43 results about "Graph sequence" patented technology

A method, system and computer readable article of manufacture for displaying visualization from an event log. According to the method, sequences are extracted from the event log. The event log is a record of a plurality of events that occurred. A sequence graph is generated from the sequences. The sequence graph includes nodes representing the events of the event log and the events of each node have a common property. A sequence graph visualization is displayed from the sequence graph. The sequence graph visualization includes (i) node icons representing the nodes of the sequence graph, (ii) node transition icons representing relationships between nodes, and (iii) an event icon for each of the one or more events of a corresponding node.

The invention discloses a video human body interaction motion identification method based on an optical flow graph depth learning model. The method mainly comprises steps that step 1, deframing of a test set video and a training set video is carried out, an optical flow sequence graph is calculated through utilizing two adjacent frames; step 2, the optical flow sequence graph is pre-processed, and optical flow graphs with relatively few information quantity are deleted; step 3, a residual error neural network is trained through utilizing the training set optical flow sequence acquired in the step 2, the test set and training set optical flow graph sequences are taken as input, and spatial domain characteristics are acquired; step 4, a long and short memory model is trained through utilizing training set characteristics, test set characteristics are inputted to acquire each type of probability output; and step 5, a classification result is acquired through employing voting model statistics. The method is advantaged in that relevant patent blanks are filled through utilizing the depth learning model to carry out human body motion identification, identification accuracy is high, and the method is applicable to multiple occasions.

The invention discloses a video behavior identification method based on an Attention-LSTM network. The method includes transforming the input RGB image sequence through an optical flow image sequencegeneration module to obtain an optical flow image sequence; inputting the optical flow graph sequence and the original RGB graph sequence into a time domain attention frame taking module, and respectively selecting non-redundant key frames in the two graph sequences; inputting the key frame sequences of the two images into an AlexNet network feature extraction module, respectively extracting timesequence features and spatial features of the two frame images, and performing a feature weight increasing operation with strong action correlation on the feature image output by the last convolutional layer through a feature weight increasing module; and inputting the feature maps output by the two AlexNet network feature extraction modules into an LSTM network behavior identification module, respectively identifying the two pictures, and fusing the two identification results in proportion through a fusion module to obtain a final video behavior identification result. According to the invention, the function of identifying the behavior from the video can be realized, and the identification accuracy can be improved.

The invention provides a telecommunication fraud event detection method and a telecommunication fraud event detection system. The constructed telecommunication fraud event detection model is used forpredicting a fraud mode. The construction of the fraud event detection model comprises the following steps: establishing a dynamic communication graph reflecting different time series communication modes based on existing fraud behavior data; an abnormal sub-graph sequence with fraud behaviors is mined based on the dynamic communication graph, and the abnormal sub-graph sequence is composed of abnormal nodes, nodes associated with the abnormal nodes and edges; and training a multi-classification model based on the abnormal subgraph sequence, and obtaining a telecommunication fraud event detection model for detecting multiple fraud modes. By utilizing the detection method and the detection system, the fraud behavior can be quickly and accurately predicted.

A frequent changing pattern extraction device (100) which extracts a frequent changing pattern from an ever-changing network structure includes: a conversion unit (12) which converts each of a plurality of graph sequences into an operator sequence by expressing changes, from a first graph included in the graph sequence to a second graph which is temporally adjacent to the first graph, using operators indicating operations necessary to transform the first graph into the second graph, each graph sequence including a plurality of graphs that show temporal changes in the graphs and each of the graphs including a vertex corresponding to a data piece and an edge corresponding to a link between data pieces; and an extraction unit (18) which extracts an operator subsequence that appears at least a predetermined number of times in the plurality of operator sequences corresponding to the plurality of graph sequences, based on the anti-monotonicity used in the Apriori algorithm.

The invention designs a sequence recommendation method fusing a dynamic knowledge graph. The item sequence of the user in the latest session is utilized to predict the next item which is possibly interacted by the user, so that the short-term preference of the user is better captured according to the item sequence in the session. The feature representation of the item sequence and the operation sequence is learned by adopting a gating graph neural network and a recurrent neural network respectively; and meanwhile, related project knowledge (such as commodity brands, rating, categories and the like) is utilized to construct a dynamic knowledge graph, a dynamic knowledge graph method is fused, a graph convolutional network (GCN) is adopted for modeling in the time dimension to learn embedding of nodes, and parameters (weight matrixes) in the graph convolutional network are evolved by using long short-term memory (LSTM) so as to capture the dynamic state of a graph sequence, therefore, the next interaction item of the user can be predicted more effectively.

The invention discloses a newborn pain expression recognition method based on a two-channel three-dimensional convolutional neural network. The method comprises the following steps: (1) collecting video segments of newborns under different states, dividing the videos into n types of expressions according to pain degrees, and establishing a newborn facial-expression video library; (2) cropping andediting each video segment in the facial-expression video library to form a frame sequence of l-frame length, carrying out grayscale transformation on each image, and extracting an LBP feature map; (3) constructing the two-channel three-dimensional convolutional neural network; and (4) inputting grayscale graph sequences and LBP feature map sequences into the two-channel three-dimensional convolutional neural network for training and optimization tuning on the network, and saving a trained network model. According to the method, the deep convolutional neural network is extended and applied tothe field of newborn pain expression recognition to improve accuracy of newborn pain evaluation, and technical support is provided for developing an auxiliary system of newborn pain evaluation.

The invention discloses a method for improving three-dimensional reconstruction point-clout density on the basis of contour validity. The method comprises the following steps: 1, extracting an object contour, and generating a corresponding effective area graph sequence; 2, calculating expansion dimensions of a point cloud on an x axis, a y axis and a z axis; 3, obtaining a derivative point cloud by expanding each point in an initial point cloud; 4, converting the derivative point cloud to a camera coordinate system, performing back projection on the derivative point cloud to an effective area graph, and reserving points in an effective area; 5, calculating point products of vectors from initial points of processed derivative points to the points and normal vectors of the points, and reserving points whose point product values are greater than zero; and 6, inspecting whether density of the derivate point cloud reaches a demand, and if the density does not satisfy the demand, by taking the derivative point cloud as the initial point cloud, carrying out operation after the second step until the demand is satisfied. According to the invention, the method is not restricted to a specific detour shot image sequence, does not excessively rely on parameter adjustment, and can improve the density of an effective point cloud within quite short time under the cognition of a quite small computation amount.

The invention discloses a group flow forecasting model and a method based on a focusing mechanism, The model comprises a continuity feature learning module, which is used for learning the continuity feature expression by using an Attentive Crowd Flow Machines (ACFM) to the time-sorted feature graph sequence to obtain the continuity feature graph; The periodic feature learning module is used for learning the periodic feature expression of the time-sorted feature graph sequence by using the focused group flow machine ACFM to obtain the periodic feature graph; the periodic feature learning moduleis used for learning the periodic feature expression of the time-sorted feature graph sequence to obtain the periodic feature graph. A fusion module with time variation is used for introducing external information to guide the fusion of the continuous feature map and the periodic feature map. The present invention deduces the future trend of the group traffic by learning the dynamic representation of the change of the data in the time domain.

The invention discloses an urban area tail gas pollution prediction method, which belongs to the technical field of environment detection, and comprises the steps of creating a training sample set byusing historical tail gas space-time distribution diagram sequence data and external environment feature data; training the deep space-time diagram convolutional network model by using the training sample set to obtain a deep space-time diagram convolutional network prediction model; and based on the deep space-time diagram convolutional network prediction model, predicting the regional exhaust emission at the moment t by using the external environment feature data at the current moment t and the historical exhaust space-time distribution diagram sequence data before the moment t-1. Accordingto the method, the exhaust space-time distribution is constructed into the graph structure data by utilizing the road network connectivity data, so that the regional exhaust emission prediction problem is converted into the space-time graph sequence prediction problem, and the regional exhaust prediction with higher precision can be realized on the real telemetry data by adopting the deep space-time graph convolution model.

The invention discloses a scattering-scene depth reconstruction method based on polarization transient-imaging and equipment. The method is applied to the transient-imaging equipment used for a scattering scene. The equipment includes a laser light source, a depth camera located at the same horizontal plane as the laser light source, a polarizer disposed between the laser light source and a photographed target object, and a polarization analyzer disposed between the depth camera and the photographed target object. The method comprises: setting polarization starting directions of incident lightemitted by the laser light source, and setting a reference direction of the polarization analyzer; controlling the laser light source to emit the incident light, rotating the polarization analyzer, respectively photographing the photographed target object on multiple polarization analysis directions through the depth camera, and obtaining multiple sets of transient graph sequences by reconstruction; preprocessing the multiple sets of transient graph sequences to separate reflection components and scattering components in a current scene; obtaining time section graphs of image spatial-coordinate points from the reflection components of the transient graph sequences; and reconstructing scene depth according to peak value positions of the time section graphs.

The invention provides a camera-equipment-array based dynamic scene depth restoring method, which comprises the following steps of: initializing parallax graphs to acquire initial parallax graphs of all visual angles at the t moment; performing the optimization of space consistency on the initial parallax graphs of all visual angles at the t moment to obtain parallax graphs, with the consistent space, of all visual angles at the t moment; and performing the optimization of space-time consistency to restore the parallax graphs of the dynamic scene. In the method, the stability of a deep graph sequence of each camera can be ensured effectively through space consistent constraint and space-time consistent constraint which have a uniform form, and simultaneously, the correction of single-moment error is performed by utilizing multi-moment depth information.

The invention discloses a software similarity detection method based on dynamic control flow graph sequence birthmark. The method comprises the following steps: firstly assembling a starting address of a basic block in the plug-in program record program execution process and a branch hopping address at the ending of the basic block under a dynamic plug-in platform DynamoRIO; and then analyzing a log file, constructing a program dynamic control flow graph, and endowing the weight; establishing a weight sequence birthmark set WSB, and serving the length ratio of the WSB as parameter to compute the similarity of each pair of programs. By adopting the dynamic plug-in analysis and extracting the feature of the software in operation, the problems that the source code is absent and the reverse disassembling is difficult in the software plagiarism detection can be avoided; only the basic block starting address and the branch hopping condition are recorded in the dynamic plug-in analysis, and the expenditure is less in comparison with the birthmark based on the dynamic data flow tracking and like technology; the influence by unrelated interference information in the dynamic operation can beresisted, and the program similarity can be detected even if the software encrypts by using an encryption shell.

The invention discloses a medical record knowledge graph constructing method based on a dynamic graph sequence, and a system. The constructing method comprises the steps of acquiring medical record data; wherein the medical record data comprises the characteristic data of the electronic medical record and a patient; preprocessing the acquired medical record data, and determining preprocessed medical record data; constructing a case group according to the preprocessed medical record data; wherein the case graph displays multiple entities and the relation between the entities; the entities comprise medicine, symptom, disease kind and medicament kind, and the relation between the entities comprises an ownership attribute relation and a treatment relation; introducing a time dimension, defining an evolution graph sequence model by means of the case graph, and determining a medical knowledge graph which is constructed by the medical record data of the patient at different time points. The constructing method and the system according to the invention can improve efficiency in searching a disease kind by a user.

The invention relates to the technical field of human body behavior recognition, in particular to a human body behavior recognition method based on a multi-scale attention graph convolutional network, which comprises the following steps: acquiring a to-be-recognized original 3D skeleton graph sequence; inputting the original 3D skeleton diagram sequence into a pre-trained human body behavior recognition model; firstly, extracting joint information, skeleton information and motion information from the original 3D skeleton diagram sequence through a multi-branch input module to serve as behavior feature data; then, enabling a multi-scale attention graph convolution module to learn correlation of 3D skeleton joint points based on the behavior feature data, and extracting time sequence information of various behaviors in different duration time; and finally, identifying human body behaviors corresponding to the original 3D skeleton graph sequence through a global attention pooling layer; and outputting a corresponding human body behavior recognition result. The human body behavior recognition method can give consideration to the accuracy and efficiency of human body behavior recognition, so that the recognition effect of human body behavior recognition can be ensured.

The embodiment of the invention discloses a skeletal animation posture matching method and device, electronic equipment and a storage medium. The skeletal animation posture matching method comprises the steps: receiving any frame of data in skeletal animation; and obtaining candidate matching nodes corresponding to the any frame of data from a pre-obtained motion graph model, wherein the motion graph model comprises a plurality of nodes connected through directed edges according to an excessive situation, and the plurality of nodes represent a plurality of motion frames in the motion graph sequence respectively; obtaining the similarity between any frame of data and each node in the candidate matching nodes; and based on the similarity between any frame of data and each node, obtaining a matching node of any frame of data from the candidate matching nodes. According to the embodiment of the invention, the accuracy and matching efficiency of the posture matched by the skeletal animationcan be ensured, and the smooth transition of the posture of each frame of data in the skeletal animation is improved.

The embodiment of the invention provides an event detection method and device and electronic equipment, and the method comprises the steps: extracting a multimedia frame sequence from a to-be-detectedmultimedia stream; performing optical flow analysis on the multimedia frame sequence to obtain an optical flow graph corresponding to each multimedia frame in the multimedia frame sequence; fusing each multimedia frame with the corresponding optical flow graph, and performing operation on an image sequence obtained after fusion through a convolutional neural network obtained by pre-training to obtain a spatial feature graph sequence carrying time sequence information; performing recursive operation on the spatial feature map sequence according to the time sequence information through a space-time recurrent neural network obtained by pre-training to obtain a space-time feature map sequence; and determining an event type in the multimedia stream by using a preset multi-classifier based on the time-space feature map sequence. Through the scheme, the detection of the universal event can be realized.

The invention discloses a human body behavior recognition method based on a history motion graph and R transformation. According to the method, a depth video is used as a recognition basis, firstly, the minimum enclosure rectangle of human body motion is calculated according to a foreground segmentation technology, then the history motion graph is extracted within a depth video area limited by the minimum enclosure rectangle, motion intensity constraint is exerted on the extracted history motion graph, so that a motion energy diagram is obtained, R transformation is calculated on the obtained motion energy graph, and therefore a characteristic vector used for behavior recognition is obtained. A method of a support vector machine is adopted for training and recognition processes. The minimum enclosure rectangle of human body behavior motion is adopted for preprocessing, and behavior characteristic extraction is accelerated; a method of history motion graph sequences is adopted for reducing influences of noise in depth graphs; characteristics are extracted through performing R transformation on the energy graph, so that calculation speed is high.

The invention discloses a wireless link quality prediction method in a high DOF (Degree of Freedom) underwater sensor network. The method comprises the steps of: by taking the hierarchical high DOF underwater sensor network formed by a sea surface Sink, underwater anchor nodes and underwater common sensor nodes as a research object, establishing a network weighted graph sequence model of the hierarchical high DOF underwater sensor network and a corresponding link quality adjacent matrix; based on characteristic analysis on a correlation function and a partial correlation function of a link quality sequence, carrying out time sequence model identification on the link quality sequence; by a pproximate maximum likelihood estimation method and a SBC criterion, further estimating model parameters and completing model order determination so as to implement link quality prediction between the anchor nodes and the sensor nodes; and finally, analyzing a link quality prediction result by utilizing a preset prediction accuracy threshold so as to carry out error control and model optimization. The wireless link quality prediction method is high in prediction accuracy and high in feasibility and can be widely applicable to sensor network link quality prediction of various underwater scenes.

The invention provides a dialogue prediction method and device in a tourism scene, electronic equipment and a storage medium. The method comprises the steps of collecting historical dialogue data of an online travel agency; forming a question sequence; forming an intention sequence; forming an intention network diagram; generating a weighted undirected graph of the intention network graph; inputting the weighted undirected graph into a Node2vec model; calculating the jump probability between the nodes of the Node2vec model; carrying out vertex sampling; enabling the Node2vec model to meet an optimization target; inputting vector representation of nodes output in the Node2vec model into a ComplEx model; initializing the ComplEx model; scoring positive samples and negative samples of a triple of the ComplEx model; calculating the model loss of the ComplEx model according to the scores of the positive sample and the negative sample of the triple; minimizing the loss using an Adagrad optimization algorithm. After the intention is recognized, the next intention and problem of the user are triggered and predicted, user input is standardized, and therefore the intelligent customer service effect is improved.

The invention relates to an action recognition method based on unsupervised graph sequence predictive coding and a storage medium, the action recognition method comprises model training and use, is used for recognizing various actions performed by a human body in a skeleton sequence, and aims to solve the problems that an existing action recognition method highly depends on a large amount of labeled data. The precision is low under the condition that only a small amount of labels exist, and an existing unsupervised method does not utilize topological information of a graph in overfitting and is poor in serious generalization ability. The method of the system comprises: carrying out view angle invariant transformation, resampling and block-level skeleton graph data enhancement on skeleton sequence data; enabling embedding of space-time diagram convolution skeleton sequence block to express extraction; aggregating context features by the graph convolutional recurrent neural network; constructing positive and negative sample pairs through predictive coding; and extracting features through the pre-training model, and obtaining an action category corresponding to the to-be-recognized skeleton sequence by using the classifier. Compared with the prior art, the method has the advantages of low training difficulty, high recognition precision, excellent performance and the like.

The invention provides a processing method and device of a 3D convolutional neural network on a neural network processor. The method comprises the following steps of: splitting a graph sequence in a time dimension, performing first convolution operation on the split graph sequence and a first convolution kernel of a P3D convolution layer to obtain a plurality of first 2D feature graphs, dividing the first 2D feature graphs, and splicing the divided first 2D feature graphs to a channel dimension to obtain a plurality of 2D spliced graphs; and meanwhile, splicing the data of a second convolutionkernel of the P3D convolution layer in the time dimension to the channel dimension to obtain a 2D spliced convolution kernel, and performing a second convolution operation based on the 2D spliced graph and the 2D spliced convolution kernel. Therefore, the neural network processor realizes convolution processing supporting the 3D neural network. Meanwhile, a P3D pooling layer is subjected to pooling operation step conversion, first pooling operation and second pooling operation are carried out respectively, and pooling processing of the 3D convolutional neural network supported by the neural network processor is realized.

A dynamic graph sequence recommendation system sensitive to user interaction is realized through a method in the technical field of artificial intelligence. The whole system adopts a reinforcement learning framework, data input is scoring data of a user for commodities with timestamps and attribute data of the user, output of the system is a recommended commodity sequence generated by continuous multi-round recommendation, a recommendation result of each round is obtained after an intelligent agent observes a system environment subjected to dynamic graph modeling, and an optimal recommendation decision is made based on the state representation of a dynamic graph environment, the commodity representation, the real-time interest of the user in the commodities and the attribute information of the user. The operation process of the system is divided into five modules in sequence, an off-line training mode in reinforcement learning is adopted for training, a small-batch gradient descent method is used for optimizing parameters, an environment state is modeled by using a graph neural network and a self-attention mechanism, a recommendation strategy can be generated based on the real-time global environment state to obtain recommendation, and the system has strong real-time performance, high dynamic performance and expandability.

The present invention proposes a loading and unloading equipment scheduling method that considers AGV collision avoidance. By establishing a dynamic model of container terminal loading and unloading equipment and two types of collision avoidance models; treating the discrete event dynamics of the equipment as a hybrid flow shop scheduling problem, and then using the overall graph sequence method to schedule all container tasks in the hybrid flow shop; Finally, when considering two types of collision avoidance, according to the introduction of the hierarchical control structure, the status monitoring controller sends each operation time window generated to the stage controller, and the stage controller assigns each operation time window to a specific device. The operation time of AGV is updated according to the minimum time control problem to achieve collision-free scheduling of all tasks of AGV. This scheduling method that simultaneously considers AGV collision avoidance and loading and unloading equipment can meet the optimization needs of large-scale automated container terminal scheduling and management problems.

The invention provides a network topology dynamic layout method and device of a clustering Ad hoc network. The method comprises the following steps of: obtaining an Ad hoc network topology snapshot G0at an initial moment; dividing the Ad hoc network nodes at the initial moment into k clusters according to the positions and connection relationships of the nodes in the G0, and performing initial layout based on the Ad hoc network topology snapshots G0 at the initial moment and the divided sub-graphs to obtain the layout L0 of the network topology snapshots at the initial moment; updating the node layout position based on the layout Li-1 of the network topology snapshot Gi at the current moment and the network topology snapshot Gi-1 at a previous moment, obtaining the layout Li of the network topology snapshot Gi at the current moment. According to the method, the number of nodes adjusted between adjacent layouts can be greatly reduced, and the layout sequence has good dynamic stability;graph sequence layout is realized in an incremental mode, the required calculation data volume is small, and online dynamic layout can be realized.

The invention discloses a human body physical stress evaluation method and device. A video signal of a human body finger end is acquired by a camera, and is decomposed into a graph sequence accordingto time; each graph in the graph sequence is subjected to pixel summing and averaging according to space to be converted into point data; the point data is synthesized into pulse wave signals according to time; the pulse wave signals are converted into an eigenvalue interval variable set; performing statistical variable calculation according to the eigenvalue interval variable set; Statistical variables are used to quantitative evaluation on human body physical stress is performed on the basis of statistical variables. The beneficial effect is that an external sensor does not need to be bought, and only a smartphone with a camera is needed to use in terms of hardware. The method is easy to implement, has good reproducibility, and can be used in many fields such as human body physical stress evaluation for a wide range of people, and treatment effect feedback.

The invention relates to a posture recognition method and device based on skeleton separation and unification and an attention mechanism. The method comprises the following steps: acquiring skeleton data; selecting a graph sequence from the skeleton data, and performing multi-scale learning graph convolution processing on the graph sequence based on a unified time-space operator of a time window to obtain a first skeleton feature; carrying out attention mechanism processing on the first skeleton feature and completing re-calibration on the first skeleton feature to obtain a weighted feature map, and taking the weighted feature map as a second skeleton feature; performing global average pooling processing on the second skeleton feature, and inputting a global average pooling processing result into a Softmax classifier; enabling the Softmax classifier to perform recognition and outputs a posture type. According to the invention, skeleton data is processed, multi-scale structural features and the long-term dependency relationship are extracted, attention mechanism processing is added at important joint points of limbs, and skeleton features with enhanced data are obtained, so that accurate recognition of limb actions and postures of factory workshop workers during working on a production line is realized, and skeleton recognition efficiency is improved.