Predictive Maintenance using Neural ODEs, Deep Koopman and Attribution Analysis (PANDA)
Abstract
Approved for Public ReleaseThe University of Texas at San Antonio is creating an integrated artificial intelligence solution for fault prediction and autonomous repair of maritime diesel engines and drivetrains. The PANDA (Predictive Maintenance using Neural ODEs, Deep Koopman and Attribution Analysis) project uses state-of-the-art neural ordinary differential equations (neural ODEs) to predict performance of diesel engines from high-dimensional time-series data. PANDA employs Fast Fourier Transforms (FFT) within the deeplearning framework to enable analysis of wide-band time series signals such as vibrations and sound. Region proposal networks (RPNs) are used to characterize the debris and temperature profiles in the exhaust of diesel engines to further facilitate performance predictions. The combination of neural ODEs, FFTs and RPNs allows PANDA to predict faults and other potential performance degradationsin maritime diesel engines and drivetrains while making simultaneous use ofhigh-dimensional time series data, wide-band time seriesdata, and video data sets.The PANDA project enables a new capability of long-term autonomous system deployment by automated system re-configuration using deep reinforcement learning (Deep RL) when a system failure or performance degradation is anticipated by an end-to-end deep neural network (DNN). The DNN performs prediction of the complex autonomous system using a combination of multi-scaleneural ODEs, wide-band time series analysis using FFTs and region proposal networks. Attribution analysis is used to identify one or more components of the system that may be responsible for an anticipated system degradation. The Rainbow Deep Q-Network (DQN) framework for reinforcement learning uses the space of possible system configurations as the set of possible actions and the predictionsfrom the end-to-end DNN as rewards under these possible actions to solve the sequential decision making problem of enabling a system to function even under one or more component failures. The DQN deep reinforcement learning approach uses Koopman operator theory, Generative Adversarial Networks (GANs), and Variational Auto-Encoders (VAEs) to learn the dynamics of the complex autonomous system from data and employs parameter as well as structural variations of the learned model to explore the space of possible faults in thesystem.There is an urgent need for the development of Artificial Intelligence (AI) approaches that enable the Navy to move to a flexible, adaptive, real-time, and data-driven maintenance system. The PANDA project provides an AI-based solution to this technological need of the Navy. PANDA will eventually enable long-term autonomy of underwater sea vehicles (USV) like the Sea Hunter by accepting its multi-scale data observations as an input to the PANDA system. The capabilities will be investigated using both real-world bigdata (few gigabytes) from marine diesel engines.PANDA predicts and diagnoses faults in complexmachinery using deep neural networks and a preliminary version has been successfully applied by our team to predict faults in a large gas turbine. Our proposed approach leverages several recent advances in deep reinforcement learning and deep neural networks, including Neural ODEs and attribution analysis, as well as deep learning adaptation of classical results such as the Koopman operator theory.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112332
Entities
People
- Sumit Kumar Jha
Organizations
- Office of Naval Research
- United States Navy
- University of Texas at San Antonio