54 results about "Hadamard product" patented technology

The invention relates to the technical field of electrical digital data processing, and discloses a train ice melting simulation optimization method of an electromagnetic thermal coupling model fused with a deep learning method, which comprises the following steps of: inputting a geometric representation tensor and a physical working condition parameter vector containing an electromagnetic excitation frequency and a reference environment temperature into a feature mapping neural network; outputting a dual-channel space source item tensor containing basic heat source power density and a heat source to temperature change sensitivity distribution matrix through nonlinear convolution operation; constructing a heat conduction discrete numerical value evolution operator configured with an active item linear correction interface; time stepping operation is executed according to the heat conduction time scale, a basic heat source is corrected in real time through the Hadamard product of a sensitivity distribution matrix and temperature deviation, and an operator is substituted for solution. On the premise that electromagnetic-thermal nonlinear coupling characteristics are reserved, decoupling of the time scale is achieved, and the calculation efficiency in high-frequency physical field simulation is effectively improved.

The invention belongs to the field of sensors and signal processing, and particularly relates to a wavelet kernel scale sensitivity oriented abrasive particle induced voltage signal noise reduction method, which comprises the following steps: acquiring an abrasive particle induced voltage signal, and carrying out harmonic elimination on the abrasive particle induced voltage signal to obtain a preprocessed signal; constructing a wavelet kernel function, and calculating the wavelet kernel function and the preprocessed signal to obtain a kernel scale guiding spectrum; constructing a sparse joint noise reduction model based on the kernel scale guiding spectrum; processing the sparse joint noise reduction model to obtain a convex optimization objective function; solving the convex optimization objective function by combining a self-adaptive step gradient descent method and a self-adaptive iterative shrinkage threshold method to obtain a weight vector representing the distribution of the abrasive particle characteristic signals; carrying out binarization processing on the weight vector representing the distribution of the abrasive particle characteristic signals to obtain a characteristic indication vector; carrying out Hadamard product on the feature indication vector and the preprocessed signal, and then carrying out low-pass filtering to obtain a noise reduction signal; according to the method, the abrasive particle characteristic signals can be self-adaptively subjected to non-destructive enhancement and noise reduction processing in a strong interference environment.

The invention discloses a feature enhancement-based infrared and visible light image fusion method. The method comprises the following steps of: preprocessing data; performing shallow feature extraction on the two types of modal images by using a double-branch encoder, and performing decomposition to obtain low-frequency and high-frequency features; constraining the low-frequency and high-frequency features; the low-frequency features are superposed and recombined, and cross-modal difference mining is carried out after the high-frequency features are superposed; refining the fusion features, performing Hadamard product operation on the fusion features and corresponding modal fusion features to obtain self-refining features, and performing cross-modal feature combination reconstruction to generate complementary refining features; and reconstructing to obtain a fused image. According to the embodiment of the invention, preliminary integration of cross-modal information is realized through feature superposition and difference mining, and core features of each modal are reserved; through refining and complementary enhancement processing, collaborative expression of cross-modal structure features and detail information can be effectively enhanced, it is ensured that a fused image has stable performance in the aspects of texture fidelity and saliency target presentation, network conciseness is kept, and the quality of a fusion result is high.

The invention relates to the technical field of wireless communication signal big data analysis, and discloses a LoRa unmanned aerial vehicle communication signal-oriented identification and analysis method, which comprises the steps of obtaining radio signal discrete data in a monitoring area, executing short-time Fourier transform to generate a time-frequency two-dimensional spectrogram, and identifying and intercepting a LoRa signal noisy IQ time sequence by using a convolutional neural network; mapping the sequence into a high-dimensional feature tensor, inputting the high-dimensional feature tensor into a slope statistical attention generative adversarial network, calculating a first-order differential tensor by using a generator, carrying out kernel convolution on the first-order differential tensor and preset linearity statistics to generate an attention mask, and reconstructing a de-noised IQ time sequence by using a bidirectional long-short-term memory network after Hadamard product gating; according to the method, the tensor gating operator based on the physical statistical law is embedded in the neural network, so that the common problem that physical semantic drift is easily generated by a generative model at an extremely low signal-to-noise ratio is solved.

The invention relates to the technical field of image semantic segmentation, in particular to an intelligent landslide identification method based on multi-source remote sensing image fusion. Comprising the following steps: calculating a gradient matrix and a water flow convergence index based on a digital elevation model, searching and determining a physical prior interval of cohesive force and an internal friction angle based on geological lithology data, and constructing an input tensor; extracting a surface semantic feature map of the optical image by using visual perception branches through a double-branch coupling network; based on a physical prior interval, mapping branches by using physical parameters, and carrying out nonlinear mapping to obtain a limited rock-soil mechanical parameter diagram; and calling a micro infinite slope stability layer, executing forward derivable calculation based on a limit equilibrium equation, generating a slope stability coefficient field, constructing physical attention gating, executing Hadamard product operation, and generating a semantic feature tensor. Through deep coupling integration and collaborative optimization of the visual perception branch and the physical parameter mapping branch, the true effectiveness of the obtained position data is ensured.

The invention relates to an aerospace target radar image data generation method based on low illumination enhancement correction. The method comprises the following steps: designing a content delivery decomposition network, decomposing low light characteristics into illumination-independent reflection components and adaptive illumination components through Hadamard product constraint in a submerged space, proposing a learnable intensity compression function, and dynamically generating a submerged space mask through gradient consistency constraint and sparse regularization; embedding a mask into a reverse denoising process to realize directional repair of a degraded region, and combining the generation capability of a diffusion model with the guidance of physical prior; a diffusion path is reconstructed based on a bidirectional diffusion principle, physical consistency of a generation process is constrained through end point binding, a diffusion step length is adaptively adjusted, a cyclic implicit iteration mechanism is introduced, a generation result is gradually refined through a cyclic neural network module, and accurate optical-ISAR image translation can be realized.

The invention discloses a dynamic nerve radiation field real-time three-dimensional reconstruction method and device based on window attention and gradient equilibrium, and the method comprises the steps: carrying out the stratified sampling of rays, and obtaining sampling points; projecting the sampling points to six groups of two-dimensional feature planes; querying corresponding feature vectors from the six groups of feature planes through bilinear interpolation, and generating spatio-temporal features through Hadamard product fusion; inputting the fused spatial-temporal characteristics into a decoding network of a Swin Transform, and outputting the color and the volume density of a corresponding sampling point; and carrying out volume rendering integration along the ray of the camera to generate a composite image, and optimizing and updating the model according to a multi-target loss function between the composite image and a real image. The invention aims to solve the problems of large memory consumption, poor rendering quality, long training time and poor geometric consistency in the prior art, and realizes real-time three-dimensional reconstruction of a dynamic scene by using six-plane grid representation, Swin Transform feature generation and a gradient equilibrium mechanism.

The application relates to a skin lesion area rapid detection and segmentation method, which comprises the following steps: given an input X element R C×H×W , which becomes Y element R C×HW after remodeling operation, and 4 times expanded attention map Att element R 4C×HW is obtained by calculating the relevance of the query vector and the memory unit; given an input feature map X and a randomly initialized learnable tensor P; the size of the tensor P is adjusted by using bilinear interpolation to match the size of the X, and a depth separable convolution is used on the P; the feature map is uniformly cut into four parts X1, X2, X3 and X4 along the channel dimension. The application improves and introduces Hadamard product attention module to extract multi-angle pathological features of different shaft groupings of the feature map, fuses multi-scale context information, aggregates cross-dimension information, improves the representation ability of the model, designs a new boundary loss function, and puts the boundary information into the model learning process, encourages the model to pay attention to the boundary details, is excellent in various segmentation performance indexes, and can accurately extract a lesion area.

The application discloses a kind of proton exchange membrane fuel cell parameter identification method, belong to fuel cell technical field, comprising: definition is the parameter vector to be identified, with the error between the actual measured terminal voltage corresponding to reference true value working condition data set of predicted voltage as optimization goal, establish objective function, based on Hadamard product introduces multi-dimensional independent random vector to make follower approach adjacent better solution and current global optimal solution, introduce random disturbance term to make leader carry out big step search update, obtain global updated individual position;If the objective function value corresponding to global updated individual position has not occurred decline, then introduce inertia weight with dynamic exponential decay along with iteration number, in the neighborhood of current global updated individual position, carry out local depth search and position update to all individuals, obtain local updated individual position, retain the solution corresponding to smaller objective function value.The application solves the problem that existing technology extracts semi-empirical parameter is not accurate.

The invention relates to an interpretable CNN (Convolutional Neural Network) training method and device based on Bernoulli sampling and a medium, and the method comprises the following steps: inputting a picture into a CNN, and obtaining a response feature map of a filter; performing Bernoulli sampling on the response feature map to obtain a binary distribution matrix; calculating a filter average weight matrix of each picture category according to the binarization distribution matrix, and calculating the sum of pairwise differences; calculating the Hadamard product of the distribution vector of the binary distribution matrix and the response feature map to obtain a mask feature map; respectively inputting the response feature map and the mask feature map into a CNN full connection layer to respectively obtain classification prediction probability vectors, and respectively calculating cross entropy loss with a real label; and according to the sum of the pairwise differences and the cross entropy loss, using a stochastic gradient descent method to realize network training, and obtaining an interpretable CNN for image classification. Compared with the prior art, the method has the advantages of high adaptability, high interpretability and the like.

The invention discloses a sound signal periodic feature extraction method, a network model training method, a storage medium and equipment, and belongs to the technical field of sound event detection. The objective of the invention is to solve the problems of high sensing difficulty and poor decoupling effect of overlapped acoustic events in the current acoustic detection process. The method comprises the following steps: for a sound signal i, mapping the sound signal i to a low-dimensional space through two different linear layers to obtain p and g, and respectively carrying out expansion convolution operation on p and g to obtain pconv and gconv; for p and g, feature coding is carried out based on a Fourier basis function and a gating mechanism to obtain Fourier features, for pconv and gconv, Fourier features are obtained in the same mode, and Hadamard product is carried out on the pconv and the gconv to obtain representation of periodic features. And in the training process of the corresponding model, performing reconstruction error on the sum and the original signal i, respectively calculating two norms of the sum, and adding the two obtained two norms to obtain a Fourier series regular term for training the model.

This invention discloses a method and system for analyzing and outputting TCM acupuncture cases based on machine learning, belonging to the field of medical image processing technology. The method acquires facial image sequences from two pathological cycles of the subject, extracts spatiotemporal feature maps, and obtains local receptive field feature matrix pairs through sliding windowing. It calculates the cross-cycle joint entropy gradient and its matrix, and generates a two-dimensional deformation vector using optical flow. Using the coordinates of the first cycle as a reference, the deformation vector is set to zero when the joint entropy gradient does not exceed the rigidity threshold; otherwise, the joint entropy gradient matrix is ​​used as a weight matrix through a sigmoid activation function, and a Hadamard product is performed with the deformation vector to obtain a denoised deformation vector. The denoised vectors are superimposed to obtain corrected pixel coordinates, which are mapped to the target localization coordinate matrix, and the acupuncture case analysis results are output. This scheme achieves precise spatial decoupling of local tissue evolution, eliminates the contamination of rigid regions by global deformation, and significantly improves the robustness and accuracy of cross-cycle acupuncture target localization.

The present application belongs to the technical field of computer vision processing, and particularly relates to a kind of method and system for generating adversarial samples based on deep neural network, first, the original image in sample data is converted into saliency map;The saliency map is used to circle the salient region in the original image of sample data for adding disturbance, and the saliency mask is obtained by binary processing of saliency image pixel value;The original image in sample data is input into image classification model, and the gradient information in the reverse transmission process of Nadam optimization algorithm and convolutional neural network is used to iteratively generate global disturbance adversarial sample;The difference between adversarial sample and original image is obtained, and global adversarial noise is obtained;The Hadamard product of global adversarial noise and saliency mask is used to obtain adversarial noise in salient region, and the final output salient region adversarial sample is obtained by combining adversarial noise and original image.The present application can improve the quality of generated samples, facilitate testing and improve the security and robustness of image classification model.

The application relates to the technical field of storage and transportation equipment maintenance support, in particular to a storage and transportation system equipment maintenance support capability evaluation method, system, equipment and medium, which comprises the following steps: constructing an evaluation index system of the storage and transportation system equipment maintenance support capability; defining a homologous index group and constructing a pairwise comparison matrix for each homologous index group; constructing a judgment influence factor matrix for the pairwise comparison matrix according to the storage and transportation scene of an evaluation object, performing Hadamard product operation to obtain a scene adaptation judgment matrix, and then performing analytic hierarchy process weight calculation to obtain a final index scene weight matrix; calculating the initial evaluation value of each final index, calculating the comprehensive evaluation value of the storage and transportation system maintenance support capability according to the initial evaluation value and the final index scene weight matrix, and further evaluating the equipment maintenance support capability of the evaluation object. The application can make the weights of various indexes match the actual demand in real time, overcome the unexplainable defects, and provide transparent basis for maintenance decision.

This application provides a method, device, equipment, and storage medium for small-sample ultra-short-term wind power prediction of newly built wind farms, relating to the field of wind power prediction technology. The method includes: calculating the distribution similarity coefficient based on two-dimensional wind speed-power kernel density estimation and KL divergence to select the optimal base station; constructing an LSTM model with embedded dynamic similarity alignment gating, fusing temporal features with the gating output using Hadamard product to alleviate long-term prediction decay; employing an inner and outer loop update strategy of model-independent meta-learning, combined with elastic weight solidification regularization to suppress catastrophic forgetting; and achieving 1-4 hour ultra-short-term power prediction for newly built wind farms through base station pre-training, meta-task set training, and small-sample adaptive updating. At the 1-4 hour prediction scale, NRMSE and NMAE are reduced by 4.19% and 4.73%, respectively, and R... 2 It improves by an average of 5.36% and has excellent generalization performance across different sites.

The invention relates to a machining uncertainty calculation method, in particular to a machining uncertainty calculation method of a technological process composed of multiple machining procedures. The method is characterized in that the process processing uncertainty is the synthesis of the regeneration processing uncertainty and the genetic processing uncertainty. The machining uncertainty of each process is a machining uncertainty vector composed of one or more machining uncertainty components. The genetic processing uncertainty is a Hadamard product of the processing uncertainty of the previous process and a genetic coefficient. The regeneration processing uncertainty is the combination of the uncertainty of all error sources and the uncertainty of the Hadamard product of the corresponding sensitivity. The uncertainty of the error source is the uncertainty synthesis of all the uncertainty components of the error source and the Hadamard product of the corresponding sensitivity. The method has the advantages that before the process test is implemented, whether the workpiece precision level which can be achieved by the given process meets the design drawing requirement or not is quantitatively calculated and predicted, the process reasonability and effectiveness are evaluated, the process test risk and cost are reduced, and the first workpiece trial-manufacturing success rate is increased.

This invention discloses a method for inverse kinematics modeling of a polyhedral mobile robot, comprising: S1, constructing a forward kinematics model of a polyhedral open-chain robot, defining a kinematic space with the edge lengths of a virtual polyhedron as the core, and establishing a mapping relationship between joint angles and the edge lengths of the virtual polyhedron; S2, constructing an optimization objective function based on the Hadamard product based on the error between the target edge length set and the actual edge length set of the virtual polyhedron; S3, generating an initial set of iteration points covering all quasi-convex regions in the joint space of the polyhedral open-chain robot; S4, solving the gradient of the objective function and iteratively updating the joint angles; S5, setting a convergence threshold and termination condition, performing convergence judgment on the iteration process, and selecting the optimal solution for inverse kinematics; This invention achieves accurate modeling of the inverse kinematics of a tetrahedral mobile robot by constructing an error optimization model based on the edge lengths of a virtual polyhedron and combining Monte Carlo sampling initialization and gradient descent iterative solution.

The embodiment of the invention provides a method for executing polynomial multiplication and acceleration hardware. The acceleration hardware comprises an input transformation module, a convolution kernel transformation module, a modular multiplication module, an output transformation module and a partial product processing module. The method comprises the following steps: respectively segmenting a first polynomial coefficient vector and a second polynomial coefficient vector to obtain T first sub-vectors and K second sub-vectors; the input transformation module and the convolution kernel transformation module transform the first sub-vector and the second sub-vector into an input square matrix and a convolution kernel square matrix respectively based on a two-dimensional Winograd input transformation matrix and a convolution kernel transformation matrix; and executing the first operation for K times. Any ith first operation comprises the steps that a modular multiplication module multiplies T input square matrixes and an ith convolution kernel square matrix element by element to obtain T Hadamard product square matrixes; the output transformation module transforms the Hadamard product matrix into partial product sub-vectors based on a two-dimensional Winograd output transformation matrix; a partial product processing module accumulates partial product sub-vectors to corresponding sub-vectors in the product polynomial coefficient vector in a specified manner.

The application discloses a solid stress field prediction method and device, equipment and storage medium, and relates to the technical field of computers, which comprises the following steps: inputting material parameters of a target matrix material into a target physical information depth operator network, processing material space coordinates in the material parameters through a backbone network in the network to obtain a first network output, and processing attribute parameters in the material parameters through a branch network in the network to obtain a second network output; performing Hadamard product on the first network output and the second network output to obtain a target physical field corresponding to the target matrix material; determining a target solution of the material parameters satisfying a target boundary condition in the target physical information depth operator network, and determining a target distance output obtained by processing the material space coordinates by a target distance function; and determining a solid stress field corresponding to the target matrix material based on the target distance output and the target physical field. Thus, the efficiency and adaptability of stress field prediction can be improved.

The invention relates to the technical field of medical data analysis, in particular to an immune marker analysis system based on machine learning, which comprises an entropy density rheological module, a boundary dynamic labeling module, a collaborative tensor modeling module and a marker discrimination module. According to the method, three-dimensional immune data are processed through kernel density estimation, spatial information entropy is calculated, multi-source heterogeneous medical data spatial features are integrated, a gradient map is generated in combination with a moving average method to capture a marker concentration change trend, and a boundary topology model is constructed based on directional derivative detection and morphological closed operation to recognize morphological boundary features. Three-order tensor modeling is utilized to fuse immune subtypes, time dimensions and marker expression data, a core factor matrix is extracted through dimension adaptive tensor decomposition to improve marker correlation analysis integrity, Manhattan distance is adopted to quantify collaborative expression intensity, and the Hadamard product of a residual matrix is combined to improve the correlation analysis integrity of the marker. And the crossing from single threshold judgment to multi-dimensional cooperative judgment of quantitative analysis of the immune marker is realized.

The invention relates to the technical field of chemical engineering and natural language processing, in particular to an efficient parameter fine tuning method for a large model in the chemical engineering field. According to the method, aiming at the defect that'fine-tuning granularity ', 'parameter flexibility' and'resource efficiency 'are difficult to consider in an existing parameter efficient fine-tuning technology, a double-Hadamard-product mechanism is introduced to realize fine-granularity adjustment of model parameters, and key parameters in a chemical task are dynamically identified and optimized in combination with an adaptive parameter selection strategy; therefore, a high-efficiency fine tuning frame with low parameter quantity and high precision is constructed. According to the method, finer-grained model adaptation can be realized under the condition that only a small number of parameters are trained, and important parameters are adaptively selected according to the complexity of different chemical tasks, so that the fine tuning performance is ensured, and the parameter utilization efficiency and the task adaptation flexibility are remarkably improved.

This invention belongs to the field of sensors and signal processing, specifically relating to a wavelet kernel-scale sensitivity-guided denoising method for abrasive particle induced voltage signals. The method includes: acquiring abrasive particle induced voltage signals and performing harmonic cancellation to obtain a preprocessed signal; constructing a wavelet kernel function and calculating a kernel-scale guided spectrum using it and the preprocessed signal; constructing a sparse joint denoising model based on the kernel-scale guided spectrum; processing the sparse joint denoising model to obtain a convex optimization objective function; solving the convex optimization objective function using an adaptive step-size gradient descent method and an adaptive iterative shrinking threshold method to obtain a weight vector characterizing the distribution of abrasive particle characteristic signals; binarizing the weight vector characterizing the distribution of abrasive particle characteristic signals to obtain a feature indicator vector; performing a Hadamard product between the feature indicator vector and the preprocessed signal, followed by low-pass filtering to obtain a denoised signal. This invention can adaptively and non-destructively enhance and denoise abrasive particle characteristic signals under strong interference environments.

The application relates to a space target radar image data generation method based on low-light enhancement correction. The method comprises the following steps: designing a content transfer decomposition network, decomposing low-light features into illumination-independent reflection components and adaptive illumination components through Hadamard product constraints in a latent space, proposing a learnable intensity compression function, dynamically generating a latent space mask through gradient consistency constraints and sparse regularization, embedding the mask into a reverse denoising process to realize directional repair of degradation areas, combining the generation ability of a diffusion model and the guidance of a physical prior, reconstructing a diffusion path based on a bidirectional diffusion principle, generating the physical consistency of the process through endpoint binding constraints, adaptively adjusting the diffusion step, introducing a cyclic implicit iteration mechanism, gradually refining the generation result through a recurrent neural network module, and realizing accurate optical-ISAR image translation.

The invention provides a VPN traffic identification method and system based on deep learning, and relates to the technical field of network security. According to the method, a dynamic Low-Rank BottConv in a dynamic low-rank gated convolution (GDBConv) structure is utilized, high-dimensional features are mapped to a low-dimensional space through low-rank approximation, the parameters and the calculation amount are reduced, meanwhile, key features are reserved, and the calculation complexity is remarkably reduced. The global features are subjected to dynamic low-rank bottleneck convolution, normalization and activation function processing, then a gating feature map is generated through a Hadamard product, and then through dynamic low-rank bottleneck convolution and residual error connection output, the capturing ability of the model to feature details is enhanced, so that the model can flexibly adapt to different crack modes and background interference, and the accuracy of the model is improved. And the generalization ability of the deep learning model is improved. Meanwhile, due to the fact that the gating mechanism can dynamically adjust the weight according to the input characteristics, the bottleneck convolution can greatly reduce parameters and the calculation amount through low-rank approximation, the real-time requirement under multiple scenes can be met, and the method can adapt to a new protocol.

The application relates to the technical field of medical data analysis, in particular to an immune marker analysis system based on machine learning, which comprises an entropy density flow change module, a boundary dynamic labeling module, a collaborative tensor modeling module and a marker discrimination module.In the application, three-dimensional immune data is processed through kernel density estimation and spatial information entropy is calculated, multi-source heterogeneous medical data spatial features are integrated, a gradient graph is generated by combining a sliding average method to capture a marker concentration change trend, a boundary topology model is constructed based on directional derivative detection and morphological closing operation to identify morphological boundary features, a three-order tensor modeling is used to fuse immune subtypes, time dimensions and marker expression data, a dimension self-adaptive tensor decomposition is used to extract a core factor matrix to improve the integrity of marker correlation analysis, Manhattan distance is used to quantize collaborative expression intensity and combine residual matrix Hadamard product, and single threshold determination is crossed to multi-dimensional collaborative discrimination in immune marker quantitative analysis.

This invention discloses a method and system for optimizing injection molding process parameters by fusing trend learning and mechanistic constraints. It addresses the technical problem that most existing injection molding process parameter optimization methods are based on pure numerical optimization, leading to poor practical applicability of the optimized injection molding process parameter schemes in industrial production. The method includes acquiring raw injection molding process parameter data and the time-domain signal of screw axial vibration, generating standardized process feature data through a feature recognition mechanism; then performing trend analysis on this data, outputting visualized analysis results and a parameter sensitivity matrix; subsequently, using a graph attention network to combine the aforementioned results, the parameter sensitivity matrix, and injection molding process expert knowledge to output dynamic rule strength; finally, constructing a reinforcement learning network based on Hadamard product modulation rules and a soft update strategy, fusing dynamic rule strength, the parameter sensitivity matrix, and standardized process feature data to complete the optimization of injection molding process parameters and output the optimal injection molding process parameter scheme.

The invention discloses a decision-level multi-model dynamic fusion classification method, which comprises the following steps of: establishing each base model, obtaining a data set containing a sample and a real classification category, predicting through the base model to obtain a predicted classification category, calculating a classification accuracy rate in combination with the real category, and constructing a matrix; converting the prediction category into a one-hot coding form to obtain a voting matrix; after filtering the constructed matrix, calculating the contribution weight of each base model by adopting an approximate ideal solution sorting method; and performing scalar multiplication on the voting matrix and the weight to obtain an effective category voting weight matrix, performing Hadamard product on the voting matrix and the weight element by element, performing matrix addition to generate a voting fusion category two-dimensional matrix of all samples, and finally processing the two-dimensional matrix to obtain a fusion classification category. Aiming at the problem that the performance of the existing fusion classification method depends on the sample data quality, more accurate fusion classification is realized by dynamically selecting the dominant basis model and the category which is good at prediction and dynamically endowing the weight.

The application discloses a kind of based on location information fusion's high-efficiency point cloud classification identification method and system, identification method includes: the position and feature embedding of point are extracted from coordinate;Different scales of point cloud are obtained by farthest point downsampling, and the KNN local relationship is established by searching K nearest neighbor centered on the centroid by the centroid;According to the adjacent relationship, obtain the weight matrix and element matrix with global information and local information;Two matrices are mapped, and Hadamard product operation is carried out to obtain the feature matrix;Global feature representation of centroid is obtained using transformation function and local maximum pooling;Global feature is obtained using mapping function, residual structure and maximum pooling;Point cloud classification is carried out based on global feature by classifier.The application can replace attention mechanism in point cloud, reduce computational complexity under the premise of ensuring high precision, and be more suitable for point cloud data structure to learn point cloud local shape information.

This invention proposes a method for accelerating inference based on distribution-aware sparsity using Diffusion Transformer, comprising: constructing and training a generative model based on Diffusion Transformer; obtaining the attention map of the generative model; generating multiple intermediate results through the generative model with text prompts, adding noise to each intermediate result; progressively removing noise from the intermediate results to obtain the target image; wherein, during the noise removal process, obtaining the approximate Hadamard product of query-key element pairs in the attention map; based on the approximate Hadamard product, identifying query-key element pairs with opposite signs and similar values ​​as redundant element pairs; filtering the redundant element pairs in the attention map and performing sparse self-attention computation. This invention dynamically adapts to changes in attention distribution at different layers and timesteps through a distribution-aware sparsity mechanism, avoiding the limitations of fixed sparsity patterns; effectively suppressing error accumulation by precisely removing redundant computations; achieving high sparsity while effectively reducing the computational cost of self-attention and improving generation efficiency.

The invention discloses a multi-modal cognitive load detection method, and the method comprises the steps: carrying out the independent coding processing of target multi-modal data, and obtaining different modal features after preliminary coding; dynamically learning the weight of each modal by using a channel attention mechanism based on the preliminarily coded different modal features, and realizing the dynamic reconstruction of the multi-modal features by broadcasting a Hadamard product; respectively inputting the reconstructed feature vectors into a load task sensitive encoder and an individual sensitive encoder, generating task sensitive features and individual sensitive features, and generating final features; and on the basis of the final features, classification of cognitive load levels is carried out through a multi-layer perceptron. According to the method, the characterization capability of the model on the multi-modal data can be improved, and then the performance of the cognitive load detection model based on the multi-modal data is improved.