93 results about "Matrix sequence" patented technology

The invention discloses a multi-layer heterogeneous sequential network feature alignment and clustering method and device based on tensor self-representation. The method comprises the steps that sequential feature matrix sequences {},..., {} of data views in a multi-view heterogeneous network in time windows are obtained; based on the matrix sequence of the first time window, constructing a Laplacian regular term by introducing a self-representation learning mechanism, modeling multiple views by adopting a tensor structure and introducing a Schatten p-norm regularization mode to construct an optimization objective function for optimization, and representing shared feature representation of each data view in the first time window; and solving the function and carrying out clustering processing on the obtained function, wherein the clustering result of each time window is used for carrying out characteristic analysis on nodes in the network. According to the method, on the premise that the original structure and semantics of the network are guaranteed, unified modeling can be carried out on the incomplete multi-view heterogeneous network, and feature expression is aligned.

The invention relates to the field of fluid pipeline system safety and intelligent control, and provides a water hammer pressure prediction and active suppression method and system. The method comprises the following steps: acquiring a multi-dimensional digital signal flow, and generating a window data matrix sequence through sliding window segmentation and dimensionless; image coding is carried out to form a spatiotemporal feature sequence, the spatiotemporal feature sequence is processed by a double-branch spatiotemporal prediction network, and a prediction feature spectrum sequence is output and decoded to form a prediction waveform matrix; future risk indexes and morphological characteristics are extracted and spliced with the multi-dimensional digital signal flow to form a prediction enhancement state vector, and the prediction enhancement state vector is input into a strategy network to generate valve action control parameters; and a target valve opening curve is decoded, a composite control signal is obtained through feedforward-feedback composite control, actual response data flow and historical data packaging are collected, and a prediction network and a strategy network are updated online through mixed priority screening. According to the method, prospective and self-adaptive water hammer prediction and active suppression are realized, the pressure evolution trend of a pipeline system is informed, and decision support is provided for safe operation of a fluid pipe network.

The invention discloses a large model mixing precision quantification method based on dynamic tree quantification. The method comprises the following steps: step 1, dividing input feature points into outliers and non-outliers; 2, extracting an outlier from the matrix sequence of the feature points, and extracting a row where a weight corresponding to the outlier is located; or processing and quantizing the non-outlier data used for the linear operator; non-outlier data applied to a nonlinear operator is not quantized; 3, performing matrix multiplication on the outlier data and the corresponding weight of the outlier; or, performing inverse quantization on non-outlier data stored in the linear operator in a quantized manner, and performing matrix multiplication operation; directly performing matrix multiplication on non-outlier data in the nonlinear operator; and 4, combining a non-outlier matrix multiplication operation result and an outlier matrix multiplication operation result through addition, and outputting a final result. The invention further discloses application and system of the method and a storage medium.

The embodiment of the invention provides an image stabilization method and device for a single-frame image of a video, and relates to the technical field of data processing, and the method specifically comprises the steps: decoding a to-be-processed video, and obtaining a plurality of continuous single-frame images in real time; generating a target image sequence based on a preset threshold number of single-frame images; for each two adjacent frames of single-frame images in the target image sequence, respectively extracting corresponding feature points, so as to obtain a first homography matrix corresponding to the two adjacent frames of single-frame images based on the feature points; generating a first homography matrix sequence based on the plurality of first homography matrixes, and performing Gaussian filtering processing on the first homography matrix sequence to obtain a target homography matrix; and based on the target homography matrix, performing transmission transformation on a middle single-frame image at the middle frame position of the target image sequence, and generating a target image stabilization image corresponding to the middle single-frame image. According to the method and the device, the quality of the generated stabilized image can be improved while the efficiency of obtaining the stabilized image is improved.

The invention provides a big data-based comprehensive energy production scheduling method and system, and particularly relates to a big data-based comprehensive energy production scheduling method and system, which are especially suitable for accurate load prediction and intelligent production scheduling of a complex power grid. The comprehensive energy production scheduling method based on big data comprises the following steps: acquiring historical power consumption data, and constructing a historical average load two-dimensional matrix sequence; future weather forecast data and internet event data are collected, and a weather weight map and an event weight map are generated respectively; fusing the historical average load two-dimensional matrix sequence, the weather weight map and the event weight map to generate a preliminary load prediction sequence; constructing a dynamic attribution power grid graph sequence based on the initial load prediction sequence, inputting the dynamic attribution power grid graph sequence into a space-time attention network model, and learning and predicting a residual sequence between the initial load prediction sequence and a real load value; and adding the residual error sequence and the initial load prediction sequence to obtain a final load prediction result.

The invention discloses a semiconductor wafer manufacturing factory logistics jam prediction method and system, and relates to the technical field of semiconductor manufacturing process optimization, and the method comprises the steps: collecting and preprocessing historical logistics data of a semiconductor manufacturing intelligent factory, carrying out the spatial-temporal feature fusion of the preprocessed data, generating high-dimensional representation data, and carrying out the prediction of the logistics jam of the semiconductor wafer manufacturing factory. Performing coding feature extraction processing on the high-dimensional representation data, extracting spatio-temporal features through a time sequence dynamic sensing operation and a spatial relationship self-adaptive operation, performing spatio-temporal feature fusion processing, performing cross-sequence feature integration on a processing result and complex spatio-temporal features, and performing decoding prediction processing on the integrated features, and through a causal constraint attention mechanism and cross-sequence correlation modeling, predicting a logistics state matrix sequence, and based on the logistics state matrix sequence, carrying out congestion state determination. According to the method, high-precision prediction and quick response of logistics blockage of the semiconductor factory are realized, and the problems of spatial-temporal characteristic splitting, response lag and high false alarm rate of a traditional method are solved.

The invention provides a communication method and system. The system comprises at least one first terminal, at least one second terminal and access network equipment, wherein the first terminal is used for sending first information to the access network equipment, the first information indicates a channel matrix sequence corresponding to the first terminal, and the channel matrix sequence corresponding to the first terminal comprises M + N channel matrixes arranged according to a time sequence; the second terminal is used for sending channel state information (CSI) to the access network equipment; the access network equipment is used for determining a channel matrix sequence corresponding to the second terminal according to the channel matrix sequence corresponding to the first terminal and the CSI; and sending data to the second terminal according to the channel matrix sequence corresponding to the second terminal. By adopting the communication system, the influence of the channel aging problem on the communication performance can be reduced, and the communication quality under high-speed movement can be improved.

The invention belongs to the technical field of robot control, and discloses a foot-arm multi-task control method and system based on time-varying terminal cost, and the method comprises the steps: generating a whole-body reference trajectory of a robot offline, and pre-calculating a time-varying terminal cost matrix sequence based on the reference trajectory; based on the time-varying terminal cost matrix sequence, constructing and solving a nonlinear model predictive control optimization problem to obtain an optimization sequence of a system state and control input; and the optimization sequence serves as a reference instruction and is input into a layered whole-body controller for multi-task priority optimization, and a joint control instruction is generated and output to a robot execution mechanism. According to the method, the problem of short vision of a traditional NMPC in a long-time task is effectively relieved, and the operation precision of the robot is remarkably improved.

The embodiment of the invention discloses a picture video generation method and device. According to the embodiment of the invention, a source image containing a target object is input into a first video generation model to obtain a material video, an inter-frame transformation matrix sequence is determined according to the material video, and then an object mask image corresponding to the target object is obtained from the source image; the inter-frame transformation matrix sequence is applied to an object mask image to obtain a plurality of mask images so as to form a mask image sequence, and the inter-frame transformation matrix sequence is applied to a source image to obtain a plurality of target object images so as to form a target object image sequence. And determining target input data according to the source image, the mask image sequence and the target object image sequence, and inputting the target input data into a second video generation model supporting local redrawing to obtain a corresponding target video. The video is generated through the two models, intelligent end-to-end video generation is achieved, and the diversity of motion tracks can be achieved while the target object is kept not to be diffused without introducing preset motion parameters.

The application provides a kind of acquisition method, system and photolithography machine of polarization aberration compensation parameter, it is related to polarization illumination technical field.The method includes obtaining the Mueller matrix sequence and the environment parameter sequence corresponding to photolithography machine system, the time sequence of the Mueller matrix sequence and the environment parameter sequence is aligned;The Mueller matrix sequence and the environment parameter sequence are respectively extracted and the features of the same time sequence are combined to obtain a feature tensor sequence;Determine the target compensation parameter according to the feature tensor sequence, the target compensation parameter is used to compensate the polarization aberration of the photolithography machine system.This method can improve the accuracy of acquiring polarization aberration compensation parameters, thereby further improving the overall compensation accuracy and imaging accuracy of the system.

The present disclosure relates to a method, a computing device, a computing system, and a storage medium for generating an operator for tensor operations. The method includes: parsing input information for generating a matrix sequence; generating symbolic representation information for operations based on the generated matrix sequence and an input symbolic vector; converting the generated symbolic representation information into abstract syntax tree information or intermediate representation information; and generating assembly code for an operator for tensor operations based on basic description information included in the generated abstract syntax tree information and intermediate representation information. The present disclosure can make full use of the hardware performance of each processor, and can efficiently generate an operator library for high-performance tensor operations for a GPU.

The invention discloses an RMFE-based distributed matrix multiplication implementation method. The method comprises the following steps: initializing parameters, constructing an RMFE function, splitting a to-be-processed matrix sequence into sequence vectors, sequentially carrying out phi function operation, obtaining a matrix distributed multiplication result through distributed matrix multiplication, and recovering through psi function operation to obtain a multiplication result of a to-be-processed matrix. According to the method, a distributed multiplication problem of a plurality of matrixes on a small domain / Galois ring is converted into a distributed multiplication problem of a single matrix on a large domain / Galois ring, and the calculation overhead on an expansion domain is allocated to the plurality of matrixes, so that the communication overhead required by distributed matrix multiplication is effectively reduced; the method has important application value for a high-speed communication system.

The application discloses a kind of stability sustained analysis method of underwater structure tensor information based on memory characteristics;It relates to underwater measurement technical field, the application is measured to underwater structure in N time points, obtains N three-dimensional height difference tensor matrix sequence;With the dimension of one of three-dimensional height difference tensor matrix as benchmark and as benchmark matrix, matrix alignment operation is carried out to other N-1 three-dimensional height difference tensor matrix, and alignment matrix is obtained;Difference is obtained between alignment matrix and benchmark matrix, and point difference matrix is obtained;For each point difference matrix, select edge irregular area and calculate effective selected area value;Based on the three-dimensional height difference tensor matrix obtained in current stage measurement, current memory length and historical effective selected area value set, current effective selected area value is updated by linear transformation and nonlinear activation processing function, the stability change trend of underwater structure is analyzed and early warning is carried out, and the accuracy of underwater structure tensor information stability analysis is effectively improved.

This invention discloses a method and device for controlling the lifting and lowering of sweeping discs in a sweeper truck, relating to the field of sanitation vehicle control technology. The method includes: S1, periodically collecting sweeping disc linkage monitoring data and preprocessing the data; S2, generating a sweeping disc linkage state matrix sequence, extracting matrix shift markers, and generating sweeping disc linkage judgment data based on the sweeping disc motion state; S3, matching the matrix shift markers to corresponding sweeping disc lifting action templates, performing timer combination matching, calculating the sweeping disc action duration value, and generating linkage strategy data; S4, generating corresponding side sweeping disc lifting control commands based on the linkage strategy data, and executing timing control according to the sweeping disc action duration value. This invention solves the problem of insufficient linkage between operation mode switching and side operation switching in existing sweeper truck sweeping disc lifting control, resulting in delayed sweeping disc lifting response, uncoordinated actions, and difficulty in dynamically adjusting the control duration.

The invention discloses a video editing processing method, which comprises the following steps of: extracting a noise frame from a training video frame, inputting the noise frame into a first neural network model, training the first neural network model to generate a pseudo feature matrix, inputting the pseudo feature matrix and a feature matrix extracted from the video frame after the noise frame is removed into a second neural network model, and training the second neural network model to distinguish the feature matrix and the pseudo feature matrix; after the training is finished, decoding the video data to be processed to obtain a video frame to be processed, inputting the video frame to be processed into the first neural network model to generate a feature matrix sequence, and distinguishing the feature matrix sequence by using a second neural network model to remove an illegal feature matrix; and clustering the remaining feature matrixes, and marking the video frames corresponding to the same type of feature matrixes. According to the method, the defects in the prior art can be overcome, the recognition speed and accuracy of the video features are improved, and the editing operation of editing personnel is effectively assisted.

The invention discloses an underwater structure tensor information stability continuous analysis method based on memory characteristics. The method relates to the technical field of underwater measurement, and comprises the following steps: measuring an underwater structure at N time points to obtain N three-dimensional elevation difference tensor matrix sequences; by taking the dimension of one three-dimensional elevation difference tensor matrix as a reference and a reference matrix, carrying out matrix alignment operation on the other N-1 three-dimensional elevation difference tensor matrixes to obtain an alignment matrix; performing subtraction on the alignment matrix and the reference matrix to obtain a point difference matrix; for each point difference matrix, selecting an edge irregular area and calculating an effective selected area value; based on a three-dimensional elevation difference tensor matrix obtained through measurement at the current stage, the current memory length and a historical effective selected area value set, a current effective selected area value is updated through linear transformation and a nonlinear activation processing function, the stability change trend of the underwater structure is analyzed, and early warning is carried out. And the accuracy of underwater structure tensor information stability analysis is effectively improved.

The present invention provides a method, medium and system for predicting cracks in a wall during civil construction, which belongs to the technical field of civil construction, and includes: firstly, based on known wall structure information, a three-dimensional finite element model is established, and stress concentration, structural defects and crack-prone areas are determined; next, multi-mode images of the wall are acquired regularly, and these images are pre-processed and feature fused to obtain a time series feature matrix; for these feature matrices, a difference matrix sequence between adjacent time points is calculated, and the matrix with the largest absolute value is selected as the maximum change matrix. Combined with the information of the crack-prone area, the maximum change matrix is ​​screened to obtain a focus matrix. The focus matrix is ​​input into a pre-trained crack prediction model, and the crack risk score of each crack-prone area is output. Finally, the crack-prone areas are graded according to the crack risk score, and the areas with higher risks are listed as key monitoring areas, thereby realizing accurate prediction and monitoring of wall crack risks.

The invention relates to a key identification method based on a millimeter wave radar. The method comprises the following steps: S1, obtaining distance angle spectrograms corresponding to different time frames; s2, obtaining an effective RA matrix corresponding to the distance angle spectrogram; s3, sorting the effective RA matrixes of the time frames, carrying out peak value detection based on the sum of the energy intensities of the effective RA matrixes, taking the effective RA matrix corresponding to each peak value as a seed matrix of single-bond data, and constructing a matrix sequence by using a single seed matrix and certain effective RA matrixes before and after the seed matrix; s4, respectively extracting spatial-temporal features based on each matrix sequence, and obtaining a key classification result based on the extracted spatial-temporal features and a first classifier; s5, extracting a continuous number part as a key sequence corresponding to one word according to a continuous key classification result; s6, generating a plurality of candidate words based on each key sequence; and S7, based on the candidate words of each key sequence, obtaining a text inference result based on a large language model. Compared with the prior art, the method has the advantages of high accuracy and the like.

This invention discloses a bird sound recognition method based on voiceprint recognition, comprising the following steps: collecting bird sound signals from a natural environment and preprocessing them to obtain a sound frame sequence; extracting acoustic features and calculating a frame feature vector sequence; performing segmented processing to calculate a structure consistency score and comparing it to obtain a frame weight sequence, constructing a weighted covariance matrix sequence; performing symmetric positive definite matrix constraint processing and mapping it to a symmetric positive definite matrix manifold space to obtain a manifold representation matrix sequence; combining the manifold representation matrix sequences to construct a covariance manifold trajectory; calculating shape invariants and constructing voiceprint feature vectors; inputting an improved supervised metric learning model to calculate similarity and determine the bird species category or individual bird voiceprint information corresponding to the bird sounds; and outputting the bird sound recognition result. This invention achieves bird sound recognition through a covariance manifold trajectory voiceprint recognition method, possessing the advantage of high recognition accuracy.

A domain name resolution request sent by a terminal is received. The domain name resolution request includes a domain name to be identified. The domain name includes at least one character. A sequence matrix corresponding to the domain name is determined. The sequence matrix includes at least one character vector. Each of the at least one character vector corresponds to each of the at least one character one to one. Each of the at least one character vector is sequentially input to an input and output gate. The input and output gate includes a logical operation rule between a plurality of activation functions. A feature vector corresponding to the sequence matrix is obtained by performing logical operation processing on each of the at least one character vector based on the logical operation rule between the plurality of activation functions.

The application relates to the technical field of network security, and discloses a network intrusion prediction method and device, which comprises the following steps: acquiring a first feature matrix sequence of a network to be measured, wherein the first feature matrix sequence comprises first feature matrices of multiple time snapshots of the network to be measured in a target historical time period; determining mapping information of spatial features of each node to be measured in the network to be measured in a target space according to the first feature matrix sequence; determining target space features of each node to be measured based on a query value, a key value and a mapping value of the spatial features of each node to be measured; determining a second feature matrix sequence of the network to be measured based on the target space features of each node to be measured; and determining a network intrusion prediction result of the network to be measured in a prediction time period based on the second feature matrix sequence. Thus, the technical effect of predicting the network intrusion that the network to be measured may face is achieved.

The invention provides a power node load prediction method and device, and relates to the technical field of data processing, and the method comprises the steps: obtaining a power grid weighted directed graph of a target regional power grid, and carrying out the prediction of a plurality of continuous historical time steps before a prediction time point; respectively constructing a node feature matrix and an adjacent weight matrix of the power grid weighted directed graph corresponding to each time step; stacking the node feature matrixes of the plurality of historical time steps along the time dimension to form a node feature matrix sequence, and stacking the adjacent weight matrixes of the plurality of historical time steps along the time dimension to form an adjacent weight matrix sequence; and inputting the node feature matrix sequence and the adjacent weight matrix sequence into a pre-trained space-time diagram neural network model to obtain a power load prediction value of each power grid physical node in the target regional power grid in one or more time steps in the future.

The embodiment of the invention provides a data processing method of a network communication system and related equipment, and the method comprises the steps: obtaining a to-be-decoded data vector sequence and a to-be-decoded matrix sequence from a plurality of batches of coding packets transmitted by a data transmitting end through network coding, and enabling the to-be-decoded matrix sequence to comprise a plurality of to-be-decoded matrixes; performing row simplification on all matrixes to be decoded in the matrix sequence to be decoded to obtain a simplified matrix sequence, and recording a corresponding row transformation operation sequence; and performing simplified iteration on the simplified matrix sequence until the variable columns of all the to-be-decoded matrixes are solved or decodable variable columns do not exist, and the row transformation operation sequence is used for performing transformation operation on the to-be-decoded data vector sequence to obtain decoded data, so that the solving efficiency of the to-be-decoded data vector is improved, and the decoding efficiency of the to-be-decoded data vector is improved. And the decoding efficiency and the decoding reliability of the matrix sequence to be decoded are improved.

The invention discloses a vehicle interaction decision-making method based on an unprotected intersection and a related device, and the method comprises the steps: obtaining a system state matrix, an own vehicle control matrix and an other vehicle control matrix in linear system state equations of an own vehicle and an other vehicle, first to fourth positive semi-definite matrixes and first and second positive definite matrixes are arranged in the optimization target of the linear quadratic differential game problem of the own vehicle and the other vehicle; taking the first positive semidefinite matrix and the third positive semidefinite matrix as a first intermediate matrix and a second intermediate matrix at the Nth moment respectively; according to the first intermediate matrix, the second intermediate matrix, the first positive definite matrix, the second positive definite matrix, the system state matrix, the self-vehicle control matrix, the other-vehicle control matrix, the second positive definite matrix and the fourth positive definite matrix at the Nth moment, reverse recursion is carried out on an optimal control gain matrix equation set of the two vehicles and a coupling Riccati equation of the first intermediate matrix and the second intermediate matrix at the kth moment; and obtaining an optimal control gain matrix sequence of the vehicle, and controlling the vehicle according to a control quantity sequence determined based on the optimal control gain matrix sequence.

This invention relates to an energy production forecasting method using a partial grey model with fractional derivatives, belonging to the field of energy production forecasting. It first selects the current monthly production values ​​of different energy sources as a database to construct an original matrix sequence X. (0) As input to the model; secondly, fractional derivatives and fractional accumulation operators are introduced when constructing the model to dynamically predict energy output under the grey effects of exponential and sine functions; finally, the simulated value X of the model is calculated. (r) , Restore value X (0) Furthermore, the model was compared with a control model in various indicators; the particle swarm optimization algorithm was used to find the optimal parameter vector that minimizes the MAPE value; finally, the new model was applied to energy production forecasting. This invention introduces exponential and trigonometric functions, giving the model's time response function oscillatory characteristics, thus accurately capturing and effectively mapping data volatility, significantly improving adaptability and flexibility; the integration of fractional derivatives and fractional accumulation operators into the model significantly improves prediction accuracy.

The invention relates to a fractional order partial grey model short-time traffic flow prediction method based on a two-dimensional advection diffusion equation, and belongs to the technical field of intelligent traffic. The method comprises the following steps: constructing original traffic flow data into traffic flow data in a matrix sequence form; processing the original matrix sequence, and calculating an accumulation generation sequence and a mean value sequence and a partial derivative sequence corresponding to the accumulation generation sequence; establishing a fractional order partial grey short-time traffic flow prediction model AD-FPGM (2, 1) based on a two-dimensional advection diffusion equation; constructing matrixes B and Y to estimate model system parameters, and searching a fractional order accumulation order r and a parameter lambda in a time response formula by adopting a particle swarm algorithm; calculating a simulation value and a reduction value of the AD-FPGM (2, 1) model; and carrying out error test according to each evaluation index of the comparison model, if the error test is passed, predicting a future trend by using an AD-FPGM (2, 1) model, otherwise, reconstructing traffic flow data to carry out model optimization. The prediction result can provide reliable information for a traffic management system, and the research method of traffic flow prediction is enriched.

The invention relates to the technical field of spectral measurement and data correction technologies, in particular to a reflectivity reconstruction method based on adaptive weighted multi-core support vector regression, which comprises the following steps: calling cross-aperture spectral measurement data, aligning and normalizing, comparing channel response vector direction differences, calculating neighborhood spacing and combining, and calculating the reflectivity of the cross-aperture spectral measurement data. According to the method, dynamic analysis of spectrum form indexes and local structure indexes is combined, kernel function weights are adjusted in a self-adaptive mode, spectrum form differences and local changes between different samples are accurately captured, and the reflectivity prediction curve sequence is obtained through dynamic analysis of the spectrum form indexes and the local structure indexes. According to the method, a dynamic weighting kernel function based on similarity between samples and neighborhood density is utilized to solve the problem that local complexity cannot be effectively incorporated into modeling, so that under different optical sampling conditions, the reconstruction result of spectral data is more stable and accurate and adapts to the change of complex spectral data, and the accuracy and consistency of reflectivity reconstruction are improved.

The invention provides an environment monitoring data evidence storage method based on a block chain, and relates to the technical field of data processing. The method comprises the following steps: firstly, performing matrix processing on target environment monitoring data to form an environment data matrix sequence; secondly, performing multi-angle semantic mining on the environment data matrix sequence to form a plurality of semantic mining vectors, and forming an environment data global vector based on the plurality of semantic mining vectors; then, performing anomaly recognition based on the global vector of the environmental data to form an environmental anomaly analysis result; on one hand, when an environment anomaly analysis result reflects that an anomaly exists, encrypted environment monitoring data and the environment anomaly analysis result are stored in a target block chain; and on the other hand, when the environment anomaly analysis result reflects that no anomaly exists, the target environment monitoring data and the environment anomaly analysis result are stored in the target block chain. Based on the method, the problem that data privacy and accessibility are difficult to effectively balance in the prior art can be improved.

The invention relates to the technical field of operation and maintenance, in particular to an operation and maintenance method and system based on deep reinforcement learning and a storage medium, and the method comprises the steps: constructing a comprehensive health index sequence and a multi-time scale matrix sequence according to multi-dimensional time sequence data of equipment; taking the comprehensive health index sequence and the multi-time scale matrix sequence as input, taking the risk level and the anomaly type as output, training a strategy network, and constructing an experience playback pool; empirical data are extracted from the empirical playback pool, a value network and a strategy network are trained, the value network carries out Q value updating based on an improved Q value function, parameters of the value network and the strategy network are updated, and an optimal decision model is obtained; and preprocessing real-time state data of the equipment, inputting the preprocessed real-time state data into the optimal decision model, and generating an operation and maintenance report according to probability distribution of risk levels and exception types output by the policy network. According to the method, the health change and abnormal fluctuation of the equipment are effectively captured, and the accuracy and robustness of operation and maintenance decisions are improved.

This invention relates to a data concatenation method, computing device, medium, and program product. The method includes: acquiring multiple query matrix sequences and multiple key matrix sequences; generating a first index matrix and a second index matrix to indicate the concatenation method of the sequences based on a preset concatenation size, the size of the multiple query matrix sequences, and the size of the multiple key matrix sequences. A first dimension of the first index matrix represents the number of groups of the multiple query matrix sequences, and a second dimension represents the sequence range of the query matrix sequences contained in the same group. The method further includes performing attention calculations on multiple input data using an attention operator based on the first and second index matrices to obtain attention results corresponding to the multiple input data. This method can increase the upper limit of the number of concatenated query matrices and key matrices, and improve the concatenation efficiency of query matrices or key matrices.