93 results about "Gradient projection" patented technology

The invention discloses a self-adaptation compression reconstruction method based on energy effectiveness observation in a cognitive sensor network. The self-adaptation compression reconstruction method based on the energy effectiveness observation in the cognitive sensor network comprises the steps that (1) a node carries out local detection and compression measurement on data which are actually sensed through an analog transcriber according to the characteristic that power consumption of the node of a cognitive sensor is limited, (2) a space-time relevance structure of sensing signals is used, sensing data are mapped to a wavelet orthogonal basis cascading dictionary to carry out sparse conversion and to carry out self-adaptation observation through a weighting energy subset function, appropriate observation values are obtained in an energy effectiveness mode, orthogonalization is carried out on selected observation vectors to construct a measurement matrix, (3) the sensing data after compression measurement are fed back to an aggregation node through a report channel, the aggregation node carries out self-adaptation reconstruction on the sensing data by using a gradient projection sparse reconstruction Barzilai-Borwein method based on a convex relaxation method, and effective compromise between reconstruction performance and energy consumption of the node is achieved. The self-adaptation compression reconstruction method based on the energy effectiveness observation in the cognitive sensor network can carry out accurate reconstruction on the sensing signals, ensures energy effectiveness of the sensing node, and has actual application significance.

The invention discloses a distance-speed synchronous pull-off deception jamming recognition algorithm based on gradient projection, and belongs to the field of radar anti-jamming. According to a PD radar, the radar is generally jammed through distance-speed synchronous pull-off jamming, the deception jamming can conduct pull-off jamming on distance information and speed information at the same time, and thus the difficulty for accurately choosing and tracking a target of the PD radar is increased. The method comprises the following steps that (1) signals received by the radar are sampled, and a complete atom library is established according to the time-variant characteristic of a distance-speed synchronous jamming echo; (2) when the AGC voltages and the target number of a receiver are suddenly changed, it is judged that the target is jammed; (3) the Doppler frequency of the target is obtained through a gradient projection method, then the target speed is figured out, and accordingly anti-jamming is performed according to the target speed. The distance-speed synchronous pull-off deception jamming recognition algorithm based on gradient projection is high in jamming recognition capability, easy in engineering realization, high in engineering application value and large in popularization prospect.

The invention provides a redundant mechanical arm moving obstacle avoiding algorithm. According to the algorithm, the relative position relationship of a redundant mechanical arm and a moving obstacle is used to obtain a prediction collision point of the obstacle relative to each rod element of the mechanical arm; then, the shortest prediction distance obtained by using the moving speed of the obstacle is used, and in addition, the avoiding speed of each rod element is obtained according to the relationship between the distance value and the safety threshold; the avoiding speed is converted into an angle speed variable corresponding to each joint of the mechanical arm by a Jacobian transpose matrix, finally, the angle speed variable is introduced into a gradient projection method to obtain the angle speed of each joint, and the avoiding on the moving obstacle is realized. The algorithm provided by the invention has the advantages that the accurate tracking on the tail track of the mechanical arm is ensured, the avoidance on the moving obstacle is realized, and the important significance is realized on improving the safety of an industrial mechanical arm.

The invention relates to a fingerprint image segmentation method based on gradient projection and morphology, which pertains to the technical field of image processing. Firstly, lateral and longitudinal first difference images of an original fingerprint image Gx and Gy and a gradient image Del f, are calculated; then the lateral and longitudinal projection curves of the gradient image Del f are utilized to determine a foreground area a rough fingerprint; the foreground area of the rough fingerprint image is divided and fuzzy areas are removed by utilizing the consistency threshold value at the gradient direction; the gradient magnitude of the rough fingerprint foreground area image with fuzzy areas removed is binarized and a module, the size of which is more than the distance between two adjacent ridge lines is utilized to carry out close operation and open operation over the gradient binarized image so as to obtain a fingerprint segmentation template; finally, the original fingerprint image and the fingerprint segmentation template carry out point multiplication to obtain the final fingerprint foreground area image. The fingerprint image segmentation method of the invention is established on the gradient map of fingerprint, and therefore the calculated amount is relatively small and the fingerprint foreground contour after being operated by the morphology is relatively smooth.

The invention discloses a non-cooperative repeated game-based power scheduling method in a wireless MIMO network, belongs to the technical field of wireless communication, and in particular relates to power scheduling in a wireless distributed MIMO network. In the method, power scheduling of each link is modeled as a non-cooperative game capable of maximizing a utility function. A pricing mechanism taking sending power and link quality into consideration is introduced into the utility function so as to acquire Pareto improvement of network throughput. The power scheduling of all sending windows is taken as a repeated game to establish a punishing mechanism for preventing the deviation, of the link, aiming to acquire excess earnings. Simulation results show that, compared with a gradient projection algorithm, the algorithm realizes distributed scheduling of resources at lower computational complexity without obviously sacrificing total throughput.

The invention relates to a data classification system and method based on KL divergence optimizationmethod for classifying data based on KL divergence optimization. The method comprises the steps thatdata preprocessing is conducted on original images, t, texts and other data, and objects are modeled into multi-dimensional distribution; S; selecting a certain amount of triple from the tagged training data to carry out model training; T; the selected triple serves as training data, a linear mapping A is applied to all the mean vectors, t, the optimal linear mapping is learned through iterativeoptimization, and the learning process is based on the basic assumption of metric learning, t, that is, t, the distance between samples of the same kind becomes smaller, and the distance between samples of different kinds becomes larger; A; an intrinsic gradient descent algorithm is adopted for optimization, and after the gradient of an objective function is projected to the tangent space of the same manifold, Riemannian gradient descent is executed on the manifold of an SPD matrix with given affine invariant Riemannian metric; A; and calculating the KL divergence between the test set and thetraining set, and classifying the samples by adopting a K-nearest neighbor (KNN) classifier. The method can effectively improve the classification precision of the system, and has more stable performance.

The invention discloses a parallel serial embarkation mechanism motion planning method based on sea wave active compensation. According to the parallel serial embarkation mechanism motion planning method, the structure of a parallel serial embarkation system is analyzed and decomposed, thus a six-degree-of-freedom parallel platform and a three-degree-of-freedom serial gangway ladder kinematic module are established, motion quantity distribution is conducted on two subsystems through a jacobian matrix pseudo-inverse technique, present mission situation and constrained situation of the subsystems are combined, motion quantity of the system is optimized by using a gradient projection method, and the situation of singularity postures of the system is avoided. According to the parallel serial embarkation mechanism motion planning method, the kinematics multi-solution problem of degree-of-freedom redundancy of a parallel serial mechanism is effectively solved, meanwhile, factors such as joint spacing and the mechanism singularity postures are comprehensively considered, space motion distribution of the parallel serial mechanism is optimized to great extent, input parameters are providedfor control of the parallel serial embarkation system, and demands of operation and maintenance under actual working situation are met.

The embodiment of the invention relates to a multi-cell cooperative resource allocation method based on a potential game theory. The method is provided by aiming at a wireless communication system, the interference among cooperative cells is reduced and the throughput capacity is maximized. The specific implementation method comprises the steps of taking the maximum transmission power on each same-frequency sub-channel as a constraint condition; taking maximizing the throughput capacities of all the cooperative cells as an objective function; building a utility function which takes the aim of maximizing the throughput capacities of same-frequency resource blocks on all the cooperative cells according to the traditional game theory; mapping the game problem as a complete potential game problem with a pricing factor; building a potential function which takes power distribution as an optimizing strategy; updating user channel distribution information firstly; updating power information by using an improved gradient projection law and a Jacobian iteration timing sequence according to a channel distribution result to reach a Nash equilibrium finally; and optimizing the throughput capability of the whole cooperative cells.

The invention discloses a heart rate monitoring wearable device and a monitoring method based on the photoelectric volumetric method. The invention is based on the photoelectric volumetric method to monitor the heart rate of the elderly in real time. The heart rate measurement module uses a heart rate sensor and embeds the heart rate sensor in a wearable watch middle. The main control chip is C8051F340, which is used to control the work of the sensor as well as signal processing and calculation. In order to make the measurement more accurate, there are two heart rate sensors on each wearable device, which measure the heart rate at the ulnar artery and the radial artery respectively. During the measurement process, due to the intentional or unintentional movement of the elderly, motion artifacts will be generated, which will affect the calculation results. Therefore, the output PPG signal needs to be processed (removing motion artifacts) after being amplified and input to the main control chip after analog-to-digital conversion. Perform heart rate calculations. The present invention uses a compressed sensing algorithm to filter the noise of the PPG signal, and since both the signal and the noise are unknown, the gradient projection algorithm is used to recover the signal, and finally the PPG signal is reconstructed.

An electric automobile and diesel locomotive mixed traffic flow-based jammed traffic network equilibrium method disclosed by the present invention comprises the steps of firstly converting an actual mixed road network into the abstract nodes and road section networks, constructing a traffic distribution model considering a conventional internal combustion automobile and electric automobile mixed travel mode, and solving the model by a gradient projection algorithm to obtain a conventional internal combustion automobile and electric automobile mixed travel network equilibrium mode and the roadsection traffic flow, thereby accurately analyzing the influence of the electric automobiles on the road network and analyzing how to adjust the road section traffic flow.

The invention provides a non-restraint face identification method based on HOG characteristic sparse representation. According to the method, face database pictures are firstly inputted, and HOG characteristics of the inputted pictures are extracted; multiple pictures of each category of persons are selected for training, and other pictures are for tests; HOG characteristic column vectors of each training picture of each category of persons are utilized to construct a characteristic dictionary; a gradient projection sparse reconstruction algorithm is utilized to acquire HOG characteristic sparse representation coefficients of test samples; the sparse coefficients are sequentially kept according to categories, zero setting of the residual coefficients is carried out to acquire an approximate sparse coefficient, and the approximate sparse coefficient is multiplied by a dictionary to acquire a test sample estimate; a mean square error of the test sample and the estimate is calculated, a category of the test sample is determined according to the mean square error minimum principle. The method is advantaged in that influence of the non-restraint environment on face identification performance is effectively reduced, non-restraint face identification robustness is improved, a problem of slow operation speed caused by large dictionary dimensions existing in a traditional sparse representation classification algorithm is solved, and algorithm operation efficiency is effectively improved.

The invention discloses a millimeter wave hybrid analog / digital beam forming method based on improved Riemann manifold optimization. The millimeter wave hybrid analog / digital beam forming method comprises the following steps of projecting conjugate gradients to a tangent space to obtain Riemann gradients; searching a point in the tangent space along the Riemann gradient, determining a step lengthby using a Wolfe-Powell criterion, and backtracking the search point to the Riemann manifold; and when a stop condition is met, obtaining an analog beam forming matrix and a digital beam forming matrix through calculation. According to the method, while the algorithm complexity is reduced, the problem that minimum points are missed in the search process is avoided, so that the calculation time isgreatly shortened, the bit error rate is reduced, and the problems that the full-digital beam forming spectral efficiency cannot be achieved through hybrid beam forming and the algorithm complexity istoo high are solved.

The invention discloses a foundation and satellite-borne radar rainfall data fusion method based on wavelet domain regularization. The method comprises: based on radar rainfall data wavelet domain statistical characteristics, selecting an appropriate prior model of rainfall data; and determining a regularization function of scale coefficient fusion and wavelet coefficient fusion after wavelet decomposition of the rainfall data of the foundation and the satellite-borne radar, then solving the scale coefficient of the rainfall data and the maximum posteriori estimation of the wavelet coefficientin a wavelet domain by using a gradient projection method, and finally performing wavelet inverse transformation to obtain the optimal rainfall estimation. According to the method, uncertainty of rainfall estimation of different sensors and wavelet domain statistical characteristics of rainfall data are considered in the fusion process. The fusion result reduces the uncertainty of a single sensor, and can better maintain and reconstruct detail features such as a heavy rainfall extreme value and small-scale change, thereby being more beneficial to monitoring and forecasting of heavy disaster weather such as flood monitoring and the like.

The invention provides an unmanned aerial vehicle time-frequency resource allocation method based on gradient projection. The method comprises the following steps: determining amount of unmanned aerial vehicle, setting a system authorization channel set, setting a system maximum power upper limit and other environment parameters; constructing a spectrum allocation matrix and an adjacent channel matrix to introduce the original problem, simplifying max-min hybrid planning problem which is originally complex by using problem structural characteristics. An iteration algorithm of the gradient projection is proposed to solve the problem; a ground control station performs power and channel allocation on the unmanned aerial vehicle through a remote channel according to the maximum SINR value of each unmanned aerial vehicle, thereby finishing the unmanned aerial time-frequency resource allocation method design. The method provided by the invention has the advantages that the reasonable allocation of the limited spectrum resource is realized in the premise of improving the reliability of a control signal received by each unmanned aerial vehicle and guaranteeing the maximum consumption powerconstraint of the system through the unmanned aerial vehicle time-frequency resource allocation method based on the gradient projection.

The invention relates to a comprehensive demand response method based on a dynamic energy price strategy. An integrated energy system energy management framework composed of a plurality of integratedenergy suppliers, a plurality of micro energy network operators and an independent integrated energy system operator is constructed; a distributed gradient projection iterative algorithm is provided;real-time dynamic energy price and distributed solution of optimized operation strategies of supply and demand parties are realized; the defects that a traditional centralized optimization algorithm needs to depend on complete information of supply and demand parties and supply and demand side interaction is difficult to consider in an existing energy optimization method are overcome, the privacyinformation security of the supply and demand parties is ensured, supply and demand interaction is promoted, and real-time balance of energy supply and demand is achieved. Comprehensive energy consumption is guided through the dynamic electricity price and the natural gas price, and multi-energy complementation and alternative energy consumption are achieved.

The invention relates to the field of redundant robot motion control research, in particular to a motion optimizing method for a redundant degree-of-freedom robot. The motion optimizing method comprises the steps that a gradient projection method is adopted as the basis, and a joint angle limiting position index and an operability index are adopted for optimizing the gradient projection method. The technical problems that in the prior art, existing gradient projection methods are poor in redundant degree-of-freedom robot inverse kinematics solution optimizing effect, and accumulated errors canbe generated are solved. The problem that the redundant degree-of-freedom robot inverse kinematics solution optimizing effect is poor is solved. The motion optimizing method has higher practicabilityfor the redundant mechanism inverse kinematics solution method, and the robot can obtain the more ideal joint space optimizing effect, the higher position resolving precision and smaller motion tracking errors.

The invention provides a method for calculating spectral emissivity and true temperature. A multi-spectral thermodetector is adopted for measuring the spectral emissivity and the true temperature of a luminous body to be measured. The method comprises the steps that 1, an optical splitter is adopted for dividing light emitted by the luminous body to be measured into multiple ways of light with different wavelengths; 2, multiple channels of the multi-spectral thermodetector receive the light of the corresponding ways, so that the brightness temperature of the light of each way is obtained; 3, the relation expression between the brightness temperature measured by each channel and the true temperature is obtained on the basis of the preset rule, please see the relation expression in the description, Ti represents the brightness temperature of the i channel, T represents the true temperature, Lambada i represents the wavelengths of split beams received by the i channel, C2 represents a second radiation constant, and Epsilon(Lambada i, T) represents the spectral emissivity of the i channel; 4, the iteration cut-off condition is obtained on the basis of the temperature relation expression and the preset conversion rule, and please see the iteration cut-off condition in the description; 5, based on a gradient projection method and the spectral emissivity value ranging from 0 to 1, the spectral emissivity and the true temperature are calculated.

The invention discloses a sampling grid partitioning method and a sampling grid partitioning device for a matrix two-dimensional barcode. The method comprises the following steps: taking the positions of peaks in a horizontal gradient projection of a two-dimensional barcode image in the horizontal direction and in a vertical gradient projection of the two-dimensional barcode image in the vertical direction as the positions of vertical segmentation lines and horizontal segmentation lines used to segment the two-dimensional barcode image; adjusting the vertical segmentation lines according to the column number of the two-dimensional barcode image and the module width in the horizontal direction; adjusting the horizontal segmentation lines according to the row number of the two-dimensional barcode image and the module width in the vertical direction; and partitioning the sampling grid of the two-dimensional barcode image according to the vertical segmentation lines and the horizontal segmentation lines after adjustment. The method provided by the invention makes use of the rich gradient features in the two-dimensional barcode image, and has low requirement on module orderliness and module size consistency. The method does not depend on the characteristics of a locator, and is of very high adaptability and robustness. Higher rate of decoding is achieved in the subsequent decoding process.

The invention relates to a permanent-magnet-spherical-motor position detection method based on 3-D magnetic-field sensors. A motor in which the method is suitable for use is a permanent-magnet spherical-motor, which includes a base, a spherical stator, stator coils and a spherical rotor. The rotor is located in a stator wall, and a rotor output shaft thereof extends out of an opening place on thestator wall. The motor is characterized in that the stator coils are cylindrical coreless structures, and three layers thereof are evenly distributed along an equator inside the spherical stator and on coils which are parallel to the equator, and are radially fixed on the spherical stator wall; and permanent-magnet poles are embedded in a surface of the rotor, the poles are divided into an upper layer and a lower layer along the equator, and N poles and S poles of each layer are alternately distributed. The position detection method includes: firstly, obtaining nonlinear relationships betweenmagnetic induction intensity and rotor position information according to a magnetic-field model and rotation transformation theory; integrating all the magnetic-field sensors to determine a fitness function of position calculation; and utilizing a position calculation algorithm, which combines a particle swarm algorithm and a gradient projection method, to carry out position resolution to obtain position information of the rotor.

A method, apparatus and computer program product are provided for estimating and verifying translation motion and / or a scaling factor between face regions in respective frames during face tracking. A method determines translation motion between face regions in two successive frames based upon integral gradient projections of a face region in a first frame and a corresponding window in a second frame. The method also verifies the translation motion between the first and second frames utilizing integral gradient projections. A method also determines a transfer function relating integral projection curves of a face region in a first frame that has a predefined position and a predetermined size and a co-located window of same size in a second frame, determines a scaling factor based upon the transfer function and then verifies the scaling factor utilizing integral gradient projections.

The invention belongs to the technical field of radar signal processing and relates to an imaging method for a bistatic inverse synthetic aperture radar. The invention provides an imaging method for the bistatic inverse synthetic aperture radar on the basis of a gradient projection algorithm and provides a constructing method of a sparse base of echo signals. By use of the imaging method, the problem of low imaging resolution caused by large bistatic included angles in the actual condition is solved, the sparsity caused by echo-signal insufficiency under the large bistatic included angles is fully utilized, sparse sampling is directly carried out on echoes of the bistatic ISAR, then reconstruction is carried out by the gradient projection algorithm, and high-accuracy estimated values for imaging points can be directly obtained. According to the imaging method, the numerous limitations of the traditional algorithm are avoided, and by utilization of the sampling bandwidth far lower than double signal bandwidth, the pressure of acquisition, transmission, processing and storage of data is relieved, simultaneously, the imaging quality is almost not influenced by the large bistatic included angles, the imaging result is not interfered by sidelobe clutters of the radar, and the super-resolution imaging can be realized when the echo signals are not complete.

The invention discloses a multi-focus image fusion method based on compressed sensing and an energy rule. The multi-focus image fusion method based on compressed sensing and the energy rule mainly solves the problems that the similarity of image blocks is high in the process of multi-focus image fusion under the condition of compressed sensing, and information cannot be easily extracted. The multi-focus image fusion method based on the compressed sensing and the energy rule comprises the steps of observing two multi-focus images through the Hadamard measurement matrix used for block scrambling to obtain observation vectors of the two input images, calculating the fusion weight through energy and the data similarity of the two observation vectors, calculating an observation vector of a fused image according to the fusion weight, and carrying out reconstruction on the observation vector through a gradient projection method on a CDF9 / 7 wavelet basis to obtain the fused image. Compared with a classic multi-focus image fusion method, the multi-focus image fusion method based on compressed sensing and the energy rule has the advantages that information of the two images can be included at the same time, the fusion weight can be adjusted by adjusting items according to the similarity of the image blocks, the problem that the similarity of the multi-focus image blocks is excessively high is solved, the fusion effect is improved, and the multi-focus image fusion method can be used for fusion of the multi-focus images.

The invention discloses a method for optimizing the motion trail of a redundant mechanical arm. The method comprises the steps of building a space obstacle environment model and a mechanical arm model, according to the space obstacle environment model, the mechanical arm model and a preset A * algorithm, determining an initial obstacle avoidance path of the mechanical arm moving from a preset starting point to a preset target point in the obstacle environment; establishing a mechanical arm motion optimization mathematical model on the basis of a preset gradient projection algorithm, determining track optimization vectors according to preset task requirements, and using the mathematical model to optimizing the motion speed of the mechanical arm; and according to the mathematical model, determining a mechanical arm joint motion law between any two adjacent nodes in the initial obstacle avoidance path, and according to the motion law and the trajectory optimization vector, optimizing the initial obstacle avoidance path to obtain an optimized path. According to the method for optimizing the motion trail of the redundant mechanical arm, the technical problem that in the prior art, track optimization of a mechanical arm cannot meet actual requirements is solved.

The invention relates to a frequency-domain regularization phase extraction method for a non-ideal phase-contrast imaging system. The method comprises: a cutter edge image is collected and a corresponding line diffusion function representing system characteristics is obtained; a value of MTF(u-) is obtained; an imaging object is placed and imaging is carried out to obtain an imaging result; Fourier transform is carried out on the imaging result; an imaging model in a matrix manner is constructed; iterative calculation is carried out to obtain a temporary solution of an all-frequency-range absorption item; on the basis of a rapid gradient projection algorithm, an absorption item including phase information is obtained; and a phase information image representing the internal structure of the object is obtained.

The invention discloses an MIMO radar parameter estimation method based on a truncation correction SL0 algorithm, comprising the following steps: first, improving a pathological MIMO radar sensing matrix, and getting a non-pathological sensing matrix from a corrected singular value and corresponding left and right singular matrixes thereof through SVD inverse transformation; then, estimating the parameters of an MIMO radar target using an SL0 algorithm, and replacing the pseudo-inverse of the pathological sensing matrix with the pseudo-inverse of the obtained non-pathological sensing matrix in calculation of the initial value and the gradient projection value; and finally, determining the angle, distance and Doppler information of the MIMO radar target according to the position of a nonzero element in a target scene vector estimated value obtained using the SL0 algorithm. The problem on how to estimate the parameters of an MIMO radar target under a pathological sensing matrix is solved. The complexity of MIMO radar target parameter estimation is reduced. Engineering implementation is facilitated. The method is suitable for a target detection occasion requiring high real-time characteristic in modern warfare.

The invention discloses a linear array scanning image sleeper positioning and counting method based on gradient projection and belongs to the field of rail detection. The method comprises a step of performing linear array scanning imaging on a track pavement along the forward direction of a train or a detection platform by using a linear array imaging system to obtain a line array image includingsleepers, a step of performing gradient filtering perpendicular to a sleeper direction on the line array image to obtain a gradient filtering image, a step of setting a detection area in a sleeper area of the gradient filtering image, carrying out gradient projection in a direction parallel to the direction of the sleepers in the detection area to obtain a gradient projection curve, and a step ofdetecting the positions and numbers of peaks or troughs in the gradient projection curve, taking paired peaks and troughs for sleeper positioning, and calculating the sum of the numbers of peaks or troughs in the line array image to be the number of sleepers in a track. According to the method, after positioning and counting the sleepers, the high-precision mileage counting can be achieved, accurate position information is provided for a track inspection system, the accurate positioning of a track inspection result and quick search in later maintenance are facilitated.

The invention discloses a redundant mechanical arm real-time look-ahead trajectory planning method based on an NURBS curve interpolation algorithm, and relates to the field of redundant mechanical arm application. According to the method, a tail end planning interpolation speed is obtained by giving a tail end target path point of the mechanical arm, constructing an NURBS parameter curve as a tail end expected path and adopting an acceleration and deceleration control algorithm, mechanical arm tail end path errors caused by interpolation are considered in a Cartesian space, the interpolation speed is limited according to bow height errors, a look-ahead algorithm is introduced, and the end planning interpolation speed is obtained. The interpolation speed is scaled by considering the joint speed and acceleration constraint of the mechanical arm, meanwhile, prediction of a deceleration point is achieved, NURBS curve interpolation is achieved through a pre-estimation correction method to obtain a tail end expected track, joint variables are solved through inverse kinematics, and joint limiting is achieved through redundancy characteristics and a gradient projection method. Real-time trajectory planning meeting the kinematics constraint condition of the mechanical arm is achieved, and the tail end trajectory tracking precision is guaranteed on the premise that the physical limit of the mechanical arm is not exceeded.

The invention provides a mechanical arm tail end trajectory tracking algorithm based on null space obstacle avoidance. A minimum pseudo-distance in an analytic form is adopted to replace a traditional Euclidean distance to serve as a proximity index of obstacle avoidance, and a null-space obstacle avoidance task of a mechanical arm is completed in combination with a gradient projection method. A self-adaptive positive definite coefficient matrix K and a speed error saturation function sat(e<.>) are designed to feed back a real-time tail end trajectory operation result to a redundant mechanical arm kinematics inverse solution, and the joint angular speed is self-adaptively adjusted according to the feedback result so as to reduce the tail end trajectory tracking error. According to the mechanical arm tail end trajectory tracking algorithm based on the null space obstacle avoidance, the tail end trajectory tracking problem of the redundant mechanical arm is solved, and meanwhile the null-space obstacle avoidance task is completed; obstacle avoidance action can be taken in advance by adopting pseudo-distance, stable space operation of a mechanical arm joint is ensured, the trajectory tracking precision can be remarkably improved by adopting the tail end trajectory tracking algorithm, and a tail end main task is completed to the maximum extent; and the simulation experiment result of the redundant mechanical arm verifies the effectiveness and superiority of the algorithm.

The invention relates to the technical field of integrated circuit design and in particular provides a division method of the parallel GPDT algorithm on a multi-core SOC. The parallel GPDT algorithm includes two layers of iteration, wherein the inner-layer iteration is responsible for solving a working set and the outer-layer iteration is responsible for updating the working set. As for critical paths of computing speed, the critical path of outer-layer circulation is gradient update and the critical path of inner-layer circulation is vector computing following each projection, wherein both the matrix operations need multi-core parallel processing; the other operations can only be carried out in a serial manner on a main core, including a gradient projection operation realized by using the Dai-Fletcher algorithm and the working set update realized by introducing in the quick sort algorithm. Vectors obtained after the computing is over are support vectors of training data of the GPDT algorithm.

The invention discloses an online training method and device for a constrained intelligent automobile autonomous decision-making system, and the method comprises the steps: taking the vehicle driving performance as an optimization target in each round of optimization, and enabling an updating gradient to be projected to a feasible and safe region through calculating a gradient projection matrix of a constraint function, and detecting the absolutely safe operation of the execution strategy by using the safety shield. According to the method, the safety of the intelligent automobile during online target optimization can be ensured, so that the requirement on a simulator is reduced, the application range is not limited by scenes, and the target of safe online training can be achieved.