In modern society, the management of attention has become increasingly complex and important. To understand the attention state of drivers during driving, we have been looking for effective identification and monitoring methods until the development of Brain-Computer Interface (BCI) technology. Electroencephalogram (EEG) has gradually become an indispensable tool in neuroscience research. However, traditional EEG technology usually requires many electrode channels to obtain sufficient information, which not only increases the cost but also increases the technical complexity. To overcome these challenges, this study uses EEG data of driver attention and adopts a new Convolutional Neural Network (CNN) model called EEGNet for model training and testing. Through optimizing parameter settings and appropriate feature extraction, as well as selecting necessary channels and adjusting the appropriate number of channels, we evaluate the performance of the method in terms of accuracy and practicality. This is to achieve accurate identification of the driver's brain attention under fewer channels and to explore the impact of model parameters and the number of channels on accuracy. We expect that this brain attention identification method based on the EEGNet model with fewer channels can achieve higher accuracy while reducing equipment costs and system complexity. This provides a more economical and effective solution for the development of intelligent driving systems, and also provides new possibilities for the application of BCI technology in the field of intelligent transportation.
In this study, we introduce a novel, functionally driven method for model pruning guided by sensitivity analysis. Conventional model compression techniques often rely on proxy metrics, such as weight magnitude, which may not accurately reflect a component’s true functional importance. The proposed method directly assesses this by systematically injecting controlled noise into network layers and measuring the resultant perturbation on inference output. Components exhibiting low sensitivity to this noise are identified as functionally redundant and are pruned. We validated the method on EEGNet, a compact convolutional neural network, using the MNE Sample Event-Related Potential (ERP) dataset, a widely used benchmark for electroencephalography classification. After training the baseline model, we generated a sensitivity profile by quantifying how noise injection at different layers impacts predictive accuracy. This profile then guided targeted pruning of less influential convolutional kernels and weights. Experimental results demonstrate the method’s efficacy, achieving a significant reduction in both parameter count and computational complexity. Crucially, the pruned model retains classification accuracy nearly identical to the original, heavyweight EEGNet. This confirms that sensitivity-guided pruning effectively removes redundancy without degrading performance. In conclusion, our noise injection framework provides a more direct and interpretable criterion for neural network simplification. By linking component pruning to functional impact, our method enables a more precise and efficient model reduction than traditional heuristic-based approaches. The method developed presents a practical pathway toward developing lightweight, accurate, and low-latency models essential for real-world neuro-computational applications. Future work will focus on automating the pruning pipeline and extending the framework’s applicability to diverse neural architectures.
The electroencephalogram (EEG) is crucial in both medical and research categories, providing insights into continuous and irregular electrical potential fluctuations on the scalp surface through Brain-Computer Interface (BCI). The processing time required by BCI in handling EEG data is a significant consideration, as it can impact not only researchers' result interpretations but also play a crucial role in medical professionals' assessment of patients' conditions. The EEGNet is an improved convolutional Neural Network (CNN) architecture designed to classification of specific brainwave patterns in EEG. The original version of EEGNet is one dimensional convolution kernels. Multi-channel data are gathered by electrodes sticked on scalp. However, the original EEGNet pay most attention on the horizontal and vertical features one by one channel. Some features across channels will be omit. Moreover, the processing time are varying for 2D kernels size. This study demonstrates the impact of 2D convolution kernels on accuracy and processing time. Our research shows that 2D convolution kernels enhances accuracy while reducing processing time.
In semiconductor and optoelectronics manufacturing, the reliability of turbo molecular pumps (TMPs) is vital for maintaining vacuum integrity and ensuring product yield. However, acoustic monitoring in cleanrooms faces severe challenges due to ambient noise levels routinely exceeding 80 dBA and the scarcity of labeled anomaly data. This study proposes an unsupervised acoustic anomaly detection and localization system to address these issues. The performance of the proposed framework is evaluated using an acoustic dataset collected from operational turbopumps in an industrial semiconductor cleanroom environment, encompassing both normal operations and naturally occurring failure conditions. We introduce a Convolutional Autoencoder (CAE) based on ResNet-18, integrated with a Convolutional Block Attention Module (CBAM) to adaptively suppress high-decibel environmental noise. To enhance sensitivity to structural spectral defects, a hybrid loss function combining Mean Squared Error (MSE) and structural similarity index measure (SSIM) is implemented. Experimental results, supported by rigorous hyperparameter sensitivity analysis, demonstrate that the proposed model achieves an outstanding AUC of 0.9535 and a fault recall of 99.06% under a validation-calibrated threshold, significantly outperforming standard U-Net architectures. Furthermore, the system generates anomaly heatmaps for precise time–frequency localization, enabling explainable diagnostics. With a model inference throughput of 330.28 FPS and an end-to-end processing rate of ≈73× in real time, the proposed framework provides an efficient and robust solution for real-time predictive maintenance in noisy industrial settings.
Deploying EEG classification networks on resource-constrained brain–computer interface (BCI) hardware motivates structured filter pruning, yet which importance criterion to trust in already-compact architectures is unclear. We present a leakage-free, physically materialized comparison of five filter-importance criteria—ablation-based structural sensitivity analysis (SSA), first-order Taylor, L1-norm, geometric median, and random—across four datasets, two paradigms (motor imagery and ERP), two architectures, multiple capacities, and with and without fine-tuning. On EEGNet, after fine-tuning, SSA and Taylor are statistically equivalent: pooled over the four datasets their accuracy difference falls within a percentage-point margin (TOST, ), and on ShallowConvNet SSA is never worse than Taylor; both criteria significantly outperform the weight-based and random ones. SSA’s value is therefore not a more accurate ranking but its formulation: on a weight-quantized (8-bit) EEGNet, where the loss gradient—and hence Taylor—is undefined on the quantized graph used here, SSA is the best gradient-free criterion, significantly beating L1, geometric-median, and random pruning in aggregate—extending structured pruning to the quantized, gradient-free but editable deployment regime. We further show that the pruning schedule and fine-tuning matter more than the choice between strong criteria, and that training large then pruning with fine-tuning recovers accuracy equivalent to a directly trained smaller model. The results clarify where the importance criterion does and does not matter for compact EEG decoders.