Federated learning (FL) has been recognized as a transformative approach in vehicular networks, enabling collaborative training between vehicles and preserving data privacy. However, the high mobility and dynamic topology changes inherent in vehicular environments pose significant challenges, primarily due to the increased communication overhead associated with exchanging model parameters. To mitigate these issues, a novel framework for decentralized wireless FL with soft clustering and 1-bit compressed sensing (SC1BCS-WFL) is proposed in this article. The framework considers vehicle position, vehicle attributes, vehicle speed, and model cosine similarity when grouping vehicles. It utilizes an adaptive threshold mechanism based on 1-bit compression to reduce uplink transmission load while maintaining FL performance. In addition, an early stopping strategy is incorporated into the proposed framework to avoid unnecessary waste of computational and communication resources. Simulation results show that the SC1BCS-WFL framework can enhance the efficiency of FL in vehicular settings, particularly with nonindependent and identically distributed data. Simulation results also validate the framework’s ability to reduce communication overhead while achieving high-model accuracy, indicating its suitability for distributed Internet of Vehicles (IoV) scenarios and contributing to the development of smarter and more efficient IoV applications.

With the advance of IoT technology, we are able to collect massive people' s trajectory data from various GPS services. These large amounts of trajectory records enable us to better understand human mobility patterns. Meanwhile, we are able to extract social relationships based on these digital records to provide personalized recommendation services, such as POI recommendation and friend recommendation. In this work, we propose to recommend friends for taxi drivers based on vehicular trajectory records. For this purpose, we propose to construct a vehicular ad hoc network based on co-occurrence phenomenon. Furthermore, we take advantages of the network representation learning technique on the vehicular ad hoc network for learning driver vectors. Finally, potential friends are recommended based on the similarity of driver vectors. Extensive experimental results on two real-world datasets demonstrate that our proposed method has the best performance on friend recommendation compared with several state-of-the-art methods. To the best of our knowledge, this is the first attempt to recommend friends for taxi drivers based on vehicular ad hoc network representation learning.

With the advance of artificial intelligence and communication technology in smart city, various location-based data of users can be collected via location-based social networks. How to make full use of these data for accurate point of interest (POI) recommendation is challenging because POI selection is influenced by various factors. In this paper, we propose a network representation learning enhanced multi-source information fusion model for POI recommendation in the context of location-based social networks. The proposed model jointly considers various factors, including user preference, geographical influence, and social influence for recommendation. Specifically, the social influence is modeled by performing network representation learning methods on the constructed co-visiting user networks so that the hidden complex social relationships among users can be measuring automatically. Moreover, considering the significance of user preference and geographical influence, a fusion model is design to jointly consider user preference, social influence, and geographical influence for POI recommendation. Our method is evaluated based on two publicly available datasets and extensive experimental results demonstrate that the proposed multi-source information fusion model outperforms several state-of-the-art algorithms for POI recommendation in terms of precision, recall, and F1.

Breast cancer is one of most serious malignant tumors that affect women’s health. To carry out the early screening of breast cancer, mammography provides breast cancer images for doctors’ efficient diagnosis. However, breast cancer lumps can vary in size and shape, bringing difficulties for the accurate recognition of both humans and machines. Moreover, the annotation of such images requires expert medical knowledge, which increases the cost of collecting datasets to boost the performance of deep learning methods. To alleviate these problems, we propose an attention-based active learning framework for breast cancer segmentation in mammograms; the framework consists of a basic breast cancer segmentation model, an attention-based sampling scheme and an active learning strategy for labelling. The basic segmentation model performs multi-scale feature fusion and enhancement on the basis of UNet, thus improving the distinguishing representation capability of the extracted features for further segmentation. Afterwards, the proposed attention-based sampling scheme assigns different weights for unlabeled breast cancer images by evaluating their uncertainty with the basic segmentation model. Finally, the active learning strategy selects unlabeled images with the highest weights for manual labeling, thus boosting the performance of the basic segmentation model via retraining with new labeled samples. Testing on four datasets, experimental results show that the proposed framework could greatly improve segmentation accuracy by about 15% compared with an existing method, while largely decreasing the cost of data annotation.

Gearboxes are widely used in various types of mechanical equipment and have become an essential part of connecting various components in the machine and transmitting power. A fault in the gearbox will cause the entire mechanical equipment to stop working, causing economic losses and safety hazards. Previous fault diagnosis models for gearbox used the vibration signals of the gearbox as input features. They designed a variety of feature extraction modules to learn the information contained in the vibration signals. However, the latent features learned by previous models have poor interpretability and robustness, thus affecting model performance. In addition, the prediction results of previous fault diagnosis models used for gearbox using single-scale features are often less robust. Moreover, the fault diagnosis models used for gearbox are usually end-to-end black box models, which cannot provide interpretive information about the predicted values, and it is difficult to handle the uncertain information in the vibration signals. Therefore, we propose a multi-scale fuzzy variational autoencoder (MFVAE) using a fuzzy neural network for Big Data-based fault diagnosis in gearbox. The Big Data technology can automatically collect vibration signals from sensors and provide high-quality training samples for fault diagnosis models. The MFVAE model first uses two variational autoencoders of different sizes to extract the latent features of the gearbox vibration signal. Subsequently, the MFVAE model fuses low-level latent features and high-level latent features in proportion to form the final head features, which are used to diagnose gearbox faults. Finally, the MFVAE model inputs the head features into the prediction layer to diagnose the fault of the gearbox. The prediction layer comprises a fully connected network and a fuzzy neural network. Experimental results on the real gearbox dataset verify the outstanding performance of the MFVAE model.