論文

In this paper, we propose an activity detection system using a 24 × 32 resolution infrared array sensor placed on the ceiling. We first collect the data at different resolutions (i.e., 24 × 32, 12 × 16, and 6 × 8) and apply the advanced deep learning (DL) techniques of Super-Resolution (SR) and denoising to enhance the quality of the images. We then classify the images/sequences of images depending on the activities the subject is performing using a hybrid deep learning model combining a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM). We use data augmentation to improve the training of the neural networks by incorporating a wider variety of samples. The process of data augmentation is performed by a Conditional Generative Adversarial Network (CGAN). By enhancing the images using SR, removing the noise, and adding more training samples via data augmentation, our target is to improve the classification accuracy of the neural network. Through experiments, we show that employing these deep learning techniques to low-resolution noisy infrared images leads to a noticeable improvement in performance. The classification accuracy improved from 78.32% to 84.43% (for images with 6 × 8 resolution), and from 90.11% to 94.54% (for images with 12 × 16 resolution) when we used the CNN and CNN + LSTM networks, respectively.

Forests have an important role in environmental preservation and maintenance. The primary threat is forest fires, which have disastrous repercussions. As a result, it is critical to identify and extinguish a fire before it spreads and destroys resources. To that end, we propose a forest fire detection and fighting mechanism using wireless sensor and actor networks (WSANs). Temperature sensors are utilized to detect fires, and actors (robots) are employed to extinguish them. Sensors and robots are distributed at random throughout the forest, forming clusters. Clustering, sleep/active scheduling for the sensors, and energy harvesting (EH)/moving modes for the robots, are used to extend and maximize the sensors/robots lifetime in the WSAN. In such a network, robots should move to the fire site as quickly as possible. To do this, we further propose a robot routing mechanism that focuses on determining the shortest path for each firefighting robot. In particular, each firefighting robot equipped with on-board processing uses a fuzzy Q-learning (FQL)-based trajectory mechanism to learn the shortest path to the fire zone in the least amount of time. Simulations are conducted to demonstrate the benefits of employing the proposed framework for rapid and effective fire response. When compared to the traditional Q-learning, the total approaching rate (a measure of how quickly the firefighting robots can reach the fire) to the fire spot is greater when utilizing the proposed FQL-based strategy.

With the development of the Internet of energy, more and more participants share data by different types of edge devices. However, such multi-source heterogenous data typically contain low-quality data, e.g., missing values, which may result in potential risks. Besides, resource-constrained devices incur large latency in edge computing networks. To alleviate such latency, distributed task offloading schemes are designed to share the computation burden between edge nodes and nearby servers. However, there are three main drawbacks of such schemes. First, low-quality data are not carefully evaluated by constraints under scenarios, which may result in slow convergence in distributed computation. Second, multi-source data including sensitive information are computed and shared among edge nodes without privacy protection. Third, distributed tasks on low-quality data may result in low-quality results even with an optimal offloading scheme. To address the problems above, a task offloading framework for edge computing based on consortium blockchain and distributed reinforcement learning is proposed in this paper, which can provide high-quality task offloading policies with data privacy protected. This framework consists of three key components: data quality evaluation (DQ) with multiple data quality dimensions, data repairing (DR) with a repairing algorithm based on a novel repairing consensus mechanism and distributed reinforcement learning for task arrangement (DELTA) with a distributed reinforcement learning algorithm based on a novel low-quality data distributing strategy. Numeric results are presented to illustrate the effectiveness and efficiency of the proposed task offloading framework for edge computing on low-quality data in the IoE.

High-Altitude Platform Stations (HAPSs) are attracting much attention as novel mobile communication platforms for ultra-wide coverage areas and disaster-resilient networks. Single-cell frequency reuse using multiple cells can increase capacity to cover a wide area. Thus far, we have proposed a cell configuration optimization method based on a genetic algorithm (GA). We clarified the optimal cell configuration in terms of spectral efficiency depending on the number of cells under a uniform user distribution scenario. Whereas user distributions differ depending on location. Thus, cell configuration optimization is required for a non-uniform user distribution. Each cell needs different antenna parameters for non-uniform user distributions, resulting in an increase in the number of parameters compared with a uniform user distribution, making the optimization difficult even by GA in some cases. To address this problem, we propose a co-evolutionary dynamic cell optimization algorithm. Co-evolution is one of the divide-and-conquer methods. The proposed method divides multiple cells into several groups to decrease the number of parameters to optimize at a time, and each group is optimized in order. The simulation results show that the cell configuration with the proposed method can increase the sum of the square root of throughput for non-uniform user distributions compared to that for uniform user distributions. Furthermore, the proposed method with three sub-areas can improve the sum of the square root of throughput while reducing the number of combinations performed compared to the method without subarea division for a nine-cell scenario.

Accurate downlink channel state information (CSI) is required to be fed back to the base station (BS) in frequency division duplexing (FDD) massive multiple-input multiple-output (MIMO) systems in order to achieve maximum antenna diversity and multiplexing. However, downlink CSI feedback overhead scales with the number of transceiver antennas, a major hurdle for practical deployment of FDD massive MIMO systems. To solve this problem, we propose a compressive sampled CSI feedback method based on deep learning (SampleDL). In SampleDL, the massive MIMO channel matrix is sampled uniformly in time/frequency dimension before being fed into neural networks (NNs), which will reduce the computational resource/time at user equipment (UE) as well as enhance the CSI recovery accuracy at the BS. Both theoretical analysis and normalized mean square errors (NMSE) results confirm the advantages of the proposed method in terms of time complexity and recovery accuracy. Besides, a suitable CSI feedback period is explored by link level simulations, which aims to further reduce the overhead of CSI feedback without degrading the communication quality.

Preventing COVID-19 disease from spreading in communities will require proactive and effective healthcare resource allocations, such as vaccinations. A fine-grained COVID-19 vulnerability map will be essential to detect the high-risk communities and guild the effective vaccine policy. A mobile-crowdsourcing-based self-reporting approach is a promising solution. However, an accurate mobile-crowdsourcing-based map construction requests participants to report their actual locations, raising serious privacy concerns. To address this issue, we propose a novel approach to effectively construct a reliable community-level COVID-19 vulnerability map based on mobile crowdsourced COVID-19 self-reports without compromising participants' location privacy. We design a geo-perturbation scheme where participants can locally obfuscate their locations with the geo-indistinguishability guarantee to protect their location privacy against any adversaries' prior knowledge. To minimize the data utility loss caused by location perturbation, we first design an unbiased vulnerability estimator and formulate the location perturbation probability generation into a convex optimization. Its objective is to minimize the estimation error of the direct vulnerability estimator under the constraints of geo-indistinguishability. Given the perturbed locations, we integrate the perturbation probabilities with the spatial smoothing method to obtain reliable community-level vulnerability estimations that are robust to a small-sampling-size problem incurred by location perturbation. Considering the fast-spreading nature of coronavirus, we integrate the vulnerability estimates into the modified susceptible-infected-removed (SIR) model with vaccination for building a future trend map. It helps to provide a guideline for vaccine allocation when supply is limited. Extensive simulations based on real-world data demonstrate the proposed scheme superiority over the peer designs satisfying geo-indistinguishability in terms of estimation accuracy and reliability.

Human activity recognition (HAR) is one of the most promising technologies in the smart home, especially radio frequency (RF-based) method, which has the advantages of low cost, few privacy concerns and wide coverage. In recent years, deep learning (DL) has been introduced into HAR and these DL-based HAR methods usually have outstanding performance. However, as the recognition scenarios and target change, the model performance drops sharply. To solve this problem, we propose a generalized method for cross-person activity recognition (CPAR), which is called snapshot ensemble learning based an attention with bidirectional long short-term memory (SE-ABLSTM). Specifically, by defining the cosine annealing learning rate, the models with diversity are saved and integrated in the same training process. In addition, we provide a dataset for CPAR and simulation results show that our method improves generalization performance by 5% compared to the original method. The source code and dataset for all the experiments can be available at https://github.com/NJUPT-Sivan/Cross-person-HAR.

Specific emitter identification (SEI) extracts the fingerprint characteristics of emitters according to the subtle differences of transmitted signals, to distinguish different emitter individuals and prevent unauthorized network access. Deep learning (DL) based SEI methods have been proposed to achieve a good identification performance in recent years. However, the existing methods need a massive specific emitter dataset to alleviate model overfitting during the training stage. In this paper, we propose data augmentation (DA) aided few-shot learning method and validate the proposed method using automatic dependent surveillance-broadcast (ADS-B) signals. Specifically, according to the characteristics of ADS-B signals, four DA methods, i.e., flip, rotation, shift, and noise are studied for the proposed method. Experimental results are provided to show that the proposed method improves the recognition accuracy and the model robustness.

Sentiment analysis refers to the process of automatically identifying the emotions expressed by people. Its accuracy is highly dependent on the amount of training data. However, it takes time and cost for humans to collect a large number of data. Many research works used generative models to generate a large amount of data based on a small amount of data for sentiment analysis. However, training on long texts and inaccurate sentiment information that might be generated are two severe challenges. It is difficult to improve the sentiment analysis accuracy effectively. In this paper, we propose a novel data augmentation framework based on Sequence generative adversarial networks (SeqGAN) to improve the sentiment analysis accuracy when the dataset already has a certain amount of data and contains long texts. Penalty-based SeqGAN is used to generate high-quality and diversified text data. Long short-term memory (LSTM) networks with attention mechanisms are used to conduct sentence compression for the training data of SeqGAN. A sentiment dictionary is used to retain the sentiment words for compressed data. We also propose a data screening method to obtain more accurate data from the generated data. The results of the usability, novelty, and diversity of the generated data show that the proposed sentence compression method can help SeqGAN learn more information from the long text data. The data generated by the proposed framework improve the classification accuracy of four classifiers applied on two distinct text datasets.

Scenario identification (SCI) is one of key techniques for intelligent vehicular communications (IVC) to maintain an effective and reliable operating state. Based on the deep learning (DL), it is a hotspot to identify scenarios of wireless communication using the characteristic quantity inherent in wireless channels. This paper proposes a decentralized learning-based SCI (DecentSCI) for IVC, relying on the algorithm of lightweight and model aggregation. By improving training efficiency and meanwhile reducing model complexity, the proposed method achieves low computing and communication, which is applicable for vehicular devices. Simulation results show that the training efficiency is upgraded by 97.15% and the model complexity is decreased by 90.25% at the cost of slight performance loss, i.e., 0.15%.

In a time division duplex (TDD) based massive multiple-input multiple-output (MIMO) system, a base station (BS) is required to obtain accurate estimation of channel state information (CSI) for a user terminal (UT). Because of the time-varying nature of the channel, the length of pilot signals is limited and the number of orthogonal pilot signals is finite. Hence, the same pilot signals are required to be reused in neighboring cells and thus its channel estimation performance is deteriorated by pilot contamination from the neighboring cells. The minimum mean square error (MMSE) channel estimation can be used to reduce the influence of pilot contamination. However, it needs to know the covariance matrix of channels for all the UTs, which is unknown to the BS in practice. In this paper, we propose two methods of deep learning aided channel estimation to reduce the influence of pilot contamination. One method uses a neural network consisting of fully connected layers, while the other method uses a convolutional neural network (CNN). The neural network, particularly the CNN, plays a role in extracting features of the spatial information from the contaminated signals. The former method is better in terms of the training speed, however, the latter one can estimate the channel more accurately. We evaluate the proposed methods under two scenarios, i.e., perfect timing synchronization and imperfect one. Simulation results confirm that the proposed methods are better than the LS and covariance estimation methods via normalized mean square error (NMSE). In addition, we also investigate the impact of channel aging, and show that including some expected data into training datasets can avoid the great degradation of estimation quality.

In this paper, we propose a novel Deep Reinforcement Learning Evolution Algorithm (DRLEA) method to control the antenna parameters of the High-Altitude Platform Station (HAPS) mobile to reduce the number of low-throughput users. Considering the random movement of the HAPS caused by the winds, the throughput of the users might decrease. Therefore, we propose a method that can dynamically adjust the antenna parameters based on the throughput of the users in the coverage area to reduce the number of low-throughput users by improving the users’ throughput. Different from other model-based reinforcement learning methods, such as the Deep (Formula presented.) Network (DQN), the proposed method combines the Evolution Algorithm (EA) with Reinforcement Learning (RL) to avoid the sub-optimal solutions in each state. Moreover, we consider non-uniform user distribution scenarios, which are common in the real world, rather than ideal uniform user distribution scenarios. To evaluate the proposed method, we do the simulations under four different real user distribution scenarios and compare the proposed method with the conventional EA and RL methods. The simulation results show that the proposed method effectively reduces the number of low throughput users after the HAPS moves.

Over the last years, more and more attention has been given by the researchers towards dementia diagnosis using computational approaches applied on speech samples given by dementia patients. With the progress in the field of Deep Learning (DL) and Natural Language Processing (NLP), techniques of text classification using these techniques that are derived from fields such as sentiment analysis have been applied for dementia detection. However, despite the relative success in these techniques, the two tasks (i.e., sentiment analysis and dementia detection) have major differences, leading us to believe that adjustments need to be made to make the detection more accurate. In the current paper, we use transfer learning applied on a common language model. Unlike conventional work, where the text is stripped from stop words, we address the idea of exploiting the stop words themselves, as they embed non-context related information that could help identify dementia. For this sake, we prepare 3 different models: a model processing only context words, a model stop words with patterns of part-of-speech sequences, and a model including both. Through experiments, we show that both grammar and vocabulary contribute equally to the classification: the first model reaches an accuracy equal to 70.00%, the second model reaches an accuracy equal to 76.15%, and the third model reaches an accuracy equal to 81.54%.

Polar code is one of the channel codes and is used in the 5th generation of mobile communication system (5G). This encoding scheme is based on the operation of channel polarization, and it is possible to achieve the capacity of arbitrary binary-input symmetric discrete memoryless channels. Compared with Turbo codes and LDPC (Low-Density Parity-Check) codes, the implementation of the encoder and decoder is easier. BP (Belief Propagation) decoding, one of the decoding methods of the polar codes, can be performed at high speed because it can decode in parallel. However, the disadvantages of the BP decoding are the hardware and computational complexities. The technique of quantization can be used to reduce complexities of hardware and calculation. One of the quantization methods is the information bottleneck method, which allows an observation variable to compressed one while trying to preserve the mutual information shared with a relevant variable. As a novel approach, the information bottleneck method is used in the design of quantizers for the BP decoding of LDPC codes. In this paper, we propose a discrete BP polar decoder that can use only unsigned integers in the decoding process by using the information bottleneck method. Thus, we can replace complex calculations of BP decoding with simple lookup tables. We also investigate the minimum bit width for quantization with negligible degradation and the suboptimal Eb/N0 for designing lookup tables, where Eb and N0 denote energy per bit and noise power density, respectively. The simulation results show that the proposed method can achieve almost the same error correcting capability compared with the BP decoding without compression in the range of low Eb/N0. Besides, we show that the proposed decoder can compress both channel outputs and BP messages with small loss compared with the uniform quantization decoder.

Activity detection is a key task in the monitoring of elderly people living alone. This is because it helps locate them and identify any accident that might occur to them. In this article, we propose a novel approach that uses 2-D light detection and ranging (LIDAR) and deep learning to perform activity detection. In a first step, our approach processes and interpolates the data collected using the 2-D LIDAR following an algorithm we propose to locate the person and identify the useful data points. In the next steps, the data are transformed into two types of representations: 1) a time-series type and 2) an image type. The time-series data are used to train different long short-term memory (LSTM) networks to identify the person and to recognize his/her activity, while the image type is used to fine-tune a convolutional neural network (CNN) for fall detection. Throughout our experiments, we show that our approach allows for the identification of people from their gait, and the detection of unsteady gait or unstable walk (i.e., when the person is about to fall or feeling dizzy) as well as the detection of up to four activities: 1) walking; 2) standing; 3) sitting; and 4) falling. The results obtained from our experiment show that the proposed method reaches an accuracy equal to 94.1% for multiclass activity detection, 98.6% for fall detection, 93.2% for person identification (for three different people), and 92.5% for unsteady walk detection.

Monitoring the activities of elderly people living alone is of great importance since it allows for the detection of when hazardous events such as falling occur. In this context, the use of 2D light detection and ranging (LIDAR) has been explored, among others, as a way to identify such events. Typically, a 2D LIDAR is placed near the ground and collects measurements continuously, and a computational device classifies these measurements. However, in a realistic environment with home furniture, it is hard for such a device to operate as it requires a direct line of sight (LOS) with its target. Furniture will block the infrared (IR) rays from reaching the monitored person thus limiting the effectiveness of such sensors. Nonetheless, due to their fixed location, if a fall is not detected when it happens, it cannot be detected afterwards. In this context, cleaning robots present a much better alternative given their autonomy. In this paper, we propose to use a 2D LIDAR mounted on top of a cleaning robot. Through continuous movement, the robot is able to collect distance information continuously. Despite having the same drawback, by roaming in the room, the robot can identify if a person is laying on the ground after falling, even after a certain period from the fall event. To achieve such a goal, the measurements captured by the moving LIDAR are transformed, interpolated, and compared to a reference state of the surroundings. A convolutional long short-term memory (LSTM) neural network is trained to classify the processed measurements and identify if a fall event occurs or has occurred. Through simulations, we show that such a system can achieve an accuracy equal to 81.2% in fall detection and 99% in the detection of lying bodies. Compared to the conventional method, which uses a static LIDAR, the accuracy reaches for the same tasks 69.4% and 88.6%, respectively.

Due to large amounts of available spectrum at high frequencies, millimeter-wave (mmWave) technology has gained extensive research attention in 5G communications, whereas mmWave links suffer from severe free space attenuation. Codebook-based beamforming techniques with multiple antennas can effectively alleviate this challenge with low computational complexity and low hardware cost. However, small delay and high-speed communications with beamforming techniques require beam alignment with small overhead so as to establish the wireless link quickly. In this context, this paper proposes a deep learning-based low overhead analog beam selection scheme by virtue of the super-resolution technology. To be concrete, deep neural networks are employed to conduct beam quality estimation based on partial beam measurements. Our proposed scheme can cover all the directions of arriving signals with low overhead by utilizing codebooks with different beam widths. Furthermore, for the purpose of further reducing the overhead, we formulate the beam quality prediction model based on the past beam sweepings. With these beam quality estimation and prediction model, the beam that achieves large signal-to-noise-power-ratio (SNR) can be selected based on partial beam measurements. Simulation results show that the proposed scheme can accurately estimate beam qualities and give high probability of optimal beam selections with low overhead.

A Multiple-Input Multiple-Output (MIMO) Frequency-Modulated Continuous Wave (FMCW) radar can provide a range-angle map that expresses the signal power against each range and angle. It is possible to estimate object locations by detecting the signal power that exceeds a threshold using an algorithm, such as Constant False Alarm Rate (CFAR). However, noise and multipath components often exist over the range-angle map, which could produce false alarms for an undesired location depending on the threshold setting. In other words, the threshold setting is sensitive in noisy range-angle maps. Therefore, if the noise is reduced, the threshold can be easily set to reduce the number of false alarms. In this paper, we propose a method that improves the CFAR threshold tolerance by denoising a range-angle map using Deep Image Prior (DIP). DIP is an unsupervised deep-learning technique that enables image denoising. In the proposed method, DIP is applied to the range-angle map calculated by the Curve-Length (CL) method, and then the object location is detected over the denoised range-angle map based on Cell-Averaging CFAR (CA-CFAR), which is a typical threshold setting algorithm. Through the experiments to estimate human locations in indoor environments, we confirmed that the proposed method with DIP reduced the number of false alarms and estimated the human location accurately while improving the tolerance of the threshold setting, compared to the method without DIP.

Fetal heart rate monitoring is an enormous challenge since the observed fetal electrocardiography (ECG) signal is typically characterized by a very low signal-to-noise ratio (SNR). In this letter, we aim to improve the accuracy of heartbeat detection by proposing an adaptive template for removing the maternal cycle. The template is formed by a matrix, each row of which consists of an abdominal recorded signal (ADS). It can be updated by integrating the incoming cycle while removing the contribution of the previous recording. This process is conducted by considering a discriminator to adapt the non-stationarity of each incoming cycle. Furthermore, to suppress the morphological change caused by noise, we propose a novel multi-layer hybrid network to reconstruct the chest maternal ECG (chest mECG) morphology from a set of templates. The approach has a deep structure of each layer consisting of a reservoir layer and an encoder layer. The reservoir layer explores multi-scale dynamics by transforming the input series into a high-dimensional space. The encoder layer achieves the collection of the encoder features from the output of the reservoir layer. Once the model is built, the output weight of a direct connection is trained by solving a regression problem. Experimental results show that the proposed method has a better performance compared with some classical approaches.

© 2018 IEEE. This paper divides antennas at the transmitter into several transmit antenna groups (TAGs) and extends the generalized spatial modulation (GSM) idea to a new precoding-aided multiple-input multiple-output system, which is referred to as precoding-aided GSM (PGSM) system. To mitigate the influence caused by correlated channels, we consider the interleaved grouping scheme which leads to a minimum antenna correlation within the same TAG. To select the receive antenna subset with low complexity, we propose an iterative greedy (IG) algorithm that contributes to a maximum equivalent channel gain with low-computational complexity. To reduce the complexity of maximum likelihood (ML) detector, a low complexity IG algorithm-based ML (IG-ML) detector is also derived. Simulation results show that the bit error rate (BER) performance gain up to 3.5 dB is obtained by the proposed PGSM with IG-ML detector over the PGSM with an ordered block minimum mean-squared error detector at a given BER of 10-4. It is also shown that the proposed PGSM with the IG-ML detector achieves a similar BER performance in a high signal-to-noise power ratio region compared with that of an ideal precoding for the GSM with an optimal ML detector.