With the rapid development and popularization of artificial intelligence technology,convolutional neural network(CNN)is applied in many fields,and begins to replace most traditional algorithms and gradually deploys to...With the rapid development and popularization of artificial intelligence technology,convolutional neural network(CNN)is applied in many fields,and begins to replace most traditional algorithms and gradually deploys to terminal devices.However,the huge data movement and computational complexity of CNN bring huge power consumption and performance challenges to the hardware,which hinders the application of CNN in embedded devices such as smartphones and smart cars.This paper implements a convolutional neural network accelerator based on Winograd convolution algorithm on field-programmable gate array(FPGA).Firstly,a convolution kernel decomposition method for Winograd convolution is proposed.The convolution kernel larger than 3×3 is divided into multiple 3×3 convolution kernels for convolution operation,and the unsynchronized long convolution operation is processed.Then,we design Winograd convolution array and use configurable multiplier to flexibly realize multiplication for data with different accuracy.Experimental results on VGG16 and AlexNet network show that our accelerator has the most energy efficient and 101 times that of the CPU,5.8 times that of the GPU.At the same time,it has higher energy efficiency than other convolutional neural network accelerators.展开更多
The blob properties in I-mode and ELM-free H-mode plasmas compared to L-mode have been investigated on the EAST tokamak,including the blob detection rate Nb,sizeδb,lifetimeτb and radial velocity vr,b.The blob proper...The blob properties in I-mode and ELM-free H-mode plasmas compared to L-mode have been investigated on the EAST tokamak,including the blob detection rate Nb,sizeδb,lifetimeτb and radial velocity vr,b.The blob properties in L-mode and I-mode are similar,and those in ELM-free H-mode are different to them.The blob Nbis smaller whileτbis larger in H-mode.The experimental blob sizeδband velocity scalings vr,b-δb show a good agreement with the theoretical models.The variation in blob properties during the L-I and H-L transitions,and their relations to the scrape-off layer(SOL)density,edge and SOL turbulence,and SOL collisionality are discussed.The suppression of the edge(inside the last closed flux surface)turbulence is not reflected in the blob behavior,while the blob detection rate shows a correlation with the SOL density and its low-frequency(3–50 kH z)fluctuations.In addition,the blob detection rate is found to increase with the divertor collisionalityΛdiv,indicating a dependence of blob behavior onΛdiv.The differences in blob detection rates among the three operating regimes might be due to their different SOL densities and collisionalities.The investigation contributes to understanding the influences of edge and SOL plasma parameters on the blob behavior.展开更多
With the extensive integration of the Internet,social networks and the internet of things,the social internet of things has increasingly become a significant research issue.In the social internet of things application...With the extensive integration of the Internet,social networks and the internet of things,the social internet of things has increasingly become a significant research issue.In the social internet of things application scenario,one of the greatest challenges is how to accurately recommend or match smart objects for users with massive resources.Although a variety of recommendation algorithms have been employed in this field,they ignore the massive text resources in the social internet of things,which can effectively improve the effect of recommendation.In this paper,a smart object recommendation approach named object recommendation based on topic learning and joint features is proposed.The proposed approach extracts and calculates topics and service relevant features of texts related to smart objects and introduces the“thing-thing”relationship information in the internet of things to improve the effect of recommendation.Experiments show that the proposed approach enables higher accuracy compared to the existing recommendation methods.展开更多
To interpret the common symmetric peaks caused by the large-scale structure in the complex S(f)spectrum from the heterodyne Doppler reflectometry(DR)measurement in EAST,a 2D circular-shaped O-mode full-wave model base...To interpret the common symmetric peaks caused by the large-scale structure in the complex S(f)spectrum from the heterodyne Doppler reflectometry(DR)measurement in EAST,a 2D circular-shaped O-mode full-wave model based on the finite-difference time-domain method is built.The scattering characteristics and the influences on the DR signal from various scales are investigated.When the structure is located around the cutoff layer,a moving radial or poloidal large-scale structure k_(θ)k_(θ),match(k_(θ),match is the theoretic wavenumber of Bragg scattering)could both generate an oscillation phase term called‘phase modulation’,and symmetrical peaks in the complex S(f)spectrum.It was found that the image-rejection ratio A_(−1)/A_(+1)(A_(±1)represents the amplitudes of±1 order modulation peaks)could be a feasible indicator for experiment comparison.In the case when the structure is near the cutoff layer with the same arrangement as the experiment for the edge DR channel,the curve of A_(−1)/A_(+1)versus kθcan be divided into three regions,weak asymmetrical range with k_(θ)/k_(0)0.15(k_(0)is the vacuum wavenumber),harmonics range with 0.15k_(θ)/k_(0)0.4,and Bragg scattering range of 0.4k_(θ)/k_(0)0.7.In the case when the structure is located away from the cutoff layer,the final complex S(f)spectrum is the simple superimposing of modulation and Bragg scattering,and the modulation peaks have an amplitude response nearly proportional to the local density fluctuation,called the‘propagationroute effect’.Under the H-mode experiment arrangement for the core DR,a critical fluctuation amplitude Amp(n_(e,Mod.@route))/Amp(n_(e,Tur.@MSA)∼1.3–4.1(Amp(n_(e,Mod.@route))refers to the pedestal large-scale structure amplitude and Amp(ne,Tur.@MSA)refers to turbulence amplitude at the main scattering area)is needed for the structure in the pedestal to be observed by the core DR measurement.The simulations are well consistent with the experimental results.These effects need to be carefully considered during the DR signal analyses as the injecting beam passes through the plasma region with large-scale structures.展开更多
The Doppler reflectometer(DR),a powerful diagnostic for the plasma perpendicular velocity(u⊥)and turbulence measurement,has been widely used in various fusion devices.Many efforts have been put into extracting the Do...The Doppler reflectometer(DR),a powerful diagnostic for the plasma perpendicular velocity(u⊥)and turbulence measurement,has been widely used in various fusion devices.Many efforts have been put into extracting the Doppler shift from the DR signal.There are several commonly used methods for Doppler shift extraction,such as the phase derivative,the center of gravity,and symmetric fitting(SFIT).However,the strong zero-order reflection component around 0 kHz may interfere with the calculation of the Doppler shift.To avoid the influence of the zerofrequency peak,the asymmetric fitting(AFIT)method was designed to calculate the Doppler shift.Nevertheless,the AFIT method may lead to an unacceptable error when the Doppler shift is relatively small compared to the half width at half maximum(HWHM).Therefore,an improved method,which can remove the zero-frequency peak and fit the remaining Doppler peak with a Gaussian function,is devised to extract the Doppler shift.This method can still work reliably whether the HWHM is larger than the Doppler shift or not.展开更多
基金supported by the Project of the State Grid Corporation of China in 2022(No.5700-201941501A-0-0-00)the National Natural Science Foundation of China(No.U21B2031).
文摘With the rapid development and popularization of artificial intelligence technology,convolutional neural network(CNN)is applied in many fields,and begins to replace most traditional algorithms and gradually deploys to terminal devices.However,the huge data movement and computational complexity of CNN bring huge power consumption and performance challenges to the hardware,which hinders the application of CNN in embedded devices such as smartphones and smart cars.This paper implements a convolutional neural network accelerator based on Winograd convolution algorithm on field-programmable gate array(FPGA).Firstly,a convolution kernel decomposition method for Winograd convolution is proposed.The convolution kernel larger than 3×3 is divided into multiple 3×3 convolution kernels for convolution operation,and the unsynchronized long convolution operation is processed.Then,we design Winograd convolution array and use configurable multiplier to flexibly realize multiplication for data with different accuracy.Experimental results on VGG16 and AlexNet network show that our accelerator has the most energy efficient and 101 times that of the CPU,5.8 times that of the GPU.At the same time,it has higher energy efficiency than other convolutional neural network accelerators.
基金supported by the National Key R&D Program of China(Nos.2022YFE03020004,2017YFE0301300,2018YFE0303104 and 2019YFE03030000)the Major Science and Technology Infrastructure Maintenance and Reconstruction Projects of the Chinese Academy of Sciences+3 种基金National Natural Science Foundation of China(Nos.12275313,11922513,12005004 and U19A20113)the Institute of Energy,Hefei Comprehensive National Science Center(No.GXXT-2020-004)the Anhui Provincial Natural Science Foundation(No.2008085QA38)the Users with Excellence Program of Hefei Science Center,Chinese Academy of Sciences(No.2020HSC-UE009)。
文摘The blob properties in I-mode and ELM-free H-mode plasmas compared to L-mode have been investigated on the EAST tokamak,including the blob detection rate Nb,sizeδb,lifetimeτb and radial velocity vr,b.The blob properties in L-mode and I-mode are similar,and those in ELM-free H-mode are different to them.The blob Nbis smaller whileτbis larger in H-mode.The experimental blob sizeδband velocity scalings vr,b-δb show a good agreement with the theoretical models.The variation in blob properties during the L-I and H-L transitions,and their relations to the scrape-off layer(SOL)density,edge and SOL turbulence,and SOL collisionality are discussed.The suppression of the edge(inside the last closed flux surface)turbulence is not reflected in the blob behavior,while the blob detection rate shows a correlation with the SOL density and its low-frequency(3–50 kH z)fluctuations.In addition,the blob detection rate is found to increase with the divertor collisionalityΛdiv,indicating a dependence of blob behavior onΛdiv.The differences in blob detection rates among the three operating regimes might be due to their different SOL densities and collisionalities.The investigation contributes to understanding the influences of edge and SOL plasma parameters on the blob behavior.
基金supported by National Key Research and Development Program of China (2019YFB2102500)China Postdoctoral Science Foundation (2021M700533)+1 种基金Natural Science Basic Research Program of Shaanxi Province of China (2021JQ-289,2020JQ-855)Social Science Fund of Shaanxi Province of China (2019S044).
文摘With the extensive integration of the Internet,social networks and the internet of things,the social internet of things has increasingly become a significant research issue.In the social internet of things application scenario,one of the greatest challenges is how to accurately recommend or match smart objects for users with massive resources.Although a variety of recommendation algorithms have been employed in this field,they ignore the massive text resources in the social internet of things,which can effectively improve the effect of recommendation.In this paper,a smart object recommendation approach named object recommendation based on topic learning and joint features is proposed.The proposed approach extracts and calculates topics and service relevant features of texts related to smart objects and introduces the“thing-thing”relationship information in the internet of things to improve the effect of recommendation.Experiments show that the proposed approach enables higher accuracy compared to the existing recommendation methods.
基金supported by Shenzhen Clean Energy Research Institute,the National MCF Energy R&D Program of China(Nos.2018YFE0311200 and 2022YFE03070004)National Natural Science Foundation of China(Nos.U1967206 and 12075155)+1 种基金the China Postdoctoral Science Foundation(No.2021M702270)the Fundamental Research Funds for the Central Universities(No.WK3420000018)。
文摘To interpret the common symmetric peaks caused by the large-scale structure in the complex S(f)spectrum from the heterodyne Doppler reflectometry(DR)measurement in EAST,a 2D circular-shaped O-mode full-wave model based on the finite-difference time-domain method is built.The scattering characteristics and the influences on the DR signal from various scales are investigated.When the structure is located around the cutoff layer,a moving radial or poloidal large-scale structure k_(θ)k_(θ),match(k_(θ),match is the theoretic wavenumber of Bragg scattering)could both generate an oscillation phase term called‘phase modulation’,and symmetrical peaks in the complex S(f)spectrum.It was found that the image-rejection ratio A_(−1)/A_(+1)(A_(±1)represents the amplitudes of±1 order modulation peaks)could be a feasible indicator for experiment comparison.In the case when the structure is near the cutoff layer with the same arrangement as the experiment for the edge DR channel,the curve of A_(−1)/A_(+1)versus kθcan be divided into three regions,weak asymmetrical range with k_(θ)/k_(0)0.15(k_(0)is the vacuum wavenumber),harmonics range with 0.15k_(θ)/k_(0)0.4,and Bragg scattering range of 0.4k_(θ)/k_(0)0.7.In the case when the structure is located away from the cutoff layer,the final complex S(f)spectrum is the simple superimposing of modulation and Bragg scattering,and the modulation peaks have an amplitude response nearly proportional to the local density fluctuation,called the‘propagationroute effect’.Under the H-mode experiment arrangement for the core DR,a critical fluctuation amplitude Amp(n_(e,Mod.@route))/Amp(n_(e,Tur.@MSA)∼1.3–4.1(Amp(n_(e,Mod.@route))refers to the pedestal large-scale structure amplitude and Amp(ne,Tur.@MSA)refers to turbulence amplitude at the main scattering area)is needed for the structure in the pedestal to be observed by the core DR measurement.The simulations are well consistent with the experimental results.These effects need to be carefully considered during the DR signal analyses as the injecting beam passes through the plasma region with large-scale structures.
基金supported in part by the National MCF Energy R&D Program(Nos.2018YFE0311200 and 2017YFE0301204)National Natural Science Foundation of China(Nos.U1967206,11975231 and 11922513)supported by the Users with Excellence Program of Hefei Science Center CAS(No.2020HSC-UE009).
文摘The Doppler reflectometer(DR),a powerful diagnostic for the plasma perpendicular velocity(u⊥)and turbulence measurement,has been widely used in various fusion devices.Many efforts have been put into extracting the Doppler shift from the DR signal.There are several commonly used methods for Doppler shift extraction,such as the phase derivative,the center of gravity,and symmetric fitting(SFIT).However,the strong zero-order reflection component around 0 kHz may interfere with the calculation of the Doppler shift.To avoid the influence of the zerofrequency peak,the asymmetric fitting(AFIT)method was designed to calculate the Doppler shift.Nevertheless,the AFIT method may lead to an unacceptable error when the Doppler shift is relatively small compared to the half width at half maximum(HWHM).Therefore,an improved method,which can remove the zero-frequency peak and fit the remaining Doppler peak with a Gaussian function,is devised to extract the Doppler shift.This method can still work reliably whether the HWHM is larger than the Doppler shift or not.
