68 results about "Attention model" patented technology

This application provides a runoff prediction method, apparatus, electronic device, and computer-readable storage medium. The method includes: acquiring the forecast meteorological time series of a target watershed during the prediction period, historical meteorological time series, and historical runoff time series for historical periods; inputting the historical meteorological time series and historical runoff time series into an attention model to extract global contextual features of the target watershed; inputting the forecast meteorological time series, global contextual features, and initial noise data into a conditional diffusion probability model, performing multiple backdiffusion processes to obtain multiple predicted runoff time series of the target watershed during the prediction period; and calculating a specified quantile for each moment in the prediction period based on the multiple predicted runoff time series to construct a confidence interval, thereby obtaining runoff prediction information containing a risk probability distribution. This method avoids gradient vanishing when processing long-sequence data and outputs the probability distribution of the prediction results.

This invention discloses a vehicle sticker recognition method, an attention model training method, and related equipment. The vehicle sticker recognition method includes: acquiring an image of a vehicle sticker to be recognized, as well as a trained attention model and a classification network; inputting the image of the vehicle sticker to be recognized into a feature extraction layer for feature extraction processing to obtain multi-scale features of the image; inputting the multi-scale features into an attention optimization layer to adjust the weights of each feature in the multi-scale features to obtain attention-optimized multi-scale features; inputting the attention-optimized multi-scale features into a feature fusion layer for feature fusion processing to obtain fused features; and providing the fused features to a classification network so that the classification network outputs the type label corresponding to the vehicle sticker image to be recognized. By extracting multi-scale features and optimizing attention, key image information is highlighted and fused to generate comprehensive features, which are then used by the classification network to determine the type of vehicle sticker, thereby achieving high-precision vehicle sticker recognition in complex scenes.

A closed-loop selective attention system for resolving conflicts in multi-source or multi-speaker environments, including a plurality of internal attention models, each outputting a probability distribution over candidate sources and an associated confidence score, a fuser detecting conflicts when two or more of said attention models output high-confidence predictions that disagree, a selective sampling policy querying one or more external agents, wherein each external agent possesses a knowledge base, a reliability model, and a communication protocol, a trust and reliability module assigning and updating dynamic trust scores for internal and external agents based on past performance, an efficiency optimizer minimizing communication overhead and decision delay by balancing token usage cost and latency cost, and a dynamical system formulator ensuring convergence of the conflict resolution process under bounded trust, decaying step size, and limited sampling.

This invention provides a method and system for early warning of line structure health. By collecting multi-source heterogeneous monitoring data of the line structure, and based on the reliability differences of different types of monitoring data, an adaptive weight fusion algorithm is designed to fuse the multi-source heterogeneous monitoring data to obtain comprehensive feature values. Historical health data, historical minor hazard data, and historical serious hazard data are selected as training samples to train an improved attention ELM model, completing model training. The comprehensive feature values ​​collected and fused in real time are used to construct feature vectors according to time windows, which are then input into the trained improved attention ELM model to output the probability distribution of the line health status level. The probability of the line health status level and the comprehensive feature values ​​are compared with the corresponding probability thresholds and warning thresholds to perform graded warnings, achieving accurate identification and graded warning of early hazards in the line structure, improving the reliability and real-time performance of the warnings.

The application discloses a prefix caching method and device suitable for a mixed attention model architecture. In view of the heterogeneous management problem of key-value caching and state snapshot, a unified prefix index mechanism based on chain hashing is proposed, so that the two types of caches share a hash space. The method determines the boundary of the two types of caches consistent and effective through the longest prefix matching and forward traversal verification, combines local snapshot query and remote confirmation, realizes the collaborative recovery and reuse of key-value caching and linear attention snapshot, and reduces the heterogeneous state step-out risk. In the reasoning process, the two types of states are captured synchronously and unified hash index is adopted, and a bidirectional mapping between cache block identification and snapshot hash is established through a double cache coordination manager; when the expulsion condition is met, the linkage cleaning of the two types of caches is realized based on the mapping. The application realizes the unified storage and life cycle collaborative management of heterogeneous caches under the mixed attention architecture, guarantees the whole process consistency, and improves the reasoning performance and system robustness.

The application discloses an accurate and efficient infrared weak small target detection method and device, and belongs to the field of computer vision, and comprises the following steps: step one, constructing an attention model based on infrared target characteristics, which is used for realizing multi-scale fusion feature extraction of the infrared weak small target; and constructing a target positioning model based on reinforcement learning, which is used for realizing image target positioning; step two, constructing a target detection model combining the attention mechanism and the reinforcement learning, which is used for completing detection and identification of the infrared weak small target by combining the attention model based on the infrared target characteristics, the target positioning model based on the reinforcement learning and a target detection algorithm. The application not only improves the detection accuracy, but also reduces the calculation amount, so that the algorithm is more efficient and reliable.

This invention provides a method for predicting the health trend of hydropower units based on the TCN-BiLSTM-Attention model, relating to the field of hydropower generator sets. The method involves: S1: acquiring the original vibration signal of the hydropower generator set and denoising the original vibration signal using a TTAO-VMD-EWT joint denoising model; S2: obtaining multiple operating state parameters of the hydropower generator set and calculating the Pearson correlation coefficient, Spearman correlation coefficient, maximum information coefficient, and distance correlation coefficient between these parameters and the vibration signal; S3: calculating the fusion evaluation index of each operating state parameter and the vibration signal, selecting key features, and constructing a multi-dimensional feature vector together with the reconstructed vibration signal; S4: constructing a TCN-BiLSTM-Attention prediction model to predict and output the predicted trend of the hydropower generator set's vibration signal. This method improves prediction accuracy, increases prediction efficiency, and enhances the accuracy, stability, and timeliness of hydropower generator set vibration trend prediction.

The application discloses a method and device for calibrating an optical phased array, a storage medium and an electronic device, wherein the method comprises: obtaining light field data generated by the optical phased array; inputting the light field data into a pre-trained composite attention model to generate phase noise, wherein the composite attention model comprises a multi-core attention module, a spatial attention module and a channel attention module, the multi-core attention module is used to capture global dependency by extracting features under different receptive fields, the spatial attention module is used to capture global spatial context information by local pooling, and the channel attention module is used to capture global channel context information by global pooling; and calibrating the optical phased array based on the phase noise. Through the application, the technical problems that the prediction accuracy of the phase noise is difficult to guarantee and the data dependency and operation complexity are relatively high are solved.

This application relates to an air conditioner and its control method, storage medium, and program product. The method includes: acquiring multi-source information of a target animal; the multi-source information includes radar data, infrared thermal imaging data, and image data; extracting features from the multi-source information to obtain a first feature vector; the first feature vector is composed of a radar feature vector, an infrared feature vector, and an image feature vector; the three feature vectors have the same number of dimensions; inputting the first feature vector into a pre-trained attention weight model to obtain dynamic weight coefficients corresponding to the three feature vectors; performing weighted calculation and concatenation of the three feature vectors according to the dynamic weight coefficients to obtain a second feature vector; inputting the second feature vector into a preset spatial channel attention model to obtain a third feature vector; and identifying the current abnormal state of the target animal based on preset recognition rules and the third feature vector. This method enables the air conditioner to more accurately identify the current abnormal state of the target animal.

The application discloses a power operation and inspection multi-label image recognition method and system based on adaptive graph convolution, and belongs to the technical field of power operation and inspection.The application discloses a power operation and inspection multi-label image recognition method based on adaptive graph convolution, which processes sample images by constructing a preprocessing model, a network extraction model, a semantic attention model, a static GCN network model, a dynamic GCN model and a binary classifier, and completes multi-label image recognition in a power operation and inspection scene, and the scheme is scientific, reasonable and feasible.Further, the application utilizes the adaptive graph convolution neural network combining the static GCN network model and the dynamic GCN model to learn the correlation between multi-labels, so as to improve the accuracy of multi-label image recognition in the power operation and inspection scene, improve the power operation and inspection efficiency, promote the safe and stable operation of the power system, and facilitate the popularization and use of the multi-label image recognition method in the field of power operation and inspection.

A method can include receiving (i) a visual representation of a scene or image data and (ii) user input specifying one or more goals for modifying the visual representation. The method can also include executing a visual attention model on the visual representation and testing combinations of scene modifications. The method can also include evaluating the impact of the scene modifications on achieving the received one or more user goals and outputting one or more recommended sets of one or more visual representation modifications that satisfy the received one or more user goals upon receiving confirmation that the one or more user goals have been achieved.

The application discloses a power component life prediction method and system of adaptive double attention Transformer, and the method comprises the following steps: acquiring time series data of an electronic power component to be predicted at a plurality of latest continuous historical moments; inputting the time series data into an input embedding and position coding layer of the adaptive double attention Transformer model to obtain a first feature matrix; an adaptive mixed attention module in an encoder of the adaptive double attention Transformer model calculates global attention features and sliding window attention features in parallel based on the first feature matrix, and the corresponding attention weights of each kind of attention features, and the global attention features, the sliding window attention features and the corresponding attention weights are weighted and fused; an attention pooling layer of the adaptive double attention Transformer model aggregates the encoder output features into a global feature vector along the time dimension, and maps the global feature vector through an output layer to obtain a normalized remaining life percentage. The application can improve the life prediction accuracy of the electronic power component.

The application discloses a security-constrained unit commitment fast solving method, system, equipment and medium based on a space-time graph attention network, which comprises the following steps: inputting graph structure input data into a pre-trained space-time graph attention model for multi-task joint reasoning to obtain start-stop state pre-judgment probability and activity prediction score; fixing the binary decision variable of unit start-stop based on the start-stop state pre-judgment probability, screening the line transmission constraint based on the activity prediction score, constructing a reduced mixed integer linear programming model and solving the model, verifying the feasibility of the whole network power flow after obtaining a candidate scheduling scheme, outputting a final scheduling scheme if the verification is passed, and modifying and re-solving the reduced mixed integer linear programming model based on the violation information if the verification exists constraint violation. The application overcomes the problem of slow solving speed of the traditional method under a large-scale system, and can meet the time requirement of future virtual power plants and distributed power sources participating in power market clearing.

This invention provides a multimodal information popularity prediction method and system inspired by group consensus, applied in the field of artificial intelligence technology. The method includes: inputting first content and first user features into a popularity prediction model, and outputting a first predicted value representing the popularity of the first content; the popularity prediction model is trained by: acquiring second content and second user features; extracting consensus opinions from individual opinions in a common group to obtain group consensus information; using the group consensus information as prior information, extracting joint consensus information representing the relationship between the second content and individual opinions using an attention model to be trained; fusing the joint consensus information and the second user features; and determining a second predicted value of the popularity of the second content based on the fusion result; and adjusting the parameters of the attention model based on the second predicted value and the label value to obtain the popularity prediction model. Applying this scheme can improve the accuracy of popularity prediction.

This disclosure provides an automatic translation method, apparatus, device, and medium for deep learning code. The method includes: acquiring source code written on a deep learning framework based on a first heterogeneous computing hardware; parsing the source code to extract its lexical sequence and program dependency graph; inputting the lexical sequence and program dependency graph into a graph-aware cross-attention Transformer model to generate a set of candidate target code that satisfies constraints of lexical consistency, structural consistency, and behavioral consistency; performing API-level replacement and reconstruction on the set of candidates through operator alignment mapping and symbol synthesis search; performing multi-dimensional consistency verification on the reconstructed candidate code; and outputting the verified target code. According to embodiments of this disclosure, the migration of deep learning code between different heterogeneous computing hardware eliminates the need for manual reconstruction and debugging, significantly shortening the adaptation cycle and reducing human resource investment and professional knowledge threshold.

The application provides a model training method, medium, equipment and program product of a transverse mixed attention mechanism. The method comprises: acquiring a data set containing multiple sample sequences, each sample sequence in the data set being composed of multiple Tokens arranged in sequence after tokenization processing; constructing a to-be-trained model based on a pre-trained full attention model and adding new parameters for linear attention calculation; in the same transverse mixed attention layer, performing full attention calculation on a Token set located within a preset full attention calculation range, performing linear attention calculation on all Tokens, and fusing the results of the two to obtain transverse mixed attention output for forward reasoning and loss calculation; based on the output and a prediction result, only updating the new parameters to optimize the to-be-trained model until the to-be-trained model converges. The application reduces the computational complexity and memory occupation of long text sequence processing, and improves the reasoning speed and resource utilization rate.

This invention belongs to the field of data processing, and particularly relates to a deep learning-based visual feature extraction and matching method, comprising: acquiring temporal images and point cloud data to obtain a modal confidence mask with dynamically determined environmental perception indicators; inputting the mask, images, and point clouds into a cross-modal attention model to generate a fused spatiotemporal feature sequence; inputting the sequence into a shared weight network to output the current frame pose and implicit map tokens; selectively adding the tokens to a keyframe set according to a first preset condition; extracting environmental feature vectors to update the mask according to a second preset condition; inputting all keyframe tokens back into the shared weight network for inversion optimization according to a third preset condition to output a globally consistent map reconstruction state quantity; feeding back its intermediate network cache to the forward propagation link to update the historical context; iterating until the path is completed, and generating the final implicit map based on the map reconstruction state quantity output by the last global optimization.

A system and method of providing an interpretable prediction of a condition of a subject may include receiving data including values of patient parameters, representing a subject's cunent physiology, and calculating an embedding matrix, representing said data in an embedding space. The embedding matrix may be processed through a cascade of stages. Each stage may include a respective Machine-Learning (ML) based, context-aware attention model, configured to generate an ad-hoc significance matrix pertaining to that stage. Embodiments may aggregate the ad-hoc significance matrices of these stages, to obtain a global significance matrix, representing contribution of each of said parameters in predicting the subject's condition. Embodiments may subsequently apply an ML-based regression model on the global significance matrix, to predict the condition of the subject, and provide the global significance matrix as an interpretation of that prediction.

The application discloses a low-light image enhancement method and system based on an illumination-guided convolution attention model, and belongs to the field of computer vision. The method and system solve the problems that the existing low-light image enhancement technology is difficult to establish global correlation, leading to noise amplification and detail loss, complex calculation and weakening of local details, slow reasoning and poor local recovery effect, and thus restricts the quality and efficiency of image enhancement. The method comprises the following steps: S1: inputting a low-light image and constructing illumination prior information; S2: using an illumination estimator to estimate an illumination feature map and an illumination map according to the illumination prior information, and constructing an illumination image; S3: using an illumination restorer to process the illumination image to generate an illumination restoration image; and S4: generating an enhanced image according to the illumination restoration image and the illumination image. The application is suitable for a low-light image enhancement scene.

This disclosure provides a method and system for online intent recognition of unmanned swarms based on a self-attention model. First, depth images collected by the onboard equipment of the unmanned system are acquired, and the 3D bounding box of each target at the current moment is determined as the current detection target. Prior predictions are obtained based on the historical trajectories of the tracked unmanned swarm targets. The similarity between the current detection target and the prior predictions is measured from the target position, 3D shape, point cloud quantity, and spatial location to achieve cross-frame data association. A Transformer model based on a multi-head self-attention mechanism is used to process the associated historical trajectory sequence, jointly outputting the probability distribution of multiple intents of the swarm targets, and the short-term predicted position and bounding box size for each intent. This invention can achieve joint intent-trajectory prediction when facing concurrent multi-target swarms; and through an effective matching mechanism, it reduces the impact of interference on cross-frame data association, effectively reducing the false positive and false negative rates.

The application provides a multi-view relationship network chart question and answer method based on an attention mechanism, and the method comprises the following steps: S1, acquiring a data set required to be processed; S2, inputting a chart image and a corresponding text question in the data set as input items respectively; S3, configuring a fusion reasoning algorithm model to output a final result. By improving the image encoder model of the traditional relationship network, an effective transformer attention model CoT module is introduced, and the problem that the extraction capability of RN for image feature information is limited is solved; a novel multi-view relationship module is proposed, the pairing process is improved based on pixel points and channel information, the problem that RN regards all feature vectors as equally important without highlighting more effective relationship image feature pairs is solved, and the problem that the overall information of each channel of the image is ignored in the traditional RN pairing process is solved.

The application relates to an AIGC-based sports commentary generation method, which comprises the following steps: data cleaning, standardization and feature extraction on multi-source sports event data; integrating a sports knowledge base and constructing a dynamically updated knowledge graph; based on a pre-trained large language model, knowledge enhancement fine-tuning is carried out by fusing the extracted features and the knowledge graph; the knowledge enhancement fine-tuning adopts a multi-head cross-modal attention model to fuse semantic features, visual features and time sequence features, and optimizes the large language model through an optimization objective function; real-time event data is input, then the large language model is guided to generate commentary text through prompt word engineering, and style adaptation and emotion adjustment are carried out; and the large language model is optimized by adopting a three-level pipeline cascade architecture. The application realizes self-adaptation of event rhythm, improves the accuracy of emotion simulation, realizes ultra-low delay output of a large-parameter large language model in a sports commentary scene, and ensures the real-time performance of live broadcast.

This invention discloses a real-time optimization method for multimodal photoacoustic tomography based on ultrasound image features and deep learning, comprising: acquiring ultrasound image data; inputting the ultrasound image data into a ResUNet network for training to obtain a pre-trained residual model, wherein the pre-trained residual model is used to output a real-time photoacoustic image after light intensity correction; acquiring raw sine wave data; inputting the raw sine wave data into a U-net network for training to obtain a pre-trained attention model; and inputting the real-time photoacoustic image into the pre-trained attention model to obtain optimized photoacoustic image data. This invention solves the problems of artifacts and noise interference, insufficient contrast of deep blood vessels, inability to restore the true size of blood vessels, and inability to provide real-time imaging optimization in existing linear photoacoustic tomography.

The present application belongs to the field of data processing, and particularly relates to a visual feature extraction and matching method based on deep learning, comprising: collecting time sequence images and point cloud data, and obtaining a modal confidence mask dynamically determined according to an environment perception index; inputting the mask, the images and the point cloud into a cross-modal attention model to generate a fused space-time feature sequence; inputting the sequence into a shared weight network to output a current frame pose and an implicit map token; selectively adding the token into a key frame set according to a first preset condition; extracting an environment feature vector to update the mask according to a second preset condition; inputting all key frame tokens into the shared weight network again for inversion optimization according to a third preset condition to output a globally consistent map reconstruction state quantity; feeding back an intermediate cache in the network to a forward propagation link to update a historical context; iterating until a path is completed, and generating a final implicit map according to a map reconstruction state quantity output by the last global optimization.

This invention provides a general EEG signal representation learning system based on a dynamic multi-scale mechanism, comprising a data preprocessing and enhancement module, a multi-scale block embedding module, and an encoder incorporating a dynamic decomposition attention mechanism, connected sequentially. Training is performed using a hybrid generative-discriminative self-supervised pre-training framework. The data preprocessing and enhancement module segments and standardizes the input continuous EEG signal, and applies a probability enhancement strategy to generate two enhanced views of the same signal. The multi-scale block embedding module processes the input signal using at least two sets of parallel convolutional kernels of different scales to generate an initial latent representation. The encoder incorporating the dynamic decomposition attention mechanism sequentially performs temporal causal modeling and dynamic channel grouping spatial attention modeling on the initial latent representation. This technical solution is adaptable to different acquisition devices, captures multi-scale time-frequency features, and possesses excellent linear separability.

A method of compositional feature representation learning for video understanding is described. The method includes individually processing a sequence of video frames received as an input of a feature map network to generate a plurality of feature maps. The method also includes binding the plurality of feature maps to a fixed set of slot variables using an attention model according to a motion segmentation signal. The method further includes combining slot states corresponding to the fixed set of slot variables into a combined feature map. The method also includes decoding the combined feature map to form a reconstructed sequence of video frames, in which objects discovered in the reconstructed sequence of video frames are identified.

A method and system for constructing a time-series attention model for electricity prices that integrates grid congestion information is disclosed. The method includes: collecting grid state, market state, and electricity price data, and constructing a dataset using window partitioning; building a grid congestion detection and vector fusion module to determine grid congestion status and generate congestion state feature vectors, fusing these feature vectors with grid state vectors and market state vectors to form exogenous variables; constructing a time-series attention network to perform feature mapping on exogenous variables, embedding vectors into endogenous and exogenous variables, and obtaining electricity price prediction results after fusion using a dual attention mechanism; training and validating the grid congestion detection and vector fusion module and the time-series attention network; and testing the grid congestion detection and vector fusion module and the time-series attention network validated in step 4 using a test set. This invention can improve the accuracy and stability of electricity price prediction under complex operating scenarios.

The application relates to the technical field of network marketing, in particular to an online marketing promotion system and method based on the Internet, which comprises the following steps: real-time monitoring of a first interaction behavior sequence of a user on a browsing interface, wherein the first interaction behavior sequence comprises a cursor moving track, a page scrolling speed and a scrolling direction; prediction of a current attention focus area and a potential migration path of the attention focus of the user through a predefined attention model based on the first interaction behavior sequence; selection of adapted target advertisement content from an advertisement library based on the potential migration path, and control of the target advertisement content to appear at a preset position on the potential migration path. Through the current focus area position and the historical scrolling direction, candidate content blocks that the user may pay attention to next are predicted, and a visual path between the two is taken as a core basis for advertisement delivery, so that the advertisement can be arranged on the path in advance in the process of natural migration of the user's attention, and the effective reach rate of the advertisement is improved.

This invention discloses a prefix caching method and apparatus suitable for linear attention model architectures, introducing a linear attention state snapshot mechanism based on hash indexes. During the capture phase, the system extracts the convolution and loop states of linear attention at block boundaries as snapshots and stores them in a mapping table associated with the block prefix hash value. During the recovery phase, runtime slots are allocated for new requests that hit the prefix, and the snapshots are written in parallel to the global state tensor by index using a state scattering kernel engine, thereby skipping the linear attention computation for already hit prefixes. The snapshot pool uses a least recently used strategy for capacity management and automatic eviction, and performs binding validity checks before pre-filling computation. If a snapshot is asynchronously evicted, reference stripping and state clearing are used to fall back to the full pre-filling mode to ensure correctness. This method effectively avoids redundant computation, reduces first-word latency, and improves inference throughput, while balancing system stability and resource utilization efficiency.

This application relates to the field of biometric recognition and image processing technology, and discloses a method, model, device, and medium for finger vein recognition based on a lightweight attention model with few samples. The method includes: acquiring raw finger vein images and extracting regions of interest (ROIs); preprocessing the ROI images; performing data augmentation on the training set images; constructing an LA-VGG network based on a VGG-19 network, introducing CA attention, PReLU, and cross-layer fusion to enhance subtle textures; introducing a loss function with additive angular margins during the training phase; training the LA-VGG network to obtain a finger vein recognition model; inputting the image to be recognized into the model, and outputting the identity recognition result. This invention can effectively alleviate the overfitting problem under small sample conditions, reduce the number of model parameters and computational complexity, improve the extraction capability of key finger vein texture features, and is suitable for deployment and application in resource-constrained devices and complex environments.