A Blended System Dynamics-Discrete Event Physics-Based Model for Anomaly Detection in Cyber-Physical Manufacturing Systems

Well-instrumented modern factories with local computing power and network connectivity can be regarded as cyber-physical manufacturing systems (CPMS). Detecting anomalous behavior in CPMS is a major research challenge of strong commercial interest to factory stakeholders. Deviations to normal operations, including unplanned downtime, inefficient throughput rate, and excessive energy consumption, contribute to large losses in resources, time and money. A single hour of unplanned downtime in automotive, semiconductor, petrochemical, or other high-value/high-volume factories can lead to >US $1M of operational losses.

Traditionally, the identification of anomalous events in production lines has relied heavily on domain expertise. The experience of the expert factory engineers and equipment specialists has often formed the basis for manual failure analysis. As modern manufacturing systems grow more complex with increasing automation offering dynamic reconfiguration capabilities of sophisticated machines, it becomes less manageable for human experts to identify anomalies in a timely manner. The detection task is especially difficult for anomalous behavior from a system perspective where each component in the system does not necessarily show abnormal symptoms.

In recent decades, advances in machine learning and big data analytics have opened the possibilities to identify rare events from a huge amount of data. This approach applies well to situations where large volumes of labeled anomaly data are available. However, in factory applications, it is usually challenging or costly, if at all possible, to acquire such labeled anomaly data. A physics model-based approach, on the other hand, makes attempts to utilize the knowledge about the production line such as plant topology, process model, and energy consumption profiles. 

The data recorded by CPMS has heterogeneous sources containing both continuous variables such as temperature, air flow rate, and energy consumption and discrete variables such as machine status, number of products and number of defects. The heterogeneous nature of the production lines presents a challenge for physics-based system modeling.

This paper presents a methodology for a blended physics model-based anomaly detection for CPMS. The blended model consists of discrete event simulation (DES) components for the discrete manufacturing process modeling, and system dynamics (SD) components for continuous variables. Parameterized fault models are then introduced to consider the impact of the unreliability of machines. The methodology strikes a balance between the computational demand and the level of details required to perform anomaly detection tasks. The implementation of models takes an object-oriented approach, allowing multiple components of a smart factory to be robustly described in a modular, extendable and reconfigurable manner.

A prototype is developed and validated with a real-world dataset from a manufacturing plant. It is demonstrated that the developed model can be used for anomaly detection of system performance, and can be easily extended to consider different configurations of production lines.