44 results about "Matrix addition" patented technology

Provided is a radar device capable of accurately calculating directions and the number of targets. A direction calculation unit includes a correlation matrix addition prohibition unit that prohibits, when a peak frequency in a plurality of modulation periods of a target is in the vicinity of 0, addition of a correlation matrix generated from a peak frequency spectrum having the peak frequency in the vicinity of 0, and calculates the direction of the target on the basis of a summed correlation matrix in which correlation matrices generated from peak frequency spectra having peak frequencies out of the vicinity of 0.

The invention discloses a large-scale MIMO (Multiple Input Multiple Output) precoding method based on an improved newton iteration method. The method comprises the following steps of firstly estimating a channel matrix, and calculating an RZF precoding expression through the obtained channel matrix; then adopting the newton iteration method for estimating an inverse matrix in an RZF precoding algorithm, and converting matrix inverse operation into matrix addition and matrix multiplication operation; finally utilizing an obtained precoding matrix for precoding a sending signal. The improved newton iteration method is characterized by building a high-order iteration formula, transforming a characteristic value located near 0 to closer to 1 , and remaining a characteristic value near 1 unchanged, so that the convergence rate of newton iteration is accelerated. An experimental result shows that when the iteration times exceeds four times, the performance of a traditional newton iteration method is superior to an inverse matrix estimation algorithm based on taylor series expansion. When the iteration times is 2, the improved newton iteration optimization algorithm can acquire around 95 percent of RZF precoding average client arrival rate.

The invention discloses a large-scale MIMO signal detection method based on Jacobi iteration, and belongs to the technical field of wireless communication. The method comprises the steps: converting amatrix inversion process into an iteration process of matrix multiplication and matrix addition, providing the search direction for a Jacobi algorithm through a gradient algorithm and an overall correction acceleration method, and determining a correction coefficient of an iteration equation. According to the invention, the method performs the estimation of a high-dimensional matrix inversion process through an improved Jacobi iteration method, and converts the matrix inversion process into the iteration process of matrix multiplication and matrix addition, thereby greatly reducing the computing complexity. The gradient algorithm and the overall correction acceleration method are used for providing the search directions for the Jacobi algorithm and determining the correction coefficient of the iteration equation, thereby enabling the iteration convergence to be better, and enabling the convergence speed to be higher.

The invention discloses a large-scale MIMO pre-coding method based on a Chebyshev iterative method. The method comprises the following steps that firstly, a base station estimates a channel matrix through obtained channel state information and computes an expression of RZF precoding according to the obtained channel matrix; secondly, the base station adopts the Chebyshev iterative method to carry out iterative estimation on an inverse matrix in an RZF precoding matrix and transforms a process of solving the inverse matrix into matrix addition and matrix multiplication; and lastly, the base station utilizes the obtained precoding matrix to carry out precoding on a transmitting signal. The experimental result shows that the Chebyshev RZF precoding algorithm can obtain an average user arrival rate approximate to the RZF precoding algorithm under the same initial conditions through two iterations, and the computation complexity is low. When the same average user arrival rate is obtained, the complexity of the Chebyshev RZF precoding algorithm is smaller than that of a Newton RZF precoding algorithm, and the rate of convergence is faster.

The invention discloses a Flink based large-scale matrix parallelization computing method. The method mainly comprises the following steps: storing large-scale matrix data by utilizing FlinkDataSet and accelerating matrix calculation by adopting BLAS during matrix calculation in a single computer; designing and realizing a series of matrix operations of matrix addition, subtraction and the like; and designing a parallelization scheme and designing three optimization means in combination with Flink and algorithm characteristics for improving the performance of matrix multiplication operation of different shapes, wherein the matrix multiplication includes square block division mode based matrix block multiplication, CARMA division mode based matrix block multiplication and broadcast mode based matrix block multiplication. According to the method, the problem that conventional large-scale matrix calculation in a single computer has high overhead and even cannot be executed is solved; and the method has very high expandability.

The invention belongs to the technical field of infrared image display processing, and particularly relates to a high-dynamic rage infrared image digital detail enhancement method with noise suppression ability. The infrared image digital detail enhancement method comprises the following steps of: respectively filtering a 14-bit original image by using two wizard filters with different smooth scale coefficients; obtaining detail images through matrix subtraction; and adding the detail image with a converted 8-bit image through matrix addition. According to the method, noises of infrared images can be effectively suppressed and digital details can be enhanced.

The invention discloses a multi-temporal remote sensing images change detection method based on fusion strategy and FCM, and the method comprises the steps: a detection range can be divided into a determined region and an undetermined region according to a change vector amplitude of multi-temporal remote sensing images and a multi-temporal spectral angle map; in the undetermined region, the change vector amplitude and SAM information are fused based on fuzzy division matrix addition; results of the determined region and the undetermined region are combined to obtain a final change detection result; wherein a fuzzy index in an FCM objective function in the undetermined region can be selected according to conflict indexes of MCV and SAM in the undetermined region to obtain a change detection result which is more stable and has higher precision than before. According to the invention, the change detection region is divided into the determined region and the undetermined region, detection results of two regions can be separately acquired through adopting a fusion strategy, so that the change detection results can be more reliable and more stable than before.

The invention discloses a data processing method for a wireless sensor real-time monitoring system based on compressive sensing. A measurement matrix constructing mode in the compressive sensing is changed, a Gaussian random matrix is replaced by a two-dimensional random matrix, the original measurement matrix is subjected to block operation, and the position with the element value of 1 in a sparse two-dimensional matrix is determined by using a coordinate position matrix, so that the dimension of a reduction matrix is greatly reduced; and meanwhile, the original matrix multiplication is converted into matrix addition operation, the operation speed is increased, and the obvious effects are that due to the algorithm improvement, the compressive sensing algorithm can be realized in actual application, the requirement that a general wireless sensor node has low memory is met, the calculation amount is reduced, the problems that the storage of the sensor overflows and the calculation time is long are solved, and the data monitoring real-time property of the system is improved.

The invention discloses a hardware acceleration system for an LSTM network model, and belongs to the technical field of deep learning hardware acceleration. The invention discloses a hardware acceleration system for a deep learning long short-term memory (LSTM) network model. The hardware acceleration system comprises a network reasoning calculation core and a network data storage core. The network reasoning calculation core is used as a calculation accelerator of an LSTM network model, calculation units are deployed according to the network model, and calculation acceleration of the calculation units such as convolution operation, matrix dot multiplication, matrix addition and an activation function is achieved; and the network data storage core serves as a data cache and interaction controller of an LSTM network model, deploys an on-chip cache unit according to the network model, and realizes a data interaction link between the computing core and an off-chip memory. According to the invention, the calculation parallelism of the LSTM network model is improved, the processing delay is reduced, the memory access time is shortened, and the memory access efficiency is improved.

The invention provides a method and a device for training a linear/logistic regression model. The method comprises the steps of: executing the following iterative processes till a preset condition ismet: acquiring a current prediction value of the linear/logistic regression model through employing secret sharing matrix addition based on a current sub-model of each training participant and a corresponding feature sample subset; determining a prediction difference value between the current prediction value and a corresponding mark value, and transmitting the prediction difference value to eachsecond training participant so as to update the respective current sub-model at each second training participant; and updating the current sub-model of the first training participant based on the current sub-model of the first training participant and the product of the corresponding feature sample subset and the determined prediction difference. When the iteration process is not finished, the updated current sub-model of each training participant is used as the current sub-model of the next iterative process. According to the method, the model training efficiency can be improved under the condition of ensuring the data security of each party.

The invention discloses a large-scale multi-operation floating point matrix calculation acceleration implementation method, which comprises the following steps: S1, receiving an external input signal and judging a matrix operation mode according to an operation type of a to-be-processed matrix: when the operation mode is matrix addition and matrix subtraction, turning to execute a step S3, and when the operation mode is matrix subtraction, turning to execute a step S4; when the operation mode is matrix multiplication, matrix-vector multiplication and matrix-scalar multiplication, turning to execute the step S2; s2, initializing an on-chip RAM (Random Access Memory) to be zero, and turning to execute a step S4; s3, the data source C is loaded into the on-chip RAM through the RAM channel, and the step S4 is executed; s4, pre-loading a part of the data stream A through an RAM channel, and loading the data stream A and the data stream B while calculating; s5, after calculation is completed, a calculation result is transmitted to the off-chip memory. The device is used for implementing the method. The method has the advantages of low storage requirement, high calculation efficiency, high reusability, wide application range and the like.

The invention relates to an arbitrary-order Kalman filtering system, comprising: a configurable memory array comprising a plurality of memory banks, wherein the memory banks are globally shared; a configurable computing array, including single-precision floating-point number multiplier, a single-precision floating-point number adder and a single-precision floating-point number divider; a matrix basic operation module, completing matrix addition, matrix subtraction, matrix transposition and matrix inversion; and the global configurable computing array is shared by time-sharing multiplexing; a state machine, according to the recurrence equation of the Kalman filter algorithm, the matrix basic operation module is called step by step, the intermediate result of the matrix basic operation module is stored in the memory array, and then the intermediate result is called according to the recurrence equation. The invention multiplexes the computing resource array and the storage resource arraythrough time-sharing and folding mode, thereby effectively reducing resources and area, and reducing power consumption. Multi-path parallel method is used to design the basic matrix operation, which can effectively improve the real-time performance and data processing ability of the system design.

A SLAM rear-end optimization-oriented Schur elimination accelerator comprises an algorithm control module used for completing maintenance of a control signal; a prefetching control module used for completing prefetching operation of input projection error data and Jacobian matrix data; a data cache access control module used for completing address generation and read-write requests of data cache, including address maintenance of different matrix data and advanced initiation of read requests; an operation module used for completing matrix multiplication, matrix inversion, matrix addition and subtraction and matrix and constant multiplication; a data rearrangement and control unit used for completing data recombination; an input and output caching unit used for completing advanced caching of prefetch input and data caching during output; and a data caching unit used for completing caching of intermediate data in the operation process. According to the accelerator, hardware acceleration is carried out on the Schur elimination process by proposing an FPGA accelerator scheme, so that a traditional embedded platform can execute the BA optimization process with higher performance.