Against the backdrop of rapidly surging computing demand driven by large-scale artificial intelligence model training and inference, the photonics-electronics convergence is accelerating the transformation of communication network performance and advanced computing architectures within intelligent computing centers. This convergence has emerged as a key technology for enabling high-bandwidth, low-latency interconnection and energy-efficient computing among computing resources. Focusing on switching, interconnection, and computing scenarios for advanced computing within intelligent computing centers, this paper systematically introduces the architectural principles, research progress, and development trends of four representative photonics-electronics convergence technologie. The technologies include Optical Circuit Switching (OCS), Co-Packaged Optics (CPO), Optical Input/Output (OIO), and photonic-electronic hybrid computing. Based on an analysis of the opportunities and challenges facing photonics-electronics convergence, the future evolution trends of these technologies are further discussed.
ObjectiveTo address the problems of low detection accuracy and slow inference speed caused by complex backgrounds and significant scale variations in small object detection of Unmanned Aerial Vehicle (UAV) aerial images, this paper proposes an improved algorithm named YOLOv11n-DPD, based on the YOLOv11n model. The goal is to enhance the detection accuracy of small objects in UAV aerial images while achieving model lightweighting and inference acceleration, making it suitable for deployment on resource-constrained embedded devices.MethodsFirst, Deformable Convolution (DCN) is introduced into the CSP Bottleneck with 2 Convolutions (C3k2) module to enhance the model's ability to extract features of multi-scale objects in complex backgrounds, especially for small objects with significant deformation. Second, according to the high proportion of small objects in UAV aerial images, a new Prediction Layer 2(P2) is added as the small object detection layer, while the Prediction Layer 5(P5) for large object detection is removed. This improves the capture accuracy of small object features and reduces redundant computation in large object detection. Then, a Focal-DIoU loss function combining Focal Loss and Distance Intersection over Union (DIoU) Loss is proposed to alleviate class imbalance in small object detection and optimize the localization accuracy of bounding boxes. Finally, Layer-Adaptive Magnitude-based Pruning (LAMP) is employed to lighten the model by adaptively evaluating each layer and removing redundant connections, thereby reducing the overall computational load and accelerating the inference process.ResultsExperimental results on the VisDrone2019 dataset show that YOLOv11n-DPD outperforms the original YOLOv11n model, with a 6.5% improvement in mAP50, namely the mean average precision at an intersection over union threshold of 0.5, reaching 40.7%. This significantly improves the detection accuracy of small objects in UAV aerial images. In addition, YOLOv11n-DPD increases the inference speed by 32 and 6 FPS on the Ubuntu and Jetson platforms, respectively. The model size and parameter count are reduced by 3.87 and 2.02 MB, respectively, demonstrating the advantage of the proposed algorithm in lightweighting. Furthermore, compared with existing mainstream small object detection methods, YOLOv11n-DPD achieves better overall performance in detection accuracy, inference speed, and computational resource consumption.ConclusionThe proposed YOLOv11n-DPD algorithm improves detection accuracy, accelerates inference, and optimizes model size by introducing DCN and model pruning, optimizing the small object detection layer structure, and improving the loss function. It is especially suitable for UAV aerial small object detection tasks and deployment inference on embedded platforms.
ObjectiveTo address the problems of inter-class similarity, significant intra-class variation, and insufficient representation capability of single-modal features for external vibration signals in complex long-distance oil and gas pipeline scenarios, a multimodal fusion-based pipeline external safety monitoring method is proposed to improve the event recognition performance of Distributed Acoustic Sensing (DAS) systems in practical operating environments.MethodsReal-world vibration signals were collected by a DAS system deployed on a provincial smart pipeline network platform. Each signal sample was constructed into three modalities, namely raw waveform, short-time Fourier transform spectrogram, and Gramian angular field image. According to the data representation characteristics of different modalities, a multi-branch feature extraction network was developed. A one-dimensional convolutional neural network was used to extract the local temporal features. A two-dimensional convolutional neural network was used to extract time-frequency texture and energy distribution features, and a vision Transformer was used to extract global correlation structure features from encoded images. On this basis, a cross-modal attention fusion mechanism was designed to align, interact, and fuse heterogeneous modal features. Meanwhile, considering class imbalance and modal fusion bias in real-world scenarios, a cross-modal collaborative focal loss function was constructed to enhance the learning capability of the model for hard-to-classify and minority-class samples.ResultsExperimental results on a real pipeline external event recognition dataset show that the proposed method achieves stable convergence during training and good overall performance in the pipeline external event recognition task. The average precision, average recall, and average F1-score reach 0.968 3, 0.972 8, and 0.970 5, respectively, outperforming the comparison models. In addition, the comparison between single-modal and multimodal inputs further demonstrates that multimodal input provides more sufficient discriminative information and significantly improves the representation capability of the model for complex vibration signals.ConclusionThe proposed method can effectively integrate raw temporal information, time-frequency distribution, and global correlation structure, thereby improving the separability of similar events under complex backgrounds. It provides an effective technical solution for DAS-based pipeline external safety monitoring and intelligent early warning.
As short-reach optical interconnects evolve toward data rates of 200 Gbit/s/λ and beyond, the capacity scaling of conventional silica single-mode-fiber interconnect systems is facing increasingly physical constraints. This paper focuses on the application value of Anti-Resonant Hollow-Core Fiber (AR-HCF) in high-capacity and low-complexity Intensity Modulation/Direct Detection (IM/DD) short-reach optical interconnect systems. Following the logic of single-wavelength high-speed support, cross-band transmission validation, Wavelength Division Multiplexing (WDM) capacity expansion, and Dense WDM (DWDM) high-density capacity validation, this paper summarizes representative experimental progress in AR-HCF-based short-reach optical interconnects in recent years. It also analyzes its technical contributions in terms of the transmission channel, device packaging, modulation formats, and low-complexity digital signal processing. This review shows that AR-HCF extends the physical boundary for high-density DWDM transmission over the C+L bands in IM/DD systems by reducing dispersion and nonlinear impairments, thereby providing a low-complexity scaling path for terabit-class short-reach optical interconnects.
ObjectiveThe aim of this paper is to generate broadband chaotic signals based on Integrated Mutually Coupled Semiconductor Lasers (IMCSL) and expand their applications in chaotic secure communication, key distribution and other fields.MethodsThe nonlinear dynamical characteristics of IMCSL were studied by building a simulation model in VPIphotonics optical simulation software. The device is composed of two mutually coupled Distributed FeedBack (DFB) sections.ResultsFirstly, the static characteristics of the two DFB sections were analyzed, and their threshold currents and slope efficiency of optical power versus bias current were explored. The center wavelengths and relaxation oscillation frequencies of the two sections under 4.5 times the threshold current, as well as the working wavelength variation range within 1 to 6 times the threshold current, were determined. Secondly, by fixing the working current of DFB2 and changing the working current of DFB1, the evolution path of IMCSL into chaos was investigated. Results show that with the increase of DFB1's working current, IMCSL enters chaos through steady state and dual-mode state, and finally returns to the steady state after mode jump. By analyzing the influence of the two DFB sections' working currents on the dynamic state of IMCSL's output signal, the parameter range for generating dual-mode and chaotic signals was obtained. Finally, by fixing the working current of DFB1 and changing the working current of DFB2, it is found that the bandwidth of chaotic signals gradually widens with the increase of working current. Through analyzing the influence of the two DFB sections' working currents on the effective bandwidth and flatness of chaotic signals, the parameter range for generating chaotic signals with effective bandwidth greater than 20 GHz and flatness less than 3 dB was finally determined.ConclusionThe integrated laser composed of two mutually coupled DFB sections, which can generate chaotic signals with bandwidth exceeding 20 GHz. The chaotic bandwidth can be further expanded by introducing external optical disturbance. The results provide a reference for the application of this integrated mutually coupled laser in chaotic secure communication, key distribution and other fields.
