Collaborative and Distributed Data Analytics for Smart and Connected Systems

Abstract

Unprecedented connectivity across tangible physical units enabled by Internet of Things technologies has set forth a new collaborative paradigm for predictive analytics. In this paradigm, physical units can borrow knowledge from each other, allowing for significantly enhanced model learning capabilities. Yet, many crucial challenges remain unsolved. First, the amount of collected data is beginning to overwhelm traditional centralized computing resources. Second, personalized analytics is instrumental as physical units often operate under different environments. Third, real-time data collection necessitates adaptive models in collaborative analytics. Last but not least, incomplete data collection greatly hinders the development of a unifying analytics framework.

This report sets research objectives to develop predictive analytics frameworks addressing the challenges above and highlight their applications in reliability engineering and healthcare informatics.

1. The tremendous increase in computing power of edge units sets the stage for a new paradigm of analytics where edge computing power is exploited to process some of the data where it is created. This new paradigm allows federated analytics where model learning is distributed across system components, and only relevant information extracted from local processing is shared. This research develops federated predictive frameworks to facilitate model inference, personalized prediction, and uncertainty quantification in a distributed fashion.

2. This research is to achieve two critical goals in data analytics for smart and connected systems: personalization and collaboration. Personalization is challenging yet essential for individual-level predictions as devices operate under different environments. At the same time, connectivity provides a significant opportunity for collaboration where shared knowledge across units is extracted and transferred to help improve individual-level predictions. The research aims to establish personalized and collaborative approaches tailored for reliability engineering and healthcare informatics applications.

3. As smart and connected systems often collect data in real-time, the research is to design learning approaches that enable continuous accumulation of knowledge from newly obtained data and fast updates of predictions, with application to the detection of engineering system failure.

4. Smart and connected systems for healthcare applications often involve sensitive personal data. For example, health monitoring data collected from wearable devices is analyzed along with the patient’s personal information, which may be unreported in practice. The research aims to establish a robust collaborative analytics framework that can achieve competitive predictive performance despite incomplete information.

This report develops four predictive analytics frameworks that achieve the objectives above, with application to reliability engineering and healthcare informatics. (i) Predictive analytics for connected systems with units equipped with multiple sensors: the proposed approach achieves significantly improved prediction of the future trajectory of multi-stream data from an in-service unit, by extracting commonality across units and updating the prediction using online observations. (ii) Transfer learning across groups in the presence of observations without group membership: the proposed transfer learning framework improves predictive performance by exploiting information from ungrouped observations, often arising in healthcare informatics that collects sensitive data. (iii) Federated predictive analytics for heterogeneous condition monitoring (CM) signals: this approach builds a learning framework to enhance the generalization of neural networks for predicting heterogeneous CM signals, by leveraging the local computational resources of the units that collect CM signals. (iv) Federated analytics with multi-output Gaussian processes (MGP): this approach develops a framework to learn MGPs in a way that distributes computing and storage demands, reduces communication burden, fosters privacy, and personalizes predictions to each unit.