PWPengyu Wang

Pengyu Wang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Pengyu Wang has a total of 26 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • DFPR-BEV: Cross-Modal Knowledge Distillation of Decouple Features and Projection Regions for Multicamera 3-D Object Detection

    by Jiaxin Liu, Wei Liu, Pengyu Wang, Yongming Li, Jun Hu, Shuai Cheng, Wenxing Yang

    Multi-camera 3D object detection receives extensive attention in autonomous driving because of its advantage in efficiency and cost-effectiveness. However, a performance gap still exists between camera-based detector and LiDAR-based detector. Because camera-based detector cannot extract precise depth and geometric information from the input image data, resulting in obvious projection deviation when transforming features from the perspective view to the bird’s eye view (BEV). We propose a cross-modal knowledge distillation framework (DFPR-BEV) to address this problem, making the camera-based student detector to improve its feature representation by imitating the trained LiDAR-based teacher detector during training. To prevent semantic information from being harmed by directly aligning features during distillation by modal variation, we improve the student model to decouple point clouds feature distillation and image projection position correction. A heatmap-based projection correction method is proposed to correct image feature projection deviation and eliminate conflicting features. A Gaussian mixture region-based method is introduced for comprehensive geometry and depth transmission of point clouds features. Finally, the detection performance is improved by effectively fusing the two distilled features. Extensive experiments on nuScenes benchmark illustrate that DFPR-BEV can significantly improve the performance of the student model, outperforming current distillation methods and other camera-based detectors with equal or lower computational budget. Notably, on the nuScenes val set after distillation, BEVDet’s NDS and mAP are improved by 4.9% and 4.2%, respectively, BEVDet4D’s inference speed is 5× faster than BEVFormer-S with the same mAP and higher NDS.

  • DMPG-BEV: Diffusion Model-Based Point Clouds Features Generation for Efficient Camera-Based BEV Perception

    by Jiaxin Liu, Pengyu Wang, Wei Liu, Shangwei Du, Yongming Li, Jun Hu, Shuai Cheng

    Bird’s Eye View (BEV) perception requires swift and accurate representation of real-world 3D information, and BEV-based 3D object detection and semantic segmentation are foundational for downstream autonomous tasks such as prediction and planning. Current BEV perception methods are divided into Light Detection and Ranging(LiDAR)-based and camera-based methods depending on the input data modalities. LiDAR-based methods offer precise 3D information, but are limited by high sensor costs. Conversely, camera-based methods are economical and widely researched. However, 2D image features transform into BEV space introducing feature distortion due to inaccurate 3D information, particularly in depth estimation which results in noisy projections. Denoising Diffusion Probabilistic Models (DDPM) are used to mitigate noise in BEV features. However, DDPM increases computational costs, relies on noisy images BEV features, and limits effectiveness. To address these issues, we propose a lightweight BEV perception framework (DMPG-BEV) to enhance the performance of camera-based methods. This framework employs Denoising Diffusion Implicit Models (DDIM) to generate point cloud features with 3D information, while reducing computational demands by fewer denoising steps. We also introduce a lightweight feature fusion module which integrates point cloud BEV features with image BEV features to incorporate 3D information and reduce noise. The refined BEV features are also well-suited for multi-task networks involving detection and segmentation tasks. Additionally, a lightweight view transformation module is included to further reduce computational costs. Extensive experiments on the nuScenes benchmark demonstrate better performance of our proposed model with reduced computation, highlighting its value in BEV perception.


  • A Two-Stage Earthquake Event Classification Model Based on Diffusion Probability Model

    by Fanchun Meng, Tao Ren, Pengyu Wang, Xinliang Liu, Wenjuan Xiang, Xinyu He

    Rapid and accurate classification of earthquake events is a serious challenge in seismology and disaster mitigation. Problems such as data imbalance, model interpretability, and model generalization limit the application of artificial intelligence methods in this research area. This paper introduces a real-time two-stage Diffusion Earthquake Event Classification (DiffEEC) model based on the Diffusion Probability Model (DPM). DiffEEC uses a two-stage classification approach and combines DPM and knowledge distillation (KD) techniques. DiffEEC focuses on seismic phase-related features and source mechanism-related features by combining time series features extracted by convolutional layers, DPM output, and InSAR data features to better extract core seismic data information and reduce reliance on manual feature design. DiffEEC uses focal loss to solve the data imbalance problem. Thus, DiffEEC can address data scarcity and imbalance, feature acquisition and selection, variability and complexity of seismic event processing, and model generalization through the mechanism. Experiments show that DiffEEC performs better in earthquake event classification.


  • Open Access

    A Ghost-Attention Network for Discriminating Tectonic and Non-Tectonic Events on a Small and Imbalanced Dataset

    by Xinliang Liu, Tao Ren, Hongfeng Chen, Pengyu Wang, Fanchun Meng

    Discrimination between tectonic and non-tectonic events is crucial to assess seismic hazards and manage associated risks. However, the discrimination process is challenging due to the imbalanced distribution of tectonic and non-tectonic events. In this paper, we propose a ghost-attention network (GA-Net) consisting of multiple ghost modules and convolutional block attention modules (CBAMs) to solve this problem. Ghost module allows the network to extract feature maps using cost effective operations, which are suitable for small and unbalanced training sets. CBAM emphasizes meaningful features along channel and spatial axes, effectively learning which information to emphasize or suppress. We train the proposed GA-Net using seismic data from Shaanxi Province in China. The evaluation shows that on the test set, the proposed GA-Net achieves sensitivity, specificity, and accuracy of 96.40%, 93.24%, and 95.75%, respectively. In addition, a comparative analysis with several state-of-the-art networks establishes its superiority. Our GA-Net exhibits robust generalized performance on several training sets with different imbalanced levels and an independent dataset.


  • Identifying influential spreaders in complex network based on the node’s weight and spreading probability

    by Tao Ren, Yanjie Xu, Pengyu Wang

    Identifying influential spreaders is a crucial aspect of network science with various applications, including rumor control, viral marketing and epidemic spread limitation. Despite the availability of various methods for identifying these spreaders in complex networks, there remains a fundamental question regarding their accurate and discriminative identification. To address the issues and account for each node’s propagation ability, we propose an algorithm to identify influential spreaders based on the node’s weight and spreading probability (NWSP) for identifying influential spreaders. The effectiveness of the proposed method is evaluated using the Susceptible–Infected–Recovered (SIR) model, Kendall’s Tau ([ ]) and monotonicity. The proposed method is compared with several well-known metrics, including degree centrality, K-shell decomposition, betweenness centrality, closeness centrality, eigenvector centrality and the centrality method based on node spreading probability (SPC), in ten real networks. Experimental results demonstrate the superiority ability of the proposed algorithm to accurately and discriminatively identify influential spreaders.

  • MFFnet: A Seismic Phase Picking Network Based on Multiple Feature Fusion

    by Pengyu Wang, Tao Ren, Rong Shen, Georgi M. Dimirovski, Xinliang Liu, Fanchun Meng

    With the recent improvement of deep learning (DL) techniques and computer hardware capabilities, neural networks are widely used to monitor massive sensor data and detect earthquakes in them. This makes designing fast, accurate, and generalized DL models necessary for an active field of research for automatic seismic phase picking. A seismic phase picking network called MFFnet is proposed to fuse power spectral density, expert knowledge, spectrograms, recurrence plots, and Gramian angle fields. The network uses fast Fourier convolution on 2-D representations to extract more interpretable features. Considering the high proportion of noisy signals in field applications, MFFnet uses focal loss as the loss function to improve network accuracy. Experimental results show that MFFnet achieves precision, recall, and accuracy with 0.96, 0.98, and 0.98, respectively, in seismic phase detection tasks. Shapley Value is used to evaluate the relationship between features and network predictions. Compared with other deep learning networks, the feature extraction approach used in this paper is more explanatory and provides greater confidence in the results.


  • Determination of earthquake focal mechanism via multi-task learning

    by Pengyu Wang, Tao Ren, Rong Shen, Hongfeng Chen, Xinliang Liu, Fanchun Meng


  • Open Access

    Neural Network Light Weighting Approach Using Multi-Metric Evaluation of Convolution Kernels

    by Pengyu Wang, Tao Ren, Jiaxin Liu, Fanchun Meng, Xinliang Liu, Wei Liu

    The prevalence of convolutional neural networks is hindered by their intricate structure and voluminous parameters, which consume significant processing resources during both training and inference. This study proposes a novel approach that involves treating convolutional kernels as tensors, utilizing suitable priori metrics to gauge their effectiveness, and employing a clustering algorithm to eliminate redundant convolutional kernels for network pruning. To measure the convolutional kernels, we use appropriate priori metrics and density-Based spatial clustering of applications with noise (DBSCAN) to cluster the convolutional kernels so that the cluster centroids are the reserved convolution kernels. The experimental results show that the method can effectively perform network lightweight while maintaining high accuracy. At the same time, the method can improve the efficiency of convolutional kernel utilization, thus reducing the computational resource consumption of the model. Empirical analyses conducted on datasets indicate that, in certain instances, the proposed pruning method outperforms established state-of-the-art methods.


  • IDENTIFYING VITAL NODES IN COMPLEX NETWORK BY CONSIDERING MULTIPLEX INFLUENCES

    by Tao Ren, Yanjie Xu, Lingjun Liu, Enming Guo, Pengyu Wang

    Identifying vital nodes is a fundamental topic in network science. Some methods are proposed to identify vital nodes in a complex network. These measures take into account different aspects of a node’s importance, such as its number of connections, the centrality of its connected nodes, and the distribution of its connections. Applying these measures makes it is possible to identify the nodes that play a vital role in the network and that have the greatest impact on its structure and function. However, there is still an inherent problem with identifying vital nodes accurately and discriminatively. To address the problem, for undirected unweighted networks, we propose an algorithm based on the nodes’ multiplex influences via the network structure to identify vital nodes. The effectiveness of the proposed method is evaluated by Kendall’s Tau ([ ]) and monotonicity and compared with well-known existing metrics such as degree centrality, K-shell decomposition, H-index, betweenness centrality, closeness centrality, eigenvector centrality, collective influence, and gravity model in 10 real networks. Experimental results show the superiority of the proposed algorithm in identifying vital nodes.


  • BTRPP: A Rapid PGA Prediction Model Based on Machine Learning

    by Tao Ren, Pengyu Wang, Hongfeng Chen, Xinliang Liu, Fanchun Meng, Yanlu Ma

    Peak ground acceleration (PGA) is a critical parameter in the postearthquake intensity snapshot. In this article, we propose a Bagged Tree for Rapid PGA Prediction (BTRPP) based on velocity waveforms recorded at stations around the epicenter, as well as a feature pruning algorithm based on BTRPP to predict PGA and analyze the relationship between PGA and manually predefined features. The research not only aims to discover the best model for forecasting PGA but also examines the impact of features using the interpretability of some models. As inputs to the model, BTRPP uses and mixes features in the time and frequency domains. Bayesian optimization (BO) is used to find the optimum hyperparameters of BTRPP. The results demonstrate that, using the data within 5 s after the first arrival of the P-wave, the BTRPP with BO can reduce the RMSE to 0.2076 and enhance the R-square to 0.91. It is, finally, found that the PGA is most closely related to the average positive amplitude, average negative amplitude, and spectral slope through the feature pruning algorithm in this article, which indicates that the results are plausible.