Deep Neural Networks for Reliability Modeling

TECHNOLOGY AREAS: Weapons; Nuclear OBJECTIVE: The classic abduction/action/prediction models so prevalent in today's reliability and causality models take relatively long to process and are compute intensive. Explore the benefits of implementing Deep Neural Networks in the defense industry to tackle inaccurate reliability predictions. At completion of this SBIR, the prediction model should be able to make reliability predictions that will be validated during the next year of reliability data collection and be less memory and intensive by orders of magnitude over abduction/action/prediction modeling. By utilizing cutting-edge methods for analyzing complex data patterns, organizations can identify key factors contributing to reliability issues and make data-driven decisions to enhance overall system performance. Leveraging the power of deep learning algorithms enables organizations to forecast equipment reliability with greater accuracy, ultimately improving operational efficiency and minimizing downtime. Embracing these advanced techniques can revolutionize reliability predictions in the defense industry and drive significant improvements in system reliability and maintenance practices. DESCRIPTION: The Air Force Nuclear Weapons Center Air-Delivered Capabilities Directorate (AFNWC/ND) currently utilizes a simplistic Markov chain event model to derive B61-12 Tail Kit Assembly (TKA) reliability measurements. This approach lacks the specificity required to predict and inform on future reliability trends eliminating opportunities for the program office to prevent reliability issues before they are realized. PHASE I: As this is a Direct to Phase II topic, no Phase I awards will be made as a result of this topic. To qualify for this D2P2 topic, the Government expects the applicant to demonstrate feasibility by means of a prior Phase I-type" effort that does not constitute work undertaken as part of a prior SBIR/STTR funding agreement. Applicant shall demonstrate a case study or prototype of having performed similar/applicable work in developing applications. Documentation should include all relevant information including, but not limited to: technical reports, test data, prototype designs/models, and performance goals/results. PHASE II: 1. Assess existing Reliability Models for accuracy and completeness 2. Assess Air Delivered Integration Division (NDS) testing data (production, surveillance, test telemetry, etc.) 3. Derive predictive reliability modeling based on future progression of continuous time trajectories a. Develop with intuitive user interface to ease model runs b. Ensure rapid data ingestion techniques are available to populate the model 4. Devise a twin neural network consisting of two interlinked networks, one representing the real world and the other the counterfactual world, to inform future decision making a. Develop with intuitive user interface to ease neural net execution b. Ensure rapid data ingestion techniques are available to populate/retrain the neural nets 5. All deliverables shall run on a dedicated desktop computer. Milestones 1. Initial NDS reliability model and NDS data assessments; Contract Award (CA) plus 60 days 2. Initial predictive reliability model delivery; CA plus 90 days (w/iterative monthly deliveries) 4. Validation & verification of the predictive model; CA plus 360 days PHASE III DUAL USE APPLICATIONS: Capability can transition to other USAF weapon system program offices, but for this SBIR AFNWC/NDS will transition to a Phase III using the B61-12 sustainment program and FYDP programmed funding. REFERENCES: 1. Vlontzos, Athanasios; Kainz, Bernhard; Lee, Ciaran. "Estimating Categorical Counterfactuals via Deep Twin Networks". https://www.researchsquare.com/article/rs-1684942/v1. KEYWORDS: predictive reliability modeling; twin neural network; counterfactual data analysis