233 results about "Lower dimensional space" patented technology

The invention relates to the technical field of fault prediction, and provides a fault prediction method for a multi-modal cross-attention enhancement graph neural network, and the method comprises the steps: collecting the data of equipment; performing adaptive enhancement and normalization processing on the image data, performing sliding window segmentation, standardization and noise suppression on a time sequence numerical signal, and performing semantic vectorization coding on a maintenance log text; extracting low-dimensional spatial features of image data by using the pruned lightweight convolutional neural network, connecting time sequence features of modeling time sequence numerical signals in series, extracting context semantic expressions of maintenance log texts, integrating the features into multi-modal data, alternately taking each modal feature as Query and the other modal features as Key and Value, and obtaining multi-modal data; calculating attention weight and performing weighted fusion; constructing a modal node weighted graph, and performing inter-node feature propagation through a multi-layer graph attention network; and a residual service life regression prediction module and a degradation level classification module are deployed in parallel, and fault early warning is completed through multi-task joint optimization.

The invention relates to the technical field of knowledge maps, in particular to a knowledge map method and system for multi-source heterogeneous data fusion. The method comprises the following steps: cleaning multi-source data through an auto-encoder, dynamically weighting and standardizing after low-rank decomposition and PCA denoising, aligning GNN entities and authenticating standard approval serial numbers; combining CNN / GNN to extract texts, images and sensor multi-modal features, fusing time sequence information with BiLSTM, and performing weighted aggregation; a BERT-BiLSTM-SelfAttention-CRF is adopted to identify an entity, a GCN inference relationship is adopted, a TransE is embedded into an entity relationship to a low-dimensional space, and the entity relationship is stored to Neo4j to support real-time query; dynamically updating the atlas by incremental learning; and visually displaying the constructed knowledge graph. According to the method, a complete closed loop from data cleaning, multi-modal fusion and graph construction to dynamic optimization is formed, and comprehensiveness, accuracy and expandability of the knowledge graph in a multi-source heterogeneous scene are ensured.

The invention relates to the technical field of artificial intelligence and natural language processing, in particular to a key value cache compression and sparse attention calculation method and system for large language model reasoning, and the method comprises the steps: an offline calibration stage; the online reasoning stage comprises the following steps: a pre-filling step; an autoregression generation step: for each newly generated lexical element, projecting a current query vector Q and a key vector K in a key cache to a low-dimensional space to obtain Q'and K '; calculating an approximate attention score based on Q'and K ', and selecting an index I of the first k most relevant lexical elements which are ranked from high to low; and calculating an accurate attention score based on Q and K [I], and calculating with the value vector V [I] to obtain the output of the current lexical element. According to the scheme, the memory and calculation bottleneck of large model reasoning in a scene of long text sequence input are solved, and the method has the advantages of reducing video memory occupation and calculation complexity at the same time.

The invention discloses a personalized commodity recommendation optimization method based on deep learning. The method comprises the following steps: S1, constructing a data set; s2, constructing a multi-dimensional knowledge graph, and embedding the multi-dimensional knowledge graph into a low-dimensional space by using a knowledge graph embedding algorithm; s3, constructing a user interest track graph by using the time sequence graph model, and generating a user interest vector based on capturing long-term and short-term interest changes; s4, obtaining an optimized user interest vector by using an improved moth fire suppression algorithm; s5, based on the optimized user interest vector, combining a graph neural network to learn relationships between users and commodities, between users and between commodities; and S6, generating a personalized commodity recommendation list according to a graph neural network learning result and real-time behavior feedback of the user. According to the method, a multi-dimensional knowledge graph, time sequence graph modeling, a moth fire suppression algorithm, a graph neural network and the like are fused, and optimization of personalized commodity recommendation is realized.

The invention relates to the technical field of power system fault diagnosis, and discloses a distribution network line fault detection data processing method, system and device and a medium, and the method comprises the steps: obtaining load data of a target distribution line detection point, and carrying out the first preprocessing of the load data, and obtaining a fusion feature vector; performing second optimization operation on the fusion feature vector to obtain an optimized fusion feature vector; calculating a potential feature matrix according to the optimized fusion feature vector, and mapping the potential feature matrix to a low-dimensional space to obtain a low-dimensional sample set; presetting an adaptive label propagation algorithm, and performing fault category judgment on the low-dimensional sample set based on the adaptive label propagation algorithm; and storing a judgment result in a relational database. The problems that feature extraction is not accurate in a high-noise environment, and a classification model is insufficient in new fault expansion capacity are effectively solved, and fault signal processing robustness is improved.

The invention relates to the technical field of battery system fault recognition, and provides a battery system fault recognition method and system based on a time sequence contrast learning encoder, and the method comprises the steps: employing a lightweight residual error convolution module and a time sequence Transform module which are connected in series and in parallel to extract a network, and obtaining a fault recognition result after the classification of extracted feature vectors; in the multi-layer time sequence feature extraction network training process, extracted feature vectors are mapped to another low-dimensional space through a contrast projection head to obtain contrast space features, and contrast loss is calculated through the contrast space features to enhance the separability of the feature vectors. According to the method, a light-weight convolution and Transform hybrid coding structure is fused, and a contrast projection head and a user-defined contrast loss function are introduced, so that the model can effectively distinguish normal and abnormal states in a battery system under the condition that a large number of labels are not needed, and the sensitivity and generalization ability of fault recognition are remarkably improved.

The invention discloses a data aggregation method based on multi-modal features, which comprises the following steps: collecting multi-modal data, preprocessing the multi-modal data, extracting features according to modals, dividing the features into high-dimensional data, medium-dimensional data and low-dimensional data, and positioning neighbor points by using a ball tree algorithm for the high-dimensional data; the neighborhood range of the medium-dimensional data is dynamically adjusted based on the distribution density; a neighborhood of low-dimensional data is calculated through Euclidean distance, then the low-dimensional data is mapped to a low-dimensional space by means of local linear embedding, then the low-dimensional data is traversed, discrete data value frequency and continuous data probability density are counted, marginal probability is calculated in combination with information entropy, data aggregation weight is determined accordingly, and finally features after dimension reduction are spliced according to a high-level sequence, a middle-level sequence and a low-level sequence. And probability normalization is carried out in each hierarchical block to generate aggregated comprehensive features. According to the method, through multi-dimensional differentiation processing and weight calculation based on data distribution, aggregation of multi-modal features is realized, and feature complementarity and accuracy are improved.

The invention relates to a trajectory prediction model robustness enhancement method based on dynamic subspace projection decomposition. Comprising the following steps: firstly, extracting hidden layer semantic features containing historical tracks and map topology through a multi-modal feature encoder; secondly, constructing a dynamic routing mechanism based on scene self-adaption, and calculating projection weights of input features on a plurality of expert subspaces; then, executing truncation projection operation based on orthogonal decomposition, retaining core semantics located in a low-dimensional space, and filtering out adversarial disturbance located in an orthogonal complementary space; and finally, introducing a feature consistency constraint training mechanism, taking the reconstructed features of the clean sample as anchor points, and compulsively aligning the purified features of the confrontation sample with the anchor points. Compared with the prior art, the method has the advantages that the robustness of the model in white box gradient attack, black box query attack and physical semantic deception scenes is remarkably improved through feature purification of a physical level and structured consistency constraint, and the prediction reliability of the automatic driving system is ensured.

The invention discloses a data dynamic partition storage method and system based on adaptive clustering, and relates to the field of data processing. The method comprises the following steps: S1, extracting multi-scale geometric features of a high-dimensional data set, calculating a local curvature and generating a curvature feature matrix; s2, constructing a feature distance matrix and a similar matrix based on the curvature feature matrix, and generating a low-dimensional embedding matrix; s3, executing self-optimization clustering according to the low-dimensional embedded matrix, determining a cluster number through singular value distribution, and generating an initial cluster; s4, calculating a stability factor of each cluster, and triggering cluster splitting or merging operation according to a threshold value to form updated cluster division; and S5, performing incremental processing on newly added data points, obtaining a new point local curvature through local curvature gradient correction, mapping the new point local curvature to a low-dimensional space, dynamically deciding affiliation based on a cluster radius, and updating partitions. Through multi-scale feature extraction, self-optimization clustering and incremental updating mechanisms, high-dimensional data clustering precision, robustness and calculation efficiency are improved.

The invention discloses a content fingerprint generation method and system based on a deep learning model and a related method, and belongs to the technical field of digital copyright protection. According to the content fingerprint generation method based on the deep learning model provided by the invention, the size of the digital work data is adjusted to adapt to the input requirement of the deep learning model, and the deep learning model is utilized to perform local feature extraction on the input image, so that subtle differences and key information of the work content can be captured; through the application of a self-attention mechanism, a global feature map is generated, and the global feature map is mapped to a low-dimensional space to obtain a content fingerprint in a high-dimensional vector form, so that the content fingerprint has high uniqueness, even if a file is subjected to operations such as clipping, compression and format conversion, the content fingerprint can still remain unchanged or only has small change, and the content fingerprint can be stored in a high-dimensional space. Therefore, the validity of copyright verification is ensured.

The invention discloses a light energy power station fault prediction system based on deep learning. The system comprises a data acquisition module used for reading equipment operation data from a sensor; the data preprocessing module is used for denoising, interpolating and standardizing the equipment data; the graph convolutional network construction module is used for constructing an equipment data graph structure and extracting features; the Lemap dimension reduction module is used for mapping the high-dimensional equipment features to a low-dimensional space; the time sequence modeling module is used for constructing a time sequence prediction model based on the low-dimensional features; the hyper-parameter optimization module is used for optimizing hyper-parameters of the time sequence model; the model verification module is used for evaluating the precision and response time of the fault prediction model; the model deployment module is used for deploying the prediction model to a monitoring system; the fault prediction and early warning module is used for monitoring in real time and generating fault early warning; and the continuous optimization module is used for regularly optimizing and retraining the fault prediction model. The method achieves the high efficiency of fault prediction of the light energy power station, remarkably improves the prediction precision and the reliability of equipment operation, and is widely suitable for equipment monitoring and early warning.

The invention relates to the technical field of new energy power generation prediction, and discloses a t-SNE visualization and depth time sequence attention model-based wind and light power prediction method and system, and the method comprises the steps: obtaining historical power data and corresponding historical meteorological data of a wind power station and a photovoltaic station, and constructing a historical data set; inputting the preprocessed high-dimensional meteorological data into a t-SNE dimension reduction module, and mapping the high-dimensional data to a low-dimensional space through symmetric joint probability density calculation to obtain low-dimensional visual data distribution; the low-dimensional data are input into a parallel model composed of a TCN-SENet branch and a BiGRU-GlobalAttention branch, and space-time local features and global time sequence features are extracted respectively; and fusing the feature vectors output by the two branches, generating a wind-solar power prediction result through a full-connection layer, and performing evaluation. According to the method, the limitation of traditional single energy independent modeling is broken through, the adaptability of the model to a complex power generation mode of a distributed station is optimized, and the precision and generalization ability of multi-energy joint prediction are remarkably improved.

The invention relates to the technical field of brain-computer interfaces, in particular to an electroencephalogram signal feature extraction and dimension reduction method, device and equipment for intelligent glasses. The method comprises the steps that an electroencephalogram signal data set is acquired, time domain, frequency domain and time-frequency domain features are extracted, and an original feature set is generated; decomposing each feature vector in the original feature set to obtain a feature atom set, and based on the feature atom set, performing atomic-scale recombination on each feature vector in the original feature set to generate a recombined feature set; calculating a kernel matrix corresponding to the recombined feature set, mapping the kernel matrix to a high-dimensional kernel space, and performing principal component analysis according to the original feature set to obtain a kernel principal component projection matrix; and projecting the feature set to a low-dimensional space by using a kernel principal component projection matrix, and generating low-dimensional nonlinear feature representation in combination with the electroencephalogram signal data set to complete electroencephalogram signal feature extraction and dimensionality reduction of the intelligent glasses. And multi-scale characteristic decomposition and analysis domain division are realized.

The invention discloses a deep forgery detection method based on a firefly optimization algorithm and locality sensitive hashing. The deep forgery detection method comprises the following steps: S1, converting collected document data into a standard format image document data set; s2, generating a preprocessed image document data set; s3, forming a document feature vector set; s4, obtaining an optimal feature weight combination; s5, generating an optimized document feature vector set; s6, constructing a locality sensitive Hash matching model based on the optimized document feature vector set, and mapping the document feature vectors into a low-dimensional space through a preset Hash function family to form a plurality of Hash buckets; and S7, carrying out similarity matching on the document data in each hash bucket, and outputting a final document data high-forgery detection result and a corresponding document forgery confidence score by adopting structural similarity comparison and image training comparison. According to the method, a step-by-step enhanced counterfeit identification process from preliminary screening to confirmation is realized, and the false alarm rate and the missing report rate are remarkably reduced.

The invention discloses a wafer defect real-time detection method and system based on multi-scale feature fusion, and belongs to the field of wafer defect detection, and the method comprises the steps: carrying out the preliminary coding of each scale feature through employing a convolutional neural network for a generated feature set, and obtaining a feature vector group containing the local and global information of a defect; if the saliency of the high-frequency component in the feature vector group exceeds a preset threshold value, performing weighted enhancement on the high-frequency feature through an attention mechanism to obtain an enhanced feature set; according to the enhanced feature set, constructing an index structure based on Hash coding, and generating a defect feature index table capable of being quickly retrieved by mapping feature vectors to a low-dimensional space; updating the index table according to the matching result, combining the newly detected defect features with the time sequence prediction result, and generating an expanded defect feature index table; and for the expanded index table, an incremental clustering algorithm is adopted to classify defect features, and defect type distribution updated in real time is obtained.

The present invention relates to a platform management method and system based on low-code development, the method comprising: obtaining multiple heterogeneous data sources, extracting first data from the multiple heterogeneous data sources, and preprocessing the first data to obtain standard data for constructing a knowledge graph. A plurality of different entities and entity relationships between different entities are obtained from the standard data, and a knowledge graph is constructed based on the plurality of different entities and entity relationships between different entities. The entity nodes in the knowledge graph and the entity relationships between different entity nodes are mapped to a low-dimensional space for vector representation through a graph embedding model to extract global knowledge including entity relationships between different entity nodes and network topology structures. A predefined business process and a plurality of decision points corresponding to the execution of the business process are obtained, and the execution results of the business process are evaluated by executing global knowledge and current business process variables to obtain corresponding guiding decisions.

The invention relates to a doctor searching and recommending method based on multi-behavior time sequence modeling, and the method comprises the steps: firstly generating an optimized query vector through a GAN and a doctor-patient relation knowledge graph, and determining the intention of a patient; secondly, mapping the high-dimensional optimized query vector to a low-dimensional space by adopting hyperplane projection index and partial homomorphic encryption, searching in an encrypted doctor database by using the projected optimized query vector, and screening out a candidate recommended doctor set matched with the intention and preference of the patient; after candidate doctors are screened, the system ranks a candidate recommended doctor set by using memory enhanced ranking in combination with a cross-patient recommendation strategy. The method aims to improve the search experience of the patient while improving the accuracy and safety of doctor retrieval, and provides an efficient and intelligent doctor recommendation scheme for an intelligent hospital guide platform.

A load feature optimization clustering method based on a one-dimensional convolution auto-encoder comprises the following steps: acquiring load types including daily load curve data as a data set, and performing preprocessing; the method comprises the following steps: constructing a one-dimensional convolution auto-encoder model which comprises an encoder and a decoder, pre-training load data by taking minimization of reconstruction loss as a target, extracting load features, and storing model parameters obtained by pre-training at the same time; the method comprises the following steps: firstly, decoding a network part, removing a decoding network part, retaining an encoding part of feature extraction, constructing a PCA-Kmeans space conversion model, transmitting training data in an encoder, and converting a potential space into a low-dimensional space by using a PCA algorithm to carry out K-means clustering analysis; the encoder is finely adjusted, the encoder is trained by taking minimization of a clustering loss function as a target, and after the encoder is trained for one epoch, K-means clustering is carried out on a newly generated potential space; new clustering distribution is obtained, and power load mode extraction is achieved. Compared with other traditional clustering methods, the method has the advantages that the application complexity is simplified, and the classification efficiency is improved.

The invention provides an electric power system model important parameter group identification method based on trajectory feature clustering, relates to an electric power system simulation technology, and solves the problem that modeling is not performed due to time domain global error index coarse graining and parameter coupling in traditional electric power system important parameter identification. The parameters are used as to-be-identified high-sensitivity parameters; performing time domain difference on each group of simulation tracks and actually measured tracks to obtain difference tracks, and extracting difference track features; the method comprises the following steps: mapping high-dimensional data to a low-dimensional space through a nonlinear dimension reduction method, and executing K-Means clustering on the data in the low-dimensional space to obtain a cluster label of each sample; sensitivity calculation is carried out on all samples in the clusters, and a parameter combination with the highest occurrence frequency and the maximum contribution to dynamic differences in the clusters is screened out to serve as a final dominant parameter set. According to the scheme, the simulation frequency and the calculation burden can be remarkably reduced, and high-precision and interpretable discrimination of important parameters really influencing the system performance is realized.

The invention provides an image compression and image reconstruction method and system based on a single-step diffusion model, and the method comprises the steps: carrying out the feature coding of an input image through a preset neural encoder, and determining the low-dimensional feature representation of the input image; performing compression processing on the low-dimensional feature representation of the input image according to a preset compression ratio, and determining a compression feature representation of the input image; performing a single prediction operation on the compression feature representation of the input image by using a preset single-step diffusion model, and determining an original feature representation of the input image; and decoding the original feature representation of the input image to determine a reconstructed image. According to the method and the device, image compression and image reconstruction in a low-dimensional space and compression feature representation of different compression ratios adopt the same single-step diffusion model to execute a single prediction operation to reconstruct the image, so that the calculation complexity in an image decoding stage is reduced, the image decoding efficiency is improved, and the image structure integrity and the image visual quality are kept.

The invention discloses a data management system and method based on artificial intelligence, and relates to the technical field of data intelligent processing, and the method comprises the steps: carrying out the frequency spectrum transformation of an encrypted feature vector, calculating the energy density of each frequency spectrum coefficient, constructing a cumulative distribution function, determining an effective frequency spectrum interval based on the energy distribution function, and constructing a projection matrix. Compressing the spectrum vector to a low-dimensional space by using a projection matrix to generate a compressed feature vector; and constructing a neural network model to calculate an abnormal score, setting a detection threshold, performing label detection on the abnormal score, forming a label vector, performing homomorphic decryption on the label vector, and generating a plaintext feature vector. According to the method, a frequency spectrum energy distribution function is introduced on the basis of homomorphic encryption, safe dimensionality reduction and rearrangement of feature vectors are realized by combining periodic mapping and an integer mechanism, the structural identifiability of the feature vectors in a finite field is enhanced, an offset period is optimized through a fluctuation potential function, and the sensitivity to abnormal changes is improved.

Systems and methods are provided for reducing a number of performance metrics generated by network functions to a number of reduced dimension metrics, which can be used to detect anomalous behavior and generate a warning signal of the detected anomalous behavior. The disclosed systems and methods transform raw performance metrics in a high dimensionality space to a reduced number of metrics in a lower dimensionality space through dimensionality reduction techniques. Anomalous behavior in network performance is detected in the high dimensionality space using the reduced dimension metrics. The systems and methods disclosed herein convert the reduced dimension metrics back to the high dimensionality space, such that the performance metrics from network functions can be utilized to understand and address potential problems in the network.

The invention relates to the technical field of wind power plant power control, in particular to a wind power plant dynamic power distribution optimization method based on projection dimensionality reduction. The method comprises the following steps: collecting a high-dimensional original operation data stream of a wind power plant, and preprocessing to output a high-dimensional state vector; extracting a dominant mode matrix through an intrinsic orthogonal decomposition algorithm, and projecting a high-dimensional state vector to a low-dimensional space; future modal coefficient evolution is predicted through an autoregression prediction model, and a low-dimensional rolling optimization problem is constructed and solved through a sequential quadratic programming algorithm; reconstructing the optimal low-dimensional modal coefficient vector into a target power instruction of each fan, and issuing and executing the target power instruction; and updating the dominant mode matrix through an incremental singular value decomposition algorithm and correcting parameters of the autoregressive prediction model. According to the method, projection dimensionality reduction from high-dimensional data to a low-dimensional space is realized through intrinsic orthogonal decomposition, and the problem of low optimization solution efficiency caused by high data dimensionality of a traditional method is solved.

The invention discloses a multi-source heterogeneous data integration method based on dynamic metadata modeling and a storage medium, and the method comprises the steps: collecting structured data, semi-structured data and unstructured data from multi-source heterogeneous data, extracting the format, structure and semantic information of the data, forming a metadata set, and storing the metadata set in a database; an original basis is provided for subsequently solving naming conflicts and data type differences; data features are comprehensively captured, and input is provided for dynamic metadata modeling; further, through a dynamic metadata model, uniformly describing the metadata set to obtain uniform description information; further, based on dynamic mapping and dimension reduction optimization of a neighborhood graph, high-dimensional data are mapped to a low-dimensional space, and a global structure is reserved; and finally, integrating and generating a fusion database supporting cross-source query based on the data subjected to dimension reduction, eliminating data islands, and providing a standardized view.

The invention discloses a method and system for determining an operation mode of a power system based on power flow section power, and the method comprises the steps: carrying out the discretization grading processing of the obtained active power of a plurality of power flow sections, so as to obtain high-dimensional discrete data; performing feature dimension reduction processing on the high-dimensional discrete data to obtain low-dimensional spatial data; and performing clustering based on the low-dimensional spatial data to perform classification of operation modes, and determining the operation mode of the power system according to a classification result. By clustering and dividing the actual power flow section operation data of the power system, different operation modes of the power system can be effectively identified, the time distribution characteristic and similarity of each operation mode can be determined, power system operation personnel can formulate corresponding control strategies according to different operation modes, and the operation efficiency of the power system is improved. Operation complexity of the power system is reduced and safe and stable operation of the power system is guaranteed.

The invention belongs to the technical field of power systems, and particularly discloses a scene generation method and device considering spatial-temporal correlation. According to the method, the time correlation and the space correlation in the renewable energy historical data (such as wind speed, illumination radiation intensity and runoff) are fully considered, and the improved particle swarm algorithm is adopted to perform time sequence reconstruction on the initial scene set for representing the space correlation in the renewable energy historical data; and the space-time correlation in the renewable energy historical data can be effectively captured. Considering that renewable energy has randomness and volatility and is relatively high in dimensionality and difficult to process when multiple possible scenes are generated, the method comprises the following steps: capturing space-time correlation in renewable energy historical data, mapping the high-dimensional renewable energy historical data to a low-dimensional space, and obtaining the high-dimensional renewable energy historical data; the challenge that a complementary system (such as a water-wind-light complementary system) generates a high-dimensional uncertainty scene is effectively dealt with, and short-term scheduling of the complementary system is better guided.

The invention discloses a virus gene sequence host prediction method and system, and belongs to the technical field of biological information analysis, and the method comprises the steps: dividing a virus gene sequence into ordered k-mer word sequences, and carrying out the vectorization of the k-mer word sequences through a pre-trained BERT language model, and obtaining an embedded vector; mapping the embedded vector of the high-dimensional space into a low-dimensional space with a fixed dimension by adopting an average pooling method, and reducing the embedded dimension to obtain a feature vector representing an original gene sequence; the feature vectors are input into a pre-trained classification network model to predict the likelihood that the sequence is infected with a particular host. According to the method, the complete virus gene sequence is directly used as input, so that information loss possibly caused by dependence on statistical characteristics is avoided, and the accuracy of characteristic extraction is remarkably improved.

The invention provides an MDS dimension reduction-based retrieval enhancement generation semantic matching optimization method and system and a storage medium. The method comprises the following steps of: obtaining a user problem and high-dimensional semantic vectors of a plurality of candidate text segments; constructing a distance matrix based on the two high-dimensional semantic vectors to obtain a high-dimensional semantic similarity; performing dimension reduction on the distance matrix by adopting a multi-dimensional scale analysis (MDS) method, and calculating low-dimensional semantic similarity between the user question and each text fragment based on a distance relationship after dimension reduction; performing weighted fusion on the low-dimensional semantic similarity and the high-dimensional semantic similarity to generate a mixed similarity; and sorting the mixed similarities from high to low, and selecting the text fragments corresponding to the top n of the mixed similarities for generating response output. By reducing the dimension of the high-dimensional retrieval vector to the low-dimensional space, the interference of irrelevant dimensions is reduced, the semantic matching precision and efficiency are improved, and then the quality of the text recalled by the large language model is improved.

The present disclosure relates to the technical field of artificial intelligence and natural language processing, in particular to a key-value cache compression and sparse attention calculation method and system for large language model inference. The key-value cache compression and sparse attention calculation method for large language model inference comprises: an offline calibration stage; an online inference stage, comprising: a pre-filling step; an autoregressive generation step, for each newly generated word element: projecting the current query vector Q and the key vector K in the key cache into a low-dimensional space to obtain Q' and K'; calculating the approximate attention score based on Q' and K', and selecting the top k most relevant word element indexes I in descending order; calculating the accurate attention score based on Q and K[I], and calculating the output of the current word element with the value vector V[I]. The above scheme solves the memory and calculation bottleneck of large model inference in the scenario of long text sequence input, and has the advantages of reducing the memory occupation and the calculation complexity at the same time.

The invention discloses an array antenna forming directional diagram synthesis method based on beam space mapping and dimensionality reduction optimization. According to the method, the traditional high-dimensional excitation optimization problem is converted into low-dimensional coefficient optimization by constructing the beam space basis matrix, so that the calculation efficiency of large-scale array antenna shaping directional diagram synthesis is remarkably improved. The specific implementation process comprises the following steps: firstly, establishing a steering vector parameterization model in arbitrary array arrangement, and then constructing a beam space basis matrix containing beams of a forming region and a null region to realize dimension reduction mapping; performing intelligent optimization in a low-dimensional space by adopting an alternating projection algorithm, and synchronously realizing accurate shaping of a main lobe, side lobe level suppression and multi-null control; and finally, obtaining realizable array excitation distribution through beam space inverse mapping reconstruction. Compared with a traditional method, the method can remarkably reduce the dimensionality of the optimization variable, is suitable for any geometric arrangement form such as a conformal array and a sparse array, and can be widely applied to beam optimization design in the fields of 5G communication, phased array radar and the like.