Mechanical Engineering
Permanent URI for this communityhttp://hdl.handle.net/1903/2263
Browse
4 results
Search Results
Item Towards Trust and Transparency in Deep Learning Systems through Behavior Introspection & Online Competency Prediction(2021) Allen, Julia Filiberti; Gabriel, Steven A.; Mechanical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Deep neural networks are naturally “black boxes”, offering little insight into how or why they make decisions. These limitations diminish the adoption likelihood of such systems for important tasks and as trusted teammates. We employ introspective techniques to abstract machine activation patterns into human-interpretable strategies and identify relationships between environmental conditions (why), strategies (how), and performance (result) on both a deep reinforcement learning two-dimensional pursuit game application and image-based deep supervised learning obstacle recognition application. Pursuit-evasion games have been studied for decades under perfect information and analytically-derived policies for static environments. We incorporate uncertainty in a target’s position via simulated measurements and demonstrate a novel continuous deep reinforcement learning approach against speed-advantaged targets. The resulting approach was tested under many scenarios and performance exceeded that of a baseline course-aligned strategy. We manually observed separation of learned pursuit behaviors into strategy groups and manually hypothesized environmental conditions that affected performance. These manual observations motivated automation and abstraction of conditions, performance and strategy relationships. Next, we found that deep network activation patterns could be abstracted into human-interpretable strategies for two separate deep learning approaches. We characterized machine commitment by the introduction of a novel measure and revealed significant correlations between machine commitment, strategies, environmental conditions, and task performance. As such, we motivated online exploitation of machine behavior estimation for competency-aware intelligent systems. And finally, we realized online prediction capabilities for conditions, strategies, and performance. Our competency-aware machine learning approach is easily portable to new applications due to its Bayesian nonparametric foundation, wherein all inputs are compactly transformed into the same compact data representation. In particular, image data is transformed into a probability distribution over features extracted from the data. The resulting transformation forms a common representation for comparing two images, possibly from different types of sensors. By uncovering relationships between environmental conditions (why), machine strategies (how), & performance (result) and by giving rise to online estimation of machine competency, we increase transparency and trust in machine learning systems, contributing to the overarching explainable artificial intelligence initiative.Item Physics-Based Model-Guided Machine Learning Analysis of Wrist Ballistocardiography for Cuff-Less Blood Pressure Monitoring(2019) Yousefian, Peyman; Hahn, Jin Oh; Mechanical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Cuff-less blood pressure (BP) monitoring technology is being widely pursued today. In this research we investigated the wrist ballistocardiogram (BCG) as a limb BCG, to develop a scientific basis to use the limb BCG to for cuff-less BP monitoring. In our study, we pursue two alternative approaches to the use of wrist BCG signal for BP monitoring: (1) use of the wrist BCG as proximal timing in pulse transit time (PTT) based methods; (2) use of wrist BCG wave features for BP monitoring. In this regard, the physics-based model is developed to elucidate the mechanism responsible for the generation of the BCG signal at the body’s extremity limb locations. The developed and experimentally validated mathematical model can predict the limb BCG in responses to the arterial BP waves in the aorta. The model suggests that the limb BCG waveform reveals the timings and amplitudes associated with the aortic BP waves and it exhibits meaningful morphological changes in response to the alterations in the CV risk predictors. Such understanding combined with machine learning techniques has helped us to extract viable features, and construct predictive models that can estimate BP. The findings of this study show that limb BCG has the potential to realize convenient cuff-less BP monitoring. First, it is a strong candidate to extract the proximal timing for PTT based methods. Second, BCG wave features are associated with BP and it could be used for BP monitoring. Third, we can combine the PTT with BCG wave features to achieve more comprehensive prediction models with superior performance.Item DEVELOPMENT OF DIAGNOSTIC AND PROGNOSTIC METHODOLOGIES FOR ELECTRONIC SYSTEMS BASED ON MAHALANOBIS DISTANCE(2009) Kumar, Sachin; Pecht, Michael; Reliability Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Diagnostic and prognostic capabilities are one aspect of the many interrelated and complementary functions in the field of Prognostic and Health Management (PHM). These capabilities are sought after by industries in order to provide maximum operational availability of their products, maximum usage life, minimum periodic maintenance inspections, lower inventory cost, accurate tracking of part life, and no false alarms. Several challenges associated with the development and implementation of these capabilities are the consideration of a system's dynamic behavior under various operating environments; complex system architecture where the components that form the overall system have complex interactions with each other with feed-forward and feedback loops of instructions; the unavailability of failure precursors; unseen events; and the absence of unique mathematical techniques that can address fault and failure events in various multivariate systems. The Mahalanobis distance methodology distinguishes multivariable data groups in a multivariate system by a univariate distance measure calculated from the normalized value of performance parameters and their correlation coefficients. The Mahalanobis distance measure does not suffer from the scaling effect--a situation where the variability of one parameter masks the variability of another parameter, which happens when the measurement ranges or scales of two parameters are different. A literature review showed that the Mahalanobis distance has been used for classification purposes. In this thesis, the Mahalanobis distance measure is utilized for fault detection, fault isolation, degradation identification, and prognostics. For fault detection, a probabilistic approach is developed to establish threshold Mahalanobis distance, such that presence of a fault in a product can be identified and the product can be classified as healthy or unhealthy. A technique is presented to construct a control chart for Mahalanobis distance for detecting trends and biasness in system health or performance. An error function is defined to establish fault-specific threshold Mahalanobis distance. A fault isolation approach is developed to isolate faults by identifying parameters that are associated with that fault. This approach utilizes the design-of-experiment concept for calculating residual Mahalanobis distance for each parameter (i.e., the contribution of each parameter to a system's health determination). An expected contribution range for each parameter estimated from the distribution of residual Mahalanobis distance is used to isolate the parameters that are responsible for a system's anomalous behavior. A methodology to detect degradation in a system's health using a health indicator is developed. The health indicator is defined as the weighted sum of a histogram bin's fractional contribution. The histogram's optimal bin width is determined from the number of data points in a moving window. This moving window approach is utilized for progressive estimation of the health indicator over time. The health indicator is compared with a threshold value defined from the system's healthy data to indicate the system's health or performance degradation. A symbolic time series-based health assessment approach is developed. Prognostic measures are defined for detecting anomalies in a product and predicting a product's time and probability of approaching a faulty condition. These measures are computed from a hidden Markov model developed from the symbolic representation of product dynamics. The symbolic representation of a product's dynamics is obtained by representing a Mahalanobis distance time series in symbolic form. Case studies were performed to demonstrate the capability of the proposed methodology for real time health monitoring. Notebook computers were exposed to a set of environmental conditions representative of the extremes of their life cycle profiles. The performance parameters were monitored in situ during the experiments, and the resulting data were used as a training dataset. The dataset was also used to identify specific parameter behavior, estimate correlation among parameters, and extract features for defining a healthy baseline. Field-returned computer data and data corresponding to artificially injected faults in computers were used as test data.Item SWARM INTELLIGENCE AND STIGMERGY: ROBOTIC IMPLEMENTATION OF FORAGING BEHAVIOR(2003-12-16) Edelen, Mark Russell; C�rdenas-Garc�a, Jaime F; Baz, Amr M; Mechanical EngineeringSwarm intelligence in multi-robot systems has become an important area of research within collective robotics. Researchers have gained inspiration from biological systems and proposed a variety of industrial, commercial, and military robotics applications. In order to bridge the gap between theory and application, a strong focus is required on robotic implementation of swarm intelligence. To date, theoretical research and computer simulations in the field have dominated, with few successful demonstrations of swarm-intelligent robotic systems. In this thesis, a study of intelligent foraging behavior via indirect communication between simple individual agents is presented. Models of foraging are reviewed and analyzed with respect to the system dynamics and dependence on important parameters. Computer simulations are also conducted to gain an understanding of foraging behavior in systems with large populations. Finally, a novel robotic implementation is presented. The experiment successfully demonstrates cooperative group foraging behavior without direct communication. Trail-laying and trail-following are employed to produce the required stigmergic cooperation. Real robots are shown to achieve increased task efficiency, as a group, resulting from indirect interactions. Experimental results also confirm that trail-based group foraging systems can adapt to dynamic environments.