83 results about "Direct computation" patented technology

The present disclosure relates to methods and systems that preserve privacy in a secure multi-party computation (MPC) framework in multi-agent reinforcement learning (MARL). The methods and systems use a secure MPC framework that allows for direct computation on encrypted data and enables parties to learn from others while keeping their own information private. The methods and systems provide a learning mechanism that carries out floating point operations in a privacy-preserving manner.

The invention discloses an intention recognition method, device and equipment for a large language model dialogue and a storage medium. According to the scheme, a weight set corresponding to a current dialogue data set is calculated through a preset weight strategy, various dialogue forms and scenes are considered to set different weights respectively, different business scenes can be adapted, and the user experience is improved. The accuracy of weight setting is improved, so that the accuracy of subsequent intention recognition is improved; in addition, the intention matching result set corresponding to the current dialogue data set is directly calculated through the preset intention matching sub-model, and the accuracy and efficiency of calculating the intention matching result are improved.

The invention relates to the technical field of wind power generation, and discloses a numerical simulation inflow turbulence scale conversion method and device.The method comprises the steps that a turbulence inflow system is decomposed into an analysis part and a simulation part through wind spectrum time scale analysis and target time scale division, direct calculation of full-scale turbulence in a traditional method is avoided, and the calculation efficiency is improved; and the calculation cost is obviously reduced. The analysis part adopts instantaneous wind speed time history direct representation, the simulation part is solved through a turbulent flow transport equation, and calculation precision and efficiency are both considered. Smooth connection of an analysis scale and a simulation scale is realized by adopting a low-pass filtering technology, manual truncation errors are reduced, and the simulation precision of turbulence pulsation characteristics is improved.

The invention provides a mechanical product assembly performance simulation uncertainty analysis method. According to the method, the influence of each uncertainty influence factor on the assembly performance in different links of the simulation process is considered, the uncertainty influence factors in the simulation process are divided into two types, and quantitative analysis of uncertainty is carried out respectively. The first class comprises a measurement link, a modeling link and a simulation analysis link in the early stage of simulation, and the influence factors of the class can be used for directly calculating the uncertainty u or error; the second type of influence factors are external loads, boundary conditions and assembly performance uncertainty caused by non-linear links caused by process parameters mainly including geometric error distribution, and for the type of influence factors, the method combines optimal SVR proxy model construction and Sobol global sensitivity analysis on the basis of a simulation analysis result of a physical digital twinborn model, so that the assembly performance uncertainty caused by the non-linear links caused by the process parameters mainly including the external loads, the boundary conditions and the process parameters mainly including the geometric error distribution is obtained. And quantitatively analyzing the sensitivity of the assembly performance to the process parameters. In conclusion, the method provides guidance for the product assembly process.

According to the low-power-consumption task distributing and scheduling method for the heterogeneous multi-core processing system, the minimum power consumption serves as the primary factor for optimizing the task distributing and scheduling scheme, the task processing efficiency and reliability of a heterogeneous multi-core processor are improved, and power consumption is reduced; a power consumption calculation function is fitted by using a Gaussian process model in Bayesian optimization, so that direct calculation of most multivariable nonlinear power consumption is avoided, and the complexity and calculation amount of an optimization algorithm are greatly reduced; the crossover mutation probability of the genetic algorithm is defined by using an adaptive function, so that the genetic algorithm can be dynamically adjusted according to the population characteristics of the existing scheduling scheme, and the method has considerable superiority in the field of task allocation and scheduling; the adaptive population updating strategy is used to balance exploration of a new scheduling scheme and maintenance of a scheduling scheme with low power consumption, population updating can be performed according to the characteristics of the current population scheduling scheme, and the efficiency of finding a globally optimal solution by the genetic algorithm is greatly improved.

The invention discloses a rapid forward modeling method for a gravity field of a rectangular coordinate system based on an isoparametric transformation global discrete grid, and belongs to the technical field of geophysical exploration. According to the method, a measuring network composed of dggrid grids is manufactured to cover the whole measuring area range by selecting the measuring network forming form and the side length, the exploration depth is selected as D, the dggrid grids extend to the ground by an n < z + 1 > layer to form dggrid units of an n < z > layer, and the corresponding thickness of the dggrid units is D / n < z >. And calculating the abnormal response of all the dggrid units at each observation point by utilizing isoparametric transformation based on a dggrid unit interpolation type function, an integral point and an integral weight coefficient obtained by a hexagonal prism integral formula. Compared with a square grid, the dggrid grid has excellent angular resolution, the discrete distance measurement of the dggrid grid is better close to the Cartesian distance, and the dggrid grid also has the advantage of displaying uniform and clear adjacency. Compared with conventional direct calculation, the rectangular coordinate system gravitational field forward modeling based on the isoparametric transformation global discrete grid has extremely high acceleration potential.

An example apparatus includes interface circuitry, machine-readable instructions, and at least one processor circuit to be programmed by the machine-readable instructions to initialize a full hot expert buffer to store entire weights of an expert used with a first frequency, initialize a partial hot expert buffer to store partial weights of an expert used with a second frequency, wherein the first frequency is higher than the second frequency, identify a selected expert associated with a Mixture of Experts (MoE) layer of a Large Language Model (LLM), and perform a direct computation or a partially direct computation, the direct computation performed when the selected expert is stored in the full hot expert buffer, the partially direct computation performed when the selected expert is stored in the partial hot expert buffer.

The invention discloses a multi-task resource allocation method based on deep reinforcement learning of a strategy gradient, and the method comprises the steps: carrying out the intelligent agent training and perception of a resource allocation decision under the environment and state of a target history based on the deep reinforcement learning of the strategy gradient; according to the strategy gradient algorithm, an intelligent agent directly interacts with the environment according to a current strategy, the gradient of strategy parameters is directly calculated through track data obtained through sampling, and then the current strategy is updated, so that the current strategy is close to a target expected to be reported by the maximum strategy; in a strategy gradient algorithm, a Monte Carlo method is adopted to sample a track to estimate an action value, and an unbiased gradient can be obtained, so that the strategy can still dynamically adjust a resource allocation decision in a new environment and state, and the benefit maximization is achieved.

This application provides a method for uncertainty analysis in mechanical product assembly performance simulation. This method considers the impact of various uncertainty factors at different stages of the simulation process on assembly performance, classifying these factors into two categories for quantitative analysis. The first category includes the measurement, modeling, and simulation analysis stages in the early stages of simulation; the uncertainty u or error of these factors can be directly calculated. The second category consists of assembly performance uncertainties caused by external loads, boundary conditions, and nonlinear processes primarily based on geometric error distribution. For this category, this method uses simulation analysis results from a physical digital twin model, combined with the construction of an optimal SVR surrogate model and Sobol global sensitivity analysis, to quantitatively analyze the sensitivity of assembly performance to process parameters. In summary, this method provides guidance for the product assembly process.

The invention provides a robust design method and system for a fractional order PI controller of an inertia and delay fractional order object. The method comprises the following steps: describing a controlled object by adopting an inertia and delay fractional order transfer function; the feedback controller adopts a fractional order PI controller; a crossing frequency and a phase margin are given, and a proportional gain, an integral gain and a controller order of the fractional order PI controller are solely solved based on three constraint equations of the given crossing frequency, the given phase margin and zero slope of an open-loop phase of a closed-loop system at the crossing frequency; if a null set is solved, the values of the crossing frequency and the phase margin need to be given again, and solving is carried out; and traversing the range of the traversing frequency and the phase margin to obtain a feasible region of the fractional order PI controller. According to the method, the parameters and the feasible region of the fractional order PI controller can be directly calculated, it is guaranteed that the fractional order PI controller has high robustness, and the method has practical application value.

The application provides a Petri net supervisory control method and a controller applied to a complex manufacturing system, and the method comprises the following steps: obtaining system information of the complex manufacturing system, and constructing a Petri net model of the complex manufacturing system according to the system information; constructing a reachable graph of the Petri net model according to the Petri net model; taking the current state and resources of each device as input, and predicting transition enabling values of each device in all reachable states through a trained neural network model; and performing supervisory control on the complex manufacturing system according to the transition enabling values of each device in all reachable states. For the complex manufacturing system, the application can avoid the state explosion problem caused by the calculation of the reachable graph of the complex manufacturing system, greatly reduce the time complexity and calculation complexity caused by the calculation of the reachable graph, and directly realize the deadlock control of the complex Petri net and has high behavior permissibility.

The invention discloses a method for analyzing time-frequency characteristics of collision vibration by using a segmented B-spline wavelet collocation method, which comprises the following steps of: (1) establishing a dynamic equation for wavelet discretization, discretizing a dynamic control equation of a collision-vibrator system by using a 4-order B-spline wavelet basis function, and establishing a nonlinear algebraic equation set about a wavelet coefficient; (2) carrying out segmented iterative solution on the time domain, dividing the whole time domain into a plurality of sub-time periods which are connected in sequence, and carrying out iterative solution on the nonlinear algebraic equation set in each sub-time period to obtain a global wavelet coefficient; and (3) analyzing the time-frequency characteristics, and directly calculating energy distribution based on the wavelet coefficient to realize time-frequency joint characterization. The invention aims to provide a method which can fuse solution of system response and extraction of time-frequency characteristics in a unified mathematical framework, and realizes high-precision time-frequency joint characterization by directly solving wavelet coefficients, thereby overcoming inherent defects of an existing two-step method.

The invention discloses an improved RNN-based power system transient stability evaluation method and device, and relates to the technical field of power system evaluation. According to the method, the transient stability boundary constraint can fully adapt to the characteristics of a power system through the rated operation data and the maximum angle-of-attack difference data of the generator, and through the rated operation data and the maximum angle-of-attack difference data which are convenient to collect, the construction process of the transient stability boundary constraint is simplified, and the calculation complexity is reduced; and finally, the solving efficiency and accuracy of the transient stability boundary are improved. By combining the accurate fluctuation data of the power system with the transient stability evaluation model, the transient stability-related characteristics in the fluctuation data can be fully and quickly mined, the accuracy of transient stability evaluation is improved, and the low efficiency and the low cost caused by a large amount of direct calculation are avoided. Therefore, the evaluation efficiency and accuracy of transient stability evaluation of the power system are finally improved, and safe and stable operation of the power system is ensured.

The invention discloses a near-field dynamics matrix-free finite element method based on a partial assembly strategy, and the method comprises the steps: decomposing a finite element stiffness matrix into a continuous product of a plurality of operators based on near-field dynamics operator decomposition; the element stiffness matrix is implicitly expressed as a function of a plurality of parameters, real-time calculation and reduction are carried out when needed, and partial assembly of the element stiffness matrix is achieved; and performing iterative solution on the partially assembled element stiffness matrix, directly calculating the element stiffness matrix and the element vector in the iterative solution process, and accumulating the calculation result back to the global degree of freedom. According to the method, a matrix-free method is used for replacing a traditional strategy based on an assembled matrix, and finite element numerical simulation of the near-field dynamic model is achieved.

The embodiment of the invention provides a training method for a machine learning model, and the method comprises the steps: constructing the machine learning model which comprises an input layer, a hidden layer, and an output layer; the hidden layer comprises at least one mixed feature extraction module; the mixed feature extraction module comprises a first transmission path for non-parametric feature extraction; the first transmission path directly calculates a non-parameterized feature based on the first input feature, and transmits the non-parameterized feature backwards as a first output feature; constructing an index mapping relationship between the first output feature and the first input feature; obtaining training set data to perform multiple iterations on the machine learning model to complete training of the machine learning model; and in the back propagation process of each iteration, the first transmission path maps the gradient parameter of the first output feature to the first input feature based on the index mapping relationship, and the gradient parameter is used as the gradient parameter of the first input feature, so that back propagation of the non-parameterized feature is realized. Through combination of mathematical calculation and trainable parameters of deep learning, the flexibility of a network architecture is promoted, the model calculation efficiency and the feature extraction precision are improved, and the training stability of the model is enhanced.

The invention discloses a five-axis linkage machine tool RTCP parameter measuring method and device, and provides a universal geometric mathematical model, a rotating shaft is deflected by the same angle in the forward direction and the reverse direction, the relative height of a workpiece tray surface is changed, an execution point falls on the workpiece tray surface, and the RTCP parameter is measured. The method comprises the following steps of: calculating the height change of an execution point relative to a reference plane in the process, and directly calculating the distance r of the tail end of a laser / cutter deviating from a rotating center in a five-axis machine tool by utilizing a formula, and the core advantage of the model is that parameters which are difficult to directly measure are solved through parameters which can be directly measured by utilizing mathematical model analysis; according to the method, the height error introduced by manual judgment is counteracted in a formula skillfully through a symmetric measurement design, so that the authenticity and reliability of a measurement result are greatly improved, micron-level-precision RTCP parameter calibration is realized, and the five-axis machining precision is remarkably improved; the same set of core mathematical model and process is suitable for two different types of five-axis machine tools, namely a cutter and a laser, and the universality is high.

The present application relates to the field of scientific computing and GPU technology, and particularly relates to a GPU scan path optimization method.The method comprises: constructing a three-dimensional Stencil calculation GPU framework based on an optimal scan traversal path; based on the three-dimensional Stencil calculation GPU framework, constructing a scan traversal path mathematical model, abstracting the scan traversal path into an integer programming problem, taking the total memory access amount as an optimization target, and optimizing to obtain an optimal two-dimensional window scan traversal path.Compared with the traditional direct calculation, the method of the present application can significantly reduce the data transfer in the global memory, improve the utilization rate of the accessed data, i.e., improve the data reuse rate, reduce the invalid waiting caused by the data blocking and high memory delay of the calculation unit, and thus improve the calculation efficiency.

The embodiment of the present invention provides a method and device for linear resonant motor parameter identification. The LRA parameter identification uses the zero-input response discrete-time system model of the second-order linear resonant system to construct a second-order forward prediction error filter, and uses the augmented Wienerhof equation to solve the coefficients of the discrete-time system model to obtain the minimum mean square error estimate, which has the advantages of strong anti-noise ability and good robustness. Through the estimated discrete-time system model, the impulse response invariant method can directly calculate the poles, free oscillation frequency and damping coefficient of the continuous-time system without iterative calculations.

The invention discloses an off-axis diffraction calculation method based on the Rayleigh-Sommerpheny theory, the method integrates convolution kernel translation and adaptive zero filling technologies, electromagnetic field off-axis diffraction calculation based on the Rayleigh-Sommerpheny theory is realized, and the method specifically comprises the following steps: step 1, obtaining input parameters and defining; step 2, executing self-adaptive zero filling; 3, calculating the translation amount of the convolution kernel; step 4, generating an off-axis diffraction convolution kernel; and 5, performing diffraction calculation and result extraction. According to the method, off-axis diffraction calculation is achieved in a convolution kernel translation mode, serious calculation redundancy caused by the fact that a calculation domain is greatly expanded in a traditional method is avoided, meanwhile, calculation distortion introduced by FFT is eliminated through the theoretical minimum filling amount, the method can be equivalent to direct calculation of a Rayleigh-Sommerfeld diffraction formula in numerical value, and the calculation accuracy is improved. And extremely high and stable calculation precision can be maintained at any diffraction angle.

This invention discloses a lithography simulation method and system based on circular sampling function decomposition and superimposed pupil shift matrix. The method includes: sampling the shape function and pupil function of the extended light source to obtain the corresponding superimposed pupil shift matrix; processing the superimposed pupil shift matrix, the frequency domain circular sampling function matrix of the extended light source and the pupil plane to obtain the frequency domain cross-transmission coefficient matrix and the projection matrix; performing singular value decomposition on the projection matrix to obtain the singular value decomposition result; processing the spatial domain circular sampling function matrix based on the singular vectors to obtain the convolution kernel matrix; constructing a spatial domain image simulation model based on the singular value decomposition result, the convolution kernel matrix and the initial Hopkins model; and obtaining the spatial domain image simulation result based on the spatial domain image simulation model. This invention utilizes the superimposed pupil shift matrix method and analytical circular sampling function decomposition, which can reduce the direct calculation and storage of the TCC matrix and significantly reduce the computational cost required to obtain the kernel.

A model training method and a node device are applied to federated learning in the field of artificial intelligence. This method is applied to a federated learning system, which at least includes a first node device and N second node devices. The method includes: the first node device obtains a first training data set; then inputs the first training data set into a first model for training to determine a first parameter value of the target parameter of the first model; encrypts the first parameter value to obtain a first encrypted representation of the first parameter value; receives a second encrypted representation of the second parameter value of the target parameter sent by each second node device; finally, calculates a parameter sum value of the first parameter value and the second parameter value according to the first encrypted representation and the second encrypted representation, and the parameter sum value is used to update the first model, so as to obtain an updated second model. The node device directly calculates the parameter sum value of all nodes and does not know the weight value of each node device, thereby protecting the privacy data of each node device.

The invention discloses an agent dynamic obstacle avoidance method based on a probability velocity obstacle method, and belongs to the technical field of agent dynamic obstacle avoidance. According to the invention, the collision probability of the intelligent agent and the obstacle is combined with the traditional speed obstacle method, so that the intelligent agent reliably avoids the obstacle in an environment with incomplete speed information or noise; an analytic integral method and a hierarchical optimization strategy are designed and utilized, direct calculation and intelligent risk tradeoff of the collision probability are realized in a complex environment, and the calculation complexity is greatly reduced. The intelligent agent can make a decision in time, and can realize a real-time obstacle avoidance function for a single obstacle or a plurality of obstacles in actual movement.

The application provides a regular wave slamming construction platform stress analysis method and system and a computer readable storage medium thereof. According to platform basic dimensions and wave basic parameters, a stress condition of the regular wave slamming construction platform is directly, quickly and efficiently calculated based on a large amount of test measurement data. The defects of the direct calculation complexity, measurement cost and the like of the slamming pressure load prediction in the existing theoretical calculation and model test are solved.

The invention discloses a privacy protection federated learning security aggregation method, apparatus and device, and a medium, a double-server federated learning architecture is constructed based on a double trap door homomorphic encryption technology, and it is ensured that the privacy of client-side related information is not disclosed in a poisoning attack detection and global model aggregation process. On the premise of strictly guaranteeing the data privacy of a client, the difference value between elements of the gradient is directly calculated in a ciphertext state, and the difference of the gradient in element dimensions is effectively captured, so that the accurate detection of the encryption model poisoning attack in a heterogeneous data scene is realized, and finally, the detection accuracy is improved. Through a dynamic threshold Byzantine robust aggregation scheme based on an average value, the technical problem of difficult poisoning gradient elimination caused by gradient threshold dynamic change in a heterogeneous data scene is solved.

The application provides an irregular wave slamming construction platform stress analysis method and system and a computer readable storage medium thereof. According to platform basic dimensions and wave basic parameters, a stress condition of the irregular wave slamming construction platform is directly, quickly and efficiently calculated based on a large amount of test measurement data. The defects of the direct calculation complexity, measurement cost and the like of the slamming pressure load prediction in the existing theoretical calculation and model test are solved.

The present application relates to a kind of rock drilling method, rock drilling device and computer readable storage medium, including establishing OXYZ coordinate system, and origin O (0,0,0), O1 (x1,y1,z1), O2 (x2,y2,z2) and O3 (x3,y3,z3), and determine the dynamic coordinate of O2 and O3 in OXYZ coordinate system;Wherein, dynamic coordinate does not exist OO1 The length L1 of and O1O2 The length L2 of, and the freedom degree of the first arm body, second arm body and the limiting parameter between execution and corresponding connecting piece;S2: based on the dynamic coordinate of O2 and O3 in OXYZ coordinate system, calculate S3: make perpendicular to YOZ plane, its beneficial effect is that the dynamic coordinate of O2 and O3 in OXYZ coordinate system can be directly calculated by mathematical operation, application this method, can throw away specific structure difference, without structure limit to the analysis and control of the posture of arm support, and provide the specific control method of the arm support posture of control overbreak, it is convenient to implement.

A light-color computation chip, belonging to the field of data computation technology, includes: an instruction scheduling unit, a decoding unit, an optical transmitting unit, an optical receiving unit, a register buffer unit, a pre-stored code library unit, and a computation result integration unit. The optical transmitting unit, optical receiving unit, register buffer unit, and pre-stored code library unit corresponding to the same data bit collectively constitute a single-data-bit computation unit. The instruction scheduling unit receives external instructions and distributes them to the decoding unit, and receives the final computation result and outputs it uniformly. The decoding unit converts the electrical signal instructions issued by the instruction scheduling unit into corresponding bit-specific light-color codes. The optical transmitting unit outputs data light signals of corresponding colors or wavelengths according to the light-color codes. Advantages: It adopts a novel architecture combining direct wavelength encoding and lookup table operations, achieving a simplified, low-latency, and modularly expandable direct computation architecture.

A method for processing discontinuous finite element discrete viscosity terms based on an internal penalty method belongs to the field of computer-aided engineering. On the basis of the internal penalty method, the present invention converts the Dirichlet boundary into a Newman boundary by introducing a penalty coefficient when solving an elliptic equation with a Dirichlet boundary condition. It is extended to the boundary of an internal unit, and by introducing a bilinear operator as a penalty function, information exchange between adjacent units is established. Based on a high-order precision discontinuous finite element format, an improved internal penalty method is proposed, a new lifting operator is defined, and the tensor is eliminated, which is used to solve the compressible Navier-Stokes equations and numerically simulate fluid behavior. This method eliminates the derivative of the gradient of the viscosity term to the variable, and directly calculates the lifting operator using the expression of the viscosity term, thereby improving the computational efficiency; it can be conveniently extended to other types of equations, such as heat conduction equations, structural mechanics equations, and electromagnetic field equations.

This invention relates to the field of police information technology, specifically disclosing a police information recommendation method and system based on facial recognition. After successful facial recognition, multi-dimensional security policy parameters, including the operating entity, task type, and environmental tags, are collected. These parameters are then combined with a target identifier to construct a joint arbitration request and homomorphically encrypted. The encrypted request is distributed to at least two independent data access control nodes. Each node performs direct computation on the ciphertext based on its local private rules, generating and returning an encrypted partial arbitration result. A secure arbitration service aggregates all encrypted results to form a final encrypted arbitration result. This result is decrypted to obtain the plaintext arbitration conclusion, which is then used to filter and output a customized information list matching the current scenario from the associated database. This invention achieves secure collaborative adjudication while fully protecting the private policies and data of all parties, ensuring the accuracy and security of information recommendations.

The application discloses a kind of hard real-time Gamma implementation method and medium for machine vision based on approximate compensation method, and the implementation method is transformed by Gamma algorithm basic formula, and the exponential operation is transformed into multiplication and shift operation suitable for processor such as FPGA implementation, the precision lost in transformation process is compensated using Taylor expansion, and lookup table is used to assist calculation for some fixed subitems not suitable for direct calculation, a kind of gamma coefficient real-time variable and change range unlimited Gamma algorithm implementation mode is provided.The RAM and LE resource consumption of the algorithm are moderate, the lookup table used in the algorithm does not change with input bit width, and it is an implementation mode based on current gamma coefficient real-time calculation result to solve the problems existing in the prior art Gamma algorithm, and provide an effective new implementation means of Gamma algorithm for machine vision field data processing and data display.