Cross-Task Mental Workload Recognition Based on EEG Tensor Representation and Transfer Learning

Guan, Kai; Zhang, Zhimin; Liu, Tao; Niu, Haijun · 2023 · Crossref

DOI: 10.1109/tnsre.2023.3277867

archive: archived pipeline: cataloged

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

**Research Problem and Motivation** Accurate evaluation of mental workload (MWL) in human-machine systems is critical for operator safety and task execution. While Electroencephalography (EEG) is a sensitive physiological indicator for MWL, existing methods often fail in cross-task scenarios. This limitation arises because EEG response patterns vary significantly across different tasks due to distinct brain information processing mechanisms, leading to feature distribution shifts that degrade recognition accuracy to random levels. Classical machine learning methods, which assume independent and identically distributed data, are ill-suited for these cross-task variations. Furthermore, traditional feature extraction methods typically process EEG as single-dimensional vectors, ignoring the structural relationships between time, frequency, and channel dimensions that are sensitive to workload levels. **Methodology** This study proposes a cross-task MWL recognition framework combining EEG tensor representation with transfer learning. The experimental design involved 16 participants (8 male, 8 female) performing four types of N-Back working memory tasks (verbal, object, space-verbal, space-object) at three difficulty levels (N=1 to 3), resulting in 12 distinct task conditions. EEG signals were recorded using a 60-channel NeuroScan system. The method constructs three-way EEG tensors (time-frequency-channel) using wavelet transform for time-frequency analysis. To address cross-task distribution shifts, a tensorized feature transfer learning algorithm is employed. This algorithm optimizes a cost function that simultaneously minimizes distribution alignment (using Maximum Mean Discrepancy) and maximizes class-wise discrimination (using General Tensor Discriminative Analysis). The optimization is solved via an alternating iteration strategy using High Order Singular Value Decomposition. Finally, Fisher Score-based feature selection is applied, and a Support Vector Machine (SVM) is used to construct the 3-class MWL recognition model. **Results** The proposed method was validated against classical feature extraction methods in both within-task and cross-task scenarios. The results demonstrated that the EEG tensor representation and transfer learning approach achieved significantly higher accuracy than baseline methods. Specifically, the model achieved a recognition accuracy of 91.1% for within-task evaluation and 81.3% for cross-task evaluation. These findings indicate that the tensor-based transfer learning effectively reduces the distance between feature distributions of different tasks while maintaining discriminative power for workload levels. **Significance** The study demonstrates that representing EEG data as tensors and applying transfer learning is a feasible and effective strategy for cross-task mental workload evaluation. By preserving multi-dimensional structural information and aligning feature distributions across tasks, the method overcomes the generalization limitations of traditional approaches. This provides a theoretical basis and practical reference for developing robust, real-time physiological monitoring systems that can accurately assess operator workload across diverse operational contexts.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.8-27b-gittensor on 2026-08-23; verification: pending re-verification.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices