509 results about "Attention model" patented technology

The invention relates to an immersive VR psychological detection system and method based on multi-modal AI. The system comprises a data acquisition and processing module which is used for acquiring a multi-modal data set of a user in a virtual reality scene based on unified clock synchronization, performing time-space alignment and noise reduction standardization processing on the multi-modal data set, and extracting key biological characteristics. And the correlation model construction module performs space-time correlation mapping through a spatial transformation network, constructs a three-dimensional space attention model, and generates a real-time fluctuation curve after inputting the key biological characteristics into the trained model. And the state report generation module identifies a real-time fluctuation curve by using a time sequence analysis model, performs backtracking analysis in combination with the psychological state conversion node and a multi-modal cross validation result, and finally generates a three-dimensional interactive report. By adopting the method, multi-modal data fusion can be realized, the dynamic change of the psychological state of the user can be effectively captured, the psychological state of the user can be comprehensively and deeply analyzed, and a scientific basis is provided for psychological health assessment and intervention.

The invention discloses a multi-language object illusion relieving method based on a cross-language attention mode, solves the problem of how to relieve multi-language object illusion during detection of a large visual language model under non-English questions, and belongs to the technical field of multi-mode questions and answers. The method comprises the following steps: identifying a cross-modal attention head set of which a visual language model shows obviously different behaviors for English and a target language when same semantic questions in different languages are processed; constructing image description queries of English and target languages for the same image, respectively inputting the image description queries into respective visual language models for reasoning, obtaining attention output under the English and target languages, and taking an average difference between the attention output as a language migration vector of the target languages; and in the reasoning process of the target language question, intervening the attention heads in the attention head set by using the language migration vector, so that the visual understanding ability of the visual language model under the non-English question is closer to the English question.

The invention relates to a dexterous hand multi-mode motion trail prediction model generation method and a related device, which are used for a dexterous hand simulation grabbing process, and the method comprises the following steps: generating a motion trail of a dexterous hand according to initial position information of the dexterous hand and end position information of a target object; based on the plurality of motion track points, a multi-modal data set of motion of the dexterous hand in the motion track is obtained, and the multi-modal data set comprises environment image data, first multi-modal data and second multi-modal data; performing multi-modal feature fusion on the environment image data, the first multi-modal data and the second multi-modal data to obtain an input sequence; and based on a preset sparse self-attention mechanism, inputting the input sequence into a to-be-trained self-attention model for training to obtain a motion track prediction model. According to the scheme provided by the invention, the real-time performance and the accuracy of the motion trail prediction of the dexterous hand can be improved through the fusion of the multi-modal data, and the grabbing success rate of the dexterous hand is improved.

The invention relates to a geological analysis technology, and discloses a multi-head attention model training method fused with geological rules, which is used for improving the scientificity and accuracy of geological body model prediction and improving the interpretability of a geological analysis process. According to the scheme, firstly, feature vectors of geological attribute data and geological coordinate data are spliced to obtain a fusion feature vector, then spatial locality and directivity are considered at the same time, geomorphic multi-head attention calculation is conducted on the fusion feature vector, and an attention feature sequence is obtained; then extracting multi-scale features from the input geological map, constructing a cross-fault weight mask matrix based on fault constraints, and obtaining comprehensive features through fusion; calculating a loss value by adopting a joint loss function containing prediction loss and fault non-penetrability constraint terms, and updating model parameters through a back propagation algorithm to complete model training; and finally, drawing an attention thermodynamic diagram for visual display, and superposing the attention thermodynamic diagram with the three-dimensional geologic model.

The invention discloses a code generation method based on dynamic hierarchical sparse attention, and belongs to the technical field of data processing, and the method comprises the following steps: S1, collecting an input data set; s2, constructing a dynamic hierarchical sparse attention model; s3, optimizing the dynamic layered sparse attention model by using the input data set, and generating a final dynamic layered sparse attention model; and S4, generating a code by using the final dynamic hierarchical sparse attention model. According to the method, the model can still keep high-quality code generation capacity at extremely low labeling cost, and a flexible and efficient solution is provided for agile development and automatic testing.

The invention relates to a cultural scenic spot people flow prediction and scheduling method based on machine learning, and belongs to the field of scenic spot management. 2, constructing and training a mixed attention model, wherein the constructed mixed attention model comprises a spatial-temporal feature extraction module and a mixed attention mechanism module; step 3, people flow prediction and result output: inputting real-time data into the trained mixed attention model to carry out people flow prediction, and outputting a future time period passenger flow volume value; a quantile regression technology is combined to output a confidence interval of the passenger flow volume; step 4, scheduling strategy implementation: constructing a dynamic thermodynamic diagram based on a prediction result, and marking crowd density levels of different areas in the thermodynamic diagram; early warning information is pushed in real time, and touring suggestions are sent to tourists in advance.

A method estimating a depth of an environment includes generating, via a cross-attention model, a cross-attention cost volume based on a current image of the environment and a previous image of the environment in a sequence of images. The method also includes generating, via the cross-attention model, a depth estimate of the current image based on the cross-attention cost volume, the cross-attention model having been trained using a photometric loss associated with a single-frame depth estimation model. The method further includes controlling an action of the vehicle based on the depth estimate.

The invention belongs to the related technical field of image processing, and discloses a texture surface defect generation method and system based on a mask perception image redrawing network, and the method comprises the steps: (1) obtaining a low-frequency structure feature based on a defect sample image of a to-be-processed product category, a binary mask for marking a defect position, and a convolutional coding module; (2) a context sliding window attention modeling module carries out long-range dependence modeling on the feature sequence, and attention calculation is carried out between effective feature marks to obtain high-frequency detail features; (3) inputting the low-frequency structural features and the high-frequency detail features into a collaborative feature fusion module for fusion to obtain fusion features; (4) a style modulation decoding module performs up-sampling on the fusion features based on the comprehensive style tensor; and (5) generating a defect image based on the trained generator network model, the defect-free sample image of the to-be-processed product category and a user-defined defect mask. According to the invention, the authenticity of defect generation is improved.

The embodiment of the invention provides a multi-modal semantic understanding method and device based on multi-order progressive alignment, computer equipment and a storage medium, and the method comprises the steps: inputting each kind of received data into a corresponding preset encoder, and grouping at least two kinds of output data features, inputting each data feature group into a preset coarse alignment module, determining a matching score corresponding to each data feature group, and screening the data feature groups of which the matching scores exceed a preset threshold value; constructing a graph attention network, inputting each screened data feature group into the graph attention network, and iterating the output first fusion feature; inputting the iterated features into a preset semantic correction module, inputting the output aligned features into a preset multi-modal attention model, and generating second fusion features of all feature groups, through a three-stage dynamic alignment mechanism of a coarse alignment module, a graph attention network and a semantic correction module, the problem of serious semantic information loss in existing cross-modal semantic understanding is solved.

The invention provides a data center interconnection link fault self-healing and path switching method, which comprises the following steps of: acquiring multi-dimensional link performance monitoring data, establishing a standardized time sequence database, realizing link health degree multi-index trend prediction by combining an LSTM-Attention model, and determining the link health degree according to the LSTM-Attention model. Risk-aware dynamic path selection and switching for multiple business scenes are realized by combining business service sensitivity modeling, path inherent stability quantification, nonlinear adaptive switching criteria and a debounce switching process, the response accuracy and the business adaptation degree of path switching are improved, false triggering and service interruption risks can be remarkably reduced, and the path switching efficiency is improved. And the high availability and robustness of routing in a complex network environment are enhanced.

The invention discloses a power consumption time sequence prediction method and system based on a self-attention mechanism, a medium and a processor, and relates to the technical field of power consumption time sequence prediction. The method comprises the steps of obtaining original data, preprocessing the original data, dividing the original data into a training input set and a label set, extracting time and power unit features through convolution operation, calculating and fusing the features through double attention, obtaining a prediction value through convolution operation, optimizing model parameters, and finally predicting a future power consumption value based on an optimization model. The system comprises a preparation module, a first operation module, a fusion module, a second operation module and a prediction module. A computer readable storage medium and a processor store and run programs for executing the method, respectively. According to the method, the nonlinear features and the long-term dependency relationship of the data are effectively captured through a self-attention mechanism, the feature extraction capability is improved through a double attention model, and the prediction precision is improved.

The invention relates to the technical field of information security management, in particular to a multi-modal data dynamic desensitization method and system based on machine learning, multi-modal features such as texts, images and time sequences are jointly extracted through CNN-Transform-LSTM, 1-5 sensitive levels are evaluated by adopting a cross-modal attention model, a self-adaptive desensitization strategy is generated through an improved MOEA / D-DE algorithm, and the dynamic desensitization of the multi-modal data is realized. And realizing cross-modal collaborative desensitization by combining GNN. The system dynamically adjusts parameters through TD3 reinforcement learning, the response delay is less than or equal to 100ms, the Bi-LSTM-AE model evaluates privacy and utility in real time, and the block chain evidence storage whole process is realized. The method solves the problems of precision and utility imbalance, poor dynamic adaptability and insufficient multi-mode collaboration in the traditional technology, and is suitable for the fields of finance, medical treatment and the like.

The invention relates to a GIS (Gas Insulated Switchgear) equipment mechanical defect diagnosis method based on a Gramb angle field and a dual-channel PCNN-Attention neural network, and belongs to the technical field of gas insulated switchgear mechanical vibration defect diagnosis. The method solves the problems that traditional diagnosis depends on artificial feature extraction, so that subjectivity is high, information mining is insufficient, and defect severity evaluation is missing. According to the technical scheme, the method comprises the steps that a one-dimensional vibration signal is converted into a GASF two-dimensional image and a GADF two-dimensional image through a GASF field so as to completely reserve time sequence topological features; carrying out data enhancement by adopting an image geometric transformation technology so as to improve the generalization ability of the model; and a dual-channel PCNN-Attention model is constructed, and synchronous intelligent identification of defect types and severity is realized through parallel feature extraction and dynamic weight optimization of an attention mechanism. According to the method, the diagnosis accuracy, reliability and adaptive capacity are improved, and support is provided for equipment state operation and maintenance.

The invention provides a power grid data anomaly detection method and system based on a hybrid deep learning model, and belongs to the technical field of power grid data processing. According to the method, a convolutional neural network is combined with a recurrent neural network, a gating circulation unit, a long-short-term memory network and an attention model to carry out hybrid modeling, an initial detection model is obtained, and meanwhile, a model collaborative optimization mechanism is adopted to carry out distributed parallel training on the initial detection model; constructing an adaptive optimizer through an adaptive optimization algorithm, and dynamically updating model parameters and a model structure by using the optimizer and a dynamic training mechanism to obtain a hybrid detection model; and inputting the real-time data stream of the power system into the hybrid detection model for anomaly detection, and executing an active security protection operation on the power system according to the detected abnormal data and potential attack behaviors according to a preset defense strategy. According to the invention, accurate detection and real-time early warning can be carried out on complex abnormal behaviors of the power system.

The invention discloses a method and equipment for eliminating network environment defects and self-healing a closed loop based on RAG and a medium, and the method comprises the steps: constructing a multi-modal knowledge base, and carrying out the regular updating based on a large-scale language model; performing fusion analysis by using a long-short term memory attention model and a graph neural network model, capturing time sequence and topology anomalies, generating a structured early warning report, and associating the structured early warning report with a knowledge base; a three-level retrieval strategy is adopted, and emergency, radical treatment and prevention schemes of multi-objective optimization are extracted from the knowledge base; utilizing an improved time sequence directed acyclic graph algorithm to dynamically select parallel self-healing operation; and automatically adjusting the knowledge base based on the self-healing execution result and the feedback index. According to the method, through deep fusion of the dynamic multi-modal knowledge base and the retrieval RAG technology, normal form transition of equipment defect elimination from passive response to active prediction is achieved, compared with a traditional method, the defect positioning accuracy is greatly improved, and the problem of misjudgment caused by single data source analysis is solved.

Systems and methods are disclosed for efficient memory allocation for processing large language model encoders and decoders based on an attention model. The system can utilize a plurality of two types of processors suitable for different types of LLM processing. These include Neural Processor Units and Graphic Processing Units. Each NPU processor has dedicated DDR memory coupled to each NPU. The DDR memory caches the neural network weights used in the generation of neural network activations. A plurality of GPUs provides KVQ token processing. LLM tokens processing can be performed in parallel batches or sub-batches to utilize idle NPU processors within the neural network. In some embodiments, the NPUs are structured in a matrix with a bus between adjacent processors. In another embodiment, NPUs provide both KVQ processing and neural network processing. The system can be integrated on a silicon substrate using chiplets in a 2.5 or 3-D architecture.

The invention relates to the field of forged image detection, and provides a forged image detection method and device, a medium and a product, and the method comprises the steps: dividing an input image into a plurality of local regions, and constructing a region dependence graph of the local regions; extracting local style features from the region dependence graph by using a local network model; extracting global style features of the input image by using the global network model; performing attention fusion on the local style features and the global style features by using an attention model to obtain a fusion vector; performing image forgery detection on the fusion vector by using a classifier; wherein the local network model, the global network model, the attention model and the classifier are optimized through training. According to the invention, high identification accuracy and robustness can be maintained in detection scenes of scarce samples, local detail forgery and cross-domain forgery.

The invention provides a product recommendation method and device, electronic equipment, a medium and a program product, and can be applied to the technical field of artificial intelligence, the technical field of big data, the technical field of block chains and the technical field of privacy computing. The method comprises the following steps: acquiring structured data and unstructured data of a target user; extracting preference features based on the unstructured data, and obtaining a time sequence preference portrait by using a time sequence attention model; obtaining a target user portrait based on the time sequence preference portrait and the structured data; obtaining a multi-dimensional sequential relation graph, and reasoning the multi-dimensional sequential relation graph by using a graph neural network to obtain relation representation; executing collaborative filtering scoring based on the target user portrait and the relationship representation, and generating an intermediate product score; taking the generated intermediate product score as an input, and outputting a candidate product score based on a recommendation optimization function containing a space-time weight; and performing product recommendation based on the candidate product scores.

The invention provides a model training method for a transverse mixed attention mechanism, a medium, equipment and a program product, and the method comprises the steps: obtaining a data set containing a plurality of sample sequences, each sample sequence in the data set being formed by arranging a plurality of Token sequences obtained through word segmentation; constructing a to-be-trained model based on the pre-trained full attention model, and adding newly added parameters for linear attention calculation; in the same transverse mixed attention layer, executing total attention calculation on a Token set in a preset total attention calculation range, executing linear attention calculation on all Tokens, and fusing results of the total attention calculation and the linear attention calculation to obtain transverse mixed attention output used for forward reasoning and loss calculation; and based on the output and prediction result, only updating the newly added parameters to optimize the to-be-trained model until the to-be-trained model converges. According to the method, the calculation complexity and video memory occupation of long text sequence processing are reduced, and the reasoning speed and the resource utilization rate are improved.

The invention relates to a port container dispatching management method and device and a medium. The method comprises the steps that space-time joint features of a storage yard are acquired in real time, and the space-time joint features comprise space state features and time sequence features of the storage yard; inputting the spatio-temporal joint features into a self-attention model, and outputting key spatio-temporal features of the storage yard through the self-attention model; and inputting the key spatial-temporal characteristics and the current state information of the storage yard into a deep learning model, and outputting a target scheduling action about the storage yard through the deep learning model which is obtained based on dynamic reward mechanism training optimization. The scheduling strategy of the port container yard can be optimized by integrating a multi-target dynamic reward mechanism.

The invention provides an intrusion detection method based on boundary sensitive federated expert multi-modal detection, which comprises the following steps of: splicing, synthesizing and fusing seven types of discriminative characteristics based on original traffic characteristics, designing a multi-modal collaborative attention model MultiModalFusion, dividing the characteristics into four modals, namely a protocol state, a traffic behavior, statistical distribution and a connection relationship, and realizing cross-modal information interaction by utilizing dynamic weight learning. In order to solve the problem of data imbalance, a boundary sensitive condition generator BSGenemator is developed to guide generation of minority class samples through a dynamic boundary strategy in combination with a composite loss function method. And finally, constructing a federal element strategy expert committee, dynamically fusing decisions of four experts by adopting a learnable strategy network, and verifying the characteristic contribution degree through an SHAP interpretable module. And finally, the efficiency of the scheme is verified by using a data set UNSW-NB15, through comparison of multiple schemes, the scheme has significant accuracy, the weighted average F1 score is improved, and a new normal form is provided for a real-time intrusion detection scheme.

The invention relates to the technical field of user portrait information processing, and particularly provides a user real-time portrait updating method and system, and the method comprises the steps: collecting a time sequence interaction track sequence generated by a user in a current interaction scene in real time, and constructing a heterogeneous behavior data flow based on a cross-modal behavior synchronous collection technology; detecting the deviation degree of the historical behavior sequence in the unified feature space, and generating a conflict state mode label; analyzing a behavior evolution path by applying a time sequence self-attention model, and outputting multi-intention probability distribution; dynamically adjusting a confidence coefficient threshold value based on a conflict state and filtering out a noise portrait label; and triggering cross-modal joint recoding and intention re-reasoning for a prediction deviation time period, and locally updating the portrait through a feedback closed loop. Static portrait updating lag limitation is broken through, second-level intention tracking and dynamic behavior mode capturing are achieved, and portrait adaptability and storage efficiency in a high-noise scene are remarkably improved.

The invention relates to the technical field of intelligent traffic systems, in particular to a highway network global traffic flow prediction method, device and equipment, and the method comprises the steps: constructing an initial feature sequence based on the multi-source data of an ETC portal system; inputting the data into a multi-head attention model driven by a query mechanism, and outputting a spatial dependency relationship between ETC door frame nodes; inputting the spatial dependency relationship into an extraction fusion model fusing a causal convolutional network and a selective state space network, extracting short-term time sequence features through the causal convolutional network, modeling long-term time sequence features through the selective state space network, and fusing the long-term time sequence features into fused time sequence features; and a global traffic flow prediction result is generated based on the fusion time sequence features, so that the problems of high traffic flow prediction cost, data coverage missing and the like in related technologies are solved.

The invention relates to an external parameter online calibration method for a laser radar and a camera in a cross-country environment, and the method comprises the steps: constructing an external parameter calibration network MLLANet based on a Mama-like linear attention model, taking a camera RGB image and a laser radar initial depth map as input, extracting multi-scale features through multistage convolution and a Mama-like attention module in sequence, and carrying out the multi-scale feature extraction; and feature matching is realized through multi-resolution feature fusion and cost volume construction, and translation and rotation parameters are respectively predicted by a global regression layer. According to the network, regression loss and depth map loss based on Huber are introduced into a loss function part, and a mixed loss system is constructed, so that training robustness is enhanced, and the influence of point cloud outliers is inhibited. And gradually reducing the external parameter search range through multi-stage iterative optimization, and finally realizing high-precision estimation of the six-degree-of-freedom external parameters. The method does not need a manual calibration board, can still keep high precision and real-time performance in a cross-country unstructured environment, and can be widely applied to environment perception and autonomous navigation tasks of an unmanned system.

The invention provides a vehicle path planning method and system based on a dynamic attention model, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining instance data containing warehouse nodes, client nodes and vehicle capacity information; a dynamic encoder-decoder architecture model is constructed, and an encoder recalculates feature embedding of unaccessed nodes every time a vehicle returns to a warehouse based on a graph attention network so as to respond to dynamic changes of instance states in the path construction process; the decoder selects a next node to be accessed according to the current path state, residual node embedding and vehicle real-time capacity; a REINFORCE algorithm is adopted to take the minimum path length as a target training model, and a greedy strategy baseline is introduced to reduce variance; and finally, the initial path is optimized by combining the local search of the 2OPT, and a final planning scheme is output. According to the method, the problem that a traditional attention model cannot adapt to a dynamic environment due to fixed node embedding is effectively solved, the total cost of a path is reduced, and good generalization and expandability are achieved.

The invention relates to the technical field of aviation safety prediction, and discloses a bumping prediction correction method and system based on aircraft geometric configuration and shipping space information, electronic equipment and a storage medium, and the method comprises the steps: obtaining input features, carrying out the cross-aircraft-type feature enhancement through comparative learning, generating an enhanced feature vector, and carrying out the recognition of the enhanced feature vector; and performing structured combination on the enhanced feature vector, the distance from the shipping space to the lift center, the normal overload value of the shipping space and the distance from the target seat to the lift center, fusing model embedded vectors and spatial position codes, generating a plurality of feature marks, and inputting the feature marks into an attention model adjusted by a geometric perception mechanism to perform spatial dependency relationship modeling. Outputting morphological parameters for representing an overload distribution form; through parameterized curve constraint calculation, obtaining a preliminary normal overload prediction value of the target seat; and performing adaptive weighted fusion based on the rule constraint prediction value, and outputting a normal overload prediction correction value of the target seat. According to the invention, the overload value of any target seat in the cabin can be accurately predicted.

The invention discloses a step-type landslide-oriented multi-source fusion deep learning deformation intelligent prediction method, and belongs to the technical field of geological disaster monitoring and early warning, and the method comprises the following steps: S1, constructing and preprocessing a landslide data set of a research region; s2, decomposing displacement data based on a CEEMDAN algorithm; s3, screening key influence factors through grey correlation analysis; s4, constructing a TCN-Attention model to respectively predict a trend term and a period term; and S5, superposing prediction results and carrying out precision evaluation. According to the multi-source fusion deep learning deformation intelligent prediction method for the step-type landslide, the problems of modal aliasing and energy loss of a traditional method are reduced, and the signal decomposition effect is optimized; and the long-term deformation trend is accurately captured, the displacement abrupt change point is identified, and the prediction precision is improved.

The invention relates to an HRRP unsupervised identification method, device and equipment based on a comparative learning fusion model. The method comprises the following steps: constructing an adaptive dynamic attention model based on a comparative learning framework; and inputting the HRRP signal sample into a dynamic mixed attention network, and performing shallow feature extraction, middle attention processing and fine-grained modeling in the dynamic mixed attention network to output multi-scale features. And designing multi-scale statistics according to the multi-scale features to adaptively adjust the attention weight so as to output the residual error of the dynamic mixed attention network. And inputting the residual error into a self-adaptive comparison loss module for loss calculation, evaluating the calculated loss through a preset evaluation module to obtain a prediction category of the self-adaptive dynamic attention model, and training the model according to the prediction category of the self-adaptive dynamic attention model. And identifying a to-be-identified HRRP signal by using the trained model to obtain a target image. By adopting the method, the target identification precision can be improved during label-free processing.

The invention relates to an AI dynamic repayment behavior monitoring and early warning method and system, and belongs to the field of financial risk control and artificial intelligence. Multi-dimensional time sequence features are constructed by fusing repayment behaviors, home decoration progress and text interaction information, dynamic risk prediction is realized by using an LSTM-attention model, grading early warning is triggered in combination with a home decoration stage adaptive threshold, a scenarized intervention strategy is generated according to a risk grade, intervention feedback is used for optimization of the model and the threshold, and the risk prediction efficiency is improved. And a closed-loop mechanism of post-loan risk identification and intervention is formed. According to the invention, real-time risk monitoring and accurate intervention of the home decoration staged loan can be realized, and the early warning accuracy and the post-loan management efficiency are effectively improved.

The invention discloses an environment gas leakage intelligent detection method based on big data analysis. The method comprises the following steps: collecting monitoring data and constructing a sensor network space-time diagram fused with meteorological conditions; predicting the background concentration of each node by using a sequence decomposition attention model; calculating a concentration residual error and constructing a space-time residual error map; extracting an abnormal feature vector through a physical information enhanced graph attention network; calculating an abnormal score in combination with the normal mode memory bank and generating a leakage source probability distribution diagram; when the conditions are met, calling a computational fluid dynamics model to simulate a theoretical concentration field; and determining a leakage event through spatial correlation analysis. According to the invention, the false alarm rate of gas leakage detection in an open environment is obviously reduced, and the leakage source positioning precision is improved.