SBIR-STTR Award

Goal is to develop intelligent system tools that learn normal patterns of life from energy consumption auditing of both cyber and manufacturing devices in manufacturing systems, and use a hybrid machine-learning (ML) and a digital-twin (DT) approach to learn and correlate changed patterns from physical and cyber threats. Unknown anomalies in a manufacturing machine will be detected and characterized by the Energy Consumption Abnormality Detection (ECAD) prototype system based upon the DF&NN Goal-Driven Condition-Based Predictive Maintenance (GCPM) baseline Condition-Based Maintenance (CBM). The DF&NN-QSI team will apply our ISA tools to generate temporally overlapping known and unknown manufacturing system and energy consumption abnormality detection and historical abnormality categorization event tracks. We will apply our Smoking Gun and TEAMS tools to discover correlation relationships in these events and use these to improve cause diagnosis and determine the effectiveness of using energy consumption data to detect cyber and physical attacks. The anomalies form inputs to QSI's TEAMS® models that capture system-agnostic functional failure-cause and effect dependency relationships. The TEAMS® model facilitates mapping these anomalies to the causal model thereby allowing TEAMS® runtime reasoning engines to perform failure root-cause isolation and corrective/preventive action determination when such anomalies are detected.

The focus problem addressed in this effort is to automatically learn historical normal Computer Numerical Control (CNC) system behavior to produce the following: unmodeled unknown CNC abnormal behavior detections with historical classification characterizations if seen historically abnormal behavior and similar abnormal class characterization flags with abnormality and classification scores categorization of similar normal unmodeled from historical behavior in real-time CNC behavior with category trust scores to inform the user how similar the current real-time abnormal declared class signature is to signatures that the categorization neural networks were trained on recommended responses to the abnormality detections and valid categorizations We plan to apply the ECAD Deep Multi-Start Residual Training (D-MSRT) NNs, Smoking Gun, and maintenance condition categorization D-MSRT NNs capabilities for as many aviation systems as available. We will train D-MSRT abnormality detection NNs to learn the labeled repair conditions that were used for each categorization NN to provide a categorization NN result trust score to the user. We will incorporate into ECAD our existing goal-driven turnkey NN capabilities that determine when to retrain, what data to retrain on, what data to test on, how to evaluate, and when to promote to on-line operations. This allow ECAD to automatically evolve and improve its performance based on progressing user goals. We will adapt the ECAD graphical user interfaces (GUI) for user-tier roles with a standardized software deployment approach designed for ease of deployment. ECAD will provide early detection and characterization of system anomalies and component failures. DF&NN will deliver ECAD Docker REpresentational State Transfer (REST) API services which support sharing of NNs and results across distributed operations. We will use these to validate ECAD performance and increase user trust in ECAD results. We will work closely with the sponsor to identify operational transition opportunities. ECAD will not be a black box solution. ECAD will provide performance sensitivity analyses and declaration confidences. ECAD will provide a system that develops trust with operators and provides CBM capabilities. A sample use case is for historical T4 measurand behavior to be learned by TrnSatDP. Historical abnormal signatures are automatically clustered, named, and tracked. These are shown in ADCV for the user to flag those of interest. Categorization D-MSRT NNs are trained to be able to flag these signatures in real-time when they occur again. Trust D-MSRT NNs are trained to provide a confidence in the categorization NNs class declarations. All 3 of these NNs are run in real-time to detect the unknown unexpected abnormal behaviors, classify signatures that need to be found, and provide trust scores.