Search

The prediction of traffic flow has emerged as a pivotal element within the domain of intelligent transport systems, garnering considerable interest and attention from various quarters. SpatioTemporal Graph Neural Networks (STGNNS) have been extensively employed to develop traffic representations. However, the extant research is constrained by several limitations: 1) The majority of STGNNs fail to account for the spatial heterogeneity of traffic data, with the distribution of traffic flows in different regions potentially being biased. This makes it challenging to capture comprehensive traffic flow data features. 2) In the case of complex spatial relationships, the loss of spatial correlation makes it challenging to accurately capture the dynamically variable spatial dependence of traffic flow data. In order to address these challenges, this paper proposes a new coordinate attention enhanced adaptive spatiotemporal convolutional network prediction model (CAAS), which introduces coordinate attention with the objective of modelling spatial heterogeneity and capturing complex spatiotemporal correlations. Moreover, a novel spatial multi-head attention mechanism is introduced with the objective of capturing the intricate interdependencies of multi-scale dynamic spatiotemporal data. This is accomplished by augmenting the input matrix of the softmax function in the attention mechanism, which enhances the differentiation of attention weights and facilitates the capture of spatial relationships at a finer granularity. Ultimately, the prediction of traffic flow is accomplished through the adaptive fusion of temporal and spatial features via a gated fusion mechanism.Experiments conducted on the publicly available PEMS04 and PEMS08 datasets show that the proposed CAAS model significantly surpasses existing cutting-edge methods.

Drawing anime characters with facial features of different ages is a challenging task. The characters’ facial features vary significantly with age, making it especially difficult for beginners to depict age-specific anime characters accurately. In this paper, we propose AgeFace, an interactive drawing interface designed to help users draw high-quality anime faces for characters of multiple age groups. AgeFace can provide a combination of local and global user guidance in the drawing process to enhance both detailed facial features and the overall aging features. Local guidance assists users in drawing detailed facial features, while global guidance provides hints for the overall layout of the face and additional features, such as wrinkles. During the local guidance stage, we apply an image retrieval approach to provide detailed instructions on facial features. In the global guidance stage, we propose the Transformer-based sequential generation model to create entire anime faces from drawn stroke sequences. The proposed framework of AgeFace combines a data-driven retrieval method and the generation model to provide users with inspiration during the drawing process. To verify the effectiveness of our guidance, we conducted user studies and comparison experiments with existing sketch generation models. The results demonstrated that AgeFace can significantly help users create multi-age anime faces and validate the effectiveness of our proposed generative model.

There is a growing demand for tea harvesting robots due to the harsh harvesting environment and rising in labor costs. Despite considerable efforts by the research community, the robustness to hilly field environment and the harvesting efficiency of the device in mechanical harvesting of bulk tea remain unimproved. To lay the foundation for automated tea harvesting, this paper proposes an integrated system of bulk tea harvesting robots with autonomous profiling logic based on computer vision. A saliency detection algorithm is applied to detect tender leaves, and a target localization system is designed with it. Moreover, a depth cue-based profiling logic and the corresponding pose adjustment strategy for cutting tool are detailed. And an actuator driven by a combination of motors is developed to provide a precise and flexible motion control to the cutting tool. Additionally, the standard position of the RGB-D camera, accuracy and timeliness of the profiling operation are confirmed by the field experiment. The results of field experiment show that the average harvesting accuracy is 87.7%, and failure rate is controlled within 15%. An analysis of failure causes reveals that damage and cutting failure are the primary reasons for the unsuccessful harvesting.

The bus conversion system in the hybrid AC/DC grid is usually composed of a three-phase interlinking converter and a DC-side input filter. The impedance interaction between the converter and the input filter may cause instability of the DC-link voltage. The previously proposed stabilization methods either require additional hardware or significantly impact the dynamic performance. Therefore, this paper presents a parallel virtual impedance regulator (PVIR), which adds desired virtual impedance in parallel with the input impedance of the converter. The proposed PVIR can precisely modify the impedance within the specific frequency range to minimize the impact on the system dynamics. The concept and detailed design procedure with the redundant design of the PVIR are introduced. The effect of the PVIR injection position is discussed. Finally, the experiment validates the effectiveness of the proposed PVIR.

Recent literature has proposed approaches that learn control policies with high performance while maintaining safety guarantees. Synthesizing Hamilton-Jacobi (HJ) reachable sets has become an effective tool for verifying safety and supervising the training of reinforcement learning-based control policies for complex, high-dimensional systems. Previously, HJ reachability was restricted to verifying low-dimensional dynamical systems primarily because the computational complexity of the dynamic programming approach it relied on grows exponentially with the number of system states. In recent years, a litany of proposed methods addresses this limitation by computing the reachability value function simultaneously with learning control policies to scale HJ reachability analysis while still maintaining a reliable estimate of the true reachable set. These HJ reachability approximations are used to improve the safety, and even reward performance, of learned control policies and can solve challenging tasks such as those with dynamic obstacles and/or with lidar-based or vision-based observations. In this survey paper, we review the recent developments in the field of HJ reachability estimation in reinforcement learning that would provide a foundational basis for further research into reliability in high-dimensional systems.

In this study, a pattern reconfiguration liquid metal printed antenna is proposed. The antenna, based on the well-known Yagi-Uda antenna, is obtained by directly printing gallium-based liquid metal on a perfluorosulfonic acid ionomer (PFSA) substrate. Since PFSA undergoes two thermally reversible transitions above room temperature, thus exhibiting a triple shape memory effect, the bending state of the antenna substrate can be changed by adjusting the temperature. By changing the operating state of the antenna, the main beam of the antenna can be reconfigured in the E-plane. The proposed E-plane pattern reconfigurable antenna is simple in structure and easy to realize, and its measured results are in good agreement with the simulation results, with the gain and radiation efficiency of 4.4 dBi and 81.2%, respectively. The antenna has the advantages of simple fabrication, light weight, strong E-plane beam switching capability and high gain. These advantages make it possible to be applied in the field of pattern reconfiguration antennas.

The carrier multiplication phenomenon involves hot carriers, which gain kinetic energy while accelerating to equilibrium with the established avalanching electric fields, and is typically explained via the local avalanche model. This work presents two vertical Ge-on-Si avalanche photodetectors fabricated in a separate absorption, charge, and multiplication configuration. Uniformity in materials, doping densities, and device dimensions is maintained, except for the multiplication width, which is used as a control parameter to manipulate avalanching fields under identical electric biasing and illumination schemes. Nonlocal carrier multiplication model is implemented during analysis of the extracted current-voltage signatures under small and large reverse biasing arrangements. For such an APD characterized by thinner multiplication region $\left ({{W_{m}=0.1 ~\mu m}}\right)$ , reduced linear and Geiger-mode multiplication regimes are perceived to be at play, outperforming the device having thicker multiplication region in almost all related figures of merit, e.g., responsivity $(22.58 \mathrm {A/W})$ , photo-to-dark current ratio $(\sim {10}^{5})$ , normalized photo-to-dark current ratio $(2.5\times {10}^{9} \mathrm {W}^{-1})$ , specific detectivity $(7.45\times {10}^{12}\mathrm {Jones})$ , and noise equivalent power $(\sim 2.42\times {10}^{-15} \mathrm {W/}\sqrt {\mathrm {Hz}})$ . The enhanced performance characteristics are due to excessively strong avalanching fields, reduced thermal charge density, and negligible dead space compared to its counterpart characterized by thicker multiplication width.

Although deep learning methods based on single sensors are widely applied in fault diagnosis, leveraging multi-sensor data to learn useful information remains a challenge. To fully utilize multi-sensor information, this paper proposes a lightweight improvement of the EfficientNetV2 architecture, combined with sensor fusion technology and transfer learning techniques, to develop an efficient and reliable new method specifically for motor fault diagnosis. First, the continuous wavelet transform is utilized to convert the signals from various sensors into time-frequency images, and the Mallat algorithm is employed to decompose each image into sub-band coefficients at different levels. Secondly, a fusion reconstruction method is constructed using coefficient absolute maximum and weighted average fusion rules to integrate the sub-band coefficients of multi-sensor time-frequency images at different levels. Subsequently, EfficientNetV2 is improved to enhance the model’s feature extraction capabilities, computational efficiency, and achieve lightweight effects. The EfficientNetV2-M0 network modifies the model’s depth and width multiplicity factors, reducing parameters and computational complexity. Furthermore, this network incorporates Diverse Branch Block (DBB) and Multidimensional Collaborative Attention (MCA) to enhance feature extraction under complex backgrounds, and the maximum cross-entropy loss function is improved by using label smoothing and focal loss to dynamically adjust the classification weights for improved accuracy. The network leverages pre-trained models obtained through transfer learning techniques for deployment, combining multi-sensor information fusion and the improved lightweight model for fault diagnosis applications. Finally, a fault diagnosis experiment is conducted using a motor state dataset. The experimental results demonstrate that the proposed method outperforms the control method in terms of diagnostic performance and robustness, with an accuracy of 100%, and it exhibits excellent performance even under conditions of small sample data, with an accuracy of 98.81%.

The identification of optimal landing sites is a critical first step for successful missions to the Moon and other extraterrestrial bodies, necessitating the integration of various environmental factors over large spatial scales. At the lunar south pole, site selection must balance engineering safety with areas of high scientific interest, requiring extensive analysis of potential locations. Although intelligent algorithms have been increasingly investigated for this purpose, the application of deep learning techniques in landing site selection remains unexplored. In this study, we employ one-dimensional convolutional neural networks (1D-CNNs) to quantitatively assess potential landing sites for exploration and lunar base construction, considering both scientific and engineering criteria. We also evaluate the influence of various factors on site selection using Shapley additive explanations (SHAP) values. The 1D-CNN model demonstrates robust performance across training, validation, and testing phases. Potential landing sites identified comprise less than 1% of the total study area, with factors such as visibility, volatile distribution, topography, and geological characteristics playing crucial roles. By applying operational constraints, we delineate sites suitable for direct landings and further refine this subset for base construction based on stringent requirements for resource utilization and energy sustainability. The combined use of CNN and SHAP enables more effective potential site screening and a deeper understanding of the factors influencing selection. Our findings offer a valuable framework for future lunar south pole expeditions, potentially minimizing manual survey efforts and enhancing the precision of landing site selection.

Goal: To establish Pulse2AI as a reproducible data preprocessing framework for pulsatile signals that generate high-quality machine-learning-ready datasets from raw wearable recordings. Methods: We proposed an end-to-end data preprocessing framework that adapts multiple pulsatile signal modalities and generates machine-learning-ready datasets agnostic to downstream medical tasks. Results: a dataset preprocessed by Pulse2AI improved systolic blood pressure estimation by 29.58%, from 11.41 to 8.03 mmHg in root-mean-square-error (RMSE) and its diastolic counterpart by 26.01%, from 7.93 to 5.87 mmHg in RMSE. For respiration rate (RR) estimation, Pulse2AI boosted performance by 19.69%, from 1.47 to 1.18 breaths per minute (BrPM) in mean-absolute-error (MAE). Conclusion: Pulse2AI turns pulsatile signals into machine learning (ML) ready datasets for arbitrary remote health monitoring tasks. We tested Pulse2AI on multiple pulsatile modalities and demonstrated its efficacy in two medical applications. This work bridges valuable assets in remote sensing and internet of medical things to ML-ready datasets for medical modeling.

To address the issue of low detection accuracy and high false positive rate in existing network intrusion detection methods, this paper proposes an intrusion detection model based on CNN-BiLSTM-Attention. Firstly, CNN is used to extract the spatial features from the intrusion data; Secondly, BiLSTM is used to mine the temporal features from the intrusion data further; Thirdly, the attention mechanism is used to assign different weights to the extracted spatiotemporal features and then enhance the role of important features in the calculation process, which can improve the classification accuracy of the model. In addition, for the problem of class imbalance existing in network intrusion data, Equalization Loss v2 is introduced as the loss function of the CNN-BiLSTM-Attention model, making the model pay more attention to minority class data during the training process, thereby improving the detection rate of the model for the minority class data. Finally, comparative experiments are conducted on NSL-KDD, UNSW-NB15, and CIC-DDoS2019 datasets. The experimental results show that the CNN-BiLSTM-Attention model outperforms the other models in terms of accuracy, detection rate, and false positive rate.

This work presents a novel InGaAs/InP avalanche photodiode, fabricated in the separate absorption, grading, charge, and multiplication configuration operated at non-cryogenic conditions under low-frequency ramp gating. An optimized three stage InP multiplication layer of $1\mu m$ thickness offers an extended linear mode operation by reducing the punch-through voltage, and indefinitely increasing the avalanche threshold voltage. A large background dark current is observed following steady, and linear multiplication in approximately direct relationship with the ramp gating. For 1310 nm short-wave infrared, normal incidence pulsed illumination at instant-to-peak voltage ratios of $(0.11,0.2,0.6, 0.89, 0.98, 0.9)$ , a sort of negative differential resistance is incorporated into the device in a qualitative sense, owing to the illumination induced switching/variations in the intrinsic values of electron, and hole avalanche coefficients in the multiplication region. Under fixed illumination, an interesting deduction from the transient photo response is the slow quenching phenomenon prolonging $\sim 120 \mu s$ for all the electrical field establishments in the device. The related measurement scheme paves the way for futuristic ramp-driven InGaAs/InP APDs for detecting SWIR wavelengths under required low power consumption environments.

Recently, significant developments have been made in Long Short-Term Memory (LSTM) networks within the realm of synthesis music. Notwithstanding these advancements, several challenges persist warranting further research. Primarily, there exists an absence of dedicated research on the application of LSTM networks for the synthesis of Original Sound Tracks (OST). Secondly, in general, people can only judge whether the synthesized music meets their expectations based on the model output. However, due to the time-consuming of training the model may need to try multiple times to obtain successful training results. Moreover, the subjective of music quality evaluation relying on human perception, not only the result of model training. To address these multifaceted challenges, this paper concentrates specifically on OST and proposes a framework termed the OST Synthesis Framework (OSTSF) utilizing LSTM. This framework accepts various OST types as input, processed through LSTM to yield innovative OST. Additionally, a novel preprocessing algorithm is proposed to screen input OST elements such as notes and chords, enabling control over music type and quality before the training phase. This algorithm serves to mitigate training uncertainties and reduce situations that require repeated training. Besides, a postprocessing approach, leveraging mathematical formulations facilitates the evaluation of synthesis OST also proposed. This approach aims to quantify subjective evaluations, providing a more intuitive representation through scoring metrics. Experiment results reveal that the OSTSF synthesized OST received favorable rate among a cohort of 100 surveyed respondents attaining 78.8%, demonstrating the efficacy of the proposed framework in the realm of music synthesis utilizing LSTM.

The reliable classification of motor unit action potentials (MUAPs) provides the possibility of tracking motor unit (MU) activities. However, the variation of MUAP profiles caused by multiple factors in realistic conditions challenges the accurate classification of MUAPs. The goal of this study was to propose an effective method based on the convolutional neural network (CNN) to classify MUAPs with high levels of variation for MU tracking. MUAP variation was added artificially in the synthetic electromyogram (EMG) signals and was induced by changing the forearm postures in the experimental study. The proposed overlapped-segment-wise EMG decomposition method and the spike-triggered averaging method were combined to obtain the MUAP waveform samples of individual MUs in the experimental study, and the MUAP profile classification performance was tested. Since the ground-truth of MU discharge activities was known for the synthetic EMG, the MU tracking performance was further verified by mimicking the tracking procedure of MU discharge activities and the spike consistency with the true spike trains was tested in the simulation study. The conventional MUAP similarity index (SI)-based method was also performed as comparison. For both the experimental and the synthetic EMG signals, the CNN-based method significantly improved the MUAP tracking performance compared with the conventional SI-based method manifested as a higher classification accuracy (93.3%±5.4% vs 56.2%±13.9%) in the experimental study or higher spike consistency (71.1%±10.2% vs 29.2%±11.0%) in the simulation study with a smaller variation. These results demonstrated the efficiency and robustness of the proposed method to distinguish MUAPs with large variations accurately. Further development of the proposed method can promote the study on the physiological and pathological changes of the neuromuscular system where tracking MU activities is needed.

Federated Learning (FL), as a promising distributed learning paradigm, is proposed to solve the contradiction between the data hunger of modern machine learning and the increasingly stringent need for data privacy. However, clients naturally present different distributions of their local data and inconsistent local optima, which leads to poor model performance of FL. Many previous methods focus on mitigating objective inconsistency. Although local objective consistency can be guaranteed when the number of communication rounds is infinite, we should notice that the accumulation of global drift and the limitation on the potential of local updates are non-negligible in those previous methods. In this article, we study a new framework for data-heterogeneity FL, in which the local updates in clients towards the global optimum can accelerate FL. We propose a new approach called Slingshot. Slingshot's design goals are twofold, i.e., i) to retain the potential of local updates, and ii) to combine local and global trends. Experimental results show that Slingshot helps local updates become more globally favorable and outperforms other popular methods under various FL settings. For example, on CIFAR10, Slingshot achieves 46.52% improvement in test accuracy and 48.21× speedup for a lightweight neural network named SqueezeNet.

Multitemporal change detection (CD) plays a crucial role in the remote sensing application field. In recent years, supervised deep learning methods have shown excellent performance in detecting changes in very-high-resolution (VHR) images. However, these methods require a large number of labeled samples for training, making the process time-consuming and labor-intensive. Unsupervised approaches are more attractive in practical applications since they can produce a CD map without relying on any ground reference or prior knowledge. In this article, we propose a novel unsupervised CD approach, named transformer-based multivariate alteration detection (trans-MAD). It utilizes a pre-detection strategy that combines the compressed change vector analysis and the iteratively reweighted multivariate alteration detection (IR-MAD) to generate reliable pseudotraining samples. More accurate and robust CD results can be achieved by leveraging the IR-MAD to detect insignificant changes and by incorporating the transformer-based attention mechanism to model the difference or similarity between two distant pixels in an image. The proposed trans-MAD approach was validated on two VHR bitemporal satellite remote sensing datasets, and the obtained experimental results demonstrated its superiority comparing with the state-of-the-art unsupervised CD methods.

Recently, some research has utilized machine learning methods to identify critical nodes in complex networks. However, existing approaches often lack a comprehensive consideration of network structural features during node feature extraction. Benefiting from the powerful feature extraction capability of network representation learning methods, a simple and effective algorithm for identifying key nodes in complex networks, termed Network Representation Learning and Key Node Identification (NRL_KNI), is proposed. The NRL_KNI algorithm utilizes network embedding techniques for learning node feature representations, followed by clustering and the utilization of quota-based limited sampling to obtain sampled nodes. Subsequently, these sampled nodes are employed to train a regression model for predicting the diffusion capability of unsampled nodes. To rank node influences, a Local Structure Influence Score (LSIS) based on the local structure is introduced to evaluate nodes’ final impact. Experimental results on eight real-world datasets demonstrate that the NRL_KNI algorithm generally outperforms traditional centrality methods and network representation learning-based methods in terms of the Jaccard similarity coefficient and Kendall’s Tau correlation coefficient evaluation metrics.

In power systems with high penetration of renewable energy, dynamic issues associated with multi-device interaction involving DC voltage control (DVC) and AC current control (ACC) of power electronic (PE)-interfaced devices are constantly emerging. The device participates in the interaction dynamics based on the way that DVC adjusts the current reference of ACC to regulate the internal voltage amplitude/frequency under the active/reactive power imbalance. Thus, to analyze the effect mechanism of the device on system dynamics, the device should be characterized as internal voltage in response to power imbalance stimulation. However, the existing modeling works fail to recognize the process of the device participating in system dynamics, so the regulation mechanism of the internal voltage by the device under power imbalance for the system dynamics analysis remains unclear. Therefore, this paper proposes a modeling method of the PE-interfaced device based on the recognition of the regulation process of the internal voltage amplitude/frequency by the device under the active/reactive power imbalance. The model based on power-imbalance stimulation and internal-voltage response relationships is first established to characterize the regulation by detailed controls. Since the stimulation-response relationships are dominated by DVC, to directly describe the regulation by DVC, the model with ideal ACC is further proposed through the infinity gain equivalence of ACC. Based on the model, the characteristics of the device in DVC timescale and the impact mechanism of the device on system dynamics by regulating the internal voltage amplitude/frequency are revealed. Simulation results for verification are also included.

The Tianwen-1 mission, China's first interplanetary endeavor and Mars mission, touched the surface of the Red Planet on May 15, 2021. With the successful landing of the Zhurong rover on the southern Utopian Planitia of Mars, the Mars Surface Composition Detector (MarSCoDe) on board the rover has started to analyze the material composition of the Martian surface by using the laser-induced breakdown spectroscopy (LIBS). However, changes in instrument temperature and external environment during operation will cause spectral drift in LIBS data, leading to inaccurate material composition inversion results due to shifts in elemental characteristic peak positions. To address this problem, an adaptive spectral drift correction (ASDC) approach is proposed. By considering the distribution of characteristic peaks and drift differences between different elements, the proposed ASDC method corrects the spectral drift in LIBS science data according to an adaptive spectral segmentation strategy. Experimental results obtained on 36 real LIBS science data acquired over 19 exploration activities confirmed the effectiveness of the proposed approach in terms of lower average drift values, by comparing it with two reference methods. Taking the onboard calibration data as a reference, the average spectral drift values of the LIBS science data before and after applying ASDC decreased from 0.1012 nm to 0.0179 nm, 0.0926 nm to 0.0158 nm, and 0.1697 nm to 0.0103 nm in three LIBS channels, respectively. Furthermore, the resulting decrease in the standard deviation and rms error of the average drift values also proved the robustness of the proposed method.

The multi-directional flow of energy in a multi-microgrid (MMG) system and different dispatching needs of multiple energy sources in time and location hinder the optimal operation coordination between microgrids. We propose an approach to centrally train all the agents to achieve coordinated control through an individual attention mechanism with a deep dense neural network for reinforcement learning. The attention mechanism and novel deep dense neural network allow each agent to attend to the specific information that is most relevant to its reward. When training is complete, the proposed approach can construct decisions to manage multiple energy sources within the MMG system in a fully decentralized manner. Using only local information, the proposed approach can coordinate multiple internal energy allocations within individual microgrids and external multilateral multi-energy interactions among interconnected microgrids to enhance the operational economy and voltage stability. Comparative results demonstrate that the cost achieved by the proposed approach is at most 71.1% lower than that obtained by other multi-agent deep reinforcement learning approaches.