CNN has demonstrated remarkable performance in EEG signal detection, yet it still faces limitations in terms of global perception. Additionally, due to individual differences in EEG signals, the generalization ability of epilepsy detection models is week. To address this issue, this paper presents a cross-patient epilepsy detection method utilizing a multi-head self-attention mechanism. This method first utilizes Short-Time Fourier Transform (STFT) to transform the original EEG signals into time-frequency features, then models local information using Convolutional Neural Network (CNN), subsequently captures global dependency relationships between features using the multi-head self-attention mechanism of Transformer, and finally performs epilepsy detection using these features. Meanwhile, this model employs a light multi-head attention mechanism module with an alternating structure, which can comprehensively extract multi-scale features while significantly reducing computational costs. Experimental results on the CHB-MIT dataset show that the proposed model achieves accuracy, sensitivity, specificity, F1 score, and AUC of 92.89%, 96.17%, 92.99%, 94.41%, and 96.77%, respectively. Compared to the existing methods, the method proposed in this paper obtains better performance along with better generalization.

In this article, we consider incentive mechanism designs in asynchronous federated learning (FL) systems. With the consideration of unique characteristics inherent in asynchronous FL, such as dynamic participating and multiminded IoT nodes such as mobile users (MUs), requirements of model training (i.e., training accuracy and convergence time), and limited uplink bandwidth, we formulate considered system as an online incentive mechanism design problem, where each MU is not only a buyer for communication resource but also a seller for computation service. To address the challenges involved in the design, we first derive the relationship between the number of participants and the global training accuracy in asynchronous FL. Then, based on that, we propose a novel mechanism, called the online incentive mechanism for asynchronous FL (OIMAF). To the best of our knowledge, this is the first work to design incentive mechanisms for asynchronous FL. Furthermore, in order to obtain a more robust mechanism, an improved online mechanism, called the two-shot-based online incentive mechanism (TOIM), is proposed by using OIMAF as a building block. Theoretical analyses show that our proposed online incentive mechanisms can guarantee individual rationality, truthfulness, a sound performance, and solution feasibilities. We further conduct comprehensive simulations to validate the effectiveness of our proposed mechanisms.

In the operational context of a cross-device federated learning (FL), the efficient allocation of resources, such as transmission powers, channels, and computation resources, significantly impacts overall performance. Existing research in cross-device FL has predominantly concentrated on either resource allocation to enhance training accuracy or incentivizing participation, while ignoring their integrated designs for further improving the performance in cross-device FL. Different from existing work, in this paper, we jointly integrate the power allocation, channel assignment, user selection, and allocation of computation frequency into the design of incentive mechanism, where each mobile user plays a dual role as both a buyer and a seller. Because of complex resource allocation, truthfulness guarantee in a dual role scenario, and unavailable prior information, the considered mechanism design problem is challenging. To tackle such combinatorial problem, we propose a Reinforcement Auction Mechanism (RAM), comprising two layers. The upper layer features a Hybrid Action Reinforcement Learning scheme to learn the outcomes of user selection and payments. In the lower layer, each selected mobile user optimizes its resources to maximize its utility. Theoretical analyses affirm that our proposed RAM ensures individual rationality and truthfulness. Extensive simulations have been conducted to validate the effectiveness of the proposed RAM.

LoRa is widely used in low-power Internet of Things (IoT) applications due to its low-power consumption, high-connection density, and wide coverage. How to optimize the physical layer parameter configuration of LoRa edge nodes is a bottleneck issue that restricts network performance. To reduce the energy consumption of the distributed LoRa networks, a reinforcement learning strategy based on link prior knowledge is adopted, and the resource allocation of the edge node of the LoRa networks is transformed into a multiarmed bandit problem. On this basis, a dynamic parameter selection algorithm suitable for LoRa edge nodes is proposed, namely, the link-weight-EXP3 (LI-WEX) algorithm. By defining the weight factors of parameters and rewards for nodes, the energy consumption factor in decision-making is enhanced, and the policy space at the node is processed based on link knowledge, thereby improving energy efficiency by selecting the optimal combination of parameters. The simulation results show that the LI-WEX algorithm not only converges faster and is robust in different scenarios, but also can effectively reduce network energy consumption.

Cardiac arrhythmia (CA) is a severe cardiac disorder that results in a significant number of fatalities worldwide each year. Conventional electrocardiography (ECG) devices are often unable to detect arrhythmia symptoms during patients’ hospital visits due to their intermittent nature. This paper presents a wearable ECG processor for cardiac arrhythmia (CA) detection. The processor utilizes a Hilbert transform-based R-peak detection engine for R-peak detection, a Haar discrete wavelet transform (HDWT) unit for feature extraction, and a Hybrid ECG classifier that combines linear methods and Non-Linear Support Vector Machines (NLSVM) classifiers to distinguish between normal and abnormal heartbeats. The processor is fabricated by the CMOS 110 nm process with an area of 1.34 mm² and validated with the MIT_BIH Database. The whole design consumes 4.08 μW with an average classification accuracy of 97.34%.