The security software platform based on the support body has increasingly complex application functions and also faces massive data transmission, which puts higher requirements on the performance of the system. At pre...
详细信息
ISBN:
(数字)9798350351033
ISBN:
(纸本)9798350351040
The security software platform based on the support body has increasingly complex application functions and also faces massive data transmission, which puts higher requirements on the performance of the system. At present, most secure computing platforms are based on distributed and redundant communication architectures, and lack appropriate failure management mechanisms to effectively utilize the fault tolerance of redundant communication architectures. In response to the problems faced by existing support managed secure computing platforms, this article conducted a study on the fast management mechanism of redundant communication architecture in support managed secure computing platforms. It discussed the implementation of different acceleration algorithms for system communication computing based on cloud edge collaboration, and experiments showed that the accuracy of system communication LSH (Locality Sensitive Hashing) technology acceleration was higher than other algorithms. The accuracy of cloud computing platforms based on LSH technology was 0.92. At the same time, cloud edge collaboration can improve the resource utilization of system communication.
Drones have become very common in inspections of transmission lines across the country, but the images and videos captured by drones still require manual processing, which is time-consuming and labor-intensive, not in...
Drones have become very common in inspections of transmission lines across the country, but the images and videos captured by drones still require manual processing, which is time-consuming and labor-intensive, not intelligent enough to meet real-time requirements, and the response is relatively passive; currently in the power sector artificial intelligence has been used to process drone footage, but it has not yet reached the level of human experts. In recent years, methods to improve model performance have focused on conceiving new network structures and improving the quantity and quality of sample sets. Both methods are costly. This paper proposes a training framework for the model based on self-distillation. Without changing the model structure or changing the data set, you can learn the knowledge of your own network to improve model performance and use smaller models under the same performance requirements. Based on this framework, we implemented a self-distillation system on ResNet and verified the framework by comparing original training and self-distillation training. Of course, this framework is also applicable to other networks. Experiments show that the self-distillation method can significantly improve the accuracy of recognition.
Indexing is an important technique to optimize graph database performance. However, indexing in existing graph databases focuses mainly on property-value based query, and has not drawn much attention on graph traversa...
Indexing is an important technique to optimize graph database performance. However, indexing in existing graph databases focuses mainly on property-value based query, and has not drawn much attention on graph traversal. We propose a new indexing technique, which maintains sorted indexing of vertex and weight values of the vertex's adjacent edges. With this proposed vertex centric adjacent weighted edge index, graph traversal algorithms based on weight values like Dijkstra's algorithm and Prim's algorithm have significant performance gains.
Precise force measurement is critical to probe biological events and physics processes, spanning from molecular motor’s motion to the Casimir effect, as well as the detection of gravitational waves. Yet, despite exte...
详细信息
Precise force measurement is critical to probe biological events and physics processes, spanning from molecular motor’s motion to the Casimir effect, as well as the detection of gravitational waves. Yet, despite extensive technological developments, the three-dimensional nanoscale measurement of weak forces in aqueous solutions still faces major challenges. Techniques that rely on optically trapped nanoprobes are of significant potential but are beset with limitations, including probe heating induced by high trapping power, undetectable scattering signals and localization errors. Here we report the measurement of the long-distance interaction force in aqueous solutions with a minimum detected force value of 108.2 ± 510.0 attonewton. To achieve this, we develop a super-resolved photonic force microscope based on optically trapped lanthanide-doped nanoparticles coupled with nanoscale three-dimensional tracking-based force sensing. The tracking method leverages neural-network-empowered super-resolution localization, where the position of the force probe is extracted from the optical-astigmatism-modified point spread function. We achieve a force sensitivity down to 1.8 fN Hz–1/2, which approaches the nanoscale thermal limit. We experimentally measure electrophoresis forces acting on single nanoparticles as well as the surface-induced interaction force on a single nanoparticle. This work opens the avenue of nanoscale thermally limited force sensing and offers new opportunities for detecting sub-femtonewton forces over long distances and biomechanical forces at the single-molecule ***-resolved photonic force microscopy employs the fluorescence of lanthanide-doped nanoparticles as a force probe, enabling the measurement of sub-femtonewton forces with a sensitivity of 1.8 fN Hz–1/2, approaching the thermal limit.
With the increasing scale and complexity of new power systems, ensuring the safe and optimized operation of distribution networks becomes particularly important. Due to constraints such as economic factors, the compre...
详细信息
ISBN:
(数字)9798350373318
ISBN:
(纸本)9798350373325
With the increasing scale and complexity of new power systems, ensuring the safe and optimized operation of distribution networks becomes particularly important. Due to constraints such as economic factors, the comprehensive deployment of real-time measurement devices in mediumvoltage distribution networks faces significant challenges, resulting in insufficient observability of the distribution network. To address this issue, this paper proposes a method that combines one-dimensional convolutional neural networks (1D-CNN) with transfer learning for pseudo measurement generation. This method first constructs a 1D-CNN pseudo measurement generation model using a large amount of historical measurement data, and then trains and optimizes the 1D-CNN model through transfer learning to achieve pseudo measurement generation even under topological changes. The test results of actual engineering data demonstrate that the proposed method can generate high-frequency pseudo measurement data for nodes in medium-voltage distribution networks in real time, accurately, and efficiently. Additionally, the proposed method exhibits good adaptability when the distribution network topology changes. In this way, this paper provides a new technical path for the perception of the operational status of medium-voltage distribution networks and provides technical support for the safe and optimized operation of distribution networks.
In the new types of power systems, characterized by the dominance of renewable energies, uncertain resources such as wind and solar power will play a significant role in power generation. However, the inherent unpredi...
In the new types of power systems, characterized by the dominance of renewable energies, uncertain resources such as wind and solar power will play a significant role in power generation. However, the inherent unpredictability and fluctuation of these renewable sources make the operating conditions complex and variable. To address this, intelligent online dispatch applications that integrate optimization and artificial intelligence techniques need to be implemented. In terms of online dispatch, estimating system states and optimizing resources require computing a large number of variables while adhering to operating constraints. Graph computing and graph database technologies have shown promising results in power system online analysis. This paper analyzes the applications of these graph computing techniques to enhance online dispatch and provides a graph computing based-grid online analysis architecture. This architecture offers a range of online situational awareness applications and decision-making support.
At present, artificial intelligence technology in power has achieved certain application results in the field of equipment defect recognition and *** article proposes a joint self-supervised learning method based on c...
详细信息
ISBN:
(数字)9798350316537
ISBN:
(纸本)9798350316544
At present, artificial intelligence technology in power has achieved certain application results in the field of equipment defect recognition and *** article proposes a joint self-supervised learning method based on contrastive learning and sorting learning, and applies it to pre-training of power inspection image quality feature extraction models. By effectively utilizing massive unlabeled image samples, the training of feature extraction models is achieved, effectively improving the model’s perception of image quality deterioration and scene generalization ability; Then, based on the pre-training model of joint learning, further supervised fine-tuning is carried out, and the results of general data pre-training, contrastive learning, and joint learning test sets are compared through experiments to verify the effectiveness of joint learning in intelligent detection of power image quality deterioration.
As artificial intelligence methods play an increasingly important role in the digital transformation of power grids, when applied in power grid scenarios, these methods face problems such as the inability of models to...
As artificial intelligence methods play an increasingly important role in the digital transformation of power grids, when applied in power grid scenarios, these methods face problems such as the inability of models to adaptively deploy data distribution changes, slow response speed, inability to ensure data security on the side, limited computing power on the side, and uneven distribution of computing resources. This method proposes a novel fine-tuning framework based on cloud-edge collaboration to address these issues. This method utilizes reinforcement learning to segment the model, discarding unreliable samples collected on-site to fine-tune the model’s fine-tuning parameters. Finally, validation was conducted on the application scenario of detecting violations in the power grid, and the experimental results show the effectiveness of our method.
Collecting images through drone and helicopter aerial photography is one of the main methods for transmission line inspection. Currently, the popular intelligent transmission defects recognition is mainly based on dee...
Collecting images through drone and helicopter aerial photography is one of the main methods for transmission line inspection. Currently, the popular intelligent transmission defects recognition is mainly based on deep neural network object detection algorithms. Due to the outstanding performance in various public datasets and rapid development of DETR algorithms, there is a promising application prospect in the field of transmission inspection. This article proposes a transmission line inspection model based on DETA, which is optimized by integrated improvement algorithm, which inlcudes SNIP sampling algorithm, multi-scale inference strategy, and prior knowledge based box-selection algorithm. Experiments have shown that the method proposed in this article has the advantages of accurate defect localization, high detection accuracy, and strong scalability.
In the recent decade, the development of FPGA kernels raises its level of abstraction from RTL to HLS C or OpenCL. And so are the simulation and debugging software suites. However, it makes some existing high-quality ...
In the recent decade, the development of FPGA kernels raises its level of abstraction from RTL to HLS C or OpenCL. And so are the simulation and debugging software suites. However, it makes some existing high-quality RTL simulators incompatible to use. Therefore, we propose AXIM, a programmable simulation environment that integrates existing RTL simulators for FPGA kernels with interfaces at a high-level abstraction via AXI buses. AXIM consists three modules, including AXIM hardware driver, AXIM runtime library, and AXIM debugger. Experimental results show that AXIM runs faster than the one integrated with Xilinx Vitis. Besides, another experiment demonstrates that visualization analysis of DRAM's activities generated by AXIM can help developers optimize the design of kernels. As a side product, we also identify an inconsistent behavior in hardware emulation in Xilinx Vitis 2019.2 about data transfer using AXI interfaces.
暂无评论