a b s t r a c tIn the past decade, rapid technological advances in the fields of electronics and telecommunications have given rise to versatile, ubiquitous decentralized embedded sensor systems with ad hoc wireless networking capabilities. Typically these systems are used to gather large amounts of data, while the detection of anomalies (such as system failures, intrusion, or unanticipated behavior of the environment) in the data (or other types or processing) is performed in centralized computer systems. In spite of the great interest that it attracts, the systematic porting and analysis of centralized anomaly detection algorithms to a decentralized paradigm (compatible with the aforementioned sensor systems) has not been thoroughly addressed in the literature. We approach this task from a new angle, assessing the viability of localized (in-node) anomaly detection based on machine learning. The main challenges we address are:(1) deploying decentralized, automated, online learning, anomaly detection algorithms within the stringent constraints of typical embedded systems; and (2) evaluating the performance of such algorithms and comparing them with that of centralized ones. To this end, we first analyze (and port) single and multi-dimensional input classifiers that are trained incrementally online and whose computational requirements are compatible with the limitations of embedded platforms. Next, we combine multiple classifiers in a single online ensemble. Then, using both synthetic and real-world datasets from different application domains, we extensively evaluate the anomaly detection performance of our algorithms and ensemble, in terms of precision and recall, and compare it to that of well-known offline, centralized machine learning algorithms. Our results show that the ensemble performs better than each individual decentralized classifier and that it can match the performance of the offline alternatives, thus showing that our approach is a viable solution to detect anomalies, even in environments with little a priori knowledge.
Abstract-Scanning transmission electron microscopes are indispensable tools for material science research, since they can reveal the internal structure of a wide range of specimens. Thus, it is of scientific and industrial interest to transform these microscopes into flexible, high-throughput, unsupervised, nanomeasuring tools. To do so, processes that are currently executed manually based on visual feedback (e.g., alignment or particle measurement) should be automated, taking into consideration their time dependencies. That is, these microscopes should be studied from the systems and control perspective.To the best of our knowledge, such perspective is lacking in the literature. Thus, it is provided here through a new modeling framework that facilitates the future development of control strategies based on image analysis. The progress made towards developing an image-based sensor for defocus control is also reported. Finally, the paper also aims to introduce scanning transmission electron microscopy as an important and untapped application area for control engineers.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.