Uncertainty-Aware Virtual Sensors for Cyber-Physical Systems

Abstract

Virtual sensors in Cyber-Physical Systems (CPS) are AI replicas of physical sensors that can mimic their behavior by processing input data from other sensors monitoring the same system. However, we cannot always trust these replicas due to uncertainty ensuing from changes in environmental conditions, measurement errors, model structure errors, and unknown input data. An awareness of numerical uncertainty in these models can help ignore some predictions and communicate limitations for responsible action. We present a data pipeline to train and deploy uncertainty-aware virtual sensors in CPS. Our virtual sensor based on a Bayesian Neural Network (BNN) predicts the expected values of a physical sensor and a standard deviation indicating the degree of uncertainty in its predictions. We discuss how this uncertainty awareness bolsters trustworthy AI using a vibration-sensing virtual sensor in automotive manufacturing. CYBER-PHYSICAL SYSTEMS (CPS) have been increasingly implanted in recent years withArtificial Intelligence (AI) and, in particular, Deep Learning Models (DLM). DLMs are arguably the only viable tool purported to make accurate decisions and predictions from a juggernaut of data streams generated by devices interconnected through the Internet of Things (IoT) for sensing, computation, control, actuation, and networking. Virtual sensors (also called soft sensors) [1] are DLMs that are AI replicas (software artifacts replicating the output of a physical component/sensor of a CPS by learning its correlated behavior with one or more different physical artifacts in the CPS) [2] of potentially billions of physical sensors (e.g., pressure, temperature, humidity, speed, force, vibration, and position sensors) deployed in CPS. They can kick in for degrading physical sensors operating in harsh environments [3]. For instance, a virtual sensor