A Deep Learning Framework for Real-Time Health and Security Monitoring and Diagnosis of Manufacturing Systems Based on Energy Consumption Auditing

The proposed effort aims to develop and demonstrate a unified framework based on deep autoencoders for real-time health and security monitoring and diagnosis of threats in manufacturing systems via energy consumption auditing. The salient aspects of the proposed solution are: (1) unsupervised learning for identification of nominal operating dynamics of manufacturing devices; (2) anomaly detection based on the observation of discrepancy between actual system inputs and the reconstructed system inputs; (3) root cause analysis for determining main contributors to the anomaly; and (4) innovation in software-hardware architecture to streamline neural network learning, model construction, and root cause analysis. In Phase I, key components including a deep autoencoder module, an anomaly detection module, a root cause analysis module, and a holistic architecture integrating these modules will be developed. Feasibility will be demonstrated via case studies of US Air Force interest, including real-time threat detection and diagnosis demonstrations in a controlled, laboratory manufacturing system. Nominal operational data will be collected for model training and cyber and physical threats will be simulated using a variety of techniques. The Phase II effort will focus on capability extension, algorithm optimization, software integration, extensive technology validation and demonstration, and technology insertion into Air Force’s workflow.