained optimization methods to optimize a measure of performance of the data fusion process. The second approach treats Level 4 processing as a classical control problem, invovling dynamic sensor tasking, to meet performance objectives. Both of these approaches generally ignore the human-in-the-loop user, and treat sensor as relatively unintelligent devices. Recent development in smart sensor, improved models of dynamic sensor performance, and advances in cognitive psychology suggests that a new perspective on Level 4 processing is needed. In this paper, the concept of Level 4 processing is extended to exploit these rapid evolutions, resulting in increased performance of data fusion systems.

GEOMETRIC TASK DECOMPOSITION IN A MULTI-AGENT ENVIRONMENT

ABSTRACT Task decomposition in a multi-agent environment is often performed online. This paper proposes a method for sub-task allocation that can be performed before the agents are deployed, reducing the need for communication among agents during their mission. The proposed method uses a Voronoi diagram to partition the task-space among team members and includes two phases: static and dynamic. Static decomposition (performed in simulation before the start of the mission) repeatedly partitions the task-space by generating random diagrams and measuring the efficacy of the corresponding sub-task allocation. If necessary, dynamic decomposition (performed in simulation after the start of a mission) modifies the result of a static decomposition (i.e., in case of resource limitations for some agents). Empirical results are reported for the problem of surveillance of an arbitrary region by a team of agents. This material is based on work carried out at the Applied Research Laboratory at The Pennsylvania State University under the ONR grant N000149615026.