78 results about "Sparse structure" patented technology

The invention discloses an edge computing-oriented lightweight large-model intelligent customer service deployment method, and relates to the technical field of artificial intelligence and edge computing. The method comprises the following steps: acquiring a pre-trained large language model, and pruning based on a lottery hypothesis to obtain a winner ticket sub-network; remapping the sparse structure into a V: N: M structured sparse format, and loading the V: N: M structured sparse format together with the weight and the mask to an on-chip memory; inserting a lateral branch outlet in the converter layer and carrying out token pruning; monitoring the confusion degree and the confidence degree, and dynamically recovering the pruned weight which does not exceed 5% when the load is high; and uploading the 8-bit quantization hidden state of the low-confidence or sensitive token to the large cloud model for collaborative reasoning, and combining the result with local reasoning output. According to the method, the reasoning delay and the communication overhead of the edge equipment are reduced, and the response efficiency and the deployment economy of the intelligent customer service system in a low-computing-power environment are improved.

The invention discloses an Internet of Things flowmeter data compression sensing transmission optimization method and system, particularly relates to the field of low-power-consumption wireless communication networks, and is used for solving the problem of compressed sensing signal distortion caused by sampling time sequence misalignment in an LPWAN bandwidth limited environment. Constructing a local phase reference by extracting a pipeline vibration fundamental frequency component, and receiving a cloud synchronization reference timestamp alignment signal; fundamental frequency and third harmonic components are separated to calculate arrival time difference, and a phase-locked loop is triggered to dynamically adjust the phase of a sampling clock based on a pipeline material threshold value; after the stability of the zero crossing point time variance is verified, compressed sensing sampling is performed on the flow signal by adopting a phase synchronous clock, and an optimized compressed measurement value is generated and transmitted to the cloud through the LPWAN; through closed-loop cooperation of physical characteristics and a communication clock, complete capture of a signal sparse structure is guaranteed under a low-bandwidth condition, and cloud reconstruction precision is significantly improved.

The invention discloses a three-dimensional object reconstruction method based on interactive segmentation prior and implicit surface reconstruction, which interactively extracts a semantic segmentation mask and a target image of a target object based on a pre-trained SAM segmentation model, and introduces an SfM algorithm to calculate camera attitude information and sparse structure information of a multi-view scene graph. And inputting the target image, the segmentation mask and the camera attitude information into a grid encoder for training to obtain a rough polygon grid and a texture feature, reconstructing the rough polygon grid and a texture map of the target object, and outputting a three-dimensional reconstruction result of the target object after refining the rough polygon grid and the texture map. According to the method, the interactive segmentation method is constructed through the pre-training large model, so that invalid occupation of a background region is avoided. The implicit surface reconstruction algorithm based on the neural radiation field is used for constructing the fine grid and texture information of the object to be reconstructed, three-dimensional reconstruction of the target object can be rapidly and accurately completed in a complex environment, and the method has good value and popularization prospects.

The invention provides a personalized federated learning method based on dynamic sparse structure adjustment, which is oriented to a non-independent identically distributed data scene and solves the problems of high communication and calculation overhead and low convergence speed in federated learning. According to the method, model rarefaction and personalized modeling are combined, and an initial sparse structure is constructed according to the parameter scale of each network layer under the constraint of a fixed overall sparse rate in an initialization stage; in the training process, a dynamic sparse updating mechanism based on gradient information is introduced, and adaptive pruning and regeneration are carried out on model connection so as to continuously optimize the model structure. In order to avoid falling into a suboptimal sparse mode, the pruning and connection regeneration proportion is dynamically adjusted, and the structure searching capability is enhanced. In the model aggregation stage, a mask perception parameter aggregation strategy is provided, only activation parameters are aggregated, and aggregation deviation is reduced. Experimental results show that the method significantly reduces the communication and calculation overhead while ensuring the model precision, accelerates the model convergence speed, and is suitable for a federal learning scene with limited resources.

The invention relates to the technical field of wireless communication, and discloses an adaptive LDPC code coding method, system and device based on evaporation waveguide field intensity distribution, and a medium, and the method comprises the steps: obtaining the evaporation waveguide field intensity distribution in real time, and generating a field intensity distribution diagram; calculating a channel state parameter, and dynamically selecting a low-density parity check code rate according to the channel state parameter information; dynamic low-density parity check coding is adopted, and a sparse structure is adjusted through a check matrix, so that real-time switching of multi-code-rate low-density parity check codes is carried out; and a receiving end carries out decoding by adopting an NW-RBP decoding algorithm, defines a node updating rule, and dynamically schedules a node updating sequence through a residual error. According to the method, the bit error rate under a complex channel is reduced, and meanwhile, the reliable transmission efficiency of system data is improved.

This invention discloses a monostatic symmetric nested array MIMO system and a near-field positioning method, belonging to the field of monostatic MIMO near-field positioning technology. This invention can effectively improve the positioning performance during near-field estimation. The monostatic symmetric nested array MIMO system proposed in this invention has a symmetrically distributed transmit and receive arrays. Compared to other symmetric MIMO systems, its differential array is a positionally continuous array with more degrees of freedom. With the same physical array aperture, fewer physical array elements are actually used, reducing costs. The near-field positioning method proposed in this invention calculates the fourth-order statistics of the signal and quantizes them to obtain a virtual differential array of the differential array, achieving accurate near-field positioning of sparse MIMO structure arrays. Furthermore, it can be extended to more sparse MIMO symmetric systems.

The application discloses a viscoelastic-plastic simulation method based on a Bingham model, and comprises the following steps: obtaining a geometric model, material constitutive parameters and boundary conditions of a target geological body; constructing a complementary energy functional containing a viscous dissipation potential, and converting a Bingham viscoelastic-plastic constitutive relation into a mathematical programming problem constrained by a Drucker-Prager yield criterion; discretizing the geometric model space and introducing a time step, and discretizing the mathematical programming problem into a second-order cone complementary problem with a block-sparse structure, which contains a normalized second-order cone constraint equivalent to the Drucker-Prager criterion; performing smoothing treatment and solving by block elimination using the block-sparse structure to obtain a displacement field and a stress field. The application solves the problem that the existing convex programming solving framework cannot handle viscoelastic-plastic effects by introducing the viscous dissipation potential into the complementary energy functional.

The invention discloses a sparse time sequence Bayesian network construction method and system based on power distribution network topological constraints, and the method specifically comprises the steps: constructing a power distribution network topological graph, and calculating an adjacent matrix between nodes and a k-hop neighborhood matrix Nk; based on the power distribution network topology distance information and the matrix Nk, generating a hard constraint rule, and constructing a topology dependence mask matrix M; calculating an electrical influence range of the fault point on surrounding nodes to obtain an electrical influence matrix E; if the electrical influence coefficient of one node on the other node is smaller than a set value, deleting the corresponding dependent edge; introducing a data source reliability matrix R, and performing hard deletion or soft weakening on edges with reliability lower than a threshold value; combining the matrixes M, E and R to synthesize a sparse structure matrix S; and taking the matrix S as a space skeleton, adding a time dimension autoregression edge, and constructing a complete sparse time sequence Bayesian network. According to the invention, by guiding the rarefaction of the network structure, the number of network edges and the number of parameters are effectively reduced, and the trainability and reasoning efficiency of the model are improved.

The application discloses a multi-target detection method in a complex electromagnetic interference environment for a pulse Doppler radar, a joint reverse diffusion process is introduced to jointly update interference samples and amplitude samples, and thus the problems of target signals being covered or submerged by interference signals, rising of a detection false alarm rate, and difficulty in effectively separating interference from targets can be overcome; a denoising score matching criterion is introduced to train a neural network of an interference score function, so as to solve the problems of high interference intensity and complex structure in multi-target detection in a complex electromagnetic interference environment; and a sparse Bayesian learning method is introduced to adaptively model a sparse structure of multi-target echo signals. The application further provides a multi-target detection device in a complex electromagnetic interference environment. The method provided by the application can realize modeling and suppression of multiple types of interference, and simultaneously improve the detection probability of multiple low observable targets in the distance and Doppler dimension, the interference suppression capability, and the system robustness in a complex electromagnetic environment.

The invention relates to the technical field of power line point cloud segmentation, and discloses a power line three-dimensional point cloud segmentation method, system and device and a storage medium, and the method comprises the steps: obtaining original point cloud data of a power line, inputting the original point cloud data into a multi-scale spatial feature extraction model, and obtaining spatial features; according to the spatial features, performing foreground probability prediction on the original point cloud data to obtain a foreground point set, performing multi-radius sphere neighborhood aggregation on the foreground point set to obtain local features, and performing super-point aggregation on the foreground point set to obtain global features; fusing the local features and the global features according to a feature fusion model to obtain a fused query vector; and performing classification probability prediction on the fusion query vector to obtain a point cloud segmentation result of the power line. According to the method, on the basis of keeping low calculation cost and low model complexity, the semantic recognition capability of the sparse structure target of the power line can be effectively improved, and a high-precision and high-efficiency semantic segmentation effect is achieved.

This invention provides a flexible production line control method and system based on a large vertical domain model. The method includes: constructing a pruning index combining temporal stability and spatial similarity for the original large vertical domain model to evaluate the redundancy of the activation tensors in the intermediate layers, and performing structured pruning accordingly to obtain a sparse structure model; performing mixed-precision quantization on the pruned model: allocating quantization bits according to the sensitivity of each network layer weight to task loss, and calculating scaling factors and zeros based on the range of non-zero weights to represent sparse weights; calculating the output deviation between the quantized pruned model and the original model under the same input, and training a feedforward network with the former's output as input to fit and compensate for the output deviation; combining the quantized pruned model and the feedforward network to form a composite control model; after deployment, adding the output command of the quantized pruned model to the compensation vector during inference to generate control commands.

This invention discloses a remote, non-contact heart rate estimation method, system, and device based on sparse structure representation, belonging to the field of physiological sign research applications. It addresses the problem of inaccurate heart rate signal detection in existing methods, including: Step 1: Inputting a face video and segmenting the face into multiple sub-regions; Step 2: Constructing the average RGB pulse signal for each sub-region, performing motion compensation, and calculating the chromaticity signal S using linear combinations of different RGB channels, and calculating the signal-to-noise ratio; Step 3: Solving for the coefficient matrix using a structure dictionary composed of cosine bases of different frequencies and wavelet bases of different scales, and reconstructing the heart rate signal using the structure dictionary and coefficient matrix; Step 4: Averaging the reconstructed sub-region heart rate signals and performing power spectral analysis to calculate the heart rate HR. Finding the optimal sparse representation based on the dictionary, reconstructing the heart rate signal using the dictionary and sparse coding, and representing the most realistic heart rate signal with as few dictionary atoms as possible helps improve the accuracy of the algorithm.

The invention discloses a user time delay estimation method of an IRS-assisted multicarrier ISAC system. In the method, a plurality of users send pilot signals, and each IRS reflects the user signals and transmits the user signals to a remote radio access unit (RRH); the RRH uploads the received signal to a centralized baseband processing unit pool (BBU) through a forward link; as the parameters of the link from the IRS to the RRH can be acquired in advance, the method carries out time delay estimation on the link from the user to the IRS; the BBU constructs an observation matrix of a pilot signal on a joint dimension of a plurality of subcarriers and an IRS unit and utilizes a multi-snapshot effect provided by a plurality of receiving antennas to realize high-precision time delay estimation from a user to an IRS link; according to the method, sparse structures of a frequency domain and a space domain are fully utilized, and the time delay resolution and the estimation precision are effectively improved, so that the user positioning capability in a complex environment is enhanced.

The invention relates to the technical field of phased-array antenna arrays, in particular to a random uniform sparse algorithm for a phased-array antenna. Comprising the following steps: carrying out minimum unit sub-array grid division on working characteristics of a target array antenna, and carrying out uniform grid partitioning; obtaining possible sparse structures of all basic unit antennas in the antenna sub-array by taking the sub-array as a basic unit; according to preset antenna performance parameters and screening criteria, the sparse structures are combined step by step, and all possible sparse mechanisms of the k-th-level antenna sub-array are obtained; taking the k-level sub-array as a unit, and according to a screening criterion, sequentially and randomly combining to form an antenna whole array; screening the randomly combined antenna whole array according to a screening criterion; and performing array pattern traversal simulation on the screened whole array, and screening an optimal sparse structure. The invention provides a random uniform sparse algorithm for a phased-array antenna, which reduces the sidelobe level, improves the generation speed and economy, and ensures the beam robustness.

The invention provides a structured sparse compensation method and system based on residual energy perception. The method comprises the following steps: calculating a quantized residual matrix of a weight in a quantized backbone network; constructing a compensation branch with a regular sparse structure based on the energy distribution characteristics of the quantized residual matrix; and integrating the compensation branch and the quantization backbone network in a parallel connection mode, and realizing compensation of an output error introduced by weight quantization by superposing output of the compensation branch and the quantization backbone network in a reasoning process. The structured sparse compensation branch has a high-rank expression capability, does not depend on global low-rank hypothesis, constructs a regularly arranged sparse substructure based on residual energy distribution, effectively approaches a high-frequency error, remarkably improves the detail recovery capability, and solves the problems of blurring and distortion of a generated image.

The application provides a high-dimensional time-varying signal space completion method and device based on a sparse attention mechanism, equipment and a storage medium, and relates to the technical field of high-dimensional information. The method divides original data into multiple sparse structured sub-blocks through adaptive blocking; performs low-rank-sparse structure decomposition on each sub-block, extracts dominant low-rank structure features and sparse disturbance features, and fuses to obtain sub-block-level sparse structure representation; a sparse attention mechanism with triple constraints of time sequence, missing mode and structure recoverability is constructed, a dependency relationship is established only between key block pairs, and a space-missing joint dependency representation is output; through multiple rounds of iterative feedback, the updated matrix is re-input into the foregoing operation until the error converges, and a completely completed high-dimensional matrix is output. The application can effectively suppress redundant correlation, enhance dependency modeling of the completable area, and significantly improve the completion accuracy and robustness of high-dimensional time-varying data in a complex environment.

Methods, systems, and computer program products for sparse quantization of a model are disclosed. The method comprises the steps of generating a fine tuning data set based on data associated with a target task of a pre-training model; executing multiple rounds of progressive sparse processing on the pre-training model; and executing multiple rounds of progressive quantization processing on the model after multiple rounds of progressive sparse processing. Each round of progressive sparse processing comprises the following steps: executing sparse operation once; and performing layered fine tuning on the model obtained after the sparse operation by using the fine tuning data set, and applying sparse structure constraint in the layered fine tuning process. Each round of progressive quantization processing comprises the following steps: executing a quantization operation once; and performing layered fine tuning on the model obtained after the quantization operation by using a fine tuning data set, and applying both sparse structure constraint and quantization structure constraint in the layered fine tuning process. According to the method, the performance of the model on the target task is effectively maintained while the storage occupation and the calculation complexity of the model are remarkably reduced.

A low-quality multi-view news data anchor graph regular division method based on diffusion completion belongs to the field of data division in low-quality multi-view news data, and comprises the following steps: firstly, inputting low-quality news data of each view and a corresponding similarity relation matrix into a heterogeneous relation convolutional network; to obtain a low-dimensional embedded representation of each view. Then, the low-dimensional embedded representation uses forward noise adding and reverse noise reduction processes of a conditional diffusion model to obtain predicted noise, and missing samples are complemented through the predicted noise; and then, an anchor point diagram is constructed by using the similarity between the complemented embedded sample and the anchor points, soft clustering distribution is obtained for the constructed anchor point diagram through an orthogonal normalization layer, and discriminative feature representation is obtained through anchor graph regularization constraint. And finally, soft clustering distribution is constrained by using a tensor Schatten p-norm so as to fully mine complementarity information and a sparse structure between the views.

The invention relates to a tensor mixed mode code generation method and system, computer equipment, a storage medium and a computer program product. The method comprises the following steps: receiving and analyzing a sparse tensor, a dense tensor, a tensor expression and hardware parameters input by a user; based on the intermediate representation, converting a sparse tensor into a compression format suitable for calculation, identifying an operator type and binding a corresponding iterator, performing block division on a sparse matrix, performing data packaging on a dense tensor, and generating a vectorization calculation kernel; generating a sparse iterator code, a vectorized kernel code and a parallel scheduling code; the technical problem that in the prior art, traditional sparse calculation optimization often adopts a fixed mode and cannot adapt to different types of sparse structures (such as irregular sparseness and dynamic sparseness), and consequently performance is poor is solved. The technical effects that adaptive vectorized and parallelized codes are automatically generated according to the sparse mode and hardware characteristics of the input data, and the calculation efficiency is improved are achieved.

The invention discloses an attention mechanism dynamic sparseness and quantification method, system, equipment and medium, which are corresponding schemes, and the scheme comprises the following steps: rearranging an input vector sequence through a block granularity clustering algorithm to obtain a vector sequence cluster of block granularity aggregating semantic information; the clustering center of each cluster is used as a representative element to calculate the attention score of each cluster, clustering blocks with high importance are screened out based on the attention scores to obtain sparse masks of block granularity, and vector sequence blocks needing to be calculated are selectively read in an attention calculation kernel according to the masks; and then performing block-by-block data smoothing operation on the read-in vector sequence blocks, and performing symmetric quantization. According to the scheme, under the condition that the model precision is guaranteed, a hardware-friendly sparse structure which facilitates uniform division of tasks and avoids waste of calculation and bandwidth resources is obtained, so that the calculation and bandwidth resources are saved, the calculation amount of an attention mechanism module is reduced, and then the reasoning efficiency of the video generation model is improved.

The application discloses a kind of dual-domain constraint hyperspectral image reconstruction methods based on deep learning, it is related to the technical field of artificial intelligence and computational imaging.The application simultaneously imposes constraint in spectral reconstruction domain and sparse coefficient domain, while ensuring that its inherent sparse structure conforms to physical prior, the complement and verification of dual-domain information significantly improve the fidelity of reconstruction result, reduce artifact and noise;Using deep neural network to realize reconstruction, only once forward propagation is needed for new compression measurement value to output reconstruction result, which greatly reduces the computational complexity, improves the reconstruction speed, and has the potential for real-time processing;The sparsity physical prior is explicitly integrated into the network learning goal, which can effectively resist noise interference and reduce the influence of noise on the reconstruction result;Through back propagation algorithm, the dual-head deep neural network is optimized and trained end-to-end, and the best mapping relationship is automatically learned, which greatly reduces the operation difficulty and application threshold of the method.

The application discloses a reactor sub-channel comprehensive acceleration calculation method and device and electronic equipment, and relates to the technical field of reactor thermal hydraulic numerical simulation, wherein the method comprises the following steps: abstracting a sub-channel gap topology into an undirected graph and performing edge coloring grouping, and constructing and caching a CSR sparse structure and a write position index of each gap based on the topology; then, serial scheduling is performed according to the coloring grouping, a plurality of threads in the group accumulate the coupling coefficients to a global numerical array according to the pre-calculated position, and the coefficient matrix assembly is completed; finally, the asymmetric sparse linear equation set is solved by AMG pretreatment combined with FGMRES iteration, the pressure gradient solution is obtained, and the current step update is completed. Through edge coloring and sparse storage pre-calculation coupling, the application realizes lock-free parallel assembly and solves the problem of data race in multi-core concurrent writing.

The invention relates to the technical field of matrix multiplication, in particular to a sparse matrix-dense matrix multiplication acceleration system and method. The system comprises a sparse structure analyzer, a structure weighting scheduler, a plurality of PE rows and a runtime feedback controller. The sparse structure analyzer extracts features of each row of the sparse matrix and calculates a structure cost score; the structure weighting scheduler is combined with the PE row real-time load to distribute the PE row with the minimum comprehensive cost for the matrix row; executing scalar-vector multiply-accumulate by the PE array in parallel; and during operation, the feedback controller generates a blacklist bitmap and a splitting signal through performance monitoring, and dynamically adjusts task allocation. Parallel data reading and load balancing processing and calculation are carried out through a three-stage assembly line. According to the method, the calculation efficiency and the energy efficiency ratio of sparse-dense matrix multiplication are effectively improved, and then the data processing speed of the CPU / GPU is improved.

The invention relates to the technical field of artificial intelligence, can be applied to business scenes of financial science and technology, medical health and the like, and discloses a data sequence generation method, device, equipment and medium, comprising: acquiring initial data, extracting depth features, generating a depth feature map, fusing the initial data and the depth feature map to obtain a joint feature representation, generating an initial data unit based on the joint feature representation, setting a current data unit and a preorder data unit thereof, generating a motion complexity feature, generating a dynamic threshold mask, executing sparse attention processing by applying the dynamic threshold mask, generating a next data unit, updating the preorder data unit and the current data unit, and circulating the process. And performing time sequence processing by combining the initial data unit and the plurality of iterative data units to form a data sequence. According to the method, the attention sparse structure is dynamically controlled by combining the depth feature and the motion complexity feature, and the calculation efficiency and the time sequence continuity in the data sequence generation process are improved.

The invention provides a CNN model parameter accelerated extraction method and system in a black box hard tag environment, and belongs to the technical field of artificial intelligence. Comprising the following steps: reconstructing a convolution kernel of a convolutional neural network into a matrix representation with a BTTB sparse structure based on structure information of the convolutional neural network; a double-point set is searched and determined through query and decision boundary geometrical shape analysis; and executing a clustering algorithm taking a convolution kernel as a center in the calculation unit to generate a double-point cluster, reconstructing a critical hyperplane by combining subspaces corresponding to each double point in the cluster, and extracting a normal vector of the critical hyperplane as unsigned estimation so as to determine the weight of a convolution layer. According to the method, through BTTB matrix reconstruction, clustering with the convolution kernel as the center and multi-stage optimization of covariance matrix acceleration, the memory overhead in the calculation process can be remarkably reduced, and the CNN model parameter extraction efficiency of a GPU processor is effectively improved.

The present application discloses methods, systems and computer program products for sparse quantization of a model. The method comprises: generating a fine-tuning dataset based on data associated with a target task of a pre-trained model; performing multiple rounds of progressive sparsification on the pre-trained model; and performing multiple rounds of progressive quantization on the model after the multiple rounds of progressive sparsification. Each round of progressive sparsification comprises: performing a sparse operation; performing hierarchical fine-tuning on the model resulting from the sparse operation using the fine-tuning dataset, and imposing a sparsity structure constraint during the hierarchical fine-tuning. Each round of progressive quantization comprises: performing a quantization operation; performing hierarchical fine-tuning on the model resulting from the quantization operation using the fine-tuning dataset, and imposing both a sparsity structure constraint and a quantization structure constraint during the hierarchical fine-tuning. The method achieves effective maintenance of the performance of the model on the target task while significantly reducing the model storage footprint and computational complexity.

The invention discloses a multi-target detection method for a pulse Doppler radar in a complex electromagnetic interference environment, and the method introduces a joint back diffusion process to carry out the joint updating of an interference sample and an amplitude sample, and can solve problems that a target signal is covered or submerged by an interference signal, the detection false alarm rate is increased, and the interference and a target are difficult to be effectively separated. A denoising score matching criterion is introduced to train a neural network of an interference score function so as to solve the problems of high interference intensity and complex structure during multi-target detection in a complex electromagnetic interference environment. A sparse Bayesian learning method is introduced to adaptively model a sparse structure of a multi-target echo signal. The invention further provides a multi-target detection device in the complex electromagnetic interference environment. According to the method provided by the invention, modeling and suppression of various types of interference can be realized, and meanwhile, the detection probability and the interference suppression capability of a plurality of low observable targets in distance and Doppler dimensions and the system robustness in a complex electromagnetic environment are improved.

To provide an anomaly detection apparatus, a program, and an anomaly detection method that enable improvement in accuracy of estimating causes.SOLUTION: An anomaly detection apparatus (10) includes: an acquisition unit (11) which acquires detection value data including detection values of sensors provided in equipment; an anomaly detection unit (12) which determines whether there is a sign of failure in the equipment, using a learning model generated with normal data out of acquired detection value data; and a cause estimation unit (13) which estimates a cause of the failure determined to show the sign when the sign of failure is determined. The learning model is a deep learning model using a sparse graph structure. The cause estimation unit executes cause estimation that rearranges the sensors so that a sensor estimated to cause the anomaly may be ranked at the top, using the weight of edges of the graph structure.SELECTED DRAWING: Figure 1