45 results about "Random perturbation" patented technology

A training method and a training apparatus for a neutral network for image recognition are provided. The method includes: representing a sample image as a point set in a high-dimensional space, a size of the high-dimensional space being a size of space domain of the sample image multiplied by a size of intensity domain of the sample image; generating a first random perturbation matrix having a same size as the high-dimensional space; smoothing the first random perturbation matrix; perturbing the point set in the high-dimensional space using the smoothed first random perturbation matrix to obtain a perturbed point set; and training the neutral network using the perturbed point set as a new sample. With the training method and the training apparatus, classification performance of a conventional convolutional neural network is improved, thereby generating more training samples, reducing influence of overfitting, and enhancing generalization performance of the convolutional neural network.

Methods and systems for evaluating the capability of a measurement system to track measurement parameters through a given process window are presented herein. Performance evaluations include random perturbations, systematic perturbations, or both to effectively characterize the impact of model errors, metrology system imperfections, and calibration errors, among others. In some examples, metrology target parameters are predetermined as part of a Design of Experiments (DOE). Estimated values of the metrology target parameters are compared to the known DOE parameter values to determine the tracking capability of the particular measurement. In some examples, the measurement model is parameterized by principal components to reduce the number of degrees of freedom of the measurement model. In addition, exemplary methods and systems for optimizing the measurement capability of semiconductor metrology systems for metrology applications subject to process variations are presented.

The invention relates to a seismic source inversion method based on micro seismic data and an SPSA optimization algorithm. The method comprises the following steps: building a horizontal layered velocity model according to the analysis of logging data and underground rock property data; fast modeling a micro seismic event, adjusting the angle of ray in a dichotomy approach to optimize the ray tracking path, calculating the first-arrival travel time, and then calculating the travel time difference between two adjacent detectors; adding the travel time difference to random perturbation as an observed value; using an SPSA algorithm to calculate the first-arrival travel time difference in the same way based on the location of a micro seismic source obtained through iterative updating, constantly optimizing the location of the seismic source by fitting calculated data and the observed value, and checking the difference between the calculated data and the observed value; if the precision is satisfied or the maximum number of iterations is achieved, stopping updating, outputting a result, and completing positioning; or, continuing next iteration. By calculating out a micro seismic event source rapidly and efficiently, the accuracy of seismic source locating is improved.

The invention discloses a moving target tracking method based on a pyramid appearance model. The method includes the steps of giving out the state and the observation of a target to be tracked in an initial frame, conducting pyramid appearance modeling on the tracked target through the pyramid segmentation and increment principle component analysis method, integrating texture information of the tracked target into the pyramid appearance model, integrating shape information of the tracked target into the pyramid appearance model, adding random perturbation to the tracked target in the state of the current frame so that the tracked target can serve as a particle of a next frame, conducting pyramid appearance modeling and comparison on the observation of the obtained particle and the observation of the target under the current frame, selecting the observation corresponding to the most similar particle as the state of the tracked target under the next frame, calculating reconstruction errors of all blocks from bottom to top layer by layer for the pyramid appearance model of the selected observation, comparing the reconstruction errors with a preset threshold value, and conducting appearance updating on the blocks with the construction errors smaller than the threshold value.

A system and method for material decomposition optimization in the image domain include a non-transitory computer readable medium has stored thereon a sequence of instructions which, when executed by a computer, causes the computer to access a reconstructed basis material image. For a first voxel of the reconstructed basis material image, the instructions also cause the computer to optimize a concentration of a pair of materials (a,b) in the first voxel exclusively in the image domain and based on a first probability based on random perturbations and a second probability based on random perturbations. The optimization is further based on a third probability based on known materials and a fourth probability based on concentrations of the pair of materials in a pair of voxels neighboring the first voxel.

The invention discloses a forward gain assessment system applicable to a full digital closed loop optical fibre gyro. In FPGA chip, the forward gain assessment system is realized by Verilog HDL, the FPGA chip is the central processor of the full digital closed loop optical fibre gyro. On-line assessment is carried out on the forward gain by adopting a pseudo-random perturbation identification method; a pseudo-random signal (111) and a square wave modulation signal (151) are overlaid on a fedback step wave signal (651) to be used as an identification input signal (31) of an integrated-optic modulator; a digital optical intensity voltage signal (621) demodulates the square wave during pseudorandom demodulation, so as to extract the value of the forward gain in real time. The forward gain assessment system can carry out on-line assessment under the normal operation of the full digital closed loop optical fibre gyro and monitor the effect brought about by the signal-to-noise ratio of the gyro and a dead area nonlinear effect in real time.

The invention relates to the field of satellite communication technologies and provides a binary diversity single-carrier transmission method based on a self-adaptive cyclic prefix. The method comprises the following steps that (1) a space segment provides information service forwarded based on processing through two co-orbital satellites, and a satellite receiver at an earth segment can synchronously receive signals of the two satellites to obtain space diversity gain; (2) inter-frame signal interference of a receiving end, which is brought by random perturbation of the satellites, is effectively resisted by utilizing a single-carrier transmission mechanism based on the cyclic prefix; (3) the satellite receiver carries out balancing and subsequent signal processing on the received signals, a specific statistical parameter is taken as an index, and the current length performance of the cyclic prefix is broadcasted to a satellite transmitting end in a manner of 1 bit; and (4) the satellite transmitting end adjusts the length of the cyclic prefix at the appointed step length according to feedback information. According to the method, when the satellite altitude is inaccurately known, lower complexity, low-speed feedback expenditure and higher transmission efficiency are achieved, and the signal diversity gain brought by binary transmission is obtained.

For low complexity error correction, a decoder modifies each reliability metric of an input data stream with a random perturbation value. The reliability metric comprises a weighted sum of a channel measurement for the input data stream and parity check results for the input data stream. In addition, the decoder may generate an output data stream as a function of the reliability metrics.

The invention discloses an improved random perturbation method on the basis of a repetition frequency structure vibration characteristic value of an agent model. The method comprises the following steps: firstly, carrying out perturbation analysis on the characteristic value of a repetition frequency structure, and obtaining a matrix equation in regard to the first-order perturbation amount of the free vibration characteristic value of an original repetition frequency structure after parameters including the rigidity, the quality and the like of the repetition frequency structure are subjected to disturbance and change; and then, on the basis of a polynomial chaos expansion method, constructing an agent model in regard to the first-order perturbation amount of a repetition frequency characteristic value; and combining a perturbation method with an agent model technology on the basis of the polynomial chaos expansion, putting forward an improved approximate calculation method of the free vibration characteristic value of the repetition frequency structure, and further obtaining the expressions of a mean value and a variance under the parameter disturbance situation of the repetition frequency structure characteristic value on the basis of the approximate calculation method. The improved random perturbation method solves the problem in the traditional perturbation method that the statistical characteristics of the traditional perturbation method can not be further researched since the first-order perturbation amount of the repetition frequency structure characteristic value can not be expressed by an explicit expression.

Sensitivity to incentive changes of survey participant may be computed and analyzed. Participants and one or more attributes associated with the participants may be identified. The participants may be clustered into one or more clusters according to the one or more attributes. An incentive amount to be given to a participant in a cluster of said one or more clusters may be perturbed by performing random perturbation computation. The incentive may be distributed to the participant. One or more responses of the participant may be monitored. Individual incentive sensitivity representing incentive sensitivity of responsiveness of the participant per incentive change may be computed based on the monitoring. Incentive amount computation may be dynamically adjusted responsive to determining that the individual incentive sensitivity changed by a predefined criterion.

The invention relates to a tire key parameter and vehicle running state identification method and belongs to the field of information science. The method comprises the steps of establishing a multi-degree of freedom vehicle kinetic equation, substituting lateral force determined in a tire magic formula into the equation, utilizing a Runge-Kutta method to convert the vehicle kinetic equation into a Markov chain form, achieving vehicle running state and tire magic formula parameter joint estimation through particle filter, introducing an auxiliary variable to enable particle weight change to be more stable through secondary weighting operation, adopting a polynomial sampling method to perform particle re-sampling, introducing a random perturbation method in the re-sampling process to increase particle diversity while improving particle degeneracy, and achieving simultaneous estimation of tire magic formula parameters and a vehicle running state. A good estimation result can be obtained, road adhesion coefficients can be further calculated according to estimated parameters, road surface characteristics can be determined, and a basis is provided for future intelligent vehicle running control.

The present invention discloses a data cleaning method based on functional dependency. The method is characterized by comprising: performing data conversion on raw data so as to totally convert different types of attributes into numerical attributes; extracting self-dependency function characteristics of the attributes of the raw data after data conversion; and extracting an inter-dependency function among the attributes of the raw data after data conversion; and determining the attributes which need to be cleaned and are to be cleaned as well as an sample according to the self-dependency function characteristics and the inter-dependency function, forming a relevant cleaning policy basis according to the attributes and the sample, determining whether an to-be-cleaned attribute object is cleaned by adopting the self-dependency function or the inter-dependency function, and if the self-dependency function is adopted to clean the to-be-cleaned attribute object, performing calibration and recovery on the sample which does not accord with the condition according to a polynomial determined by the self-dependency function, and using white noise as random perturbation. According to the data cleaning method provided by the present invention, the dirty data problem in mass data can be solved and high quality data is provided for subsequent analysis and mining of the mass data.

The invention discloses a double-crane system variable amplitude angle response modeling algorithm and a random response domain prediction method. The variable amplitude angle response modeling algorithm includes the steps: 1 building a system geometric model and giving coordinates of points; 2 building a system constraint equation; 3 building a variable amplitude angle response model according to the system constraint equation. The random response domain prediction method includes the steps: 1 describing uncertainty of a load according to a random parameter model; 2 building a random variable amplitude angle response equivalent equation based on the variable amplitude angle response model and the random parameter model; 3 proposing a perturbation random composite function method according to composite function characteristics and a random perturbation method, and solving random variable amplitude angle response expression; 4 further solving variable amplitude angle response expectation and variance. The prediction problem of a lower variable amplitude angle response domain with random parameters can be solved, the algorithm and the method have the advantages of high speed and precision, and system operation reliability is ensured.

The invention provides a transverse microstructure generation method of a unidirectional long fiber reinforced composite material. In the target area where the RVE model needs to be generated, the initial parameters of the RVE model are determined, the regularly distributed fiber position is taken as the initial fiber position, combined with the intersection algorithm between adjacent rows or columns, under the condition that the periodicity of the fiber at the boundary is ensured, a random perturbation method is utilized to generate the RVE with periodic repetitive fiber random distribution.Based on the obtained random fiber position coordinates, the initial position of the micro-pores is determined, and the size and shape of the pores are determined randomly. Finally, the transverse microscopic model of the composites considering the random distribution of the fiber and the micro-pores is established by random perturbation of the position of the pores. The invention considers the reconstruction technology of the transverse microstructure of the unidirectional long fiber composite material, adopts the random perturbation method for the random distribution of the fibers and the pores, and can effectively and efficiently establish the transverse RVE model considering the random distribution of the fibers and the pores.

The present invention discloses a digital audio camouflage and reconstruction method based on segmented sequences. The method comprises the steps of taking sub segment sequences obtained by dividing a published audio and a secret audio as a secret sequence and a published sequence, adding random perturbations to the sequences with the same element value, directly carrying out least squares matching on an equidistant transformation sequence obtained by the rotate right of the secret sequence and a corresponding position published sequence, finding a rotate right step length with a minimum residual error and a matched parameter, and thus camouflaging the secret audio as the published audio and further reconstructing the camouflaged audio as the secret audio. The method is easy to realize, since only the corresponding position equidistant transformation sequence matching is carried out, while the encoding time is reduced, the equidistant transformation number is raised, the matching precision is improved, the detail loss and overflow problems brought by the deformation from a smooth block to a complex block and the restoration from the smooth block to a complex texture block can be effectively avoided, and thus the auditory quality of camouflaged and reconstructed audios is improved further.

The present application relates to an image registration method, comprising: selecting a registration source image {p,} and a registration target image {q_j}; applying a random perturbation to {p_i} in accordance with a preset random perturbation control parameter σ so that it is deformed to obtain {p_i′}, and obtaining a set of closest points, i.e., {q_j′}, on {q_j} corresponding to points on {p_i′}; performing an iterative operation on {p_i} and {q_j′} in accordance with a preset initial coordinate transformation H₀ to obtain a coordinate transformation {H_l}, 0<l<L, between {p_i′} and {q_j′}, and calculating an average distance {E_l} corresponding to the coordinate transformation {H_l} in accordance with the coordinate transformation {H_l}; judging magnitudes of the average distance {E_l} and a preset ideal average distance E_x, and terminating the registration when the average distance {E_l} is smaller than or equal to the preset ideal average distance.

The present invention provides a circuit, a chip and a method against power attack for grouping algorithms, the circuit comprising: a round operation module, a round key generation module, a data random perturbation and recovery unit, a round data random perturbation and recovery unit, and a MUX2_1 module; the method comprising: a random perturbation and a random perturbation and recovery algorithms. The circuit, the chip and the method against power attack for grouping algorithms provided by the present invention, use the random perturbation and the random perturbation and recovery algorithms, to directly conceal the information of the plain text and the round data in registers with little cost, thus making defense effects better.

The invention relates to a reactive power compensation method for a doubly-fed wind power system. On the basis of a small-signal mathematical model of a static synchronous compensator, a designed sliding mode surface and a sliding mode controller designed on the basis of a reaching law arrival condition can effectively solve the problem of voltage fluctuation, caused by changes of a fan input power and a reactive power load, in an individual wind-electricity hybrid power system and the problem of uncertainty of parameters of a static synchronous compensator model; simulation is carried out through applying a step or random perturbation to the fan input power and the reactive power load in the individual wind-electricity hybrid power system; and the simulation result confirms the validity of the method provided by the invention.

The invention provides a cloud-computing-system-oriented multi-scale task quantity predicted value determining method which aims to solve the problem that task quantity prediction precision is difficult to control in a single-scale task quantity prediction method. Long-time tendency, periodic fluctuation and recent situations of the task quantity in a cloud computing system and random perturbation events bursting currently are taken into consideration comprehensively, scientific and reasonable prediction of the user task quantity at a certain moment in the future is carried out, a task quantity prediction level within a period of time in the future is given, and important supporting bases are provided for plans and strategies of cloud service providers in adjustment of resource allocation and the scale of the cloud computing system.