Federated learning emerges as a solution to the dilemma of data silos while safeguarding data privacy, particularly relevant in the consumer electronics sector where user data privacy is paramount. However, federated learning is generally employed in a heterogeneous scenario, consisting of various factors that influence the training efficiency and accuracy of the federated learning models. There are many classic references focusing on federated communications, federated robustness and federated fairness, conversely, few of them clarify and summary systematically the influence of heterogeneity on the effect of federated learning. Therefore, we provide an overview of three heterogeneous challenges faced by federated learning in practical applications: device heterogeneity, data heterogeneity and model heterogeneity, and analyze their influence on federated learning. This is especially crucial in consumer electronics, where heterogeneity directly influence the performance and user experience of AI-driven features. And then, we highlight current solutions, ideas and challenges to compare different strategies for facing heterogeneous problems and outline several directions of future work that are relevant to a wide range of research communities.

When the electromagnetic suspension (EMS) type maglev vehicle is traveling over a track, the airgap must be maintained between the electromagnet and the track to prevent contact with that track. Because of the open-loop instability of the EMS system, the current must be actively controlled to maintain the target airgap. However, the maglev system suffers from the strong nonlinearity, force saturation, track flexibility, and feedback signals with network time-delay, hence making the controller design even more difficult. In this article, the minimum levitation unit of the maglev vehicle system has been established. An amplitude saturation controller (ASC), which can ensure the generation of only saturated unidirectional attractive force, is thus proposed. The stability and convergence of the closed-loop signals are proven based on the Lyapunov method. Subsequently, ASC is improved based on the radial basis function neural networks, and a neural network-based supervisor controller (NNBSC) is thus designed. The ASC plays the main role in the initial stage. As the neural network learns the control trend, it will gradually transition to the neural network controller. Simulation results are provided to illustrate the specific merit of the NNBSC. The hardware experimental results of a full-scale IoT EMS maglev train are included to validate the effectiveness and robustness of the presented control method as regards to time delay.

This paper provides an approach for recognizing human activities with wearable sensors. The continuous autoencoder (CAE) as a novel stochastic neural network model is proposed which improves the ability of model continuous data. CAE adds Gaussian random units into the improved sigmoid activation function to extract the features of nonlinear data. In order to shorten the training time, we propose a new fast stochastic gradient descent (FSGD) algorithm to update the gradients of CAE. The reconstruction of a swiss-roll dataset experiment demonstrates that the CAE can fit continuous data better than the basic autoencoder, and the training time can be reduced by an FSGD algorithm. In the experiment of human activities’ recognition, time and frequency domain feature extract (TFFE) method is raised to extract features from the original sensors’ data. Then, the principal component analysis (PCA) method is applied to feature reduction. It can be noticed that the dimension of each data segment is reduced from 5625 to 42. The feature vectors extracted from original signals are used for the input of deep belief network (DBN), which is composed of multiple CAEs. The training results show that the correct differentiation rate of 99.3% has been achieved. Some contrast experiments like different sensors combinations, sensor units at different positions, and training time with different epochs are designed to validate our approach.

Digital Twins (DTs) can generate digital replicas for mobile networks (MNs) that accurately reflect the state of MN. Machine learning (ML) models trained in DT for MN (DTMN) virtual environments can be more robustly implemented in MN. This can avoid the training difficulties and runtime errors caused by MN instability and multiple failures. However, when using data from various devices in the MN system, DTs must prioritize data privacy. Federated learning (FL) enables the construction of models without data leaving devices to protect DTMN data privacy. Nevertheless, FL’s privacy protection needs further improvement for it only guarantees device-level data ownership but ignores that models may retain private information from data. Therefore, this paper focuses on data forgetting in privacy protection, and proposes a novel FL-based unlearning framework (FedME ² ), which contains MEval and MErase modules. Guided by memory evaluation information from MEval and employing MErase’s multi-loss training approach, FedME ² gets accurate data forgetting in DTMN. In four DTMN virtual environments, FedME ² achieves an average data forgetting rate of approximately 75% for global models under FL and kept the influence on global models’ accuracy below 4%. FedME ² has better data forgetting and improves DTMN data privacy protection while guaranteeing model accuracy.

This paper proposes a novel continuous sparse autoencoder (CSAE) which can be used in unsupervised feature learning. The CSAE adds Gaussian stochastic unit into activation function to extract features of nonlinear data. In this paper, CSAE is applied to solve the problem of transformer fault recognition. Firstly, based on dissolved gas analysis method, IEC three ratios are calculated by the concentrations of dissolved gases. Then IEC three ratios data is normalized to reduce data singularity and improve training speed. Secondly, deep belief network is established by two layers of CSAE and one layer of back propagation (BP) network. Thirdly, CSAE is adopted to unsupervised training and getting features. Then BP network is used for supervised training and getting transformer fault. Finally, the experimental data from IEC TC 10 dataset aims to illustrate the effectiveness of the presented approach. Comparative experiments clearly show that CSAE can extract features from the original data, and achieve a superior correct differentiation rate on transformer fault diagnosis.