34 results about "Sparse model" patented technology

The invention provides a large model lightweight reasoning deployment method under limited hardware resources, and the method comprises the steps: quantifying the weight importance of a large model through a composite index of gradient sensitivity and activation frequency, and carrying out pruning operation in combination with an improved index weighted moving average strategy, thereby obtaining a structured sparse model; the sparse model is divided into sub-networks by adopting double rules, a routing decision network is trained, and an adaptive feature shunting architecture model is constructed; a multi-precision weight set is generated through a nested quantization technology, quantization bit width is dynamically adjusted, and edge equipment hardware parameters are adapted to complete reasoning environment initialization; after a reasoning request is received, an optimal sub-network is selected based on the trained routing decision network, corresponding weights are loaded in parallel, and a reasoning result is fused and output; and converting a reasoning result format, and dynamically optimizing a scheduling strategy based on a system real-time monitoring index. The method is compatible with a mainstream large model and a hardware platform, and an efficient and universal deployment scheme is provided for end-side AI engineering landing.

The invention provides a sparse Transform model reasoning acceleration method based on a GPU platform, and the method comprises the steps: building a first storage format in a sparse Transform model, and carrying out the coding of a sparse mask of any type through the first storage format; wherein the first storage format comprises an inner layer block and an outer layer block; an outer layer block adopts a block compression sparse row structure, and row pointers and column indexes of a completely effective block and a partially effective block are recorded; the inner layer block adopts an unsigned 64-bit integer value bitmap and is used for representing masks in the block; based on the sparse mask matrix coded by the first storage format, performing sparse calculation on the query, the key and the value, and outputting a corresponding context vector; compared with a traditional sparse mask storage structure in the prior art that calculation sparsity and memory access regularity are difficult to consider when the masks present random, global or mixed distribution characteristics, the method has the advantages that support for any type of masks is achieved, storage overhead is reduced, memory access continuity is improved, good GPU parallelism is kept, and thread divergence and memory access conflicts are avoided.

The invention belongs to the technical field of GIS fault detection, and discloses a GIS latent insulation fault diagnosis method and system under strong noise, and the method achieves the omnibearing signal capture through synchronously obtaining optical, mechanical, electrical and acoustic signals, and solves a problem of weak signal leak detection. For strong noise interference, noise and fault characteristic frequency bands are effectively separated by using multi-modal signal difference and constructing a time domain and wavelet domain cascaded double-layer dictionary. A discriminative sparse model is combined with a K-SVD algorithm to carry out end-to-end iterative optimization, sparse features with higher discriminative force are automatically mined, and the problem of deep feature mining is solved. For the problem of weak generalization ability of small samples, a classifier error term is introduced into an objective function, a joint optimization objective function fusing reconstruction and classification errors is constructed, dictionary learning and classifier training are combined into one, the generalization ability of the model is significantly enhanced, and the method is suitable for large-scale popularization and application. Therefore, timely and accurate identification of the latent insulation fault under the conditions of strong noise and small samples is realized.

The invention discloses a sparse reasoning method, system, equipment and product for large language model reasoning, and relates to the technical field of artificial intelligence. The method comprises the following steps: firstly, acquiring a weight matrix of a large language model, and then carrying out pruning sparsification processing on the weight matrix to obtain a sparse matrix; the storage format of the sparse matrix is converted into a bitmap coding storage format which is suitable for tensor calculation core perception and adopts a multi-level block structure to respectively correspond to different calculation granularities in a GPU / NPU hardware architecture so as to obtain a sparse model, and then the sparse model is deployed to respond to a reasoning request to perform reasoning service; sparse matrix multiplication is completed through a sparse matrix multiplication kernel based on on-chip storage bitmap coding so as to perform reasoning, so that through an innovative sparse matrix storage format and calculation optimization, the storage efficiency and the calculation performance in the reasoning process are remarkably improved, and especially in a low-sparseness scene, the reasoning efficiency is greatly improved. And the performance blank of the existing sparse reasoning framework in the field is filled.

The application relates to a deep learning channel estimation method and device suitable for an underwater acoustic sensor network. The method utilizes OFDM pilot data to establish a sparse signal model, and obtains a real-value sparse model by performing real-value transformation on the sparse signal model. An approximate message passing (AMP) estimation framework is introduced to realize preliminary recovery of channel information. Further, a model-driven deep learning framework is proposed. An ST-LAMP network based on AMP without considering sparse prior and a GGM-LAMP network considering sparse prior are respectively established. Network training is respectively performed according to a predetermined strategy. Optimal matching parameters are learned through data learning and are updated and fixed. Correct channel estimation values can be output through input of new measurement values. The problem that traditional empirical parameter setting cannot be applied to high complexity of an underwater acoustic channel, leading to deviation of channel estimation from an actual channel, is solved. The precision of channel estimation is improved. The channel estimation method has the characteristics of low complexity and strong self-adaptability.

Examples including a system described herein implement a novel liquid time constant neural network (LTC-NN) based architecture to recover an underlying model of physical dynamics from real world data. The automatic differentiation property of LTC-NN nodes overcomes problems associated with low sampling rate, the input dependent time constant in the forward pass of the hidden layer of LTC-NN nodes creates a massive search space of implicit physical dynamics, the physics model solver based data reconstruction loss guides the search for the correct set of implicit dynamics, and drop out in dense layer ensures extraction of the sparsest model. Further, to account for perturbation timing error, the LTC-NN based architecture of the system utilizes dense layer nodes to search through input shifts that results in the lowest reconstruction loss.

The embodiment of the invention provides a data transmission method and device, a data processing method and device and electronic equipment, and aims to improve the processing efficiency during data transmission and parallel computing. The method comprises the following steps: for each to-be-transmitted data packet, performing structured sparse processing on the to-be-transmitted data packet through a data sparse model deployed at an initiating node; evaluating the data quality of each processed to-be-transmitted data packet to obtain the quality score of each to-be-transmitted data packet; the quality score is used for representing the proportion of effective data in the to-be-transmitted data packet; according to the quality score corresponding to each to-be-transmitted data packet, sequentially transmitting each to-be-transmitted data packet to the corresponding path node; and for each path node, determining respective semantic importance of each to-be-transmitted data packet received by the path node, and transmitting each to-be-transmitted data packet to a corresponding destination node according to each semantic importance.

The application relates to the technical field of aero-engine aerodynamic acoustic detection, and discloses a pipe-in-sound mode integrated detection method considering abnormal gain of a microphone, which comprises the following steps: a microphone array unit, a rotating speed detection unit and a signal acquisition module are built to realize data acquisition; a core algorithm unit is built, a joint sparse model is constructed based on a blade passing frequency calculated from rotating speed data, and super parameter adaptive updating is completed through an EM algorithm until the super parameter meets a convergence condition; a mode coefficient matrix and an abnormal gain matrix are extracted based on the converged super parameter, elements of the abnormal gain matrix are arranged in ascending order, and the position of an abnormal microphone is determined according to a preset threshold; subsequently, channel data corresponding to the abnormal microphone is removed, and the mode coefficient is re-inverted by using a Bayesian compressive sensing method; and a standardized detection report is output by the system. The method realizes high-precision mode recognition and abnormal detection, does not need manual intervention, and improves detection efficiency and robustness.

The application discloses a kind of training after sparse benchmark evaluation method and device.The method constructs a precise evaluation system by five indexes: sparse distribution mode, reconstruction technology, model architecture accuracy, size robustness and different task accuracy. After the original model is sparsified using the standardized sparse paradigm, the sparsification performance of the model is evaluated using these indexes. This method can compare different post-training sparse models horizontally, comprehensively evaluate the performance of the model from both the algorithm and the model, and distinguish the impact of sparsification and reconstruction on performance, as well as the impact of architecture, size or task on sparsification performance. Using the present application, not only the model compression efficiency is improved, but also an important reference is provided for model selection and optimization.

The invention relates to a motor fault diagnosis method based on sparse Bayesian and a one-dimensional neural network. The method comprises the following steps: S1, collecting a motor stator current signal, and carrying out filtering and preprocessing; s2, performing Hilbert transformation on the current signal to obtain an analysis signal; s3, constructing a sparse signal model in the truncated frequency domain and performing frequency grid division; s4, solving the sparse model based on sparse Bayesian learning, and correcting the deviation between the fault frequency and the grid by adopting an off-grid model to obtain an envelope spectrum; and S5, inputting the envelope spectrum into a one-dimensional convolutional neural network to output a motor fault diagnosis result. The method has higher frequency recovery precision, stronger robustness and fault recognition accuracy exceeding 98% in a complex environment.

The invention discloses a distributed sparse model fingerprint method based on dual-key driving, and the method comprises the following steps: inputting an original model, an owner secret key and an identity label into a computer system; generating a seed based on an owner secret key, calculating a global index set in one or more stable target convolution layers pre-selected in the model, and selecting a weight subset as a fingerprint carrier; constructing a random projection matrix based on the identity label; the weight vector and the random projection matrix generate an original model fingerprint; verifying the suspicious model, repeating the steps to obtain a fingerprint to be detected, and calculating the cosine similarity between the original model fingerprint and the fingerprint to be detected; and judging whether an illegal derivative relationship is formed based on a calibrated threshold value, and outputting an ownership judgment result. According to the method, identity and pseudo-random processes are bound through double keys, and distributed sparse sampling and normalized random projection are combined, so that the uniqueness and robustness of fingerprints are improved, various attack scenes can be effectively resisted, and the ownership of the model is accurately judged.

This invention describes methods and systems for utilizing blockchain-verified sparse language models (SLMs) in enterprise applications. Sparse models address the limitations of large language models (LLMs) by reducing overfitting, bias, and complexity, while blockchain integration ensures result verification and maintains immutable records of model operations. The system employs smart contracts for automated validation and distributed consensus mechanisms to verify model outputs. This comprehensive approach leads to improved transparency, trust, and cost-efficiency through both model sparsification and blockchain-based accountability, making these verified SLMs particularly suitable for high-risk applications. The invention's combination of sparse modeling techniques with blockchain verification creates a robust framework for deploying trustworthy and efficient language models in critical enterprise environments.

This invention discloses an image restoration method based on a hybrid structured sparse model, specifically including the following steps: initializing the restored image and setting the number of iterations; constructing a matrix of similar image patch groups; establishing a hybrid structured sparse model; using the hybrid structured sparse model to sparsely encode each similar image patch group; reconstructing each similar image patch group; and restoring the entire image based on all reconstructed similar image patch groups. The image restoration method of this invention better reconstructs details such as edges and textures, and effectively suppresses unwanted visual artifacts, further improving the image restoration effect.

This invention proposes a multi-target detection method based on group sparsity OTFS communication sensing integration. The implementation steps are as follows: initializing parameters; constructing a sparse signal recovery problem based on the OTFS communication sensing integrated group sparsity model; solving the sparse signal recovery problem; and obtaining multi-target detection results. This invention constructs an OTFS communication sensing integrated group sparsity model and uses this model to construct a sparse signal recovery problem for N channels. During the iterative solution of the sparse signal recovery problem, the channel gain parameters and error terms are gradually updated, avoiding the impact of forced convergence on the channel gain estimate. Furthermore, the previously estimated channel gain parameters are used to gradually eliminate interference from other targets, thereby significantly reducing interference between multiple targets and avoiding the defect that interference between multiple targets increases with the number of targets, further improving target detection accuracy and robustness.

The application discloses a direct positioning method and system based on a single-bit signal, and the method comprises the following steps: a single-bit model and a space sparse model of a received signal are established; a probability density function of a single-bit received signal in the single-bit model of the received signal is obtained according to a prior distribution of noise, a prior distribution of observation data and a Laplace prior distribution; a mean value and a variance expression of a space sparse signal and a maximum posterior probability density function of a target radiation source position are obtained; an optimal target function of a hyperparameter is obtained; the optimal target function of the hyperparameter is solved to obtain an updating expression of the hyperparameter; the updating expression of the hyperparameter and the mean value and the variance expression of the space sparse signal are alternately iterated and solved to obtain the mean value and the variance of the space sparse signal, and the space position of the target radiation source is obtained according to the mean value of the space sparse signal. The application can reduce signal transmission cost and effectively improve the positioning speed and precision of multi-target radiation sources.

The invention discloses a sparse model parameter-oriented target task adaptive parameter optimization transfer learning method, which is suitable for medical image classification and model optimization training. The classification contribution degree of each convolution kernel in a pre-trained CNN model to each target category of a target domain is calculated based on a feature attribution method, the convolution kernels are sorted according to the classification contribution degree, the convolution kernels with the correlation degree larger than a threshold gamma are screened out for optimization and fine tuning, and other parameters are frozen; and self-adaptive low-rank fine tuning: performing fine tuning on the frozen convolution kernel parameters by adopting low-rank adapters, including calculating function values to determine the rank of each adapter, and performing self-adaptive low-rank fine tuning on the low-correlation convolution kernels by allocating different ranks.

The invention provides a two-stage fine-grained structured sparse training traffic signal control method based on reinforcement learning, and the method comprises the following steps: S1, constructing a traffic signal control model based on deep reinforcement learning, and the traffic signal control model comprises a strategy network and a value network; s2, performing two-stage sparse model training on the traffic signal control model, and outputting a customized sparse model suitable for local features of each intersection; and S3, deploying the customized sparse model suitable for the local features of each intersection in the corresponding edge control device, and enabling each edge control device to operate the customized sparse model suitable for the local features of each intersection according to the real-time traffic observation data so as to generate a traffic signal control instruction. According to the invention, through structure recombination and a training mechanism, the resource efficiency is obviously improved innovatively.

The application discloses a state-related convolution sparse model-based electroencephalogram signal pattern recognition method, aiming at the problem that existing convolution sparse coding is difficult to recognize state-related biomarkers, state-shared and state-specific waveforms are used for modeling, and the waveform characteristics of electroencephalogram signals in different states are clearly distinguished. Based on this, the application adopts waveform inconsistency constraints to efficiently identify potential biomarkers related to certain states, and the identified potential biomarkers have good interpretability and can be used as an effective tool for auxiliary medical diagnosis.

The invention provides an exogenous radar target detection method and system based on a sparse model, and relates to the technical field of radar signal processing, and the method comprises the steps: receiving a time domain monitoring signal of an orthogonal frequency division multiplexing waveform, carrying out the discrete Fourier transform, obtaining a carrier domain monitoring signal, and extracting data at a pilot frequency position; constructing a carrier domain reference signal based on known transmitting end pilot frequency information; carrying out clutter suppression processing on the carrier domain monitoring signal; performing inverse discrete Fourier transform on the carrier domain monitoring signal and the carrier domain reference signal, reconstructing a new time domain reference signal and a new time domain monitoring signal, and performing distance processing to obtain distance domain data; according to the method, only pilot frequency information is utilized to construct reference signals, direct waves do not need to be accurately estimated, and the system complexity is greatly reduced; and clutter suppression is introduced before sparse modeling, so that the problem that strong clutters cover weak targets is effectively solved.

The invention belongs to the technical field of model compression, and discloses a lightweight pruning method and system based on channel dominant modeling, and the method comprises the steps: executing one-time forward reasoning under the condition of not modifying the model weight, and extracting each linear layer input activation matrix; calculating a basic scoring matrix according to the mechanism; channel dominance indexes are calculated, and dominant channels are sorted and identified; selecting channels of a front proportion to form a dominant set, applying a scaling factor to a dominant channel column, and generating a correction scoring matrix; independently sorting according to output channel dimensions, and cutting low-score weights to a target sparse rate; and storing the sparse model and directly performing reasoning evaluation. According to the method, on the premise of not increasing extra training and calculation overhead, the pruning quality and the model robustness under the high sparse rate are remarkably improved.

This application relates to a method for accelerating the training and push of a large-scale ultrasound model based on sparse tensor optimization using next-generation supercomputing. The method involves fine-tuning the large-scale ultrasound model to generate a structured block sparse pattern in the weight tensors of each expert model, resulting in a sparse ultrasound model. A sparse block index table and an expert index table are constructed and linked to form a joint encoding table. The sparse ultrasound model is deployed to each parallel group of a next-generation heterogeneous supercomputing cluster. Multiple rounds of iterative training are performed on the sparse model. In each round of iterative training, the ultrasound data of the current batch are divided into parallel groups. In each parallel group, the target expert model to be activated is determined based on the feature data of each sample ultrasound data, and the target computing node is located based on the joint encoding table. The target expert model is activated on the target computing node, and the corresponding non-zero weight blocks are loaded to perform calculations on the feature data. The model weight parameters are updated based on the output results. This method can accelerate the processing efficiency of large-scale ultrasound models.

The invention discloses a programmable metasurface two-dimensional off-grid DOA estimation method based on sparse Bayesian learning, and relates to the field of signal processing. Establishing a DOA estimation mathematical model according to a data measurement mechanism of the programmable metasurface; converting the DOA estimation mathematical model into a two-dimensional off-grid sparse model by using Newton binomial expansion and two-dimensional linear Taylor expansion; and carrying out Bayesian inference on the two-dimensional off-grid sparse model by adopting an expectation maximization technology through a sparse Bayesian probability model to obtain an estimated value of the DOA. According to the method, a DOA estimation problem is converted into an off-grid sparse signal recovery problem through a derived two-dimensional off-grid sparse model, and sparse Bayesian learning (SBL) is adopted for solving. According to the method, an SBL basic framework is followed, a Bayesian probability model is constructed, iteration is performed through Bayesian inference, DOA estimation efficiency and precision are greatly improved, and a theoretical template is provided for a sparse signal recovery estimation algorithm under a compressed sensing framework.

The application discloses an image registration method based on grouping motion estimation and neighborhood refinement sampling. The method comprises the following steps: constructing an image registration network comprising a shared feature extraction module, a sparse motion module and a dense motion module; obtaining a source image and a target image, and training the image registration network, extracting shared features, grouping motion estimation and neighborhood refinement sampling are sequentially performed until the loss function converges to complete the training; and finally obtaining a registration motion flow by processing the source image and the target image to be registered through the network, and registering the target image to the source image. The method introduces grouping motion estimation to perform sparse modeling on input features, combines an adaptive fusion mechanism to enhance local deformation expression capability, designs a neighborhood refinement sampling method, introduces a learnable local weighting correction mechanism in the motion flow sampling process, effectively improves the boundary definition and detail accuracy of the dense motion field, and significantly improves the accuracy and robustness of image registration.

The invention provides an unmanned aerial vehicle inspection terminal data anomaly detection method and system based on behavior coding cooperating with Transform-RVM, and relates to the technical field of artificial intelligence, the method comprises the following steps: preprocessing sensor original data of an unmanned aerial vehicle inspection terminal, and generating a time sequence sample set; the time sequence samples are converted into behavior coding vectors, and space-time fusion features are generated by fusing equipment space topology information through a graph convolutional network; inputting the space-time fusion features into a Transform editor, and extracting deep features through a multi-head self-attention mechanism, residual connection, layer normalization and a feedforward network; constructing a probability classification model by adopting a relevance vector machine RVM, modeling a parameter optimization process by utilizing a neural differential equation, and performing efficient training in combination with a dynamic sparsity control and adjoint sensitivity method to obtain a highly sparse RVM model; and performing anomaly detection on the unmanned aerial vehicle inspection terminal data acquired in real time by using the highly sparse RVM model. According to the scheme, the anomaly detection precision of the unmanned aerial vehicle inspection terminal data can be improved.

The invention discloses a fine adjustment method and device of a model, a storage medium and electronic equipment. The method comprises the steps that firstly, a weight matrix of a sparse model is obtained, then an adapter matrix and a sparse mask matrix are generated based on the weight matrix, the former enables the model to adapt to a new task through parameter adjustment, and the latter keeps a sparse structure of the model. Multiple rounds of training are carried out based on the matrix, in the training process, adapter matrix parameters are controlled to be adjusted based on gradients, weight matrix parameters are kept unchanged, a target sparse model adaptive to a target task is finally determined, video memory use is optimized, and training cost is reduced. According to the method and the device, the technical problem of relatively low model fine tuning efficiency caused by excessive model parameters due to model fine tuning in related technologies is solved.

The invention discloses an off-grid target direction finding method and system based on array amplitude linear approximation and a medium. According to the method, a high-power reference signal with a known angle is introduced to the edge of an observation area, and a nonlinear array amplitude signal model is approximately linearized into a linear signal model about sine difference frequency of a target and a reference signal; then, under a sparse representation framework, first-order Taylor expansion about sine difference frequency is carried out on column vectors in the frequency steering matrix, and an off-grid sparse model containing a grid deviation item is constructed; and finally, through a two-step iteration strategy, an initial target orientation is determined by using a coarse grid, and then signal difference frequency and off-grid deviation are alternately corrected through a closed-form solution. According to the method, phase information is not needed, the influence of array phase errors on the estimation precision can be effectively eliminated, the problem that the precision is limited due to grid mismatch is solved through off-grid correction, and the estimation precision of an amplitude-only direction finding system is remarkably improved under the condition that the calculation complexity is not increased.

The application relates to the technical field of intelligent construction, and discloses a method for intelligently predicting the construction progress of a bent cap, which comprises the following steps: collecting multi-source real-time data of equipment, structures, personnel and environments in the construction process, and calculating construction entropy for measuring the uncertainty of a system; establishing a causal state evolution model to predict the evolution trend of the construction entropy in a future period; and when a construction progress disturbance risk is predicted, carrying out counterfactual reasoning based on the model or a sparse model obtained by dynamically pruning the model, and generating and recommending an optimized adaptive correction strategy. The application quantifies the overall uncertainty of a construction system as construction entropy, and combines the forward-looking reasoning capability of the causal state evolution model, so that the change from experience-dependent passive response to risk to data-driven proactive early warning and optimal decision recommendation is realized, and the scientificity, foresight and timeliness of construction risk management are improved.

The application provides a signal spectrum sensing and DOA estimation method based on cooperative MWC. The method is based on a non-uniform array MWC sampling structure with arbitrary array element positions, and each receiving channel has a different mixing sequence. In combination with the different mixing information and spatial phases in each receiving channel, a sparse model is constructed to realize simultaneous joint estimation of the carrier frequency and the DOA. Compared with the traditional subspace-based method, the method does not need to make a uniform array assumption, and relaxes the restriction on the array manifold. Through the space-time distribution characteristics of the spectrum and the DOA, the parameter estimation problem can be solved when the coherent signals and the source ambiguity exist, and the method has a wide application prospect.

The invention relates to the technical field of communication, and provides a low-overhead sparse beam direction determination method and device, the method is applied to a communication system, the communication system comprises P to-be-observed beam directions, and the method comprises the following steps: controlling a transmitting end to transmit irregular beams in M time slots respectively, and obtaining actually measured amplitude gains obtained by observation of a receiving end, M being smaller than P; taking the weight matrix and the gain matrix corresponding to K beam directions in the P to-be-observed beam directions as variables, and taking the minimum difference value between the actually measured amplitude gain and the predicted amplitude gain as a target to construct a sparse model; the predicted amplitude gain is obtained according to the weight matrix and the gain matrix corresponding to the K beam directions; carrying out K times of iterative solution on the sparse model to obtain K target beam directions; and under each iteration round, determining the candidate beam direction with the maximum correlation coefficient under the current iteration round as the target beam direction under the current iteration round. According to the invention, the calculation overhead of beam training is reduced.

This invention discloses an indoor target localization and tracking method based on dynamic CSI 3D sparse recovery, comprising the following steps: acquiring CSI data and preprocessing it; pruning the Doppler parameter space by static background estimation and stripping, combined with prior constraints on target motion velocity, discretizing the time delay and angle parameter spaces, and constructing a 3D dynamic sub-dictionary; establishing a CSI 3D sparse model, and extracting dynamic path parameters using an alternating search 3D orthogonal matching method; introducing physical consistency constraints and time continuity verification to screen effective target paths; mapping the time delay, angle, and Doppler frequency shift parameters of the effective path to target distance, azimuth, and radial velocity observations; and achieving recursive estimation of target state and trajectory tracking based on a constant velocity state transition model and a nonlinear observation model through particle filtering. This invention effectively suppresses static multipath and noise interference, and improves dynamic path resolution and positioning and tracking accuracy.