108 results about "Linear transform" patented technology

The application discloses an eye movement tracking method and system for the elderly based on eye-face double-flow timing interaction, and the method comprises the following steps: extracting a full-face region image and an eye local region image for an image frame of a face video sequence, extracting multi-scale face features for the full-face region image, performing layer-by-layer linear transformation, projection operation, splicing and fusion to construct a full-face token sequence; defining an eye prior dictionary based on the eye local region image, and encoding the eye local region image to obtain an eye feature representation, and further obtaining an enhanced eye feature representation; generating an eye query vector, obtaining a face key vector and a face value vector based on the full-face token sequence, and further obtaining an updated eye feature representation and a face feature representation optimized by gaze constraint; obtaining a differentially enhanced face feature representation; inputting the differentially enhanced face feature representation into a time sequence module to obtain a gaze point prediction result, and outputting an overall objective function.

This invention provides a method for modeling a cylinder motion system based on a physical information neural network, comprising: acquiring time-series data of the cylinder motion system, including time-series control quantities and system state quantities; constructing a physical information neural network model with a dynamic feature input layer and a physical parameter input layer, receiving the time-series control quantities and physical parameter vectors respectively; performing a nonlinear transformation on the physical parameter vectors through a physical parameter encoding network to obtain encoded physical feature vectors; fusing the time-series control quantities and physical feature vectors to form a joint input vector; inputting the joint input vector into the backbone neural network, synchronously predicting the system state quantities through multi-layer nonlinear transformations; constructing a composite loss function including a data loss term and a physical loss term; training the model using time-series data, updating the model parameters by minimizing the composite loss function until the model converges, thereby achieving high-precision modeling using only single motion data.

A method for detecting the cliques of evaluators in a decision-making process includes collecting a data matrix Sij to a computer system using an automatic input interface. The elements of the matrix Sij are the real numbers in a predefined range. Each of the elements of the matrix Sij represent a numerical evaluation provided by an evaluator j for an evaluated entity i. For each pair of evaluators j, where j1 and j2 (j1≠j2) from the data matrix Sij, calculating by the computer system the Pairwise Adjusted Distances EDj<sub2>1< / sub2>j<sub2>2< / sub2>. Applying by the computer system a nonlinear transformation to the Pairwise Adjusted Distances EDj<sub2>1< / sub2>j<sub2>2< / sub2>. Identifying cliques of evaluators by comparing by the computer system the transformed distancesEDj1⁢j2*to a robust lower threshold.

Reduce circuit size and improve coding efficiency. [Solution] The encoding device (100) comprises a circuit (160) and a memory (162). The circuit (160) derives a prediction residual that shows the difference between the target block and a prediction image generated using matrix operation type intra prediction, which generates a prediction image by performing matrix operations on the pixel sequences obtained from the left and upper pixel values ​​of the target block. A linear transformation is performed on the prediction residual, a quadratic transformation is performed on the result of the linear transformation, quantization is performed on the result of the quadratic transformation, the result of the quantization is encoded, and a common transformation set is used as the transformation set for the quadratic transformation for multiple prediction modes.

The application provides a color cast enhancement method and device of a display device, and belongs to the technical field of image processing. The method comprises the following steps: acquiring a three-channel image of a display area of the display device when displaying a test image, and splitting the three-channel image into three single-channel gray scale images; based on an image area division mode of the three-channel image, calculating a gray scale mean value of a gray scale sub-image corresponding to each gray scale image under a current image area, so as to determine a maximum gray scale sub-image and a minimum gray scale sub-image of the current image area; subtracting the minimum gray scale sub-image from the maximum gray scale sub-image to determine a region difference value image of the current image area; performing linear transformation on the region difference value image to generate a region enhancement image; and traversing all the image areas, and constructing a target enhancement image according to the generated region enhancement image. The application can process various color cast conditions on the display area, has high adaptability, and improves the color cast processing efficiency and precision of the display device.

The application discloses a kind of gas insulated switchgear gas chamber pressure leakage degree judging method and system, applied to insulating switchgear gas chamber detection technical field, method includes first obtaining the pressure value of the gas chamber to be evaluated in preset time period, constructs pressure value sequence, and then the sequence is weighted average processing, obtains first final pressure value sequence, again to first final pressure value sequence Pressure mutation identification, locates pressure mutation interval and removes it, obtains second final pressure value sequence.Then, construct pressure state space model, input model to predict second final pressure value sequence, obtain optimal estimation pressure value sequence, calculate pressure decay rate according to optimal estimation pressure value sequence, and the pressure decay rate is segmented linear transformation, finally output gas chamber pressure leakage severity value.Through the above method, the accuracy of pressure leakage evaluation is effectively improved.

Example implementations related to next state prediction and action selection are disclosed. In an example, an initial state is estimated based on a set of priors including at least one engagement opportunity. A diminishing effect of the at least one engagement opportunity is determined using a non-linear transformation and a set of likelihood estimates for the initial state is generated using a Bayesian steady state filter model that receives the initial state and the diminishing effect of the at least one engagement opportunity. A next state is predicted based on the initial state and the set of likelihood estimates.

The application discloses a machine learning-based retrospective clinical data management method. The method comprises the following steps: obtaining an electronic medical record log file and parsing unstructured text to generate a word vector; calculating the Hamming distance between the word vector and the knowledge graph standard entity vector to extract standardized test indicators; establishing a two-dimensional physiological correlation mask matrix according to the pathology dependent edge; truncating the time series data into a sequence feature matrix and performing linear transformation to generate a query matrix, a key matrix and a value matrix; calculating the dot product value of the query matrix and the transpose of the key matrix, adding the scaling factor and the mask matrix after exponentiation, and then multiplying the value matrix to obtain the context feature representation; and generating a dense feature tensor data through a feedforward neural network to predict the interpolation value. The application can effectively curb the overfitting of the time series model under sparse data, improve the physiological authenticity of the feature array, and reduce the generalization error of the downstream prediction model.

This invention provides a method and apparatus for reducing quantization errors in neural network models using regularization techniques, belonging to the field of artificial intelligence technology. The method includes: analyzing the model's target interval to determine the target interval; applying regularization techniques and a linear transformation algorithm to quantize the model to be quantized; training the quantized model; and deploying the trained model using quantization, thereby enabling the vehicle to perform environmental perception tasks in intelligent driving scenarios based on images of the external environment captured by an onboard camera. This invention can improve the model's resistance to quantization errors, thus improving the accuracy of the quantized model and achieving high-precision environmental perception in intelligent driving scenarios.

This application relates to a sensor data processing method and electronic device. The method includes: identifying effective movements during the movement of the measured object based on a displacement data stream and the movement type of the measured object; analyzing the effective movements to obtain motion parameters; performing a linear transformation on the motion parameters; and generating an effectiveness analysis result. The displacement data stream is obtained by sequentially filtering, differentiating, and integrating the original data stream. The original data stream includes raw data obtained by the target sensor based on sampling during the movement of the measured object, and the raw data is correlated with the displacement in a preset direction. Using this application, the problems of baseline drift and noise in sensor data can be solved, enabling the evaluation of motion effectiveness and providing reliable data support for the quality analysis of athletes' core movements.

The application discloses a spectrum graph convolution network heterogeneous graph representation learning method based on path collaborative graph enhancement, relates to the technical field of graph neural networks, and comprises the following steps: obtaining to-be-processed heterogeneous graph data, projecting node features of different types into a unified latent feature space by using type-aware linear transformation; calculating structure prior weights; constructing a path collaborative graph, learning path interaction weights by using a graph attention network; constructing a collaborative polynomial spectrum filter; performing spectrum graph convolution based on path collaborative graph enhancement, introducing a learnable positive definite diagonal matrix into the collaborative polynomial spectrum filter, performing weighting and feature transformation on the filtered node features, and obtaining final node representation. The application can learn the importance of meta-paths in a fine-grained manner and capture semantic interaction between meta-paths, solves the problems that existing spectrum heterogeneous graph convolution cannot distinguish the importance of fine-grained paths and lacks semantic collaboration, effectively improves the classification performance of heterogeneous graph nodes, and has a good application prospect.

A fault diagnosis method based on feature enhancement via linear transformation is provided, comprising collecting signal samples from an experiment table for a rotating device, and constructing a tra

The application discloses a flue gas desulfurization emission concentration control method, related device and system, relates to the flue gas purification technical field, and will run process variable time series data be input into the pre-trained deep learning prediction model, eliminate sequence non-stationarity and variable scale difference by the normalization layer, the embedding layer converts the time series data into the time series slice embedding vector which retains semantic information, the decoding layer extracts the causal characteristics between the running variables and the emission concentration by dynamically modeling the dependency relationship between each time series slice through the attention mechanism, and the output layer outputs the emission concentration prediction value through the pooling and linear transformation. The prediction value replaces the current measured value and is input into the PID control module, so that the control loop knows the change trend of the emission concentration in advance, the absorbent addition amount is adjusted before the actual change of the concentration, time delays introduced by the reaction process, equipment execution and measurement points are effectively compensated, the timeliness and accuracy of the absorbent adjustment are improved, standard flue gas emission is ensured, and the absorbent consumption is reduced.

The application relates to a method and device for identifying a modification site in an RNA sequence, computer equipment and a medium. The method comprises the following steps: obtaining sequence features, structure features of the RNA sequence to be identified, and a pre-trained modification site identification model; inputting the structure features into a structure convolution layer for convolution to obtain convolution structure features; in each attention encoding unit: inputting the sequence features into a query weight matrix, a key weight matrix and a value weight matrix respectively for linear transformation, thereby obtaining query features, key features and value features; fusing the query features, the key features and the convolution structure features to obtain first fusion features; fusing the first fusion features and the value features to obtain second fusion features; splicing the second fusion features output by all the attention encoding units to obtain spliced features; and performing feature adjustment on the spliced features based on a feature adjustment module to obtain an identification result. The method can improve the accuracy of the identification result.

The present application belongs to the technical field of road surface information detection, and particularly relates to a road surface crack segmentation method based on channel decoupling and cascaded state space, which comprises: acquiring a to-be-segmented road surface image and inputting the image into an encoder to extract multi-scale features; a split-channel feature fusion module is arranged at each stage of the encoder, the input features are decoupled into a global branch used for crack cross-region semantic modeling, a local branch used for edge feature extraction and an identity branch used for detail reservation; the global branch establishes cross-axis semantic dependence in horizontal and vertical directions through a cascaded structure perception scanning, and the local branch extracts directional edges through asymmetric spatial convolution; an adaptive grouping transformation module is arranged at each stage of the decoder, adaptive gate fusion of a learnable nonlinear transformation and a depth separable convolution is combined with a skip connection to gradually reconstruct, and a pixel-by-pixel crack probability map is generated. The present application effectively repairs crack topological discontinuity, improves segmentation accuracy and reduces calculation redundancy, and is suitable for real-time road surface inspection scenes of edge devices.

The application relates to the field of ethylene industry diagnosis, in particular to an ethylene furnace tube coking thickness prediction method and system.The method comprises the following steps: collecting ethylene furnace tube parameter time sequence data, extracting time sequence convolution features, obtaining bidirectional time sequence features through bidirectional time sequence dependence processing, and obtaining time sequence pooling features through attention pooling; linearly mapping the features to obtain a furnace tube coking thickness linear prediction value, combining a learnable nonlinear transformation with the linear prediction value to obtain a furnace tube coking thickness nonlinear prediction value; finally, fusing historical furnace tube coking thickness prediction error features, the furnace tube coking thickness linear prediction value and the furnace tube coking thickness nonlinear prediction value to calculate the furnace tube coking thickness prediction value of the ethylene furnace tube.Compared with the prior art, the linear behavior and the nonlinear behavior are combined, and the fusion strategy is adaptively adjusted based on the historical prediction error, so that the ethylene furnace tube coking thickness can be effectively predicted.

The application discloses a horizontal double-runner impulse water turbine generator state monitoring method and system, relates to the technical field of intelligent state monitoring, and comprises the following steps: collecting double-runner multi-modal data of a water turbine generator for preprocessing, generating two groups of standardized data, calculating a difference vector, and constructing a multi-dimensional feature matrix; performing linear transformation based on the multi-dimensional feature matrix, generating a query, key and value matrix, performing near and far field division, obtaining a near field and a far field, calculating near field attention on the basis of the near field, constructing a binary tree on the basis of the far field, calculating far field attention, fusing the near and far field attention, obtaining a final output feature vector, performing state prediction by using a full connection layer, outputting a state label for parameter iterative training and feedback, and obtaining a prediction result; and the application improves the accuracy of state monitoring and identification through effective feature fusion.

The invention relates to a petrochemical engineering emergency early warning method and device based on large model data distillation. The method comprises the steps of obtaining original data, preprocessing and standardizing the original data, and constructing a data set. The method comprises the following steps: mapping a data set from a high-dimensional space to a low-dimensional space through linear transformation, and performing eigenvalue decomposition on a covariance matrix of data in the data set to obtain eigenvalues and eigenvectors; and sorting the feature vectors according to the feature values from large to small, selecting the feature vectors of which the feature value ranking is not lower than a first threshold value to form a new feature space, and extracting important features from the feature space through LASSO regression. And calling a CNN deep learning model to extract key features from the important features, distilling representative features in the original data, and outputting a feature set after distillation. And performing anomaly detection on the petrochemical engineering system through a clustering algorithm according to the feature set after distillation, and triggering early warning when a detection result does not meet a preset condition.

The application discloses a network security situation prediction method and device based on machine learning and a medium, relates to the technical field of network security, and comprises the following steps: acquiring multi-modal network security data, and constructing a multi-scale time sequence feature tensor; performing weight adjustment based on threat feature importance, and generating a threat reinforced feature tensor; performing self-adaptive feature selection and time sequence correlation analysis on the threat reinforced feature tensor, and constructing a threat correlation feature tensor; performing feature-entity mapping and situation evolution modeling in combination with the threat correlation feature tensor and a network entity identifier, and generating an entity situation feature tensor; inputting the entity situation feature tensor into a deep neural network, and generating a situation representation vector sequence through multi-layer nonlinear transformation; and performing multi-step situation deduction and threat propagation path analysis based on the situation representation vector sequence, and outputting a network security situation prediction result. Through multi-scale time sequence feature construction and self-adaptive weight adjustment based on threat importance, the application improves the accuracy of situation prediction.

The application provides an image adaptive quantization communication method and system, relates to the technical field of artificial intelligence and semantic communication, and comprises the following steps: performing feature extraction on an input image to form an initial code word representation; adopting an adaptive dynamic quantization mechanism to perform quantization processing on the initial code word representation to generate a code word index as a quantization feature; after the receiving end obtains the code word index of the sending end, performing inverse quantization by using a codebook knowledge base, and reconstructing a recovered image by using a decoder. The method and system provided by the application avoid representation collapse and realize adaptive matching and dynamic updating between codebook elements and semantic features through multi-codebook cooperation, a learnable linear transformation and a dynamic selection mechanism based on attention; and through quantization representation and transmission of the image and then decoding and recovery, the efficiency of image transmission is greatly improved, and an efficient and intelligent basic framework is provided for a next-generation semantic-oriented image communication system.

The invention discloses a parallel message authentication code construction method and system for quantum security. The method comprises the following steps: 1) selecting an adjustable block cipher algorithm key K, nonlinear transformation theta and the length tau of a message identification code by a participant; 2) the sender executes an MAC generation scheme, takes the key K, the temporary value N and the message M as input, outputs a message identification code and sends the message identification code to the receiver; the flow of the MAC generation scheme is as follows: a) dividing a message M into a plurality of n-bit groups; b) encrypting the ith group Mi by using an algorithm, and inputting theta to obtain a state Si after the output and the state Si-1 are subjected to XOR; c) intercepting the leftmost side tau bit of the state Sm corresponding to the last group as a message identification code; and 3) the receiver executes an MAC verification scheme, and verifies whether (MAC, M) is correct or not by taking the key K, the temporary value N, the message M and the message authentication code MAC as input.

The present application relates to the field of financial technology and the field of model data processing, and provides a financial business-based model data processing method and apparatus, a device, and a medium. The method comprises: acquiring a financial business model to be processed and a preset number of stages; parsing a decoder layer into virtual layers, and on the basis of the vocabulary of an embedding layer and the dimension of the decoder layer, determining the number of virtual layers corresponding to the embedding layer and the number of virtual layers corresponding to a linear transformation layer; and on the basis of the preset number of stages and the total number of virtual layers in said financial business model, determining the number of virtual layers in each model block, and on the basis of the number of virtual layers in each model block, deploying each model block into a compute card in a corresponding stage. The method of the present application improves the computational efficiency of model training and reduces the computational consumption of model training.

This application discloses a matrix operation processor and data processing method for quantum-resistant cryptographic algorithms, belonging to the field of quantum-resistant cryptography. The matrix operation processor includes a control module and an operation module. The control module is configured to receive an operation request containing a target matrix, allocate a storage address, control a matrix generator to generate and store data to the corresponding address; in response to a trigger signal indicating that a row of data in the target matrix has been generated, it reads the data from the corresponding address and sends it to the operation module, and controls the matrix generator to generate the next row of data in the target matrix; the operation module includes multiple parallel multiply-accumulate units, configured to perform multiplication and accumulation operations on the data sent by the control module according to the operation rules corresponding to the linear transformation operation type between the target matrix and the first matrix; the number of multiply-accumulate units is determined by the bit width of a single data read. Applying this application aims to improve the efficiency of matrix operations.

This application provides a natural language processing method, apparatus, device, and computer-readable storage medium; relating to the field of artificial intelligence technology, the method includes: obtaining a hyperbolic word vector sequence based on a preset word vector table; performing attention encoding on the current hyperbolic word vector to obtain attention encoding features corresponding to the current hyperbolic word vector; performing a linear transformation on the attention encoding features through a preset Lorentz transformation function to obtain linear encoding features corresponding to the attention encoding features; the preset Lorentz transformation function is obtained by training a linear transformation matrix model that satisfies preset constraints; the preset constraints are used to ensure that the output feature vector obtained by the linear transformation matrix model after processing the input feature vector in the hyperbolic space is located in the hyperbolic space; and obtaining the target processing result corresponding to the sentence to be processed based on the linear encoding features. This application can improve the stability and efficiency of natural language processing using hyperbolic neural networks.

This invention relates to a student classroom action detection method and system in the field of big data processing technology, comprising the following steps: stacking and grouping aggregated features, and calculating the behavioral feature center of each individual in the aggregated features; constructing a temporal representation of each individual based on the behavioral feature center of each individual, and after evaluating the importance using a scaled dot product attention mechanism, concatenating and linearly transforming the outputs of all attention heads to obtain the final output features of multi-head attention; reconstructing the final output features to obtain reconstructed features; performing original classification prediction on the reconstructed features to obtain initial prediction results, and performing co-occurrence matrix enhancement to obtain co-occurrence enhancement results; performing nonlinear modeling based on the co-occurrence enhancement results, calculating the final loss function, and detecting individual student actions in the video to be detected, thus solving the problem that existing classroom behavior recognition methods are highly subjective and difficult to accurately identify in multi-person scenarios.

This invention relates to the field of electronic real estate records management technology, and discloses a method and system for intelligent classification and structured storage of electronic real estate records. The method involves collecting image data from electronic real estate records, extracting and fusing features, performing preliminary classification based on multimodal fusion features, and storing the records in a structured manner when the classification score FL > classification threshold FY. When the classification score FL ≤ classification threshold FY, the method performs a time-series update of the knowledge graph and stores the records according to the reclassification results. This invention sets up a semantic module, a visual module, and a structural module, and uses a fusion unit to use the semantic embedding vector as a query. It independently performs scaling dot product attention calculations on visual and structural features to generate visual alignment vectors and structural alignment vectors. Finally, these vectors are concatenated and fused, and then subjected to linear transformation and layer normalization to output a multimodal fusion feature vector. This fully mines cross-modal correlation information, thereby improving the accuracy of record classification.

The application discloses a short-term power load prediction method and device based on a space-time neural network model, adopts a novel space-time neural network model, and the space-time neural network model comprises a preset number of adaptive space-time feature extraction modules; in the adaptive space-time feature extraction modules, space features and time features are respectively extracted through adaptive graph convolution branches and gated recurrent unit branches, and are fused to obtain space-time features. Through the preset number of adaptive space-time feature extraction modules, the features output by the last adaptive space-time feature extraction module are connected in residual connection with input features, linear transformation is performed, and a prediction result is obtained. The model disclosed by the application can extract potential non-Euclidean space features by using an improved learnable adjacency matrix, and does not need to use a predetermined adjacency matrix as prior knowledge, so that the precision of short-term power load prediction can be further improved.