Protograph Neural Decoder Training for Faster LDPC Convergence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing protograph low-density parity-check (LDPC) codes face high complexity and low training efficiency due to the use of single-stream structures and issues with convergence during iterative decoding, particularly with variable points of degree 1 leading to poor network training efficiency.
Innovation Solution
An improved protograph neural decoder is developed, comprising an initial variable sub-network layer, an initial check sub-network layer, and a preset shuffled belief-propagation (BP) sub-network layer, trained using a mean square error loss function and decoder objective function to optimize the network structure, increasing shunt and accelerating convergence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single-stream structure is used for decoding non-regular LDPC codes, then the decoder can handle complex code structures, but the complexity increases and hardware implementation becomes difficult
Solution Approach 1:
The decoder is divided into multiple parallel streams, each handling a subset of variable nodes. This segmentation allows the complex decoding task to be distributed across multiple simpler processing units, reducing the complexity of each individual unit while maintaining the ability to handle non-regular LDPC codes.
Solution Approach 2:
The patent transitions from a single-stream sequential processing approach to a multi-stream parallel processing architecture. This dimensional change from 1D to 2D (multiple streams) enables simultaneous processing of multiple code blocks, improving throughput and reducing per-unit complexity.
2Reliability
If variable points of degree 1 are included in LDPC code construction, then the code can achieve better performance, but convergence failures occur during iterative decoding and network training efficiency deteriorates
Solution Approach 1:
Variable nodes of degree 1 are extracted and removed from the code construction process. By taking out these problematic elements that cause convergence failures and training inefficiency, the patent achieves stable iterative decoding convergence and improved network training efficiency while maintaining overall code performance through alternative code design strategies.
3Device complexity
If traditional NMS or OMS decoders are used, then computational complexity is reduced, but decoding performance deteriorates compared to BP decoder
Solution Approach 1:
The patent applies different processing qualities to different parts of the decoding process. Within each stream, simplified operations (like NMS/OMS) are used for regular processing, while neural network-based processing is applied selectively at critical decision points. This local differentiation maintains low overall complexity while improving performance at key stages.
Solution Approach 2:
The decoder combines multiple approaches into a composite structure: traditional NMS/OMS decoding algorithms are integrated with neural network components. This composite architecture leverages the low complexity of traditional methods while incorporating the performance benefits of neural networks, achieving a balance between complexity and performance.
Data Source
AI summary
A training method based on an improved protograph neural decoder includes the following steps: a to-be-trained decoding network is constructed based on an initial variable sub-network layer, an initial check sub-network layer and a preset shuffled belief-propagation (BP) sub-network layer; the initial variable sub-network layer, the initial check sub-network layer and the preset shuffled BP sub-network layer are updated and trained by calculating log-likelihood ratio (LLR) based on a preset mean square error loss function and a preset decoder objective function to obtain a target protograph neural decoder; and the preset mean square error loss function is configured to calculate a loss value between output information of the check sub-network layer and the preset shuffled BP sub-network layer. The target protograph neural decoder includes an optimized variable sub-network layer, an optimized check sub-network layer and an optimized shuffled BP sub-network layer. A training device is also provided.