85 results about "Quasi-Newton method" patented technology

The present invention discloses a parallelized human body behavior identification method. According to the method, skeleton data of Kinect is used as input; a distributed behavior identification algorithm is implemented based on a Spark computing framework; and a complete parallel identifying process is formed. Acquisition of the skeleton data of a human body is based on scene depth acquisition capacity of Kinect and the data is preprocessed to ensure invariability of displacement and scale of characteristics; and a human body structural vector, joint included angle information and skeleton weight bias are respectively selected for static behavior characteristics and a dynamic behavior searching algorithm for a structural similarity is provided. On the identification algorithm, a neural network algorithm is parallelized on Spark; a quasi-newton method L-BFGS is adopted to optimize a network weight updating process; and the training speed is obviously increased. According to an identification platform, a Hadoop distributed file system HDFS is used as a behavior data storage layer; Spark is applied to a universal resource manager YARN; the parallel neural network algorithm is used as an upper application; and the integral system architecture has excellent extendibility.

The invention relates to a transient stability distributed simulation method of an electric power system, belonging to the transient stability analysis technical field of the electric power system; the transient stability distributed simulation method is characterized in that: variable dimension of boundary coordination equation is reduced to one half of that of a partitioning method based on connecting line by using the partitioning method based on a node decomposing method, so as to improve the simulation efficiency; the accurate initial Jacobi inverse matrix is obtained by using a boundary perturbation method, so as to reduce iterations; the calculation process is accelerated by comprehensively adopting the methods that the boundary node voltage value of the next time step is pre-estimated and the Jacobi inverse matrix is coordinated and dynamically updated simultaneously in multiple time steps; by adopting coordination algorithm based on inverse Broyden quasi-Newton method, the solving times of the boundary coordination equation are reduced greatly, so as to improve the simulation efficiency. In the technical proposal, the data interface is simple, the data transmission capacity is low, the system model is simplified, superlinear convergence speed close to the newton method is provided, and hyper real-time simulation can be realized in a wide area communication network, therefore, the method is conveniently applied to the practical electric power system.

The invention provides an oral cavity soft and hard tissue CT sequence and three-dimensional grid model registration method comprising the following steps of: firstly manually dividing a biting impression model to obtain the part without a pedestal in the biting impression model; secondly, according to rough tooth outlines of user editing and calibrating, dividing input CT sequences to obtain the coordinates of the tooth outlines in the CT sequences, and constructing a hard tissue point set with the point coordinates of tooth hard tissue outlines of input layers of CT sequences; thirdly using iterative closest point algorithm for registration of the point set coordinates of the biting impression model and the point set coordinates of the hard tissue outlines to obtain the rough registration position in three-dimensional space of the biting impression model; fourthly, using the rough registration position obtained through the ICP algorithm as the initial result of accurate registration, constructing initial energy function values for the given CT sequences and the position of a three-dimensional grid model of initial registration, and optimizing the energy function through the quasi-Newton method to obtain an accurate registration result. The method enables superposition display of registration of CT sequences and biting impression models and can measure the thickness of soft tissue in CT sequences.

The invention relates to a collaborative analysis reentry guidance method considering multiple no-fly zone constraint, which comprises the steps of 1, describing a reentry guidance problem; 2, providing an analytic solution of flight time based on a rotational globe model; 3, carrying out an analytic reentry guidance process considering multiple no-fly zone and arrival time constraint; 4, providing a multi-aircraft arrival time collaboration scheme. The method herein has the advantages that an exact analytic solution of hypersonic gliding trajectory flight time based on the rotational globe model, and a prediction error is kept less than 3% to adapt to conditions with non-constant vertical lift-drag ratio profile; the problem with collaborative flight allowing multiple hypersonic gliding aircrafts to be guided to arrive at a target at the same time in a multiple no-fly zone environment is solved on the basis of the analytic solution of three-dimensional reentry trajectory and the analytic solution of flight time; a multi-target numeric iteration plan based on online trajectory simulation considering terminal time, speed and height requirements is designed. A directional derivativeis sued to modify the quasi-newton method, and the quantity of trajectory simulations is decreased.

The invention provides a method for determining five stand tandem cold mill depressing distribution. The method comprises steps of designing rolling strategy modes and various grading forms which are suitable to a tandem cold mill in accordance with process requirements of the tandem cold mill and original data of coils, selecting data of grading forms and device parameters of the mill of a device capability form in accordance with selected depressing distribution rolling strategy modes and depressing distribution proportionality coefficients, setting up a depressing distribution calculation model, calculating rolling process parameters by using a rolling process mathematical model, solving a non-linear equation model through a quasi-Newton method, calculating obtained satisfied depressing distribution, storing the depressing distribution data to a database, and preparing data for setting a calculation model. According to process properties of the tandem cold mill, the proper rolling strategy modes are provided, the depressing distribution calculation model is set up, the principle is simple, the implementation is easy and the operation is convenient.

The invention discloses a method using an improved regularization method to restrain difference global positioning system (DGPS) integer ambiguity ill-condition. The method using the improved regularization method to restrain the DGPS integer ambiguity ill-condition comprises the steps of (1), collecting observation data of a global positioning system (GPS) carrier phase, building a DGPS carrier phase double-difference observation equation; (2) obtaining a floating point solution of DGPS integer ambiguity and a corresponding variance-covariance matrix based on a least square method according to the DGPS carrier phase double-difference observation equation; (3) and using a two-step solution to build a regularization matrix in a Tikhonov regularization algorithm, obtaining a corresponding regularization parameter according to a DFP quasi-newton method, processing the variance-covariance matrix by using the obtained Tikhonov regularization algorithm, restraining ill-condition of the DGPS integer ambiguity, and at last obtaining accurate integer ambiguity. According to the method using the improved regularization method to restrain the DGPS integer ambiguity ill-condition, the improved Tikhonov regularization algorithm is used for restraining the problem of the ill-condition in the DGPS integer ambiguity, the method using the improved regularization method to restrain the DGPS integer ambiguity is beneficial for obtaining the accurate integer ambiguity, and high-accuracy positioning and attitude measurement of DGPS can be achieved.

The invention discloses a skip distance correction WSN three-dimensional space target positioning method based on coplanarity. The method comprises the following steps: first, the RSSI distance measurement technology is employed to measure a distance between adjacent nodes; second, beacon node information is received and transmitted through a distance vector exchange protocol, and minimum skip values and cumulative distance of all the nodes and a beacon node are obtained; third, a correction parameter is calculated, and the cumulative distance is corrected according the correction parameter; fourth, the coplanarity value of each positioning unit is calculated, and the best positioning unit is selected based on the coplanarity values; fifth, a four-face measuring method is employed to calculate a coordinate of an unknown node according to the selected best positioning unit; sixth, a quasi-Newton method is employed to optimize the coordinate of the unknown node; seven, the node positioned successfully is raised to a beacon node to participate in secondary positioning. The best positioning unit suitable for positioning is selected based on coplanarity values, the quasi-Newton method is employed to raise the positioning result precision, and positioning precision and the positioning coverage rate of three-dimensional space positioning can be raised.

The invention relates to a multi-regional fully distributed active distribution network economic dispatching method and belongs to the field of electric power system operation and control technology, and the method comprises the steps: 1) establishing an active distribution network economic dispatching model for considering linearizing transmission losses and minimum light discarding cost; 2) realizing localization quasi-Newton matrix iteration initial value construction, localization quasi-Newton direction iterative computation and localization step size searching calculation for regions through neighbor boundary node communication; 3) judging whether global convergence is achieved through the neighbor communication after regional convergence. According to the invention, the method aims at operating characteristics of an active distribution network, point-to-point communication technology are fully utilized, on the basis of fully distributed quasi-Newton method, the method complete active distribution network economic dispatching multi-regional fully distributed superlinear convergence computation without a coordination center, the reliability is improved, privacy of the regions is protected, and plug-and-play is realized.

The invention discloses a non-rigid multimode medical image precise registering method. The non-rigid multimode medical image precise registering method includes the following steps that normalization processing is performed on a reference image and a floating image. A pulse sending cortex model is used for processing so as to obtain two ignition mapping images. A Sobel operator is used for respectively abstracting edge features of the two ignition mapping images to obtain corresponding gradient images. A sum of squareddifferences of the two gradient images is calculated out. Finally, based on a free grid deformation model and the sum of squareddifferences a target function is determined. A quasi-Newton method is used for performing iterative optimization on the target function to obtain optimal deformation parameters. According to the non-rigid multimode medical image precise registering method, the pulse sending cortex model and the sum of squareddifferences are used for performing non-rigid multimode medical image registering and registering precision is greatly improved.

The invention relates to an industrial unit dichloroethane cracking furnace coupled modeling method and application. In the method, during modeling, a dichloroethane cracking furnace is divided into a hearth model and a furnace tube model, a one-dimensional partition method in combination with method is used for the hearth, the smoke temperature distribution condition of the hearth is solved by a heat balance equation, for the furnace tube model, the furnace tube is partitioned according to the amount of straight tube sections of a radiant section, and the reaction and the heat absorption capacity in the tube are calculated section by section. The hearth model and the furnace tube model are mutually iterated through a quasi-newton method according to the calculated smoke temperature and heat transfer rate information, until meeting the convergence precision and the model precision. The model is based on the dynamical model and the energy conservation relation, and is corrected by real-time data so as to further measure the relevant parameters of the cracking furnace in a soft sensing manner. The accurate model is the basis of optimization, provides reliable model basis for optimization of field important economic indicators, such as the outlet temperature of the furnace tube of the cracking furnace, the dichloroethane cracking conversion rate, the selectivity, the unit consumption and so on, and realizes real-time soft sensing of relevant parameters.

The invention discloses a FPGA system and implementation method based on an on-line training neural network of the quasi-newton method. The FPGA system comprises modules of LS, GC, HU, NNE, CSC and PNG. The implementation method includes 1 analyzing c++ codes of the quasi-newton method, dividing the algorithm into three calculation modules, and converting each calculation module into a hardware block by editing Verilog; 2 determining the hardware structure of a neural network evaluation module NNE by editing Verilog according to the topological structure, training method and excitation function of the neural network; 3 generating a module PNG by realizing the random number on the basis of a 32-bit linear shift register; 4 adopting a FPGA on-chip memory as a buffer to link the five hardware modules, storing the middle calculation results, and determining the operation order of the five modules and implementing data delivery between memory and corresponded modules in a manner of a finite state machine; 5 carrying out performance test through hardware design. The speed of neural network training can be increased through the FPGA, and the real-time capability requirements of neural network on-line training can be met.

The invention discloses a method for solving motion control of a redundant mechanical arm. The method comprises the following steps of modeling a motion control problem of the mechanical arm accordingto a given tail end track of the mechanical arm, and converting the motion control problem into a linear time-varying equation; converting a motion control solving problem of the mechanical arm intoa solving problem of the linear time-varying equation, designing an error differential equation with exponential type variable parameters for the linear time-varying equation, and establishing an exponential type variable parameter convergence differential neural network model; dispersing an exponential type variable parameter convergence differential neural network by adopting an Euler forward differential formula and a BFGS quasi-Newton method; solving the linear time-varying equation by using the exponential type discrete variable parameter convergence neural network; and transmitting a result obtained by solving to a given linear time-varying system, and solving the motion control problem of the given mechanical arm. According to the method, the motion control problem of the given mechanical arm can be quickly solved with high precision, and a discretization result can be output after discretization, so that the method is suitable for controlling a discrete input system by a digital circuit.

The present invention discloses a turboshaft engine self-adaptive model building method based on a modified Broyden algorithm to solve an equation set. The method comprises: matching a self-adaptive model with a real engine by forming and solving the equation set. A non-linear equation set is established according to a principle in which a model outputs a tracked real engine output and in combination with a coacting equation between engine parts, and the equation set is solved by using a modified Broyden algorithm, to acquire a part performance degradation parameter and a turboshaft engine part level model guess value. The modified Broyden algorithm is based on a Broyden quasi-newton method and is combined with a computation divergence determination and correction mechanism, so that computation accuracy and real-time performance are improved. The performance degradation parameter acquired by solving the equation set is introduced into a turboshaft engine part level model as an adjustable parameter, so that the model has a self-adaptive capability.

The present invention discloses a harmonic current estimation method based on an improved FastICA algorithm. The harmonic current estimation method is characterized in that a quasi-Newton method is adopted to replace a Newton method as an optimization method of FastICA. The present invention has the beneficial effects that (1) harmonic currents in a power grid can be accurately estimated in the condition of unknown harmonic impedance; (2) the problems that a harmonic measuring device is difficultly obtained by the actual harmonic impedance of a system, and the cost is great are overcome; and (3) compared to the original method, the harmonic current estimation method is capable of improving the accuracy of the algorithm and better solving the problem that the harmonic impedance is difficult to determine.

The invention belongs to the power system short-term load prediction technology field and discloses an electric power system short period load prediction method based on the combined neural network. The method comprises steps that (1), historical power grid load and meteorological data are collected, and a database is built for backup; (2) abnormal data acquired in the step (1) is removed, and normalization processing of the residual data is carried out; (3), a model structure having feedforward and feedback functions is determined; (4), the historical data is utilized to train a prediction model, and model parameters and weights are determined; and (5), the prediction model is utilized for actual load prediction to acquire a prediction load value. The method is advantaged in that the combined neural network is composed of an AM-NN sub model based on the additional momentum method and a QN-NN sub model based on the Quasi-Newton method, the two models are fused through a time-varying comprehensive weight coefficient, meteorological factor data is introduced to the model, the rolling optimization strategy is employed, the model is made to have relatively good generalization and convergence property, and actual scheduling prediction requirements can be more precisely satisfied.

The invention discloses an intranet transaction identification method based on a random forest and a naive Bayes model. The method comprises the following steps: acquiring an internal screen transaction sample data set under different time window periods, screening and constructing a characteristic index set by adopting a random forest model, constructing a Bayesian identification model of the internal screen transaction according to the screened characteristic index set, and performing internal screen transaction identification by adopting the Bayesian identification model to obtain a resultof whether the internal screen transaction exists or not; and after the event, supervising and verifying whether the internal transaction recognition result is correct, and training and updating the Bayesian recognition model according to the recognition result. According to the invention, the stock intranet transaction identification model is established, so that whether the test target is subjected to intranet transaction or not is accurately identified; a quasi-Newton method and a genetic algorithm are combined, so that parameters of the random forest model are quickly optimized to an optimal solution with high precision, and the solution of the optimal solution has small dependence on an initial value; the method is easy to implement and stable in performance, and robustness and accuracy can be further improved along with increase of sample data.

The invention belongs to the technical field of direct positioning in passive positioning, and relates to a moving target direct positioning method based on a quasi-newton method. According to the moving target direct positioning method based on the quasi-newton method, net points of searching target possible positions are simplified by mainly using the quasi-newton method; and a gradient is solved through an objective function, then iteration is carried out on an unknown quantity by using the quasi-newton method, the searching process is simplified, and rapid positioning is realized.

The invention discloses a method for reconstructing a sound velocity image by optimizing the full-wave inversion method, which comprises the following steps: using an ultrasonic data acquisition device to obtain a measured sound pressure signal at the location of a sensor; completely simulating the transmitting and receiving process on a computer as a forward function, the input variable of this function is the matrix of the sound velocity map, and the output variable is the norm of the difference between the sound pressure signal recorded at the corresponding position of the sensor and the measured sound pressure signal; enter the iterative process: give a sound velocity image as the initial image, make it Bring it into the above forward function as an independent variable to obtain the output norm; calculate the first-order gradient map of the forward function at the current position of the sound velocity map; solve the optimization problem by quasi-Newton method, update the original sound velocity map, and obtain a new The speed of sound map; use the new speed of sound map as the initial image to repeat the iterative process until the conditions are met to exit the iterative process and get the current speed of sound image.

The invention discloses a rapid quasi-newton method-based full-waveform inversion method, belongs to the technical field of seismic exploration speed modeling, and aims at improving the convergence rate of a quasi-newton method so as to improve the calculation efficiency of seismic full-wave inversion and reduce the calculated amount. The method comprises the steps of defining a group of weightingcoefficient sequences to construct a parameter model iterative formula; importing a middle iterative parameter model and calculating a quasi-newton iterative direction and an optimum step length on the basis of the parameter model; and updating an original parameter model and the middle iterative parameter model by utilizing weighting coefficients, the iterative direction and the optimum step length. According to the method, calculation is carried out to obtain parameter models closest to optimum solutions, and the parameter models are used for the calculation of the next iteration directionand optimum step length, so that the convergence rates of target functions are improved. In the iterative calculation of each time, more calculated amount is not added in the method, and only a smallamount of multiplication and addition with relatively low complexity are carried out, so that the calculated amount of the full-waveform inversion is effectively reduced.

The invention relates to a multi-user detection method for a non-orthogonal multiple access system on basis of a gradient tracing and multi-step quasi-Newton method technology. Compared with the priorart, the defects that multi-user detection is heavy in computing burden, low in efficiency and poor in precision are overcome. The multi-user detection method comprises the following steps: preprocessing of an uplink base station, wherein the base station receives signals sent by user devices to acquire equivalence channel coefficients for iterative detection initialization; and detection of active user devices, wherein the active user devices which send signals at T continuous time slots in a signal frame and the sent signals of the devices are iteratively detected out. According to the multi-user detection method, the gradient tracing thought is introduced into multi-user detection of the base station, orthogonal projection computing when the user devices send signal estimation values is avoided, and a Hessian matrix that multi-step gradient information approximates a target function is computed by adopting a multi-step quasi-Newton method; and on one hand, the second-order convergence is introduced, the precision of multi-user detection is improved, and on the other hand, the convergence speed is increased, and the computing burden of multi-user detection is reduced.