This paper proposes a long period fiber grating (LPFG) pressure sensor based on mode transition, consists of LPFG and two polymer films. The inner layer of polyvinyl alcohol (polyvinyl alcohol, PVA) with high refractive index properties can be tuned to operate within the LPFG’s mode transition region. The outer layer of porous polydimethylsiloxane (PDMS) can be used as a pressure-sensitive material for the sensor. Experimental results show that the sensor achieves a sensitivity of 612 pm/kPa when the PVA thickness is 244 nm, which is approximately 50 times higher than that of a conventional LPFG sensor without excitation mode transition. To address temperature sensitivity inherent in both the PVA and PDMS, a fiber Bragg grating (FBG) is connected in series for temperature compensation. The proposed LPFG pressure sensor based on mode transition exhibits high sensitivity and a simple fabrication process, making it a promising candidate for applications in high-precision pressure sensing.

In band-limited digital communications, square-root Nyquist (SR-NYQ) filters are commonly applied at both the transmitter and receiver to effectively mitigate sampling inter symbol interference (ISI). This paper proposes a novel design method for linear-phase SR-NYQ filters based on genetic algorithms (GA), in which the fitness function consists of key performance metrics including ISI, passband ripple and stopband ripple. Due to the excellent global optimization capability of GA, the proposed method enables a closer approximation to the ideal Nyquist condition while providing additional design flexibility. To evaluate performance, the proposed SR-NYQ filter is compared with the conventional root raised cosine filter within an orthogonal frequency division multiplexing (OFDM) system. Simulation results demonstrate that the SR-NYQ filter designed using the proposed method achieves a superior frequency response and significantly reduces the symbol error rate (SER).

Data-driven Gappy proper orthogonal decomposition (POD), namely GP algorithm, is a method to solve inverse problems in physical phenomena such as flow and heat transfer. However, the actual data is usually polluted by various noises, thus affecting the accuracy of the GP algorithm. The database was built on the Burgers equation because it contained some important forms of the flow and heat transfer equations. The reconstruction accuracy and stability of the GP algorithm for Gaussian noise and random noise based on ordinary least squares (OLS), weighted least squares (WLS), total least squares (TLS), and L1 regularization were studied. The results show that the GP algorithm can reconstruct the one-dimensional Burgers equation with high accuracy with only a small amount of incomplete data. Compared with those of GP-OLS, the root-mean-square error and maximum error of GP-WLS, GP-TLS, and GP-L1 are significantly reduced, and the correlation coefficient is closer to 1. Under the noise condition, the root-mean-square error of GP-WLS is reduced to 1/27 that of GP-OLS, with improved reconstruction accuracy. The root-mean-square error and maximum error of GP-TLS reconstruction are the smallest, which are 0.014 1 and 0.013 0, respectively. The reconstruction performance is the best when the data matrix and observation vector are noisy. The correlation coefficient of GP-L1 reconstruction is close to 1, which improves the trend prediction ability of the algorithm. Before and after adding noise, the reconstruction ability of GP-L1 does not change much, indicating that the GP-L1 algorithm has strong anti-interference ability against outliers and noise and improves the robustness of the model.

Click-through rate (CTR) prediction is one of the fundamental tasks in recommendation systems. Dual-stream models have been widely adopted in mainstream recommendation frameworks due to their superior flexibility, scalability, and efficiency in information interaction and fusion. To further enhance CTR prediction performance, this paper proposes the FJ hybrid network (FinalBlock-JRC hybrid network, FJHN), which integrates the factorized interaction block (FinalBlock) and the joint ranking and calibration loss optimization algorithm (JRC) based on the structure of the dual-stream model. First, a feature gating layer is introduced to enable differentiated feature inputs, thereby enhancing the representation of important features. Then, FinalBlock is combined with a multilayer perceptron (MLP) to strengthen high-order feature interaction learning. Furthermore, an enhanced interaction aggregation layer is employed to fuse the outputs of each tower, deepening the degree of feature interaction. Finally, an improved JRC mechanism is applied to compute the loss function, which effectively improves the model’s prediction accuracy and adaptability across diverse application scenarios. Experimental results on three publicly available benchmark datasets demonstrate that compared with several mainstream models including self-attention model (SAM), the FJHN model achieves noticeable performance gains in CTR prediction.

Traditional point of interest (POI) recommendation methods often fail to fully mine the relationships between users and POIs, resulting in limited capacity to capture user preferences. Although graph enhancement method offers improved relational modeling, it is susceptible to noise, which reduces the recommendation precision. To address these challenges, this paper proposes a POI recommendation method named combining universal trajectory maps and multiple preferences. Firstly, a weighted bipartite graph between users and POIs is constructed. Graph convolutional network (GCN) is used to extract the interactive relationship between users and POIs to learn users’ interest preferences. Users are clustered using the obtained interest preferences. The general trajectory maps of the same type of users are built to reduce the impact of noise information. GCN is further used to mine the group features of different types of users and enrich feature representation. Secondly, the group features are combined with time-aware category information and spatio-temporal context from the user’s current trajectory. Transformer model is used to capture deep behavioral preferences. Finally, a non-linear additive function is used to dynamically combine interest preferences with current behavior preferences to fully capture user preferences and generate POI recommendations. The validity of the proposed method is verified on real data sets.

In multi-process lithium battery manufacturing, data analysis faces challenges of multi-source heterogeneity and missing data. To address these issues, this article proposes a filling method that combines autoregressive integrated moving average model (ARIMA) with interpolation techniques. The proposed method extracts trends and periodic features of time series data through ARIMA models, and integrates interpolation techniques to repair missing values caused by equipment failures or incomplete data collection, enhancing the ability to capture complex data change patterns. Multiple experiments have shown that the proposed ARIMA-interpolation method outperforms traditional techniques such as mean filling, K-nearest neighbor filling, and standalone interpolation in terms of filling accuracy and data integrity. The results confirm that the proposed method effectively improves the quality of data preprocessing in lithium battery manufacturing, providing a reliable data foundation for subsequent feature extraction and analysis.

Existing transfer-based attacks primarily focus on maximizing the empirical risk while ignoring the expected risk, which often leads to suboptimal transferability. To address this issue, we propose a transferability-gap-aware attack framework. First, we formulate the objective of transfer-based attacks as an expected risk and introduce the notion of the transferability gap, which quantifies the absolute discrepancy between the empirical risk and the expected risk. Our analysis reveals that when the transferability gap is small, maximizing the empirical risk becomes approximately equivalent to maximizing the expected risk, thereby leading to highly transferable adversarial examples. Based on this insight, the proposed method min-max the transferability gap while maximizing the empirical risk. Such min-max problem allows the attack algorithm with the strongest transferability to be found in the case of the hardest transferability. Experimental results show that the proposed method outperforms the recent state-of-the-art transfer-based attacks and achieves fast generation of highly transferable adversarial examples.

Medical image segmentation based on semi-supervised learning has attracted extensive attention because of the high cost and scarcity of annotated images. Effectively leveraging unlabeled data remains a challenging task. This paper proposes a two-stage segmentation model based on a multi-scale feature fusion network to make use of unlabeled data, and address the empirical distribution mismatch between labeled and unlabeled data. The model uses labeled data to train a teacher model in the first stage and both labeled and unlabeled data are used to co-train a student model in the second stage. To improve the robustness of the teacher model, a copy-paste strategy is employed to increase data diversity. To alleviate the misguidance problem caused by the pseudo-labels generated in the second stage, confidence learning based on an assumption of classified noised process is introduced, thereby reducing the potential bias caused by pseudo-labels. Extensive experiments and ablation studies on two publicly available organ datasets demonstrate that the proposed model achieves high-precision segmentation.

To enable automatic and accurate quantitative evaluation of hand function rehabilitation in stroke patients, this paper proposes a skeleton-based gesture recognition and evaluation method. First, the MediaPipe framework is used to extract hand keypoints and connect them to form a hand skeleton, converting traditional RGB video datasets into hand skeleton datasets. Then, a 3D convolutional neural network (C3D) model is employed to train on and recognize hand functional movements. Based on correct recognition, further evaluation is conducted using the dynamic time warping (DTW) algorithm. The DTW distance between the motions of the healthy and affected hands performing the same action is calculated, aligning both temporally and spatially to represent the similarity in action execution. Experiments establish optimal DTW thresholds for distinguishing different rehabilitation ratings for each action, which serve as the criteria for quantitative evaluation.Results show that using skeletal data instead of traditional video improves gesture recognition accuracy to 99.01% and reduces training time. With the DTW algorithm, automatic hand function rehabilitation assessment is achieved.

With the rapid development of telemedicine, the secure management and transmission of patient privacy data face significant challenges. To achieve unified management and secure storage of electronic medical records (EMRs) and medical images, this paper proposes an encoding-domain hidden encryption scheme for multimodal medical data. Specifically, a string-to-image transformation algorithm based on UTF-8 encoding and positional numeral decomposition is designed to convert EMRs into encoded images, ensuring data privacy and security. To facilitate integrated management of multimodal data, an improved HiNet reversible network is introduced to embed medical images into encoded images. By incorporating the Kullback-Leibler (KL) divergence to constrain the distribution distance, the scheme enhances the accuracy and robustness of image embedding and reconstruction. Furthermore, to strengthen the security of the encoded images, a bit-level encryption algorithm based on the logic-sine-cosine chaotic system is designed, leveraging its high sensitivity and nonlinear characteristics for robust encryption. Experimental results demonstrate that the proposed encoding-domain hidden encryption scheme effectively ensures data security while enabling lossless access to EMRs and high-quality recovery of medical images, offering enhanced confidentiality for secure management of multimodal data in telemedicine.

Diabetic retinopathy (DR) and diabetic macular edema (DME) are the major causes of blindness in humans. This study proposes to leverage the ResNeSt network to focus on the close relationship between DR and DME, and use it to achieve their joint grading to improve grading accuracy. Specifically, common features and unique features of DR and DME are extracted by using different stages of ResNeSt. Subsequently, a mutual attention module for feature fusion is employed to realize their respective grading. On the Messidor dataset, by using both ResNeSt network and the mutual attention module, the grading accuracies of DR and DME were 95.6% and 95.0%, respectively, and the joint accuracy was 86.7%. In contrast, when only the ResNeSt network was used, the grading accuracies of the two were 95.0% and 90.8%, respectively. On the IDRiD dataset, the joint accuracy of the two reached 66.3%. The results on the datasets indicate that the proposed joint grading of DR and DME based on the ResNeSt network can improve grading accuracy.

Sharding technology is a crucial method to improve the scalability of blockchain systems. However, current sharding designs often overlook the issues of load imbalance between shards and the heterogeneity of nodes, leading to resource waste and a decline in system performance. To solve this problem, this paper proposed BalChain, a sharded blockchain system based on reputation and load, aimed at enhancing system throughput and security. BalChain employed a dual-chain architecture, consisting of a transaction chain and a reputation chain, utilizing the Raft protocol and a collective signature Byzantine fault tolerance (CSBFT) mechanism, respectively, to ensure efficient transaction processing and robust system security. This paper also introduced a reputation-load matching sharding algorithm, which dynamically allocated computing resources based on shard load, fully utilizing the heterogeneity of nodes. Moreover, the system reduced cross-shard transactions through Metis graph partitioning algorithm, further improving transaction processing efficiency. Experimental results demonstrate that BalChain outperforms existing sharded blockchain systems in terms of throughput, latency, and cross-shard transaction processing efficiency, which proved the effectiveness and scalability of the design in real-world applications.

Blockchain technology leverages distributed ledgers and cryptographic techniques to offer novel solutions for data privacy and transaction security. Among various cross-chain solutions, the notary group-based model has been widely adopted for data exchange across various domains. However, the centralized architecture of traditional notary group models introduces inherent risks and remains susceptible to network fluctuations. To address the centralization issue in notary-based cross-chain technologies, this paper proposes a novel transaction protocol jointly designed by notaries. First, an improved PageRank algorithm is introduced to reduce centralization during notaries selection process. Then, a sub-secret segment table scheme is proposed to optimize the threshold signature algorithm, thereby enhancing the protocol’s robustness under network fluctuations while preserving decentralization. Experimental results show that with 50 participating notaries, the optimized node selection algorithm improves the overall balance of node reputation values by 47.3% compared to traditional approaches. Moreover, under a network packet loss rate of up to 70%, the optimized threshold signature algorithm achieves approximately three times the robustness of standard threshold signature schemes, with negligible additional time overhead and Gas costs. Overall, the proposed protocol significantly enhances decentralization and robustness of cross-chain transactions with minimal impact on transaction costs, demonstrating its practical feasibility.