140 results about "Temporal modeling" patented technology

The invention relates to the technical field of network security, and particularly discloses an edge device network threat detection method and system based on an electric power large model, and the method comprises the steps: capturing a network message sequence in real time, extracting a time sequence randomness feature and a semantic deviation feature from a time dimension and a protocol dimension, and carrying out the fusion to form a comprehensive threat feature vector; performing multi-dimensional feature analysis and time sequence modeling by adopting a lightweight electric power large model to realize millisecond-level threat assessment; establishing a multi-level response mechanism, dynamically triggering a differential protection strategy according to the threat level, and ensuring the reliability and consistency of response actions through digital signature and collaborative verification; according to the method, the complex network attack in the power edge equipment can be effectively identified, the threat detection accuracy and the system defense capability are improved, and the strict requirements of a power system on real-time performance and reliability are met.

The invention discloses a continuous time dynamics prediction method, system, terminal and medium fusing a diffusion model and a graph frequent differential equation, and relates to the technical field of dynamics prediction.The method comprises the steps that a multi-node time sequence is obtained, network structure inference is conducted on the multi-node time sequence through the diffusion model, and a potential graph structure between nodes is obtained; carrying out continuous time dynamic modeling by adopting a Scheng ordinary differential equation, and predicting a node state at any time point; diffusion reconstruction loss, dynamic prediction errors and structure sparsity constraints are constructed, total loss is established, network structure inference based on a diffusion model and dynamic modeling based on a Shenzheng differential equation are coupled based on the total loss, and collaborative training optimization is achieved. According to the method, the potential graph structure of the system can be stably recovered in a complex noise environment, high-precision and continuous prediction can be carried out on dynamic evolution of the potential graph structure, and the limitation that structure inference and continuous time modeling cannot be considered in the prior art is overcome.

The invention discloses a global dynamic perception space-time traffic flow prediction method based on graph packet representation learning, and belongs to the technical field of traffic flow prediction, and the method comprises the following steps: S1, traffic data input, S2, traffic graph packet construction, S3, graph packet initial feature extraction, S4, time sequence feature extraction, S5, spatial feature extraction, and S6, traffic flow prediction and output. Through a space-time modeling technology of graph packet representation learning and global dynamic perception, space-time characteristic elements of a traffic road network can be comprehensively covered, traditional traffic indexes such as flow and speed are concerned, elements such as road network topological association and cross-regional multi-hop association are also included, a dynamic dependency relationship between a time sequence and a spatial dimension is deeply mined, and a real-time dynamic perception effect is achieved. Therefore, the prediction result can reflect the real evolution law of the traffic flow more accurately, and a more scientific basis is provided for traffic management and decision making.

The invention provides a traffic flow prediction method combining a dynamic graph neural network and a Mama mechanism. The future short-term traffic flow is predicted by using historical traffic data. According to the method, firstly, normalization preprocessing is carried out on traffic state data collected by multiple sensors, a training sample is generated by adopting a sliding window, and a traffic flow value in the next one hour is predicted according to data in the past 24 hours. On the basis of the model structure, a time modeling module composed of multiple layers of MambaBlocks is constructed and used for capturing historical time sequence dependence; constructing a spatial modeling module of dynamic graph convolution, and combining a static adjacency matrix of the road network with a learnable adaptive adjacency structure to extract spatial association; and finally, the outputs of the modules are fused, and a prediction result is obtained through a prediction output module. In the training process, a Huber loss function is used as an optimization target, and evaluation indexes such as a mean absolute error (MAE), a root mean square error (RMSE) and a mean absolute percentage error (MAPE) are used for evaluating the performance of the model. According to the method, the traffic space-time dynamic characteristics are effectively mined, and the long-range dependence modeling capability and the prediction precision are improved.

The invention provides a multi-modal semantic fusion image and text relevance dynamic analysis method and system, and belongs to the technical field of multi-modal data processing. The method comprises the steps of image and text data preprocessing, feature extraction, fusion semantic vector generation through a bidirectional cross attention mechanism, dynamic relevance score calculation through a time sequence attention long-short-term memory network and combined loss function end-to-end training. According to the method, accurate alignment of image and text features can be realized through a bidirectional cross attention mechanism, the cross-modal matching accuracy is improved, the dynamic correlation analysis capability is enhanced by using time sequence modeling, and the problems of insufficient feature coding suitability and limited time sequence modeling capability in the prior art are solved.

The invention discloses a time-space multi-mode video understanding method based on Video LLeMA2. The time-space multi-mode video understanding method comprises the following steps: S1, extracting an image frame sequence and an audio stream sequence of video data; s2, carrying out pretreatment; s3, inputting the image frame sequence into a visual encoder to generate a visual initial feature; s4, inputting the visual initial features into a space-time convolution connector, and generating visual modal features by using three-dimensional convolution and a RegStage module; s5, inputting the audio stream sequence into an audio encoder to obtain audio modal features; s6, aligning the audio modal features with the visual modal features; s7, performing feature fusion; and S8, obtaining a video content understanding result through a language decoder. The video content semantic understanding method is based on the Video LLeMA2 model, integrates multi-modal space-time modeling and language generation technologies, realizes video content semantic understanding, and has the advantages of natural expression and high precision.

The invention provides a sensor data recovery method and system based on mask perception space-time modeling, and belongs to the technical field of Internet of Things and intelligent sensing data processing. The method comprises the following steps: modeling a missing position into a trainable potential vector representation through an adaptive missing representation module, and avoiding noise introduced by traditional zero filling or mean filling; a missing perception space-time decoding module is adopted, and a dynamic weight distribution mechanism of mask constraint is combined in the decoding process, so that information is effectively prevented from being excessively smooth; designing a space-time dual-channel feature aggregation unit, and capturing spatial dependence and time sequence dependence between the sensors at the same time; and finally realizing accurate completion of a large-scale space-time sensor matrix through an output recovery unit. According to the method, the accuracy and robustness of sensor data restoration can be remarkably improved, the reliability of subsequent monitoring, prediction and anomaly detection is enhanced, and the method has wide engineering application value.

The invention provides a business resource optimization management method and system based on computing power demand prediction, and the method comprises the steps: collecting multi-source heterogeneous data in real time, so as to construct and continuously update a topological graph which represents the dynamic association relationship among a business entity, a logic service entity and a physical resource entity; a time sequence state of a topological graph is input into a prediction model fusing a graph structure and continuous time modeling, multi-dimensional evolution calculation of a node state is carried out by the prediction model, a dynamic resource value signal is generated by a central scheduler under a hierarchical decision framework, a plurality of service unit intelligent agents are guided to carry out distributed collaborative decision, and a multi-dimensional resource value signal is generated by the central scheduler. Iteratively generating an optimal resource allocation scheme; and before the optimal resource allocation scheme is executed, potential risk assessment and credible automatic execution are carried out. Therefore, in a complex dynamic commercial environment, the computing power resource is converted from global perception, accurate prediction and intelligent optimization to risk-controllable closed-loop management and adaptive configuration.

The invention provides an action recognition method based on adaptive skeleton grouping and direction sensitive space-time modeling, and relates to the technical field of skeleton action recognition. Comprising the steps of skeleton action sequence input and feature embedding, adaptive skeleton grouping and direction weighting space-time modeling, trunk modeling and feature refinement, adaptive time down-sampling, multi-stream feature fusion and classification output and action recognition loss evaluation. The method comprises the following steps: acquiring an introduction result through skeleton action sequence input and feature embedding, acquiring an adjustment result through adaptive skeleton grouping and direction weighted space-time modeling, acquiring a processing result through trunk modeling and feature refinement, acquiring a reconstruction result through adaptive time downsampling, performing multi-stream feature fusion and classified output, and finally executing action recognition loss evaluation. According to the method, the skeleton action recognition precision is improved, and meanwhile, the complexity of long sequence modeling calculation is reduced, so that the skeleton action recognition real-time performance and deployment efficiency are improved, and the problem that the skeleton action recognition precision is not high in the prior art is solved.

The invention provides a multi-scene driving risk assessment method based on transfer learning. The method is suitable for real-time assessment and short-time early warning of driving risks in tunnel and non-tunnel environments. According to the method, a multi-source data set collected by multiple sensors is constructed based on a natural driving test, static and dynamic characteristics are extracted by using a sliding time window technology, and the driving risk is evaluated through a multi-scene driving risk evaluation model. The scene driving risk assessment model integrates a static information encoder, a variable selection network and an interpretable multi-head self-attention mechanism, has strong feature selection and time sequence modeling capabilities, and improves the adaptability and generalization capability to scarce tunnel data through a transfer learning strategy. Test results show that the method is superior to the existing mainstream model in the aspects of alarm rate, accuracy rate and false alarm rate, and is suitable for a risk assessment and early warning module in an intelligent traffic system.

The invention discloses a multi-scale time sequence prediction method based on an adaptive sparse expert selection strategy and a closed continuous time neural network. The method comprises the following steps: carrying out normalization and low-dimensional feature mapping based on RevIN; performing trend-seasonal structure enhancement processing on the feature sequence after linear mapping; constructing a multi-scale expert model based on the feature sequence after trend-season enhancement; self-adaptive sparse expert selection and load balancing loss calculation are carried out; carrying out weighted aggregation and residual fusion on multi-scale expert output; global modeling of a closed continuous time neural network based on channel weighting is carried out; and finally performing prediction generation and reverse normalization. The multi-scale time series prediction method has the multi-time-scale adaptive modeling capability, the sparse expert efficient selection mechanism and the global continuous time modeling capability, and can be applied to various multivariable time series prediction scenes such as power load prediction, weather prediction, industrial production monitoring, traffic flow prediction and financial price prediction.

The invention discloses a multi-dimensional network intrusion behavior intelligent identification method based on deep learning. The method comprises the following steps: collecting network multi-dimensional data and generating a standardized network event set and a network control path input set; establishing a neural controlled differential equation model, and generating a continuous time context representation set through hidden state evolution; establishing a neuro-hox process identification model, performing intensity function modeling, and generating an event intensity prediction sequence set and an identification intermediate representation set; forming an intrusion behavior decision rule set and a reasoning configuration set through joint training; new network multi-dimensional data are collected, the reasoning configuration set operation model is loaded, and a new event intensity prediction sequence set and a new candidate trigger time set are output; and generating a network intrusion behavior recognition result set in combination with the intrusion behavior decision rule set. According to the method, time modeling and logical reasoning are fused, and high-precision intrusion identification is realized.

The invention discloses a 3D human body posture estimation method based on a millimeter wave radar, and relates to the technical field of 3D human body postures. The invention provides a symbol perception energy separation mechanism by introducing a SignAware module and aiming at the problem of information annihilation caused by mixing of positive and negative Doppler components in the prior art, and according to the design, inductive bias consistent with a Doppler imaging physical mechanism is injected into a network structure, so that the model is constrained to follow the dynamic law of a speed symbol, and the model is more accurate and reliable. And separable positive and negative subspaces are formed in the representation space. Therefore, cross-symbol interference is reduced, the alignability and interpretability of specific joint movement are improved, and the sensitivity to speed polarity is kept in a complex multi-joint scene; a PhaseFilm module is provided, aiming at the defect that an existing method lacks periodic and sequential modeling, a phase conditional feature re-calibration mechanism is introduced, the two modules are put into an attitude network, then a feature extraction mechanism combining a time domain and a frequency domain is introduced, and a final SPTFormer network is obtained.

The invention relates to the technical field of image recognition, in particular to a low-resolution food packaging image recognition method and system and a medium, and the method mainly comprises the steps of low-sampling physical degradation modeling, pseudo-super-resolution guide structure perception, dynamic recognition trunk processing, semantic prior guide recognition and collaborative loss optimization. Through physical sampling inverse modeling, key structure information, such as characters and bar codes, in a low-resolution image is effectively recovered, and artifact interference is avoided; through dynamic time sequence modeling, the Mama-Core backbone network dynamically captures the local and global relation of the image through state space modeling, the target recognition precision of the low-resolution image is improved, based on the method, the traditional normal form of super-resolution first and then recognition is broken through, and the recognition accuracy is greatly improved.

The invention relates to the field of automatic driving, and particularly discloses a lane line detection method based on time sequence curvature and multi-scale context. The method comprises the following specific implementation steps of: giving a data set image and preprocessing the data set image; lane features of an image are extracted through a backbone network, a multi-scale feature map is generated, and the multi-scale feature map is continuously processed by a double-branch detection framework. According to the framework, two parallel branches of time sequence modeling and attention enhancement are fused, firstly, a feature map is sent into a multi-scale attention branch, multi-scale channels and spatial features are fused, and semantic information is enriched; and meanwhile, the time sequence optimization branch captures prior knowledge by using time sequence information, extracts context information and models curvature change. And finally, carrying out weighted fusion on the output features of the two branches and the original trunk features, processing and refining lane prediction through a loss function, and outputting a final result. According to the method, the problems that visual clues of lane lines are lacked in a complex scene and a long-distance dependency relationship is difficult to model in a curve scene are effectively solved, and the robustness and accuracy of the model are enhanced.

The invention relates to the technical field of information, and particularly discloses a digital human real-time construction method of a personal number, which comprises the following steps: S1, acquiring a user video stream and associated voice data in real time; s2, extracting spatial features of each frame in the user video stream to obtain a spatial feature sequence; s3, performing time sequence modeling on the spatial feature sequence, and outputting a spatial-temporal feature sequence; s4, generating a personalized preference embedding vector based on historical interaction data of the user; s5, inputting the spatial-temporal feature sequence and the personalized preference embedded vector into a generator of the generative adversarial network, and outputting a digital human frame sequence in real time; and S6, converting the voice data into lip motion parameters through a voice driving module, and fusing the lip motion parameters into the digital human frame sequence to generate a final digital human video stream. According to the method, through innovative algorithm fusion, breakthrough is made in the aspects of real-time performance, individuation, reality and efficiency, and a reliable solution is provided for large-scale application of personal number digital persons.

The application discloses a traffic flow prediction method and device based on adaptive cycle time modeling, and the prediction method comprises the following steps: constructing a multi-mode cycle base matrix based on historical traffic flow data; the historical traffic flow data comprises traffic flow data of multiple regions in a historical window length, and the multi-mode cycle base matrix is used for representing shared cycle base modes in the historical window length; the shared cycle base modes comprise at least two traffic flow cycle modes; a global cycle matrix is obtained by linearly combining the shared cycle base modes through an adaptive weight matrix, and the global cycle matrix is used for representing traffic flow modes of each region in the historical window length; and a traffic flow prediction result is obtained based on the global cycle matrix and the historical traffic flow data. The method can improve the expression capability of the model in a cross-region scene and a diversified time cycle, and effectively solves the single mode deviation problem.

The application provides a video summarization method based on multi-dimensional features and fine-grained hierarchical modeling, and relates to the technical field of video processing. In practical application, the video summarization technology can facilitate large-scale video retrieval and browsing. The method comprises the following steps: firstly, frame extraction is performed on an input video to obtain a frame sequence, and a multi-dimensional feature extraction network composed of a 2D network and a 3D network is used to extract multi-dimensional features; then, hierarchical temporal modeling is performed to complete the modeling process of the temporal dependence of the entire video sequence; finally, a regression network is used to obtain the importance score of each frame and generate a video summary. The application further explores the influence of the spatiotemporal features extracted by 3D feature extractors with different spatiotemporal complexities on the video summary result. The application shows excellent performance on the video summary datasets SumMe and TVSum. Whether from the application scene or the performance index, the application has strong practical value.

This invention belongs to the field of campus management and discloses a method for monitoring campus personnel behavior trajectories based on multi-source data fusion, including the following steps: acquiring multi-source data within the campus; preprocessing the multi-source data using a data weighted fusion algorithm; performing differentiated scene optimization processing on the collected facial images to output clear facial image frames and effective feature data; constructing real-time behavior trajectories of campus personnel based on a temporal modeling model using clear facial image frames, effective feature data, and preprocessed multi-source data; analyzing the real-time behavior trajectories through an anomaly detection mechanism to determine the trajectory anomaly status and classify the anomaly level; executing multi-channel alarm push according to the anomaly level; visually restoring the real-time behavior trajectories and historical behavior trajectories based on a campus spatial model; and absorbing new data and anomaly judgment results through an incremental learning algorithm to update the parameters of the temporal modeling model, anomaly detection mechanism, and differentiated scene optimization processing related models.

The invention discloses a reservoir slope displacement prediction method and system, and the method comprises the following steps: obtaining reservoir slope multi-modal data of a target region, and carrying out the preprocessing of the reservoir slope multi-modal data; extracting reservoir bank slope space change characteristics according to the preprocessed reservoir slope multi-modal data, and determining reservoir bank slope displacement according to the reservoir bank slope space change characteristics; based on the pre-processed reservoir slope multi-modal data and the reservoir bank slope displacement amount, extracting spatial correlation characteristics among the modal data of the reservoir slope in the target area; and reservoir slope displacement prediction is carried out according to the spatial correlation characteristics among the modal data of the reservoir slope, and a prediction result is output. According to the method, multi-modal monitoring data are fused, space and time modeling capability is provided, modal weight can be self-adapted, deep learning reservoir slope displacement prediction can be carried out in combination with visual prediction, and high-precision, real-time and credible deformation trend prediction and risk early warning are realized.

The invention discloses a motion scene human body posture estimation method driven by adaptive time modeling and a medium, and the method comprises the steps: constructing an adaptive time modeling motion scene 3D human body posture estimation network model based on the motion speed degree; the robustness of the model to a motion scene is improved by using motion prior based on physical constraints and predicting the uncertainty of 3D human body articulation points; human body motion videos with different motion speeds are used as input, a corresponding 2D human body joint point sequence is obtained on the basis of an existing high-quality 2D human body posture estimation model, the 2D human body joint point sequence is input into the network model, and a trained model is obtained; according to the method, the receptive field can be adaptively adjusted according to the movement speed, the small receptive field is used for capturing transient changes for fast movement, the large receptive field is used for capturing long-term dependence for slow movement, the method adapts to movement at different speeds, and therefore the human body posture estimation accuracy in movement scenes at different speeds is improved.

This invention discloses a method and system for predicting turbine disk temperature fields, an electronic device, and a storage medium. The method utilizes an improved Mamba-2 model with bidirectional coupling of an improved Mamba module and a spatiotemporal attention mechanism to predict turbine disk temperature fields. The spatiotemporal attention mechanism can dynamically adjust feature weights according to the temporal evolution of the temperature field, avoiding being limited to static feature weighting. The improved Mamba module performs feature filtering based on the weighted output of the spatiotemporal attention mechanism, which can selectively filter redundant temporal information and compensate for the deficiency of the Mamba-2 model in capturing features sensitive to core parameters. This forms a closed-loop collaborative mechanism of temporal modeling, feature enhancement, and gating optimization, achieving deep integration of the two. This improves both the temporal modeling accuracy of turbine disk temperature fields and the ability to capture core features, thereby improving the prediction accuracy of transient temperature fields of turbine disks.

The invention provides a communication network traffic prediction method and device and electronic equipment, and the method comprises the steps: obtaining historical traffic data and a first time sequence of a target region, and determining a similar region according to the first time sequence through employing a dynamic time warping algorithm; obtaining a second time sequence of the similar region, and determining an interval adjacent matrix according to the first time sequence and the second time sequence; inputting the interval adjacency matrix into a spatial modeling model, and outputting spatial features through processing of the spatial modeling model; inputting the historical flow data into the time modeling model, processing the historical flow data through the time modeling model, and outputting time features; carrying out fusion processing on the spatial features and the time features to obtain fusion features; and inputting the fusion features into a communication network traffic prediction model, and processing the fusion features by the communication network traffic prediction model to obtain target prediction traffic data. According to the invention, spatio-temporal joint prediction of the communication network flow is realized, and the prediction precision is effectively improved.

The invention relates to the technical field of data processing, and discloses a knowledge security evolution updating method and system oriented to newly-added information, and the method comprises the steps: constructing a multi-granularity time sequence index structure containing business cycle key nodes, marking system input timestamps for collected data, and analyzing business effective timestamps; identifying and inputting an effective time difference distribution zone based on the distribution pattern of the difference value of the two, and calculating a time sequence reversal congestion factor representing the priority reversal risk of the data near the key node; combining the factor, the input aging attenuation degree and the service aging fitting degree to calculate a security evolution sorting weight, and generating a structured suppression parameter by using the factor; and finally, performing incremental data fusion on the target node of the index structure according to the weight, and updating the knowledge base. Through two-dimensional time modeling and an anti-congestion mechanism, the problem that core knowledge is submerged due to high-concurrency window period fragmentation information is effectively solved, and orderly and safe evolution of the knowledge base is achieved.

The invention discloses a bridge seismic response prediction method based on Bayesian optimization and an SCINet-Attention-LSTM hybrid network, and relates to the field of civil engineering, and the method comprises the steps: inputting seismic oscillation data into a trained SCINet-Attention-LSTM hybrid prediction model, and obtaining a seismic response prediction result of a target bridge structure; an SCINet feature extraction module in the hybrid prediction model is used for performing feature extraction on input data through an iterative down-sampling-convolution-interaction mechanism, and outputting a feature vector sequence containing original seismic oscillation sequence time sequence information; the Attention module is used for performing dot product or weighting processing on the feature vector sequence; and the LSTM module is used for carrying out time sequence modeling on the feature vector sequence processed by the Attention module and outputting an earthquake response prediction result of the target bridge structure.

A multi-view foreground motion sensing positioning and detection method for a three-dimensional object comprises the following steps: step 1, extracting foreground features: step 1.1) performing foreground region division and scoring; 1.2) dividing a foreground region; step 1.3) performing foreground scoring; step 2, feature aggregation; 3, designing a dynamic query generator; step 4, image channel affine transformation design; step 5, three-dimensional time modeling: step 5.1) calculating a three-dimensional deformable attention feature map; and step 5.2) generating a three-dimensional target detection frame. The method dynamically adapts to the feature extraction process and improves the detection precision.

The invention discloses a tumble risk assessment method and system based on individualized information, and relates to the technical field of medical assessment and artificial intelligence, and the method comprises the following steps: a data acquisition step, a support period detection step, an individualized feature fusion step, a gait period segmentation step, a feature extraction step, and a time sequence coding step. A risk assessment step; firstly, multi-channel plantar pressure time sequence data of a subject and individualized information of the subject are obtained, and then through five-layer progressive processing, the problems of supporting period self-adaptive detection, gait cycle precise segmentation, multi-dimensional feature extraction, time sequence dynamic modeling and self-adaptive model training are solved respectively. And finally, accurate individual fall risk assessment is realized. Thresholds are automatically adapted for subjects with different physiological features, the gait deterioration trend caused by fatigue is captured through periodic-level double-flow time sequence modeling, and self-adaptive personalized feature learning is achieved through cascaded personalized information layer-by-layer fusion.

The invention provides a low-complexity human body motion reconstruction method based on a sparse inertial measurement unit, and belongs to the technical field of virtual reality and three-dimensional human body motion reconstruction, and the method comprises the following steps: obtaining a data set required for reconstruction training based on human body postures, carrying out time modeling through a time sequence encoder, updating joint features on a human body skeleton graph, and obtaining a reconstruction result; a skeleton is divided into a trunk and four limbs according to a human anatomical structure, global rotation of a root joint and local rotation of each joint are respectively predicted by a partition kinematics regression head, low-rank decomposition is introduced into a large-scale linear layer to compress model parameters, and forward kinematics is utilized to recover three-dimensional joint positions of the whole body. In the training process, a two-stage teacher-student distillation model framework is adopted, a teacher network is trained through real labels, and then joint rotation and joint positions output by a teacher are used as soft targets to jointly restrain a student network through rotary distillation and position distillation. While the parameter quantity is reduced, the reconstruction precision and the motion smoothness of the whole body are improved.

This invention discloses a time series anomaly detection method and system based on multi-scale spatiotemporal modeling, comprising: extracting multivariate time series from business data; generating multiple sub-time series at different scales using one-dimensional convolution; independently encoding each scale sub-time series using a scale-independent spatiotemporal encoder to obtain spatiotemporal enhanced features of the multi-scale time series; employing a cross-scale hybrid expert mechanism to achieve information exchange between scales, obtaining a multi-scale sequence representation after scale interaction; integrating the interaction representations of each scale and performing decoding and reconstruction; generating anomaly scores by calculating the reconstruction error between the original input and the reconstructed sequence, and obtaining anomaly detection results. This method decomposes complex time series into multiple scales, with each scale collaboratively modeling spatiotemporal dependencies, which helps to more accurately discover and verify anomalous signals, improving the accuracy and reliability of detection.

The application belongs to the technical field of behavior recognition, and particularly relates to a human behavior recognition method and system based on dynamic space-time modeling and semantic quantification, which comprises the following steps: acquiring human skeleton sequence data and converting the data into space-time feature representation; constructing a Transformer network based on dynamic space-time modeling and semantic quantification, which comprises a preprocessing layer, a progressive feature extraction layer and a classifier; the preprocessing layer generates initial space-time features; in the progressive feature extraction layer, the joint importance weight is adaptively assigned by a global dynamic joint weighting module in the shallow stage, and the global dynamic joint weighting and a semantic quantification module are simultaneously used in the deep stage, the continuous features are discretized into semantic indexes by using a learnable motion primitive codebook, a semantic graph is constructed based on hard assignment and is fused with a physical skeleton graph, and the differentiable feature reconstruction is performed based on soft assignment; finally, the semantic enhanced features are input into the classifier to complete behavior recognition. The application improves the precision, interpretability and generalization ability of behavior recognition.