Performance evaluation method for network monitoring based on separable temporal exponential random graph models with application to the study of autocorrelation effects

With the quick development of sensors and information technologies enabling Industry 3.5, network data, which represent the interactions among related entities, have extensively emerged in manufacturing and service industries. Statistical process monitoring serves as an efficient tool for supporting accurate and timely decision-making in Industry 3.5. Applying statistical process monitoring approaches for monitoring networks significantly facilitates the early detection of potential failures in complex relational systems, and therefore has been increasingly studied in recent years. Selection of an effective network monitoring method relies on the evaluation of performances of candidate methods. However, researches on systematically evaluating and comparing network monitoring methods are very few. Especially, the capability of frequently collecting data with the assistance of modern measuring devices tends to induce autocorrelations among networks. Yet performance evaluation methods for autocorrelated networks are severely lacked. This paper proposes a performance evaluation method for network monitoring based on the separable temporal exponential random graph models, which is applicable to both independent and autocorrelated networks. Further, the effects of neglecting autocorrelations on the detection power of network control charts are studied as an application of the proposed method. The simulation results show the adverse effects of autocorrelations on performances of Shewhart, EWMA, CUSUM control charts for network density, and the residual control chart is suggested in the high autocorrelation scenarios. Following the guide, a residual control chart is applied to the analysis of the Enron email networks, and anomalous events are effectively detected.