3446 results about "Rate of convergence" patented technology

A SoftRouter architecture deconstructs routers by separating the control entities of a router from its forwarding components, enabling dynamic binding between them. In the SoftRouter architecture, control plane functions are aggregated and implemented on a few smart servers which control forwarding elements that are multiple network hops away. A dynamic binding protocol performs network-wide control plane failovers. Network stability is improved by aggregating and remotely hosting routing protocols, such as OSPF and BGP. This results in faster convergence, lower protocol messages processed, and fewer route changes following a failure. The SoftRouter architecture includes a few smart control entities that manage a large number of forwarding elements to provide greater support for network-wide control. In the SoftRouter architecture, routing protocols operate remotely at a control element and control one or more forwarding elements by downloading the forwarding tables, etc. into the forwarding elements. Intra-domain routing and inter-domain routing are also included.

The invention provides a generation method for consensus blocks in a block chain and a computing device. The method comprises the steps that nodes in a peer-to-peer network voluntarily register as candidate nodes; a selected node subset is formed by selecting nodes from in-network candidate nodes, and the nodes of the subset are randomly classified into accounting nodes and audit nodes; the accounting nodes package transactions in a transaction pool as candidate sub-blocks and broadcast the candidate sub-blocks to the audit nodes; the audit nodes verify all received candidate sub-blocks according to a preset audit rule, select an unique candidate sub-block, add audit stamps to the unique candidate sub-block and broadcast the unique candidate sub-block to other audit nodes; and when the number of the audit stamps of the candidate sub-block exceeds a preset threshold value, the candidate sub-block becomes the recorded selected block, so a consensus process is finished. According to the scheme provided by the invention, as improvement of the network performance, code and manual parameter change is avoided, the convergence speed of the consensus blocks is automatically improved, and the difficulty of DDOS (Distributed Denial of Service) and Eclipse attacks is greatly improved.

An equalizer for equalizing channel multi-path distortion includes digital filters. To improve the convergence speed and tracking ability of the equalizer while lowering noise and power consumption, the digital filters are divided into sections. Various parameters of the sections, such as step-size, shutdown and update rates can be controlled. Control of the various parameters can be realized either in software on an embedded or external processor or by dedicated hardware.

Systems, methods, and a computer readable medium for performing auto exposure (AE) techniques that are beneficial in variable lighting conditions—and particularly applicable to handheld and / or mobile videoconferencing applications—are disclosed herein. Handheld and / or mobile videoconferencing applications—unlike their fixed camera counterparts—are often exposed to a wide variety of rapidly changing lighting and scene conditions, and thus face a difficult trade-off between adjusting exposure parameter values too frequently or not frequently enough. In personal electronic devices executing such handheld and / or mobile videoconferencing applications, it may be desirable to: use a small, centered, and center-weighted exposure metering region; set a relatively low brightness target value; and adjust the camera's exposure parameter values according to a distance-dependent convergence speed function. The use of such techniques, in conjunction with a relatively large stability region, may also improve the quality of a video encoder's temporal predictions—and thus video quality—in videoconferencing applications.

The present application relates to tracking a position of a device, e.g. for detecting slow and rapid earth deformation, by making use of a recursive filter having the filter characteristic adapted to a detected type of motion. If the motion of the position tracking device is rapid, the filter characteristic is set such that the rapid motion can be tracked with the necessary speed. On the other hand, if the motion is slow, e.g. during times of a normal tectonic drift, the filter characteristic is set such that the motion is slowly tracked with the advantage of efficient noise reduction, i.e. noise in the input signal is effectively barred and does not pass through the filter to the output signal. Thus, in times of rapid motion the convergence speed of the filter output signal to the input signal is set high for fast convergence and in times of slow motion the convergence speed of the filter output signal to the input signal is set low for a slow convergence. The filter may be a Kalman filter.

This invention relates to one rapid movement estimation method, which comprises the following steps: a, dividing each frame of images into several blocks and determining vector space upper and down message on each current image composed of upper and down messages to determine several prepare motion vector; computing each prepare motion vector matching error values with low motion matching error values with current pixel module designed as motion vector step.

In a burst-mode, high-speed spread-spectrum communications system, faster convergence of a receiver's automatic gain control (AGC) circuit reduces the time required to bring a received signal within the operating range of an operation amplifier and other radio-frequency and digital sections of the receiving system. A gain control circuit includes a coarse-gain feedback loop and a fine-gain feedback loop to improve convergence speed and at the same time maintain the stability of the AGC circuit. The coarse-gain feedback loop quickly brings the received signal, using a large gain signal, to the desired operating range. The fine-gain feedback loop uses a smaller gain signal to gradually smooth the received signal to avoid saturation on the A / D converters.

The invention discloses a method for autonomously localizing robots on the basis of laser radar. The method includes randomly generating N particles to form particle swarms around initial locations ofthe robots, and updating the particle swarms according to robot real-time movement distances and real-time rotation angles measured by sensors of the robots at current operation moments of the robots; computing the superposition quantity of point cloud of the laser radar and obstacles of maps for each particle to use the superposition quantity as a score of the particle, computing weighted position and posture average values of the particle swarms by the aid of the score, which is used as a weight, of each particle and utilizing the weighted position and posture average values as AMCL (adaptive Monte Carlo localization) estimation positions and posture; utilizing the AMCL estimation positions and posture as initial values, acquiring scanned and matched positions and posture by the aid ofscanning and matching algorithms on the basis of Gauss-Newton iterative processes and utilizing the scanned and matched positions and posture as the optimal positions and posture of the robots at thecurrent operation moments; re-sampling the particle swarms by the aid of AMCL algorithms to ultimately obtain the global optimal positions and posture of the robots during operation. The global optimal positions and posture of the robots are used as localization results. The method has the advantage that the localization convergence rate can be greatly increased, and the localization precision andthe localization stability can be greatly enhanced.

The invention discloses a short-term load prediction method based on similar day segmentation and an LM-BP network. According to the method, through quantitative calculation of comprehensive correlative coefficients between meteorological factors and a historical load curve corresponding to a to-be-predicted day, the load curve of the to-be-predicted day is segmented, and according to prediction load curves within different time intervals, corresponding similar days are separately figured out; selection of the similar days is conducted through comprehensive consideration of a multi-feature similarity judgement standard of tendency similarity and shape similarity which are based on historical day meteorological similarity and historical load data. A similar day sample with the highest similarity is selected form historical data of the same kind; different neural network models are constructed through different training samples within different load prediction time intervals, and therefore the prediction precision of the neural network models is further improved. By means of the method, the calculation speed and convergence speed of a prediction algorithm are increased.

The invention relates to an image classification method capable of effectively preventing a convolutional neural network from being overfit. The image classification method comprises the following steps: obtaining an image training set and an image test set; training a convolutional neural network model; and carrying out image classification to the image test set by adopting the trained convolutional neural network model. The step of training the convolutional neural network model comprises the following steps: carrying out pretreatment and sample amplification to image data in the image training set to form a training sample; carrying out forward propagation to the training sample to extract image features; calculating the classification probability of each sample in a Softmax classifier; according to the probability yi, calculating to obtain a training error; successively carrying out forward counterpropagation from the last layer of the convolutional neural network by the training error; and meanwhile, revising a network weight matrix W by SGD (Stochastic Gradient Descent). Compared with the prior art, the invention has the advantages of being high in classification precision, high in rate of convergence and high in calculation efficiency.

The embodiment of the invention provides a model training method and device based on federated learning and a server. The method comprises the steps of sending a to-be-trained model to a plurality ofworking nodes; receiving a local model fed back by the plurality of working nodes, wherein the local model is obtained by training a to-be-trained model by each working node according to own data; determining the precision of each local model according to the test data set; determining a weight coefficient of each local model according to the precision of each local model, wherein the weight coefficient is positively correlated with the precision; and updating the to-be-trained model according to the plurality of local models and the corresponding weight coefficients. According to the method provided by the embodiment of the invention, by increasing the weight coefficient of the high-precision local model and reducing the weight coefficient of the low-precision local model, the convergencerate of the to-be-trained model is increased, and the precision of the to-be-trained model is improved.

The invention provides a continuous voice recognition method based on a deep long and short term memory recurrent neural network. According to the method, a noisy voice signal and an original pure voice signal are used as training samples, two deep long and short term memory recurrent neural network modules with the same structure are established, the difference between each deep long and short term memory layer of one module and the corresponding deep long and short term memory layer of the other module is obtained through cross entropy calculation, a cross entropy parameter is updated through a linear circulation projection layer, and a deep long and short term memory recurrent neural network acoustic model robust to environmental noise is finally obtained. By the adoption of the method, by establishing the deep long and short term memory recurrent neural network acoustic model, the voice recognition rate of the noisy voice signal is improved, the problem that because the scale of deep neutral network parameters is large, most of calculation work needs to be completed on a GPU is avoided, and the method has the advantages that the calculation complexity is low, and the convergence rate is high. The continuous voice recognition method based on the deep long and short term memory recurrent neural network can be widely applied to the multiple machine learning fields, such as speaker recognition, key word recognition and human-machine interaction, involving voice recognition.

The invention discloses an image target recognition method based on optimized convolution architecture. The image target recognition method includes collecting and enhancing an input image to form a sample; training the sample on the basis of the optimized convolution architecture; performing classified recognition on an image target by using the convolution architecture after training, wherein optimization of convolution architecture includes ReLU activation function; locally responding to normalization; overlapping and merging a convolution area; adopting neuron connection Drop-out technology; performing heuristic learning. Compared with the prior art, the image target recognition method has the advantages that tape label samples can be expanded, and the image target recognition method is supportive of classification of many objects and acquiring of high training convergence speed and high image target recognition rate and has higher robustness.

The invention discloses a multisource information fusion method in an evidence high-conflict environment, which comprises the following steps of: determining the mutual support degree of evidences according to evidence distances, further calculating the relative discount factor of each evidence by taking a characteristic vector corresponding to the maximum characteristic value of a evidence support degree matrix pattern as a weighting vector of the evidence, wherein the evidence with the maximum weighing coefficient is a key evidence; and discounting each evidence information, and fusing the information of multisource sensor data by using the D-S (Dempster / Shafer) rule so as to increase the accuracy and precision of a monitoring proposition judgment under an evidence high-conflict status.The method can effectively eliminate the uncertainties of multisource information, increase the accuracy of information evaluation and overcome the limitation that correct evaluation results can not be acquired by using the traditional D-S evidence theory under the evidence high-conflict status, and has the advantages of simple algorithm, high convergence rate and the like.

The present invention discloses a method and a device for identifying a reticulate pattern face image based on a multi-task convolutional neural network. The method comprises the steps of: collecting reticulate pattern face image and corresponding clear face image pairs, then using the multi-task convolutional neural network to respectively design object functions based on regression and classification, training a face image reticulate pattern removing model, and finally inputting the reticulate pattern face image into the trained reticulate pattern removing model to obtain a face image without reticulate pattern, thereby performing subsequent face image identification tasks. According to the method, a multi-task learning frame is adopted, the task for restoring a reticulate pattern image to a clear image is expressed as two object functions which are assistant with each other, and the convolutional neural network is utilized to learn complicated nonlinear transformation referred therein. The method not only effectively improves convergence rate during model training, but also can greatly improve image restoration effect and generalization ability, thereby greatly improving identification accuracy rate of the reticulate pattern face image.

Methods and systems are provided for reliably providing a prognosis of the life-expectancy of a vehicle battery. A state of degradation of the battery is predicted based on a rate of convergence of a metric, that is derived from a sensed vehicle operating parameter, towards a defined threshold, determined based on past history of the metric. The predicted state of degradation is then converted into an estimate of time or distance remaining before the component needs to serviced, and displayed to the vehicle operator. Vehicle control and communication strategies may be defined with respect to the predicted state of degradation.

The invention relates to a smooth path planning method for mobile robots based on a dynamic complex environment, and belongs to the field of robot path planning. The method comprises the steps of starting; searching for an optimal path; conducting smooth treatment on the optimal path; ending. The step of searching for the optimal path uses a two-way RRT algorithm with the idea of gravity of an artificial potential field to perform path searching, that is, the artificial potential field gravity guides a random tree to grow toward a target direction, and the RRT algorithm is prevented from randomly sampling the global; a two-way A* algorithm after a heuristic function is improved is adopted to add angle factors based on distance factors, the units of distances and angles are normalized, thepath planning time is greatly reduced, and the convergence speed of paths is improved; smoothing treatment is conducted on the optimal path, a Freudian smoothing algorithm is adopted to remove redundant nodes, a B-spline method is used three times for smoothly connecting remaining nodes, and the smooth optimal path is finally obtained.

A method for iteratively reconstructing image data acquired by a computed tomography system is provided. The method comprises generating a calculated sinogram from an image estimate and generating an error sinogram based on the calculated sinogram and a measured sinogram. Then, one or more backprojections are performed, each based upon a reconstruction parameter. The reconstruction parameter impacts at least one of convergence speed and computational cost of each iterative step and corresponding reconstruction. A filtering step is performed prior to performing the one or more backprojections. Finally, the initial image is updated by adding corresponding results of the one or more backprojections to the image estimate to obtain the reconstructed image.

The invention discloses a non-cooperative spacecraft attitude estimation method based on virtual sliding mode control, and belongs to the technical field of non-cooperative spacecraft navigation. The non-cooperative spacecraft attitude estimation method comprises the following steps: utilizing a virtual control sliding mode controller based on the Lyapunov principle; using target satellite absolute attitude obtained by a stereoscopic vision system as a control objective; according to motion characteristics of the target satellite, establishing a virtual satellite motion model of the target satellite; using a kinetic model of the virtual satellite as a controlled member to obtain attitude parameters of the virtual satellite; using attitude parameters estimated by the virtual satellite and the target satellite absolute attitude obtained by the stereoscopic vision system as controlled input, and calculating the virtual revolving moment on the motion model of the virtual satellite through the virtual sliding mode controller, so as to realize the estimation of the target satellite attitude parameters by the virtual control sliding mode controller. The non-cooperative spacecraft attitude estimation method disclosed by the invention is low in calculated amount, and can still achieve higher convergence rate and higher precision when the initial error of the state variables is high or the system error emerges, so as to meet the requirements of the high performance navigation system.

The invention relates to a mixed measurement analysis method for a satellite antenna. The method effectively solves the problems that various measurement devices are used for co-measurement to reduce detection difficulty and improve detection efficiency during a measurement process of the satellite antenna. The method comprises the steps: cubic mirror collimating measurement is carried out by electronic theodolites during antenna installation and detection processes, scanning measurement of an antenna shaped surface is carried out by a laser radar, a space point position is measured by a laser tracker, and thus the measurement of the satellite antenna is jointly completed by the various measurement devices; union calibration algorithm of 'six freedom degree measurement station three-dimensional network' is employed, a conversion relationship between measurement station coordinates and measurement coordinates is utilized, various observed value error equations are directly listed, so as to overcome shortcomings of a traditional algorithm and improve adaptability of the algorithm. The method provided by the invention is simple, is easy to operate, enables an initial value to be fast acquired, has low requirements for precise degree of the initial value, has a few iteration times, is quick in convergence speed, theoretically is an optimal solution, and has strong algorithm adaptability, high measuring efficiency, fast speed and high precision.

The invention discloses a wind power short-term prediction method based on a long-term and short-term memory network, comprising a long-term and short-term memory neural network training algorithm, ashort-term wind power prediction error distribution algorithm and a wind turbine generator power short-term prediction model design. A long-term and short-term memory network algorithm (LSTM)-based wind pow prediction model is established based on the depth learning network, and the Gaussian mixture model (GMM) is used to analyze the error distribution characteristics of the short-term wind powerprediction. The invention can obtain different confidence intervals of two units through the GMM model. It is proved that LSTM method has higher precision and faster convergence rate, and GMM method has practical application value for wind power dispatching.

The invention discloses a feedback artificial neural network training method and a feedback artificial neural network calculating system and belongs to the field of calculation of neural networks. According to the artificial neural network training method, the synapse weight is adjusted according to a feedforward signal and a feedback signal at the two ends of each neural synapse; when the signals at the two ends of each neural synapse are an excitation feedforward signal and an excitation feedback signal respectively, the synapse weight is adjusted to the maximum value; when the signals at the two ends of each neural synapse are a tranquillization feedforward signal and an excitation feedback signal respectively, the synapse weight is adjusted to the minimum value. According to the feedback artificial neural network calculating system, each node circuit comprises a calculating module, a feedforward module and a feedback module and the node circuits are connected through the neural synapses simulated by memristors, and a series of pulse signals are adopted to achieve the feedback artificial neural network training method. An artificial neural network provided by the system and the method is high in rate of convergence, and the artificial neural network calculating system is few in control element, low in energy consumption and capable of being applied to data mining, pattern recognition, image recognition and other respects.

The invention discloses a method and system for estimating SOC (State-of-Charge) of a power battery based on dynamic parameters. The method comprises the following steps: carrying out a discharge-standing experiment on the battery, obtaining OCV (Open Circuit Voltage)-SOC characteristic curves of the battery at different temperatures, and fitting out an OCV-SOC relational expression; carrying out a constant current pulse discharge-standing experiment on the battery, recording voltage response during the experiment, and identifying the initial value of the parameter of a battery second-order RC equivalent circuit model by an offline method; carrying out dynamic parameter identification on the second-order RC equivalent circuit model by using a forgetting factor-containing recursive least squares method RRFLS; carrying out online estimation on the SOC of the battery by using an EKF (Extended Kalman Filter) algorithm. The estimation method overcomes the defects of inaccuracy and cumulative error of the initial value of SOC in an ampere-hour integral method, and adapts to the dynamic change of battery characteristics, the battery model is high in precision and convergence speed, and is stable and reliable, and the precision of SOC online estimation is improved. The method and system can be widely used in fields of electric vehicles and energy storage battery management systems.

The invention discloses a multi-band structure self-adaptive filter switching method for AEC (acoustic echo cancellation). Firstly, a far-end voice signal is acquired; a voice endpoint is detected, and a VAD (voice activity detection) flag bit and an improved envelope decision threshold are output; the voice signal is fed into a loudspeaker to serve as a desired signal and also input into a self-adaptive filter; the self-adaptive filter adopts a switchable multi-band structure and a corresponding self-adaptive algorithm, parameters of the filter are adjusted by use of the least mean square criterion according to feedback information, and the optimal solution is obtained. According to the provided switching method, voice characteristics are considered sufficiently under the condition that steady maladjustment is guaranteed, and optimized configuration of the convergence rate and the algorithm complexity is realized while advantages of the algorithm in the convergence rate are utilized. During actual application of echo cancellation, a single algorithm does not easily meet various variable demands. The variable switching algorithm provides more probability for a user and has great significance in application of self-adaptive echo cancellation.

The present invention discloses an SOC (state of charge) estimation method. According to the method, a battery OCV-SOC relationship module, a parameter acquisition module, an offline identification parameter value, a parameter discrete state space model, a battery parameter online identification module, a battery dynamic parameter update module and a battery SOC estimation module. The method includes the following specific steps that: 1, a discharge-standing experiment is performed, an OCV-SOC relationship expression is obtained through fitting, parameter values in an equivalent circuit model are identified; 2, a battery second-order RC system discrete state space model is established, and the battery model parameters are identified online and dynamically updated; and 3, the SOC of a battery is estimated online. With the method of the invention adopted, the defects of inaccuracy and accumulative error of the initial value of the SOC of a battery of in an ampere-hour integration method can be eliminated. The method is applicable to the dynamic change of the characteristics of the battery, can improve the accuracy of SOC online estimation and can be widely applied to the electric vehicles and storage battery management system field. The method has the advantages of high battery model precision, fast convergence, high stability and high reliability.

Methods and apparatus operating in a stream processing network perform load shedding and dynamic resource allocation so as to meet a pre-determined utility criterion. Load shedding is envisioned as an admission control problem encompassing source nodes admitting workflows into the stream processing network. A primal-dual approach is used to decompose the admission control and resource allocation problems. The admission control operates as a push-and-pull process with sources pushing workflows into the stream processing network and sinks pulling processed workflows from the network. A virtual queue is maintained at each node to account for both queue backlogs and credits from sinks. Nodes of the stream processing network maintain shadow prices for each of the workflows and share congestion information with neighbor nodes. At each node, resources are devoted to the workflow with the maximum product of downstream pressure and processing rate, where the downstream pressure is defined as the backlog difference between neighbor nodes. The primal-dual controller iteratively adjusts the admission rates and resource allocation using local congestion feedback. The iterative controlling procedure further uses an interior-point method to improve the speed of convergence towards optimal admission and allocation decisions.

The invention discloses a non-rigid heart image grading and registering method based on an optical flow field model, which belongs to the technical field of image processing. The method comprises the following steps of: obtaining an affine transformation coefficient through the scale invariant characteristic vectors of two images, and obtained a rough registration image through affine transformation; and obtaining bias transformation of the rough registration image by using an optical flow field method, and interpolating to obtain a fine registration image. In the non-rigid heart image grading and registering method, an SIFT (Scale Invariant Feature Transform) characteristic method and an optical flow field method are complementary to each other, the SIFT characteristic is used for making preparations for increasing the converging speed of the optical flow field method, and the registration result is more accurate through the optical flow field method; and the characteristic details of a heart image are better kept, higher anti-noising capability and robustness are achieved, and an accurate registration result is obtained. Due to the adopted difference value method, a linear difference value and a central difference are combined, and final registration is realized by adopting a multi-resolution strategy in the method simultaneously.

Configurable permutators in an LDPC decoder are provided. A partially-parallel architecture combined with the proposed permutators is used to mitigate the increase in implementation complexity for the multi-mode function. To overcome the difficulty in efficient implementation of a high-throughput decoder, the variable nodes are partitioned into several groups, and each group is processed sequentially in order to shorten the critical-path delay and, hence, increase the maximum operating frequency. In addition, shuffled message-passing decoding can be adopted in decoders according to the invention to increase the convergence speed, which reduces the number of iterations required to achieve a given bit-error-rate performance.

Amplifying magnitude metric of received signals during iterative decoding of LDPC code and LDPC coded modulation. By appropriately selecting a metric coefficient value that is used to calculate the initial conditions when decoding LDPC coded signals, a significant reduction in BER may be achieved at certain SNRs. The appropriate selection of the metric coefficient value may be performed depending on the particular SNR at which a communication system is operating. By adjusting this metric coefficient value according to the given LDPC code, modulation, and noise variance, the overall performance of the decoding may be significantly improved. The convergence speed is slowed down so that the decoder will not go to the wrong codeword, and the moving range of the outputs of the decoder is restricted so that the output will not oscillate too much and will eventually move to the correct codeword.