Patents
Literature

365results about How to "Calculation speed" patented technology

Approximate-computation-based binary weight convolution neural network hardware accelerator calculating module

ActiveCN106909970ACalculation speedReduce power consumptionPhysical realisationApproximate computingHardware acceleration
The invention discloses an approximate-computation-based binary weight convolution neural network hardware accelerator calculating module. The hardware accelerator calculating module is able to receive the input neural element data and binary convolution kernel data and conducts rapid convolution data multiplying, accumulating and calculating. The calculation module utilizes the complement data representation, and includes mainly an optimized approximation binary multiplier, a compressor tree, an innovative approximation adder, and a temporary register for the sum of the serially adding part. In addition, targeted to the optimized approximation binary multiplier, two error compensation schemes are proposed, which reduces or completely eliminates the errors brought about from the optimized approximation binary multiplier under the condition of only slightly increasing the hardware resource overhead expense. Through the optimized calculating units, the key paths for the binary weight convolution neural network hardware accelerator using the computation module are shortened considerably, and the size loss and power loss are also reduced, making the module suitable for a low power consuming embedded type system in need of using the convolution neural network.
Owner:南京风兴科技有限公司

Fully convolutional neural network-based screening face image identification method and device

The invention discloses a fully convolutional neural network-based screening face image identification method and device. The method includes the following steps that: screening face images and corresponding clear face images which form image pairs are collected, and the image pairs are utilized train a fully convolutional neural network which is used for restoring a clear face image from a screening image, namely performing de-screening; and when identification is carried out, a screening face image is inputted into a trained de-screening model, so that a clear face image can be obtained so as to be used for performing a subsequent face identification task. According to the method and device of the invention, the fully convolutional neural network is adopted as the main body of a learning frame, so that the characteristics of larger visual receptive field and faster computing speed of the fully convolutional neural network can be utilized; and according to the design of the training of a target function, pixel-level reconstruction loss and face feature-level reconstruction loss are combined, a spatial transformation module is used in a matched manner to perform precise perfect alignment on face regions in the network so as to realize the accurate extraction of the features of the face regions. With the method of the invention adopted, the clear face image can be restored from the screening image, face features can be kept relatively stable during a restoration process, and therefore, the identification accuracy of the screening face image can be greatly improved.
Owner:INST OF AUTOMATION CHINESE ACAD OF SCI

Network flow rate abnormality detecting method based on super stochastic theory

InactiveCN101286897ACalculation speedData switching networksTraffic volumeNetwork traffic control
The invention discloses a method for detecting network traffic anomaly based on super-statistics theory, which comprises the steps that: (1) a distribution model is selected according to the actual characteristic of the network traffic and the distribution meets the requirements of test for the fitting of distribution of the network traffic; (2) slow variable sequence of the time sequence of the network traffic, namely, distribution parameter sequence, is calculated according to the distribution model; (3) the network traffic anomaly is detected according to the abnormal fluctuation of the slow variable sequence. By establishing the network traffic model based on the super-statistics (statistics of statistics), the method of the invention can describe the time sequence of the network traffic which shows abruptness, non-stationarity, long-range dependence and heavy-tail and carry out anomaly detection on the network traffic. The slow variable sequence of the time sequence of the network traffic calculated by the method of the invention accurately describes the characteristics of the network traffic; the network traffic can be accurately analyzed by analyzing the slow variable sequence and calculating work is greatly reduced. The experiments indicate that the method for detecting the network traffic anomaly based on the slow variable is obviously superior to the traditional detection method.
Owner:HUAZHONG UNIV OF SCI & TECH

Spatial distribution monitoring system of whole bridge deck moving load based on dynamic weighing and multi-video information fusion

The invention discloses a whole bridge deck moving load spatial distribution monitoring system based on dynamic weighing and multi-video information fusion. A dynamic weighing system is arranged at the starting position of the bridge deck to obtain the vehicle weight information of the moving load, A plurality of cameras are arranged along the bridge to obtain video information of traffic flow covering the whole bridge deck, There is an overlap area in the field of view of two adjacent cameras, information collection and network communication system is arranged on site, and deploy the information processing and analysis master control system in the cloud, Combined with Kalman filter and visual field demarcation line method to track vehicle trajectory and position in visual field, Accordingto the time that the vehicle crosses the online slot of the piezoelectric sensor of the dynamic weighing system, the dynamic weighing and monitoring multi-video information are fused to realize the accurate and real-time identification of the magnitude and position of the moving load in the whole bridge deck. The invention has the advantages of fast calculation speed and high identification accuracy, and is suitable for monitoring the full coverage of various bridge decks or pavement moving vehicle loads in a designated area.
Owner:TONGJI UNIV

Video frequency objects recognition method and system based on supporting vectors machine

The invention discloses a video object identification method and a relevant system based on support vector machines; with structure training samples and according to resolution that is selected from the training samples, utilize a method combined with wavelet outline description symbol, shape factor and invariant torque, to describe outline characteristics of the training samples; gain a support vector machine model according to the outline characteristic training, and meanwhile, determine decision-making function parameters at the optimum category aspects in the support vector machine model; then, extract outline characteristics from the video object to be identified; the support vector machine model after training can follow the video object outline characteristics that is input, in order to perform category for the acquired video objects with a decision-making function operation for the optimum category face. The invention has advantages of high calculation speed, high identification accuracy, reliable arithmetic performance and multi-category identification; and moreover, with increase of objects to be identified, the identification performance can still be kept stable, and the identification speed can still meet real-time monitoring demands.
Owner:HUAWEI TECH CO LTD

Method for obtaining load density index based on cellular historical data

ActiveCN103258246AMethod scienceCalculation speedForecastingPhysicsFeeder line
The invention discloses a method for obtaining a load density index based on cellular historical data. The method for obtaining the load density index based on the cellular historical data is characterized by comprising the following steps: a cellular is generated, the power supply area of a 10KV feeder line serves as class cellulars, and the class cellulars contain actual measurement data; the power supply area is divided according to square meshes of the same size to form the class cellulars, and a load is to be forecasted. An electric power geographic information system (GIS) is set up, historical loads, power supply areas and land information of the class cellulars are integrated in the GIS, the cooperation index of the load density is determined, the maximum value of the load density of the class cellulars serves as a benchmark, and the load densities of other similar cellulars are normalized. The classified load densities over the years are obtained, a relation equation between the cellular load and the classified load densities is set up, and the load density is obtained through the least square method. A spatial load is forecasted, the size of the spatial load in a target year is forecasted according to the obtained index of the classified load densities, and the load value of each class cellular is further worked out.
Owner:NORTHEAST DIANLI UNIVERSITY

Method and device for dynamically analyzing big arch springing CRD stratified excavation footage of cross-section tunnel

The invention discloses a method and a device for dynamically analyzing the big arch springing CRD stratified excavation footage of a cross-section tunnel. The method comprises the following steps: stratified excavation: building a three-dimensional numerical model to achieve the numerical simulation during the excavation process; obtaining the deformation and the safety indexes of surrounding rocks, and forming a data sample using the mechanical parameters and the footage as the input and the deformation and the safety of the surrounding rocks as the output; building a surrounding rock parameter identifying sample set and a footage determining sample set; building a surrounding rock parameter identifying neural network model and a footage determining neural network model; performing the dynamic classification on the surrounding rocks; obtaining the deformation of the surrounding rocks of the tunnel; obtaining the distribution range of the mechanical parameters of the surrounding rocks; obtaining the mechanical parameter of the surrounding rocks; determining the neural network model through the footage to obtain the excavation footage in next-stage construction. According to the method and the device, the blindness of artificially selecting footage due to the uncertainty and the variability of a geologic body is overcome, and the method and the device have great economic and social meanings.
Owner:CHINA RAILWAY CONSTR BRIDGE ENG BUREAU GRP 1ST ENG

Method for predicting crude oil characteristic through near infrared spectrum

The invention relates to a method for predicting the crude oil characteristic through near infrared spectrum. The method comprises: collecting different types of crude oil samples; determining the near infrared spectrums and the physical data; carrying out second order differentiation treatment on the near infrared spectrums, and then carrying out main component analysis; taking previous w scores and four physical data, carrying out standardization treatment, and then carrying out main component analysis; taking previous w+1 scores to establish a crude oil identification database; treating the near infrared spectrum of a crude oil sample to be tested, and carrying out the same identification analysis as the database establishing to obtain the identification variable of the crude oil sample to be tested; determining the crude oil sample mostly approaching the crude oil sample to be tested in the database through the identification variable; carrying out spectrum fitting on the crude oil sample to be tested through the spectrums of the crude oil samples, and determining the fitting property through the comparison of the fitting degree and the threshold value; and for the completely-fitting sample to be tested, predicting the physical data of the crude oil sample to be tested by using the physical data corresponding to the database spectrum participating the fitting.
Owner:CHINA PETROLEUM & CHEM CORP +1

People stream trajectory tracking and area dwell time statistics method and system based on heterogeneous network

InactiveCN104217245ACalculation speedHigh precision of iBeacon technologyBiological neural network modelsNetwork modelHeterogeneous network
The invention discloses a people stream trajectory tracking and area dwell time statistics method based on a heterogeneous network. The method comprises the following steps: deploying wireless access points and iBeaocn base stations, collecting and recording AP (Access Point) RSSI (Received Signal Strength Indicator) values collected by a plurality of sampling points and the coordinate values of the corresponding sampling points by a handheld mobile terminal, taking collected data as training data to determine a BP (Back Propagation) neural network model, integrating training results as an output value which is used as a function of a current mobile terminal coordinate value, obtaining a current position coordinate through an iBeacon and RSSI-distance relational expression, and obtaining a current position according to the distance; connecting the position coordinates to form a people stream trajectory tracking route graph; and meanwhile, recording the dwell time of a people stream at each point. The invention also provides a corresponding system. The invention combines the advantages of high calculation speed of wifi technology and high precision of iBeacon technology, and realizes people stream trajectory tracking and area dwell time statistics in a complex large building by combing with the multiuser support of a smartphone.
Owner:高阳

Land-based cloud chart recognition method based on classification trees of support vector machine

The invention discloses a land-based cloud chart classification method based on classification trees of a support vector machine. The land-based cloud chart classification method comprises the steps as follows: firstly, training samples are selected from land-based cloud charts; secondly, a Gabor filter bank is utilized to perform frequency domain decomposition on the training samples; thirdly, sorting histogram spectrum characteristic vectors and interested operator characteristic vectors of each filter image are extracted, so that training sample sets can be obtained; fourthly, K types of the training samples in the training sample sets are clustered to form ni types according to the specified clustering number, and then centers of the ni types are used as training samples of the ni types, so that new training sample sets can be obtained; fifthly, a classification tree model based on a sorter of the support vector machine is established; and sixthly, the samples in T are classified, and the land-based cloud charts can be classified. The land-based cloud chart classification method considers various characteristic values among different cloud genera based on the land-based cloud charts, combines an SVM (Support Vector Machine) learning algorithm with a classification tree algorithm so as to classify and recognize a plurality of types of the cloud charts automatically, and has the advantages of stronger robustness, higher classification speed and high classification accuracy rate.
Owner:HUAZHONG UNIV OF SCI & TECH

Method for recognizing grid topology based on measurement spanning tree

The invention relates to a method for recognizing grid topology based on a measurement spanning tree, belonging to the technical field of electric power system dispatching automation. The method comprises the steps of: firstly, generating a bus-branch model of a network and a tree structure, and calculating voltages and phase angles of nodes by layers; calculating active and passive powers of two ends of the branches according to impedance parameters of the branches; checking the calculating values and the measurement values, if having larger deviation, adding a bad mark on a corresponding network element, sequencing suspicious degrees of the bus and the branches according to the quantity of the bad marks; speculating switch brakes with telecommands possibly having errors and lines with operation states possibly having errors from the suspicious bus and the suspicious branches; and determining whether the deflection of the switch brakes is reasonable or not through judging whether the quantity of the bad marks in the network after the corresponding switch brake deflects is reduced or not. The method is used for recognizing topological errors without state estimation results and convergence problems, and has the advantages of higher calculating speed and wide application range compared with the algorithm using state estimation.
Owner:江西电力调度中心 +1

Method for monitoring low-frequency oscillating source based on PMU data

ActiveCN104865474AEffectively identify the locationCalculation speedElectrical testingVoltage amplitudeDynamo
The invention provides a method for monitoring a low-frequency oscillating source based on PMU data, and relates to safety operation of an electric system. The method comprises: using a phasor measurement unit PMU to obtain information of a bus outlet side of a generator of an electric power system power plant, the information specifically including active output power, reactive output power, voltage phase angles, voltage amplitude values, and power angles of the power plant outlet bus, and according to the obtained data, calculating oscillation energy of the power plant bus, and using a TLS-ESPRIT algorithm to identify the obtained data, extracting a dominant oscillation mode of a power plant generator side, and obtaining corresponding oscillation characteristic parameters. If an oscillation source is in the power plant, oscillation energy flows out of the power plant, an oscillation energy flow is a positive value, and conversely, the oscillation energy flow is a negative value. The method is advantaged in that the PMU can be directly used to obtain input signals and monitor the low-frequency oscillating source of the electric power system in real time, and the method is convenient and fast, and is fast in computation speed, and is good in practicability.
Owner:STATE GRID CORP OF CHINA +2
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products