AUV Channel Estimation Using Self-Attention Feature Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing channel estimation methods for autonomous underwater vehicles (AUVs) face challenges in maritime environments due to high computational complexity and difficulty in implementing deep learning-based networks, which are designed for terrestrial communication and do not adequately address the complex and changing electromagnetic conditions.
Innovation Solution
A multi-dimensional feature aggregation network using a self-attention mechanism and an interpolation method based on DPA estimation is employed for channel estimation, aggregating spatial and channel features through spatial and channel self-attention blocks, and upsampling pilot signals to enhance estimation performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If traditional LS estimation is used, then the implementation is simple and easy, but the estimation performance cannot meet the requirements in complex marine electromagnetic environments
Solution Approach 1:
The patent introduces a deep learning-based channel estimation network as an intermediary between the received signal and channel estimation result. This network processes the input signal through multiple layers of feature extraction and aggregation, achieving high estimation performance in complex marine environments while maintaining reasonable computational complexity through optimized architecture design
2Measurement precision
If MMSE estimation is used, then the estimation performance is extremely high, but the computational complexity is very high involving two matrix inversions
Solution Approach 1:
The patent replaces the traditional mathematical optimization-based MMSE estimation (which requires complex matrix operations) with a deep learning-based estimation network. The network learns the channel characteristics through training data and directly outputs channel estimates, substituting the mechanical computation of matrix inversions with a trained neural network model that has lower real-time computational complexity
Solution Approach 2:
The patent performs channel estimation offline by collecting training data and training the deep learning network in advance. The trained model parameters are stored and reused during actual communication, avoiding the need to perform complex real-time calculations during operation. This preliminary action transfers the computational burden from runtime to training time
3Measurement precision
If deep learning methods such as ChannelNet and CRCENet are used, then the estimation performance is much higher than traditional methods, but the computational complexity is extremely high and difficult to implement on AUV hardware devices
Solution Approach 1:
The patent segments the channel estimation task into distinct functional modules: a feature extraction module that processes input signals, a multi-dimensional feature aggregation module that combines features from different dimensions, and an output module that generates channel estimates. This modular segmentation allows for optimized computation at each stage and reduces overall complexity
Solution Approach 2:
The patent aggregates features from multiple dimensions including spatial dimension, frequency dimension, and temporal dimension. By processing and combining features across these different dimensions, the network achieves high estimation performance while maintaining computational efficiency through dimension-wise operations rather than exhaustive full-dimensional processing
Data Source
AI summary
The invention discloses method of improving AUV wireless communications systems, including steps of: collecting the channel frequency response and receiving signal data in the electromagnetic environment where AUV works on the water surface,; constructing a multi-dimensional feature aggregation network model based on self-attention mechanism; training a multi-dimensional feature aggregation network; and preprocessing received AUV signals to obtain improved input data. The network model completed by offline training is loaded, and the input data is input into the network for channel estimation. The invention uses a self-attention mechanism in deep learning to build a multi-dimensional feature aggregation network model. The estimation performance is much higher than the traditional channel estimation method, and the space-time complexity of the network is very low, which can be applied to an AUV wireless communication system through offline training.