517 results about "Sparse matrix" patented technology

The invention discloses a big language model reasoning optimization method and system based on dynamic video memory compression and memory isomerism, and intelligent management of video memory resources is realized by integrating a dynamic compression strategy of KV Cache and a memory parallel architecture. The method comprises the following steps: 1) analyzing the spatial-temporal characteristics of the KV Cache in real time, adaptively selecting a quantization compression algorithm, a rarefaction algorithm or a low-rank decomposition algorithm, performing hierarchical storage based on attention head importance scores, keeping high precision of a core head, and implementing low-bit quantization on a secondary head; (2) the compressed inactive data are divided into a plurality of data blocks to be stored in a system memory, a parallel data channel group is established according to the number of physical channels, the compressed blocks are concurrently read through multiple channels during loading, and parallel decompression of a sparse matrix is accelerated through a GPU tensor core; and 3) constructing a KV Cache multiplexing mechanism and a parallel channel, and parallelizing a compression / decompression process and model calculation by adopting a hardware acceleration compression and asynchronous pipeline mechanism.

The invention discloses an APT attack path reconstruction method based on time sequence diagram comparison clustering and a medium. A security event standardized data set is obtained; mapping each security event into a multi-modal node through a heterogeneous time sequence diagram set construction method, and generating a directed edge to construct and complete a heterogeneous time sequence diagram set; a stage embedding time sequence diagram set is obtained through the joint attack stage set; outputting a time sequence diagram similarity matrix by adopting a multi-scale diagram similarity algorithm of time alignment perception; generating an event semantic sparse matrix based on the threat intelligence knowledge graph; obtaining an image clustering result set; uncertain samples in the graph clustering result set are obtained and processed, and an APT attack path reconstruction result is obtained. The problem that an APT attack path reconstruction method mainly depends on rule matching, single-dimensional feature comparison and manual analysis of security logs or alarm streams is solved, the interpretability of APT traceability is greatly enhanced, and subjective errors of manual research and judgment are reduced.

The invention provides an all-silicon carbide double-sided heat dissipation module packaging method based on gradient thermal resistance optimization, and belongs to the technical field of electronic component packaging. Establishing a thermal resistance characteristic curve, and calculating a gradient thermal resistance vector to determine a heat flow path; constructing a heat dissipation temperature distribution matrix; obtaining thermal diffusion data; heat dissipation path thermal resistance is calculated by combining the thermal diffusion matrix and the temperature distribution matrix, and a thermal resistance sparse matrix is established and is dimensionally reduced into a thermal resistance low-dimensional matrix; calculating an optimal heat dissipation path parameter set; inputting the parameter set into a thermal resistance prediction deep attention network model for verification and fine tuning, and generating a final thermal resistance distribution scheme; a module structure and a material ratio are designed according to the scheme to form a sandwich structure; a silver paste sintering process is adopted to realize low thermal resistance connection and complete packaging, and the technical problems of non-uniform thermal resistance distribution and difficulty in optimizing a heat flow path in a double-sided heat dissipation structure of a silicon carbide power module in the prior art are solved.

The invention relates to an accelerator and an acceleration method based on Top-K sparse moment vector multiplication. The accelerator comprises a first acceleration unit and a second acceleration unit. The first acceleration unit is used for data preprocessing and sparse matrix coding, and the second acceleration unit is used for executing Top-K sparse matrix vector multiplication so as to solve the process of the first K results with the maximum values in the operation results of multiplication of a sparse matrix and a dense vector; wherein the first acceleration unit comprises a preprocessing module and a coding module; the data preprocessing step of the preprocessing module comprises quantization, sparse reconstruction and shuffling; the broadband demand of data transmission is reduced based on the quantization step of the sparse matrix; pruning non-zero elements which have small influence on sorting based on a sparse reconstruction algorithm; carrying out shuffling operation on the sparse matrix after sparse reconstruction so as to avoid a Top-K aggregation phenomenon; and the coding module is used for recoding the preprocessed sparse matrix. The method has the remarkable effects that the end-to-end performance is improved by 153.3 times, 2.5 times and 3.4 times respectively.

The invention relates to the technical field of geophysical exploration, and particularly provides a three-dimensional finite element transient electromagnetic forward modeling calculation method based on an iterative solution method.The method comprises the steps that tetrahedron space discretization is carried out based on a vector finite element method, and an electric field is given to six edges of a space tetrahedron; constructing a back-pushing Euler finite element equation set, and assembling a linear system formed by a large sparse matrix based on the equation set; performing iterative solution on the linear system based on an FGMRES algorithm to obtain an electric field value corresponding to a spatial tetrahedron unit edge; optimizing a spatial tetrahedral mesh based on a quadratic nesting method, and optimizing a time step length based on a third-order backward Euler difference format; a grid pre-encryption method is adopted, the calculation error of iterative solution is reduced, a region decomposition method is used for achieving parallel calculation, and the calculation efficiency is greatly improved. According to the method, the requirements of high precision and high efficiency in three-dimensional transient electromagnetic forward modeling calculation can be met, and an advanced calculation tool is provided for resource exploration in a complex geological environment.

The invention provides an efficient privacy set intersection method and device based on bucket coding, and the method comprises the steps: obtaining an input set and a target value set, the input set comprises a private set of a sender and a private set of a receiver, and the target value set is generated based on the private set of the receiver; encoding the private set of the receiver into a sparse matrix composed of a plurality of buckets through a hash function; determining a linear system of the sparse matrix; performing ascending sorting and re-labeling on the linear system according to the initial bit position set to obtain a target linear system; according to a bucket-based Gaussian elimination algorithm, based on a target linear system, through exclusive-or operation, determining oblivious key value pairs for storage; an intersection of the private set of the sender and the private set of the receiver is determined based on the oblivious key-value pair storage according to an oblivious linear evaluation protocol. According to the invention, communication resources required by privacy set intersection in an actual production environment can be further reduced.

The invention belongs to the technical field of geophysics. According to the truncation model space three-dimensional magnetotelluric inversion method and system, a three-dimensional inversion objective function is constructed according to an inversion initial model, the inversion objective function is optimized by adopting a Gaussian-Newton method, and a Gaussian-Newton increment equation is obtained, a sensitivity matrix in the Gaussian-Newton increment equation is a sparse matrix which is obtained after truncation is carried out by adopting a distance truncation threshold value and a sensitivity amplitude truncation threshold value, the Gaussian-Newton increment equation is converted into an unconstrained least square form, and an inversion updating direction is determined according to the least square form; and performing iterative optimization according to the initial inversion model, the inversion updating direction and the model updating step length to obtain an updated inversion model. According to the invention, units with small contribution to inversion updating are effectively cut, and the model space corresponding to each piece of data is reduced.

The invention provides a multi-source constraint gridding atmospheric pollutant and greenhouse gas emission estimation method, and relates to the technical field of environmental information processing, and the method comprises the steps: constructing a hierarchical grid structure of an urban core domain and a peripheral background domain based on a unified data basic set, forming a gridding baseline emission list through unified activity data and emission factors, establishing a multi-source scale coordination mechanism, fusing ground monitoring and satellite observation information, and generating an observation data set; according to the method, an SMOKE sparse matrix processing chain is embedded to construct a joint estimation model, parameter inversion and consistency constraint optimization are implemented, boundary flux coordination is carried out between a core domain and a background domain, and finally an emission list which has cross-atmospheric pollutant and greenhouse gas consistency and can be used for refined emission inversion is formed. According to the method, the problem of unstable data fusion caused by scale mismatching between the multi-source observation data and the emission model in the prior art is solved.

The disclosure describes hardware compression of sparse matrix content. One embodiment provides a graphics processor, the graphics processor comprising: a base die, the base die comprising a plurality of chiplet slots; and a plurality of chiplets, the plurality of chiplets being coupled with the plurality of chiplet slots. At least one chiplet of the plurality of chiplets comprises: a graphical core cluster comprising a plurality of processing elements; a shared local memory coupled with the plurality of processing elements; a plurality of matrix engines coupled to the shared local memory; and codec circuitry coupled with the shared local memory and the plurality of matrix engines. Codec circuitry is configured to decode matrix data stored in a first format in a shared local memory into a second format for consumption by a plurality of matrix engines.

Aspects of the disclosure are directed to hardware acceleration of structured sparse workloads with block quantization. A hardware accelerator can receive compressed input matrices, for example as part of a workload for training or processing a machine learning model. The hardware accelerator can multiply the compressed input matrix with a gains matrix loaded in one or more matrix multiply units (MXUs) of the hardware accelerator. The input matrices can be further provided in a block data type format, in which blocks of mantissas are represented with a single shared scaling factor. An MXU can multiply the block data, shift or cast the block data according to a shared scaling factor to generate an output product. To that end, block data type matrices exhibiting structured sparsity patterns can be accelerated without affecting the overall accuracy or quality of the output to the workload being processed.

The invention relates to a hierarchical optimization pruning method and device for electric power multi-modal defect detection and a medium, and the method comprises the steps: collecting multi-modal data in the defect detection of electric power equipment, inputting the multi-modal data into a pre-training defect detection model, carrying out the feature vector extraction of each modal, and carrying out the perception strategy based on a Hessian matrix, and obtaining the weight importance of each modal; calculating the L1 norm of each channel weight in the pre-trained defect detection model, evaluating the channel importance of the defect detection model, determining a redundant channel according to the L1 norm and cutting off the redundant channel, comparing the weight importance with a preset threshold, marking the weight importance lower than the preset threshold as a pruning object, generating a pruning mask matrix, and performing pruning on the pruning object; performing pruning operation on the pre-trained defect detection model through the pruning mask matrix; and sparse matrix representation optimization is carried out on the pruned model, and retraining is carried out. Compared with the prior art, the method has the advantages of enhanced adaptability, enhanced stability, high resource utilization rate and the like.

The invention provides a multi-LoRA large language model deployment system based on a cloud computing platform. A cloud computing AI platform layer trains a LoRA adapter matched with a basic large language model for a reasoning process; the multi-LoRA dynamic loading layer dynamically switches LoRA adapters needing to be mounted according to the request parameters, GPU optimization configuration is carried out according to service priorities in the request parameters, a basic large language model is loaded, and a plurality of LoRA adapters stacked in a sparse matrix form are mounted; and the resource scheduling optimization layer responds to the resource scheduling application, and outputs the request parameters subjected to priority management and hardware sensing optimization to the LoRA dynamic loading layer based on the container arrangement platform. According to the method, LoRA and container arrangement are deeply fused, and an automatic assembly line of training and reasoning is achieved. A video memory sharing mechanism enables a plurality of service scenes to share the same basic large language model, a LoRA adapter is loaded as required, and video memory occupation is greatly reduced. Full-life-cycle management from data preparation to model service is supported, and the large model deployment cost is remarkably reduced.

The invention discloses a high-speed imaging method, system and device based on an SPAD array and a time counting circuit. The method comprises the following steps: acquiring original data of a photon event to obtain an original photon counting matrix of each frame; performing statistical modeling based on the original photon counting matrix to obtain a dynamic threshold matrix; performing threshold filtering based on the dynamic threshold matrix in combination with the original photon counting matrix to obtain a sparse matrix; and performing accumulation and image reconstruction based on the multi-frame original photon counting matrix to output a final image. The high time resolution of the SPAD array and the TCSPC is combined with the BNQSA algorithm, the noise problem of high frame rate imaging under extremely low illumination is solved, the system has the characteristics of high sensitivity, real-time performance and low power consumption, and a breakthrough solution is provided for a weak light scene.

A honeycomb-shaped active power distribution network topological structure selection and optimization method based on a graph theory comprises the following steps: taking line impedance and rated apparent power as constraints, constructing a multi-target weight factor, adopting an improved spectral clustering algorithm to cluster and distribute power grid nodes, and transiting a micro-grid group after cluster division into a honeycomb-shaped active power distribution network; abstracting each micro-grid as a honeycomb distribution network single node, constructing a connection relation mathematical representation of the HSPH and the micro-grid, and realizing sparse matrix method modeling of the honeycomb active power distribution network; optimizing a site selection strategy of the HSPH by taking a maximized HSPH return on investment index and a system stability index as a target function; and optimizing the energy storage capacity of the HSPH based on the net load time sequence data of the micro-grid group, and finally generating an optimal topological structure of the honeycomb-shaped active power distribution network. And through honeycomb topology and HSPH addressing and sizing optimization, the new energy consumption capability of the power distribution network is improved, and the economy, reliability and flexibility of the power distribution network are improved.

The invention relates to a risk control method for cross-border trade data. Cross-border trade behaviors are divided into three levels, namely a system type layer of national system / policy triggered transaction, a business type layer of a process chain and an execution type layer of data submission / customs clearance / payment behaviors. Modeling a behavior stability mapping relation among the three hierarchies through sparse matrix reconstruction and adversarial noise regression; introducing an interlayer fluctuation tension coefficient for each transaction to measure offset appearing in a behavior chain; when the tension peak value exceeds a training threshold value, even if single-point data is not abnormal, the event is marked as a potential risk transition event; according to the method, cross-border transaction behaviors are divided into a system type layer, a business type layer and an execution type layer, and an interlayer fluctuation tension coefficient is introduced, so that a cross-level collaborative imbalance phenomenon in a transaction chain is effectively captured. Even under the condition that no obvious field anomaly exists, the pseudo-legal transaction mode caused by behavior chain structure mutation can be recognized, and early warning of non-dominant structure anomaly is achieved.

The invention relates to a fault arc-caused fire early warning method based on multi-sensor fusion, and belongs to the field of fault arc fire. The method comprises the following steps: establishing an arc electrical fire cause fault simulation test platform, selecting a typical single load and a combined load to perform a test before normal operation of a line and the occurrence of an arc cause fire, and establishing a fault arc waveform database and an arc fire database; the power-on detection method for the fault arc detection device is based on dictionary learning, current characteristics are represented through a sparse matrix, and fault classification is carried out in combination with an SVM. The fluctuation range of a plurality of characteristic quantities of the line in a normal state and when a fault arc occurs is analyzed, a threshold value is determined according to the periodic increase ratio of the characteristic quantities in different states, a final combined load detection threshold value is formed by complementation of a plurality of threshold values, and fault arc detection is realized; and based on the arc fire database, training a double-layer LSTM model under Bayesian optimization, and identifying multi-sensor signals in a time window to realize fire early warning.

The invention relates to the technical field of image analysis, in particular to a metal composite material surface defect intelligent detection method based on machine vision, which comprises the following steps: acquiring an original image and converting the original image into an observation matrix; decomposing the matrix into a low-rank matrix and a sparse matrix through a robust principal component analysis algorithm; utilizing a nuclear norm constraint low-rank matrix to establish a background statistical model, and utilizing a norm constraint sparse matrix; positioning a defect area in the sparse matrix and extracting a gray level co-occurrence matrix feature vector; performing connected domain analysis on the sparse matrix, calculating geometrical morphology parameters and evaluating a stress concentration coefficient; and establishing a self-adaptive decision model to carry out fusion judgment so as to judge the physical damage. According to the method, through low-rank sparse matrix decoupling and multi-dimensional statistical geometric feature fusion analysis, accurate stripping of weak defects and quantitative evaluation of physical damage attributes under a complex background are realized.

The invention relates to the technical field of signal processing, and discloses a multichannel data acquisition system and method based on an FPGA (Field Programmable Gate Array), an acquisition module drives a global counter by using a global synchronous clock, writes acquired data into an annular buffer area with a physical address and a counter value in a linear mapping relationship, and establishes implicit time index storage. When a trigger event occurs, the system broadcasts the locked global trigger timestamp, and the acquisition module backtracks and reads historical data according to the global trigger timestamp, and packages the historical data into a sparse matrix type data packet in combination with the channel validity mask. And after receiving the data packet, the data processing module directly calculates a memory mapping address by using global time information in the data packet, and writes a data load into a corresponding position of the waveform reconstruction buffer area. According to the method, through strict binding of the physical address and the absolute time, independent time label redundancy is eliminated, high-precision synchronization of distributed multiple channels is ensured, and out-of-order automatic in-situ recombination and efficient waveform reconstruction of data of a receiving end are achieved.

The invention discloses an aircraft simulation method, device and equipment based on parallel block sparse matrix incomplete LU decomposition and a medium, and relates to the technical field of simulation, and the method comprises the steps: building a block sparse matrix comprising a plurality of zero sub-blocks and non-zero sub-blocks based on the topological relation of each grid point in an aircraft grid and simulation physical parameters, determining dependency pairs respectively corresponding to a plurality of non-zero sub-blocks from the non-zero sub-blocks, and storing the dependency pairs by adopting a preset data structure; wherein the dependent pair comprises a pair of matched non-zero sub-blocks; respectively distributing the plurality of non-zero sub-blocks to a plurality of GPU thread bundles for parallel processing so as to update the corresponding non-zero sub-blocks based on the updated dependency pairs after the GPU thread bundles complete updating on the dependency pairs corresponding to the corresponding non-zero sub-blocks; and after the plurality of non-zero sub-blocks are updated, updating the block sparse matrix based on the plurality of updated non-zero sub-blocks, and determining an incomplete LU decomposition result according to the updated block sparse matrix so as to perform aircraft simulation.

One embodiment provides a graphics processor comprising a base die including a plurality of chiplet sockets and a plurality of chiplets coupled with the plurality of chiplet sockets. At least one of the plurality of chiplets include a graphics core cluster including a plurality of processing elements, a shared local memory coupled with the plurality of processing elements, a plurality of matrix engines coupled with the shared local memory, and codec circuitry coupled with the shared local memory and the plurality of matrix engines. The codec circuitry is configured to decode matrix data stored in the shared local memory in a first format into a second format for consumption by the plurality of matrix engines.

The invention discloses a dynamic compression coding method, which is applied to a sparse matrix, and comprises the following steps: S1, obtaining the sparse matrix, and generating a bitmap with the same dimension as the sparse matrix, a value in the bitmap being used for indicating whether an element at a corresponding position in the sparse matrix is zero; s2, a mark sequence is generated according to the bitmap, and each mark corresponds to multiple continuous elements in the sparse matrix and is used for representing whether each element in the multiple elements is zero or not; s3, extracting non-zero elements in the sparse matrix according to the mark sequence; and S4, combining the flag sequence with the non-zero elements to form a data stream after compression coding. The invention also discloses a dynamic compression coding system, a storage medium and a device. According to the dynamic compression coding method, the index storage cost can be reduced, the data access efficiency can be improved, the storage and bandwidth occupation can be reduced by supporting parallel coding and decoding, and finally the calculation throughput can be improved.

The invention provides a high-voltage circuit breaker voiceprint feature extraction method based on PSO-VMD and multi-scale permutation entropy, and belongs to the technical field of electrical digital data processing, and the method comprises the steps: firstly obtaining circuit breaker voiceprint signals under different working conditions, carrying out the size floating function processing, and carrying out the optimal parameter combination of VMD; variational mode decomposition is carried out on the voiceprint signal to obtain a plurality of intrinsic mode functions; extracting time domain features, frequency domain features and multi-scale permutation entropy features to construct sparse permutation features; then calculating a bilateral approximation balance contribution value of each modal function to construct a product operation of a feature weight matrix and an original feature matrix to form a disorder sparse matrix; and then hierarchical processing is carried out to convert into a disordered clustering matrix, and the disordered clustering matrix is input into a double-current CNN-SVM network to realize circuit breaker fault mode identification and classification. The technical problem that voiceprint feature extraction of the high-voltage circuit breaker cannot comprehensively represent complex fault acoustic characteristics is solved.

The invention discloses an acoustic temperature field reconstruction algorithm based on sparse matrix and refraction effect consideration, and belongs to the field of acoustic temperature measurement. The algorithm introduces a sparse reconstruction technology and refraction effect optimization to solve the problems of large reconstruction error of an underdetermined equation set, sensitivity to a measurement error and the like caused by the fact that the number of sound rays is smaller than a grid sectioning number in existing acoustic tomography. The method comprises the following specific steps: dividing a temperature measurement area into N grids, constructing a compressed sensing model through an observation matrix A and a sparse base psi, reconstructing a sparse signal by using an improved generalized orthogonal matching algorithm, namely IMOMP, reconstructing a temperature field through a relationship between sound slowness and temperature, and improving precision by using cubic spline interpolation. Meanwhile, the sound wave refraction effect in the non-uniform temperature field is considered, the sound ray path layout is optimized, the algorithm effectively deals with inverse problem instability, the reconstruction error is reduced, the anti-noise capability is improved, and the method is suitable for complex temperature field reconstruction and has application value for industrial temperature measurement.

The invention discloses an infrared bidirectional thermal effect simulation method based on radiation and GPU acceleration, and belongs to the field of thermal radiation simulation and parallel computing. Dispersing the complex geometry into patch units, constructing an inter-patch radiation energy balance equation based on a radiance theory, and iteratively solving the final temperature of the patches through a radiance method; a three-level parallel strategy of a task level, a data level and an instruction level is designed, data intensive tasks such as shape factor calculation and radiance iteration are migrated to a GPU to be executed, data storage is optimized in combination with a structure array (SoA) layout and a sparse matrix compression technology, and efficient data sharing of a CPU end and a GPU end is achieved through a CUDA zero copy technology. According to the method, the dependence of a traditional method on regular grids is broken through, the calculation efficiency and precision of million-level surface patch heat radiation transmission in a complex scene are remarkably improved, and an efficient tool is provided for heat radiation indirect transmission calculation and a global illumination model in the field of three-dimensional scene infrared simulation.

The invention relates to a calculation device, a calculation method, and a program. This computing device (10) computes a matrix product of a sparse matrix and a weight matrix in a neural network, the sparse matrix having a structure in which a predetermined number of non-zero elements are included in each block of a predetermined size. The sparse matrix is defined by a local index representing the position of the non-zero element in each block and the value of the non-zero element, and the arithmetic device (10) performs: a process for acquiring, from the weight matrix, an element corresponding to the position of the non-zero element in the sparse matrix by referring to the local index; and multiplication and accumulation operation of non-zero elements of the sparse matrix and corresponding elements of the weight matrix is carried out.

The invention discloses an entropy coding compression method for high-dimensional sparse data, which relates to the technical field of data compression, and comprises the following steps: reading an original high-dimensional sparse matrix, extracting a position index set and a corresponding non-zero value set of all non-zero elements, extracting active samples from the non-zero value set, and compressing the active samples. After an active sample matrix and an optimal mean value centralization matrix are constructed, principal component projection and a self-expression structure are introduced for joint modeling, a low-rank robust optimization objective function is formed, and a principal component feature matrix is finally output by alternately optimizing mean values, projection, weights and residual errors. According to the method, the compression efficiency and the processing pertinence of high-dimensional sparse data are effectively improved, self-expression structure modeling and residual regular optimization between samples are further combined, the structure consistency is kept in the dimension reduction process, a key information structure is kept while the compression ratio is guaranteed, and the high-fidelity and low-redundancy entropy coding compression effect is achieved.

The invention discloses a simulation method and system for improving the simulation efficiency of a PIN diode, and belongs to the technical field of semiconductor device simulation, and the method comprises the steps: firstly, employing a hash table data structure to optimize the preprocessing process of a semiconductor physical model time domain numerical method, and remarkably improving the parameter preprocessing efficiency; carrying out bandwidth compression optimization on the sparse matrix in combination with an inverse Cuthill-Mckee matrix reordering algorithm; and secondly, constructing a parallel iteration solver architecture to realize efficient distributed solution of a large nonlinear matrix equation. Through cooperative application of the method, the overall calculation efficiency of a PIN diode numerical simulation program is improved by more than three times compared with that of traditional commercial simulation software COMSOL on the premise that the calculation precision is kept, the industrial problem that the simulation timeliness of a semiconductor device is limited is effectively solved, and an innovative technical path is provided for high-precision real-time simulation. The method has important application value in the fields of power device design, integrated circuit simulation and the like.

The invention discloses a multi-node concurrent transmission signal superposition diversity method. The method comprises the following steps: S1, generating a Turbo coding sequence by adopting a Turbo encoder; s2, applying phase disturbance to each Turbo coding sequence; s3, performing non-orthogonal superposition in an air channel to form a complex baseband signal; s4, constructing an equivalent observation model; s5, modeling the sending symbol matrix into a sparse matrix to obtain a symbol estimation matrix; s6, inputting the symbol estimation matrix into a Turbo decoder, executing iterative decoding by adopting an improved Max-Log-MAP algorithm, and outputting a bit estimation sequence; s7, completing initial phase estimation by using a Zadoff-Chu lead code, and outputting an estimated phase sequence through a Kalman filter; and S8, dynamically updating the equivalent observation model based on the estimated phase sequence, and completing parallel computing and scheduling. According to the method, Turbo coding, sparse reconstruction and phase disturbance modeling are fused, and multi-node non-orthogonal concurrent transmission is achieved.

The invention provides a multi-feature factor matching and interrupt processing method for storm tracking. The method comprises the following steps: calculating storm features required by tracking; carrying out discretization processing and coding on the storm features, and constructing a sparse matrix; constructing an auto-encoder model, performing data reconstruction on the sparse matrix based on the auto-encoder model, and calculating main features of the single storm; on the basis of the main features, calculating the feature vector distance between the previous and later time secondary storms, and constructing a feature space distance matrix; a bipartite graph algorithm is adopted, the incidence relation of the previous and later time-order storms is obtained based on the feature space distance matrix, and a storm matching result is constructed; identifying an interrupted storm based on a storm matching result and carrying out secondary matching; and constructing a storm trajectory based on a secondary matching result and removing a repeated storm trajectory. According to the method, the movement track of the storm can be accurately tracked from formation to extinction of the storm, reliable data support is provided for meteorological monitoring and forecasting, and the timeliness and accuracy of severe convective weather early warning are improved.

The invention discloses a method for accelerating a convolutional neural network by using matrix sparsity on a multi-GPU platform, and the method comprises the following steps: determining a data scale F and a GPU number G, and starting an MGPUSimm kernel; the input data is loaded to a multi-level cache from a global memory of the MGPUSimm; non-zero value mask traversal is carried out on input data, and a mask graph set is obtained through GPU parallel; convolution calculation is carried out through the mask graph, and zero value data input is ignored, so that calculation is reduced; if a pooling layer exists behind the convolutional layer, subsequent pooling layer calculation is completed; and completing calculation of the remaining layers and outputting a result to a global memory. The invention aims to provide a method for accelerating a convolutional neural network by using matrix sparsity on a multi-GPU platform, and aims to solve the current situations that a sparse matrix based on a CPU or a single GPU platform is low in calculation efficiency in convolutional layer calculation of the convolutional neural network and the overall calculation time of the convolutional neural network is relatively long. The method comprises the following steps: adding a mask to remove zero value calculation so as to reduce the calculation amount, and then inversely deducing an input data position from a result mask to ignore the influence of a 0 value in a sparse matrix on the calculation of a convolutional layer; and meanwhile, multiple GPUs are used in the steps of forming a mask image set, calculating a convolutional layer, calculating a pooling layer and the like, so that the overall calculation efficiency of the convolutional neural network is improved in parallel.