Edge-Fog Computing-Enabled EEG Data Compression via Asymmetrical Variational Discrete Cosine Transform Network
Xin Zhu, Hongyi Pan, Ahmet Enis Cetin · 2025
Read Edge-Fog Computing-Enabled EEG Data Compression via Asymmetrical Variational Discrete Cosine Transform Network by Xin Zhu, Hongyi Pan, Ahmet Enis Cetin (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-09961.txt.
Introduction
Xin Zhu, Hongyi Pan, and Ahmet Enis Cetin compress brain-wave recordings so brain-computer interfaces can send them over thin network links. Their 2025 paper works within edge-fog computing, where light devices near the user do simple work and a nearby fog layer does heavier lifting. Raw EEG data is bulky, and slow transmission bottlenecks connected health systems.
At the edge, trainable cosine-transform units strip redundancy and extract a compact latent form of each signal. At the fog layer, a filter bank borrows strength across channels, which carry related information. An attention-based decoder then rebuilds the signals, guided by a variational objective balancing size against fidelity. On two public data sets, the method beat rivals on compression without losing useful content.
Catalog record
| CREATOR | Xin Zhu, Hongyi Pan, Ahmet Enis Cetin |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-09961
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Edge-Fog Computing-Enabled EEG Data Compression via Asymmetrical Variational Discrete Cosine Transform Network Xin Zhu, Hongyi Pan, Ahmet Enis Cetin Abstract The large volume of electroencephalograph (EEG) data produced by brain-computer interface (BCI) systems presents challenges for rapid transmission over bandwidth-limited channels in Internet of Things (IoT) networks. To address the issue, we propose a novel multi-channel asymmetrical variational discrete cosine transform (DCT) network for EEG data compression within an edge-fog computing framework. At the edge level, low-complexity DCT compression units are designed using parallel trainable hard-thresholding and scaling operators to remove redundant data and extract the effective latent space representation. At the fog level, an adaptive filter bank is applied to merge important features from adjacent channels into each individual channel by leveraging inter-channel correlations. Then, the inverse DCT reconstructed multi-head attention is developed to capture both local and global dependencies and reconstruct the original signals. Furthermore, by applying the principles of variational inference, a new evidence lower bound is formulated as the loss function, driving the model to balance compression efficiency and reconstruction accuracy. Experimental results on two public datasets demonstrate that the proposed method achieves superior compression performance without sacrificing any useful information for BCI detection compared with state-of-the-art techniques, indicating a feasible solution for EEG data compression. I Introduction The brain-computer interface (BCI) is emerging as a pivotal technology within the framework of Industry 4.0 [ 1 ] . BCI facilitates direct communication between the human brain and machines, bypassing traditional physical interfaces [ 2 ] . This capability holds transformative potential for enhancing automation and operational efficiency in industrial environments. For instance, BCI allows operators to control robotic arms or machinery with greater precision and less physical effort [ 3 ] . Furthermore, BCI can enhance worker safety by monitoring cognitive load and stress levels [ 4 ] . Fig. 1: Edge-fog computing framework. The BCI system captures brain signals by measuring metabolic activity through functional near-infrared spectroscopy (fNIRS), magnetic activity using magnetoencephalography (MEG), or electrical activity via electroencephalography (EEG) [ 5 ] . EEG is recorded via electrodes placed on the scalp. It is widely utilized and well-suited for industrial applications due to its non-invasive nature, cost-effectiveness, high temporal resolution, and portability [ 6 ] . The integration of BCI into the industrial field is intricately linked with the Internet of Things (IoT) [ 7 , 8 ] . In smart factories, IoT enables various devices and systems to communicate and operate cohesively. BCI-equipped operators can manage multiple IoT-connected devices simultaneously to optimize workflows. However, the vast volumes of EEG data generated by the BCI system pose significant challenges to rapid transmission and real-time analysis in bandwidth-limited channels [ 9 ] . Edge and fog computing frameworks [ 10 ] address these challenges by deploying the edge gateway close to the data generators, such as biosensors, to perform real-time data compression, as illustrated in Fig. 1 . The compressed data is then transmitted to the fog gateway for decompression and feature extraction before being forwarded to BCI applications. Additionally, the reconstructed data is uploaded to the medical health cloud for storage and further analysis. Finally, both the applications and the medical cloud provide feedback to the users. The fog gateway, positioned between the cloud and the sensors, provides an intermediary layer that enhances processing capabilities and reduces reliance on cloud resources for comprehensive data analysis. In contrast, while the edge gateway is situated near sensors, it is limited in its computational capabilities, rendering it insufficient for complex tasks such as training neural network models and performing extensive analyses on large-scale EEG datasets. Therefore, an efficient low-complexity EEG compression algorithm is required at the edge gateway to eliminate redundant data, reducing the latency, data transfer costs and bandwidth. EEG signal compression techniques are generally divided into three main categories: traditional signal processing methods, neural network-based methods, and transform-based learning techniques [ 11 ] . Traditional signal processing methods, such as discrete cosine transform (DCT) and discrete wavelet transform (DWT), have been extensively employed in EEG data compression. These compression techniques can be categorized into single-channel and multi-channel schemes. In the single-channel methods, each EEG channel was processed independently using compression techniques, e.g., the fast DCT [ 12 ] , DWT [ 13 ] . In contrast, multi-channel compression algorithms perform simultaneous compression of EEG signals across all channels by exploiting inter-channel correlations. For instance, Hejrati et al. utilized the K-means algorithm to cluster EEG channels [ 14 ] . Similar channels were then grouped to eliminate inter-channel dependencies, thereby reducing the data size. However, the maximum compression ratio achieved in [ 14 ] was smaller than 3, indicating a low compression efficiency. Recently, neural networks have been increasingly utilized for data compression due to their capacity to learn complex data representations, such as the underlying patterns in EEG signals. Among these methods, autoencoders are the most frequently employed for compression tasks. For instance, the convolutional autoencoder (CAE) [ 15 ] and variational autoencoder [ 16 ] have demonstrated their efficacy in EEG data compression by applying convolutional layers, max-pooling layers, and evidence lower bound (ELBO) to optimize their models. Furthermore, Lerogeron et. al proposed a neural network-based approximation of dynamic time warping (DTW) as a reconstruction loss function, enhancing the ability of convolutional autoencoders to compress EEG signals while preserving essential information [ 17 ] . Nevertheless, these autoencoder-based models capture effective latent representations of EEG signals by incorporating numerous convolutional layers or linear layers, leading to increased computational complexity and higher hardware costs. Over the past few decades, transform-based learning methods have gained considerable attention in EEG data compression for their ability to reduce data size while maintaining critical signal information. The authors in [ 18 ] proposed a learning-based adaptive transform method, which combines the DCT with an artificial neural network (ANN) to achieve near-lossless compression. The DCT is utilized to compress the original data, and then the MLP further compresses the primary DCT coefficients. Additionally, an asymmetrical sparse autoencoder [ 11 ] was designed for compressing EEG sensor signals. This model integrates a low-complexity DCT layer within the encoder module to eliminate redundant coefficients and generate the latent representation. In contrast, the decoder module is equipped with additional linear layers to improve the quality of signal reconstruction. Nonetheless, These models function as single-channel approaches, where EEG signals are compressed and reconstructed on a per-channel basis, neglecting the correlations that exist among multi-channel EEG signals. As a result, the reconstruction accuracy is limited. To address the aforementioned issues, this paper proposes a novel multi-channel asymmetrical variational discrete cosine transform network (AVDCT-Net) to compress EEG signals across all channels simultaneously within the edge-fog computing framework. At the edge gateway, each channel is processed through a linear layer, followed by a low-complexity DCT compression unit (DCU) in the encoder module. In this unit, parallel hard-thresholding nonlinearities eliminate redundant data, while multiple scaling operators serve as an approximate filter bank within the DCT domain. The entire compression process is guided by an ELBO-based cost function, enforcing a regularization constraint on the latent space coefficients to ensure a more structured and efficient representation. At the fog gateway, the decoder module reconstructs the EEG signal from compressed data utilizing