Zhanhong Jiang

An unsupervised spatiotemporal graphical modeling approach to anomaly detection in distributed CPS

Modern distributed cyber-physical systems (CPSs) encounter a large variety of physical faults and cyber anomalies and in many cases, they are vulnerable to catastrophic fault propagation scenarios due to strong connectivity among the sub-systems. This paper presents a new data-driven framework for system-wide anomaly detection for addressing such issues. The framework is based on a spatiotemporal feature extraction scheme built on the concept of symbolic dynamics for discovering and representing causal interactions among the subsystems of a CPS. The extracted spatiotemporal features are then used to learn system-wide patterns via a Restricted Boltzmann Machine (RBM). The results show that: (1) the RBM free energy in the off-nominal conditions is different from that in the nominal conditions and can be used for anomaly detection; (2) the framework can capture multiple nominal modes with one graphical model; (3) the case studies with simulated data and an integrated building system validate the proposed approach.

An unsupervised anomaly detection approach using energy-based spatiotemporal graphical modeling

Abstract This paper presents a new data-driven framework for unsupervised system-wide anomaly detection for modern distributed complex systems within which there exists a strong connectivity among sub-systems, operating in diverse modes and encountering a large variety of anomalies. The framework is based on a spatiotemporal feature extraction scheme built on the concept of symbolic dynamics for discovering and representing causal interactions among subsystems. The extracted features from the spatiotemporal pattern network (STPN) are then used to learn system-wide patterns via a Restricted Boltzmann Machine (RBM), to form an energy based anomaly detection approach. While STPN is treated as a weak learner of system modes (in terms of difficulty in discovering true graphical representations), RBM is treated as a boosting approach to form a strong learner of system characteristics. Case studies with simulated data and real data from an integrated building system are used to validate the proposed approach. The results show that: (i) the increase in RBM free energy in the off-nominal conditions compared to that in the nominal conditions can be used for anomaly detection; (ii) the proposed framework formulates a strong learning model (STPN+RBM) from weak frequentist model–STPN, via boosting with RBM; and (iii) the STPN+RBM framework can capture multiple nominal operating modes of distributed complex systems with a single graphical model.

On distributed optimization using generalized gossip

This paper presents a generalized gossip-based algorithm to solve distributed optimization problems in multi-agent networks, especially for multiple supply-demand optimization problems. The proposed algorithm provides a generalization such that the optimization process can operate in the entire spectrum of “complete consensus” to “complete disagreement”. A user-defined control parameter θ is identified for controlling such tradeoff as well as the temporal convergence properties. Analytical results for first moment convergence analysis are presented and it is shown that with θ → 0, the formulation boils down to a classical consensus based protocol. Beyond the control parameter, the agent interaction matrix is also shown to be useful for effectively suppressing large localized uncertainties in subgradient estimation. A practical use case regarding building zone temperature control is presented as a numerical example for validation.

Understanding Wind Turbine Interactions Using Spatiotemporal Pattern Network

This paper presents a data-driven modeling framework to understand spatiotemporal interactions among wind turbines in a large scale wind energy farm. A recently developed probabilistic graphical modeling scheme, namely the spatiotemporal pattern network (STPN) is used to capture individual turbine characteristics as well as pair-wise causal dependencies. The causal dependency is quantified by a mutual information based metric and it has been shown that it efficiently and correctly captures both temporal and spatial characteristics of wind turbines. The causal interaction models are also used for predicting wind power production by one wind turbine using observations from another turbine. The proposed tools are validated using the Western Wind Integration data set from the National Renewable Energy Laboratory (NREL).Copyright

Probabilistic Graphical Modeling of Distributed Cyber-Physical Systems

Probabilistic graphical models (PGMs) have been shown to efficiently capture the dynamics of physical systems as well as model cyber systems such as communication networks. This chapter focuses on some recent developments in applying PGMs as data-driven models for jointly analyzing cyber and physical properties of distributed complex systems. Among various PGM techniques, Bayesian networks and symbolic dynamic filtering are discussed with three use cases: spatial modeling, multiscale temporal modeling, and spatiotemporal interaction modeling. While spatial modeling deals with anomaly detection and root-cause analysis in large building systems, temporal modeling involves the online identification of dynamic modes of a complex system operation. Finally, spatiotemporal modeling aims to characterize casual dependencies among wind turbines in a large wind farm, both spatially and temporally, for predicting farm-wide power generation. The chapter concludes with key observations such as enhanced robustness and scalability in cyber-physical system modeling with PGMs along with recommendations for future research directions.

Comparative Evaluation of Control-Oriented Zone Temperature Prediction Modeling Strategies in Buildings

Predictive modeling of zone environment plays a critical role in developing and deploying advanced performance monitoring and control strategies for energy usage minimization in buildings while maintaining occupant comfort. The task remains extremely challenging, as buildings are fundamentally complex systems with large uncertainties stemming from weather, occupants, and building dynamics. Over the past few years, purely data-driven various control-oriented modeling techniques have been proposed to address different requirements, such as prediction accuracy, flexibility, computation and memory complexity. In this context, this paper presents a comparative evaluation among representative methods of different classes of models, such as first principles driven (e.g., lumped parameter autoregressive models using simple physical relationships), data-driven (e.g., artificial neural networks, Gaussian processes) and hybrid (e.g., semi-parametric). Apart from quantitative metrics described above, various qualitative aspects such as cost of commissioning, robustness and adaptability are discussed as well. Real data from Iowa Energy Center’s Energy Resource Station (ERS) test bed is used as the basis of evaluation presented here.Copyright

Convergence and noise effect analysis for generalized gossip-based distributed optimization

The generalized gossip-based subgradient algorithm has been recently proposed for solving distributed optimization problems associated with multi-agent networks. The algorithm provides a generalization such that the optimization process can operate in the entire spectrum of “complete consensus” to “complete disagreement”. Beyond the existing work of first-order convergence analysis results, this paper presents the second-order convergence results and convergence rate estimates for the proposed algorithm. Moreover, this work also takes into consideration the effect of noise in subgradient estimates as well as measurements on the function value error bounds. A numerical case study based on a building energy system is presented to validate the algorithm.

Generalised gossip-based subgradient method for distributed optimisation

ABSTRACT This paper presents a generalised time-synchronous gossip-based algorithm for solving distributed optimisation problems associated with multi-agent networked systems. The proposed algorithm presents a generalisation such that the optimisation process can operate in the entire spectrum from ‘complete consensus’ to ‘complete disagreement’. A user-defined control parameter θ is identified for controlling such tradeoff as well as the temporal convergence properties. We formulate the algorithm based upon generalised time-synchronous gossip algorithm and subgradient method and provide analytical results for first and second moment convergence analysis. The proposed algorithm also provides a convergence rate estimate of in the number of iterations m when the step size is constant. We consider the effect of noise in networked systems from the perspectives of modelling uncertainty and measurement noise in subgradient estimation process and communication among agents, respectively. A numerical case study based on a multi-agent building energy system is used to validate the proposed algorithm.

Scalable supervisory control of building energy systems using generalized gossip

This paper presents a novel distributed optimization framework to achieve energy efficiency in large-scale buildings. The modular problem formulation presented in this paper decouples the supervisory optimization scheme from the data-driven micro-level modeling aspect leading to significant scalability and flexibility. Recently developed generalized gossip protocol is used as a robust distributed optimization technique. A supervisory control design problem for multi-zone temperature regulation and energy usage minimization is considered as a case study to describe the generic framework. Numerical simulation results, presented based on a physical testbed in the Iowa Energy Center, demonstrate the advantages of the distributed optimization methodology compared to a typical baseline strategy. The paper also outlines a software architecture based on the VOLTTRON platform, recently developed by the Pacific Northwest National Laboratory (PNNL), for real-life implementation of the proposed framework.