29 April 2022 in IEEE Transactions on Neural Networks and Learning Systems
Recently, electroencephalogram (EEG) emotion recognition has gradually attracted a lot of attention. This brief designs a novel frame-level teacher-student framework with data privacy (FLTSDP) for EEG emotion recognition. The framework first proposes a teacher-student network without prior professional information for automated filtering of useful frame-level features by a gated mechanism and extracting high-level features by using knowledge distillation to capture the results of EEG emotion recognition from a teacher network and student networks. Then, the results from subnetworks are integrated by using the novel decision module, which, motivated by the voting mechanism, adjusts the composition of feature vectors and improves the weight of accurate prediction to optimize the integration effect. During training, an innovative data privacy protection mechanism is applied for avoiding data sharing, where each student network only inherits weights from all trained networks and does not inherit the training dataset. Here, the framework can be repeatedly optimized and improved by only training the next student subnetwork on new EEG signals. Experimental results show that our framework improves the accuracy of EEG emotion recognition by more than 5% and gets state-of-the-art performance for EEG emotion recognition in the subject-independent mode.
4 December 2020 in IEEE Transactions on Cognitive and Developmental Systems
In real scenes, humans can easily infer their positions and distances from other objects with their own eyes. To make the robots have the same visual ability, this paper presents an unsupervised OnionNet framework, including LeafNet and ParachuteNet, for single-view depth prediction and camera pose estimation. In OnionNet, for speeding up OnionNet’s convergence and concretizing objects against the gradient locality and moving objects in videos, LeafNet adopts two decoders and enhanced upconvolution modules. Meanwhile, for improving the robustness of fast camera movement and rotation, ParachuteNet uses and integrates three pose networks to estimate multi-view camera pose parameters by combining with the modified image preprocess. Different from existing methods, single-view depth prediction and camera pose estimation are trained view-by-view, where the variations between views is gradual reduction of view range and outer pixels disappear in next view, similar to onion peeling. Moreover, the LeafNet is optimized with pose parameter from each pose network in turn. Experimental results on the KITTI dataset show the outstanding effectiveness of our method: single-view depth performs better than most supervised and unsupervised methods which contain two same subtasks, and pose estimation gets the state-of-the-art performance compared with existing methods under the comparable input settings.
25 February 2022 in IEEE Transactions on Neural Networks and Learning Systems
In multilabel images, the changeable size, posture, and position of objects in the image will increase the difficulty of classification. Moreover, a large amount of irrelevant information interferes with the recognition of objects. Therefore, how to remove irrelevant information from the image to improve the performance of label recognition is an important problem. In this article, we propose a convolutional network based on feature denoising and details supplement (FDDS) to address this issue. In FDDS, we first design a cascade convolution module (CCM) to collect spatial details of upper features, in order to enhance the information expression of features. Second, the feature denoising module (FDM) is further put forward to reallocate the weight of the feature semantic area, in order to enrich the effective semantic information of the current feature and perform denoising operations on object-irrelevant information. Experimental results show that the proposed FDDS outperforms the existing state-of-the-art models on several benchmark datasets, especially for complex scenes.
9 September 2025 in IEEE Sensors Journal
In the industrial scene, mechanical equipment often faces dynamic changes of speed, load and other working conditions, so accurate diagnosis of rolling bearing fault is still a key challenge in industrial condition monitoring, in which the scarcity of marker data and cross domain differences seriously limit the generalization of the model. Conventional unsupervised domain adaptation approaches predominantly focus on data classification and domain-specific annotation, often neglecting the extraction of inherent structural characteristics present in the data. To address this issue, we propose the implementation of a Multi-Scale Feature Fusion Graph Convolutional Network for UDA. The classifier and domain discriminator are responsible for processing class-related and domain-specific label information, respectively. Regarding structural representation, a multi-scale convolutional architecture is first utilized to derive feature representations from the input signals, producing the feature matrix (X). In the subsequent phase, the FGGL (fusion graph generation layer) investigates the latent structural patterns of samples by generating a topological matrix (A). The sample graph representation is subsequently constructed using the MRF-GCN (multi-reference feeler-graph convolutional network) architecture. During the final stage, the maximum mean discrepancy (MMD) metric is employed to reduce domain discrepancy through the alignment of structural patterns observed in the sample graphs. The experimental findings demonstrate that the MFGCDL framework outperforms existing UDA approaches in terms of both stability and precision for rolling bearing fault detection, while successfully capturing transferable cross-domain characteristics.
27 November 2025 in Journal of Integrative Neuroscience
Background: This study addressed three key challenges in subject-independent electroencephalography (EEG) emotion recognition: limited data availability, restricted cross-domain knowledge transfer, and suboptimal feature extraction. The aim is to develop an innovative framework that enhances recognition performance while preserving data privacy. Methods: This study introduces a novel multi-teacher knowledge distillation framework that incorporates data privacy considerations. The framework first comprises n subnets, each sequentially trained on distinct EEG datasets without data sharing. The subnets, excluding the initial one, acquire knowledge through the weights and features of all preceding subnets, enabling access to more EEG signals during the training process while maintaining privacy. To enhance cross-domain knowledge transfer, a multi-teacher knowledge distillation strategy was designed, featuring knowledge filters and adaptive multi-teacher knowledge distillation losses. The knowledge filter integrates cross-domain information using a multi-head attention module with a gate mechanism, ensuring effective inheritance of knowledge from all previous subnets. Simultaneously, the adaptive multi-teacher knowledge distillation loss dynamically adjusts the direction of knowledge transfer based on filtered feature similarity, preventing knowledge loss in single-teacher models. Furthermore, a spatio-temporal gate module is proposed to eliminate unnecessary frame-level information from different channels and extract important channels for improved feature representation without requiring expert knowledge. Results: Experimental results demonstrate the superiority of the proposed method over the current state of the art, achieving a 2% performance improvement on the DEAP dataset. Conclusions: The proposed multi-teacher distillation framework with data privacy addresses the challenges of insufficient data availability, limited cross-domain knowledge transfer, and suboptimal feature extraction in subject-independent EEG emotion recognition, demonstrating strong potential for scalable and privacy-preserving emotion recognition applications.