Driven by the demand to accommodate today’s growing mobile traffic, 5G is designed to be a key enabler and a leading infrastructure provider in the information and communication technology industry by supporting a variety of forthcoming services with diverse requirements. Considering the everincreasing complexity of the network, and the emergence of novel use cases such as autonomous cars, industrial automation, virtual reality, e-health, and several intelligent applications, machine learning (ML) is expected to be essential to assist in making the 5G vision conceivable. This paper focuses on the potential solutions for 5G from an ML-perspective. First, we establish the fundamental concepts of supervised, unsupervised, and reinforcement learning, taking a look at what has been done so far in the adoption of ML in the context of mobile and wireless communication, organizing the literature in terms of the types of learning.We then discuss the promising approaches for how ML can contribute to supporting each target 5G network requirement, emphasizing its specific use cases and evaluating the impact and limitations they have on the operation of the network. Lastly, this paper investigates the potential features of Beyond 5G (B5G), providing future research directions for how ML can contribute to realizing B5G. This article is intended to stimulate discussion on the role that ML can play to overcome the limitations for a wide deployment of autonomous 5G/B5G mobile and wireless communications.

In this letter, we introduce CGDNet, a cost-efficient hybrid neural network composed of a shallow convolutional network, a gated recurrent unit, and a deep neural network, for robust automatic modulation recognition for cognitive radio services of modern communication systems. Our model employs pooling layers, small filter sizes, Gaussian dropout layers, and skip connections which leads to an increase in network capacity, a reinforced process of feature extraction, and prevents the vanishing gradient problem. From our experiments, CGDNet incurs a low computational complexity and reaches the overall n-modulation recognition accuracy of 93.5% and 90.38% on two widely used Deep-Sig datasets.

Electric vehicles (EVs) have gained attention owing to their effectiveness in reducing oil demands and gas emissions. Of the electric components of an EV, a battery is considered as the major bottleneck. Among the various types of battery, lithium-ion batteries are widely employed to power EVs. To ensure the safe application of batteries in EVs, monitoring and control are performed using state estimation. The state of a battery includes the state-of-charge (SoC), state-of-health (SoH), state-of-power (SoP), and state-of-life (SoL). The SoC of a battery is the remaining usable percentage of its capacity. This mainly depends on variations of the operating condition of the EV in which the battery is applied. The SoC of a battery is reflected by its output voltage. That is, the SoC is considered to be zero when the output voltage of a battery drops below a cut-off voltage. This study proposes an SoC and output voltage forecasting method using a hybrid of the vector autoregressive moving average (VARMA) and long short-term memory (LSTM). This approach aims to estimate and forecast the SoC and output voltage of a battery when an EV is driven under the CVS-40 drive cycle. Forecasting using the hybrid VARMA and LSTM method achieves a lower root-mean-square error (RMSE) than forecasting with only VARMA or LSTM individually.

This paper focuses on visual attention , a state-of-the-art approach for image captioning tasks within the computer vision research area. We study the impact that different hyperparemeter configurations on an encoder-decoder visual attention architecture in terms of efficiency. Results show that the correct selection of both the cost function and the gradient-based optimizer can significantly impact the captioning results. Our system considers the cross-entropy, Kullback-Leibler divergence, mean squared error, and negative log-likelihood loss functions; the adaptive momentum (Adam), AdamW, RMSprop, stochastic gradient descent, and Adadelta optimizers. Experimentation shows that a combination of cross-entropy with Adam is the best alternative returning a Top-5 accuracy value of 73.092 and a BLEU-4 value of 20.10. Furthermore, a comparative analysis of alternative convolutional architectures demonstrated their performance as an encoder. Our results show that ResNext-101 stands out with a Top-5 accuracy of 73.128 and a BLEU-4 of 19.80; positioning itself as the best option when looking for the optimum captioning quality. However, MobileNetV3 proved to be a much more compact alternative with 2,971,952 parameters and 0.23 Giga fixed-point Multiply-Accumulate operations per Second (GMACS). Consequently, MobileNetV3 offers a competitive output quality at the cost of lower computational performance, supported by values of 19.50 and 72.928 for the BLEU-4 and Top-5 accuracy, respectively. Finally, when testing vision transformer (ViT), and data-efficient image transformer (DeiT) models to replace the convolutional component of the architecture, DeiT achieved an improvement over ViT, obtaining a value of 34.44 in the BLEU-4 metric.

The uncertainties and disturbances in the actual driving conditions of hybrid electric vehicles (HEVs) complicate the design of energy management strategy (EMS). To achieve better EMS performance for a battery-supercapacitor HEV, this paper proposes an improved and adaptive deep learning-based velocity prediction control EMS that can prolong the battery lifetime through efficient utilization of both the battery and supercapacitor. First, feature engineering techniques are used to extract and increase the key features from the historical driving cycle data of known driving conditions. With the extracted features, an improved long short-term memory (LSTM) velocity predictor was developed to predict future driving cycles for a real-time EMS under an unknown driving condition. Second, a real-time EMS based on the rule-based framework optimized with a neural network is proposed to optimize the power allocation online. Simulation results show that the proposed strategy smoothens battery peak power (i.e. prolongs battery life span) by approximately 26.85% on average and increases supercapacitor participation in the EMS, as evidenced by its increased energy throughput. Furthermore, compared with other EMS approaches, the proposed strategy improved the efficiency by significantly reducing total energy losses by approximately 22.25%. These results validate the reliability and robustness of the proposed strategy.

Estimating the state of health (SOH) of lithium-ion batteries is crucial for ensuring that the batteries operate safely and have a long lifespan. The existing approaches for SOH estimation on embedded systems only consider one health indicator (HI) to represent either capacity or internal resistance (IR) behavior because of limitations in the hardware devices. Nevertheless, both capacity and IR provide valuable battery health information and neither of these could be neglected. Hence, we propose the SOH estimation method that can consider both capacity degradation and IR growth by representing it with HIs that can be directly measured in embedded systems with less complex computation. The results reveal that the proposed method improves the estimation accuracy by at least 47.59% and reduced the inference time by an average of 29.20%. All tests are performed on an actual embedded system using several datasets to demonstrate and verify both the accuracy and effectiveness of the proposed method.