Neural Network Repair Agent for Radiation Damage Mitigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Radiation damage poses a significant threat to neural networks deployed in aerospace applications, leading to errors in processing, and traditional solutions like redundant hardware and radiation-hardened chipsets increase cost and complexity.
Innovation Solution
A software-based solution involving a primary neural network and a repair agent, where the repair agent is trained to detect and mitigate damage by selecting and performing repair actions, such as adjusting node weights or connections, to maintain the primary network's functionality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If redundant hardware is used to compensate for radiation damage, then reliability is improved, but device complexity and cost increase
Solution Approach 1:
The patent creates a copy of the neural network model stored in memory, which serves as a backup that can be restored if the primary model is damaged by radiation. This copying approach provides reliability without requiring redundant hardware systems, as the backup model is stored as data rather than duplicated hardware.
Solution Approach 2:
The patent replaces traditional hardware redundancy mechanisms with a software-based model restoration system. Instead of using multiple physical hardware systems that vote or replicate functionality, the system uses a stored neural network model that can be restored from memory, substituting mechanical/hardware redundancy with information-based redundancy.
2Reliability
If radiation-hardened chipsets are used, then reliability is improved, but manufacturing cost increases
Solution Approach 1:
The patent uses a stored copy of the neural network model in memory as a backup mechanism. This copying approach is significantly cheaper to implement than radiation-hardened chipsets, as it involves storing model data rather than fabricating specialized radiation-resistant hardware.
Solution Approach 2:
The patent employs a cost-effective backup model that can be restored from memory, treating the backup as a disposable data copy rather than investing in expensive radiation-hardened hardware. This approach prioritizes cost-effective protection over permanent hardware fortification.
3Reliability
If traditional hardware redundancy is implemented, then reliability is improved, but weight and resource consumption increase
Solution Approach 1:
The patent stores a copy of the neural network model in memory, providing a lightweight backup mechanism that does not add significant physical weight. The backup is stored as data (weights and architecture information) rather than as duplicate hardware, making the system much lighter than traditional redundant hardware approaches.
Solution Approach 2:
The patent replaces heavy physical hardware redundancy with a data-based restoration system. The backup model is stored as information in memory, eliminating the need for additional physical components and significantly reducing the overall system weight while maintaining reliability.
Data Source
AI summary
Aspects of the disclosure mitigate effects of damage to neural networks (NNs) onboard a platform using a primary NN trained to perform a primary task and a repair agent trained to repair the primary NN. The repair agent performs the steps of detecting a degradation of the primary NN's ability to perform the primary task and performing a repair action to repair the primary NN. The primary task is then performed by the repaired primary NN.