AI Parameter Memory Remapping for Runtime Fault Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Advanced driver assistance systems (ADAS) and autonomous driving (AD) systems are vulnerable to memory faults, particularly permanent faults, which can cause prediction or classification errors, posing safety risks due to the reliance on AI functions for vehicle control.
Innovation Solution
A system monitors for memory faults and remaps sensitive parameters from faulty areas to fault-free areas using SPM and FAM maps, ensuring continuous operation and reducing the need for costly chip replacements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If memory faults are detected and handled by replacing chips, then system reliability is improved, but device complexity and cost increase
Solution Approach 1:
The patent implements dynamic remapping of memory parameters from faulty regions to healthy regions during runtime. The system continuously monitors memory health and adapts the mapping between logical parameter addresses and physical memory locations, allowing the system to respond dynamically to detected faults without requiring hardware replacement.
Solution Approach 2:
The patent creates redundant copies of critical parameter data in healthy memory regions. When faults are detected in original memory locations, the system can switch to using copied parameter values from backup locations, ensuring continuous operation without data loss or system failure.
2Reliability
If comprehensive memory scanning is performed to detect faults, then reliability is improved, but processing time increases
Solution Approach 1:
The patent extracts and monitors only the specific memory regions that contain critical AI parameters, rather than scanning the entire memory space. By focusing surveillance on parameter-storing regions identified through data structure analysis, the system achieves effective fault detection with minimal processing overhead.
Solution Approach 2:
The patent performs partial memory scanning focused on critical parameter regions rather than exhaustive full-memory scanning. The system monitors only the portions of memory that contain AI model parameters and affect safety-critical functions, achieving sufficient reliability for the application while significantly reducing the time and computational resources required.
3Measurement precision
If AI functions use millions of parameters stored in memory, then accuracy is improved, but vulnerability to memory faults increases
Solution Approach 1:
The patent segments the large set of AI parameters into multiple groups based on their criticality and storage location. By organizing parameters into distinct categories (critical safety parameters, non-critical parameters) and monitoring them differently, the system can isolate the impact of faults to specific segments while maintaining overall system accuracy.
Solution Approach 2:
The patent prepares protective measures in advance by identifying and monitoring critical parameter regions before faults occur. The system establishes health monitoring mechanisms and remapping capabilities beforehand, so when faults are detected, the system can quickly switch to backup parameter copies or alternative memory regions, cushioning against the harmful effects of memory faults on prediction accuracy.
Data Source
AI summary
In some examples, a system monitors for a memory fault associated with execution of software. Based at least on receiving an indication of a possible memory fault, the system scans a first area of the memory based on a data structure that indicates that the first area of the memory stores a portion of parameters of the software that affect an accuracy of the software if there is a memory fault in the first area of the memory. Based at least on the scanning indicating that there is a memory fault in the first area of the memory, the system remaps the portion of parameters of the software from the first area to a second area of the memory that is determined to be free.