Satellite Data Processing Device Using Neural Network Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for mitigating software errors in digital circuits, such as triple redundancy and error correction codes, increase circuit size and may induce communication errors in bursts due to periodic correction, making them inefficient for handling software errors in satellite communication systems.
Innovation Solution
A data processing device with a restoration unit implemented using a denoising autoencoder neural network that learns to correct software errors, combined with a selection unit that compares feature quantities of input and restored signals to select the error-free signal, reducing the need for redundant circuitry and enabling real-time error correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional error correction methods (triple redundancy, error correction codes) are used, then software error correction capability is improved, but circuit size increases
Solution Approach 1:
The patent replaces traditional hardware-based error correction mechanisms (redundant circuits, ECC logic) with a software-based neural network model. The denoising autoencoder learns to correct bit errors through training on error patterns, substituting physical redundancy with intelligent software processing that achieves error correction without increasing hardware circuit size.
Solution Approach 2:
The patent changes the approach from static hardware redundancy to dynamic software-based correction. The neural network adapts its correction strategy based on learned error patterns, changing the parameter of error correction from fixed hardware logic to flexible software processing that can be updated and adapted without hardware modification.
2Reliability
If periodic scrubbing correction is used, then error correction is achieved, but communication errors in bursts may occur within correction periods
Solution Approach 1:
The patent implements continuous error correction capability through the neural network, which processes and corrects errors in real-time as data flows through the system. Unlike periodic scrubbing that operates at discrete intervals, the denoising autoencoder provides continuous correction, eliminating the vulnerability windows where burst errors could occur between correction cycles.
Solution Approach 2:
The neural network is pre-trained on various error patterns and scenarios, preparing it to recognize and correct errors before they propagate through the system. This preliminary learning action enables the network to immediately identify and correct errors as they occur, rather than waiting for periodic detection and correction cycles.
3Reliability
If dedicated space-use devices are used, then resistance to cosmic rays is improved, but cost and performance are reduced compared to consumer devices
Solution Approach 1:
The patent introduces a software intermediary (the denoising autoencoder neural network) that mediates between the vulnerable consumer-grade hardware and the requirement for high reliability. The neural network acts as a protective layer that detects and corrects cosmic ray-induced errors, allowing the use of cheaper consumer devices while achieving the reliability previously requiring expensive space-grade hardware.
Solution Approach 2:
The patent enables the use of inexpensive consumer-grade FPGAs and processors in space applications by compensating for their lower inherent radiation hardness through software-based error correction. This allows the system to use cheap, high-performance consumer devices rather than expensive space-grade components, achieving cost savings without sacrificing reliability.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
A data processing device (20) includes a restoration unit (22) that performs a conversion operation on an input signal to convert the input signal into a signal having no distortion caused by an external factor, and a selection unit (23) that selects and outputs either an unrestored signal, which is the input signal, or a restored signal, which is a signal obtained by the restoration unit (22) by performing the conversion operation, based on a feature quantity of the unrestored signal and on a feature quantity of the restored signal.