Cross-session brainprint recognition method based on tensorized spatial-frequency attention network (TSFAN) with domain adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional EEG-based biometric recognition methods face challenges in cross-session stability due to domain shifts caused by factors like electrode position changes and subject status variations, leading to unsatisfactory performance in real-world scenarios.
Innovation Solution
A cross-session brainprint recognition method using a tensorized spatial-frequency attention network (TSFAN) with domain adaptation, which maps source and target domain data to different time feature spaces, employs tensor-based attention to capture domain-invariant spatial-frequency features, and uses a low-rank Tucker format to manage dimensionality, facilitating stable feature extraction across sessions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional deep learning methods (CNN, RNN, GCNN) are used to extract EEG features, then identity authentication features can be obtained in temporal, frequency, and spatial domains, but cross-session recognition performance deteriorates due to domain shifts from electrode position changes and subject status variations
Solution Approach 1:
The patent segments the EEG signal processing into distinct temporal and spectral components using Short-Time Fourier Transform (STFT), dividing the continuous signal into time-frequency segments that can be independently analyzed and processed to capture stable cross-session features
Solution Approach 2:
The patent transforms the EEG feature extraction problem from traditional temporal-spatial domains into a tensor-based multi-dimensional space, where domain adaptation can be applied across multiple dimensions (temporal, spectral, spatial) simultaneously to achieve better cross-session generalization
2Quantity of substance
If data from multiple training sessions are mixed to improve recognition, then more training data is available, but distribution differences between sessions create domain shifts that reduce performance
Solution Approach 1:
The patent introduces domain adaptation as an intermediary mechanism that mediates between multiple source domains and the target domain. The domain adapter learns to align feature distributions across domains, enabling effective use of multi-session training data while compensating for distribution differences through learned domain-invariant representations
Solution Approach 2:
The patent dynamically adjusts model parameters through domain adaptation, learning domain-specific and domain-invariant feature representations by modifying the feature extraction process to account for session-specific variations while maintaining core identity-related features across domains
3Reliability
If domain adaptation is applied to minimize differences between source and target domains, then domain-invariant features are captured, but the domain-invariant feature is affected by the involved source domain and cannot benefit from common relationships of multiple source domains
Solution Approach 1:
The patent merges features from multiple source domains through a tensor-based fusion mechanism that captures both individual domain characteristics and common relationships across domains. The tensor representation allows simultaneous processing of multiple domains, combining their unique features while identifying shared patterns that generalize to the target domain
4Reliability
If tensor-based attention is used to capture domain-invariant spatial-frequency features, then cross-session stability is improved, but computational complexity increases due to high-dimensional tensor operations
Solution Approach 1:
The patent segments the high-dimensional tensor operations into smaller, manageable components by decomposing the attention mechanism into separate temporal, spectral, and spatial dimensions. This segmentation allows efficient processing of each dimension independently while maintaining the overall tensor structure for capturing domain-invariant features
Data Source
AI summary
The present disclosure discloses a cross-session brainprint recognition method based on a tensorized spatial-frequency attention network (TSFAN) with domain adaptation. For most of existing multi-source domain adaptation methods, domain gaps between multiple source domains and target domains are individually bridged, but a relationship between domain-invariant features in distribution alignment is ignored. The present disclosure assists performance of a target domain by modeling an important relationship of the domain-invariant features without being affected by a distribution difference between source domains. A new TSFAN is used to combine pairwise source and target and an appropriate common spatial-frequency feature across source domains. Considering of a dimension, the TSFAN is further approximated as a low-rank Tucker format, to enable the TSFAN to adapt to scale linearly in a quantity of domains, and apply the TSFAN to a case of any quantity of sessions.