56results about How to "Fast training convergence" patented technology

The invention provides a power system reactive power optimization method based on depth determination strategy gradient reinforcement learning. A deterministic depth gradient strategy algorithm is applied to the traditional power system reactive power optimization problem. The voltage state of the power system is sensed through a depth neural network, a corresponding action decision is then made by using a reinforcement learning method, a correct generator terminal voltage adjustment action, a node capacitor bank switching action and a transformer tapping point adjustment action are made to adjust reactive power distribution in the power system, the active power network loss of the power system is minimized. As the neural network is divided into an online network and a target network, association between parameter updating and adjacent training in each training process of the neural network is avoided, and the problem that reactive power optimization of the power system is caught in local optimization is avoided. On the premise of conforming to the security constraint of the power system, the economical efficiency of the operation of the power system is improved by reducing the network loss of the power system.

The invention relates to the technical field of computers, in particular to an information pushing method and device based on a graph network. The method comprises the steps of determining a sub-graphstructure taking the to-be-processed node as a central node in a graph network, the sub-graph structure comprising nodes connected with the to-be-processed node and a connection relationship; inputting the subgraph structure into a node feature extraction model to obtain an aggregation feature vector corresponding to a to-be-processed node output by the node feature extraction model; screening nodes matched with the to-be-processed nodes as information pushing nodes based on the aggregation feature vectors of the nodes in the graph network; and pushing the service information corresponding tothe information pushing node, so that the graph structure feature vector is added on the basis of the aggregation feature vector during training, namely, the graph structure information is added forconstraint, the convergence speed and accuracy of training are improved, and the sub-graph structure can be input during application. And the node representation accuracy and the information pushing accuracy in the graph network are improved.

The invention discloses a pipeline safety event identification and knowledge mining method based on an HMM model, belonging to the pipeline safety event monitoring field. The method comprises the following steps: (1) extracting the multi-domain features of signals collected by each space point and obtaining the feature vector sequence of the signals; 2, inputting a feature vector sequence into anHMM model for off-line training to complete that construction of a typical event HMM model library; 3, after obtaining that current feature vector sequence of the signal to be identified through the step 1, inputting the signal to be identified into a typical event HMM model library for identification and outputting an event judgment type, and calculate an optimal hidden state sequence as the information of the evolution process of the event state sequence for output, and completing knowledge mining; The HMM model of the invention is analyzed and identified based on the characteristic timing sequence, and the event identification rate is effectively improved. At the same time, knowledge mining is realized in the evolution process of event state series, which can be used for short-term prediction.

The invention discloses a method for improving frame rate up conversion effect by using a generated network. The method comprises the following steps of 1, comprising an original video via JM18.6 to acquire the compressed video, then reserving odd-number frames in the compressed video, and reconstructing even-number frames to acquire reconstructed frames of the even-number frames by using a framerate up conversion algorithm based on direction selection movement estimation; 2, building a generated network G, inputting the reconstructed frames IE into a neural network for training, comparing the corresponding frames IP in the original video with network output, backwards propagating an error, and acquire the optimal remediation effect after multiple times of adjustment; 3, repeatedly executing the steps 1 and 2 until an error variation rate is invariable or less than a set threshold; and 4, applying the trained network, inputting the reconstructed frames has been subjected to frame rateup conversion into the network to acquire an image with high quality. According to the method provided by the invention, psnr values and quality of the reconstructed frames are improved, and thus theframe rate up conversion effect is improved.

The invention relates to a marine weak and small moving ship real-time detection method based on deep learning. The method comprises the steps: step 1, carrying out target detection processing on an image in a marine monitoring system video; step 2, performing size unification operation on the size of the input image; step 3, extracting small target features by using the first part of the networkstructure; and step 4, for the extracted features, learning the effective features of the small targets from the shallow network by using the classifier of the second part of the network structure, and then integrating the position information of the deep network to finally classify the small targets, wherein the network structure is of a full convolutional network type and comprises two parts, and the first part is used for extracting effective features of a target; and the second part is used for detecting a target and determining a final detection result.

The present invention discloses a text comparison algorithm based on a stacked bidirectional lstm neural network, which relates to the field of natural language processing, and includes the followingsteps: Step 1, inputting a sentence segmentation and calculating a word vector, and obtaining a word vector as an input word vector; Second, input the input word vector into the lstm neural network inthe form of network stacking to obtain the input sentence vector; in step 3, obtain the sentence vector of the two input sentences according to steps 1 and 2; input the sentence vector of the two input sentences Go to the classifier and get the similarity of the two sentences. The application of the present invention enables accurate text similarity calculation.

The invention relates to a video behavior recognition method based on weighted fusion of multiple image tasks. The video behavior recognition method comprises the following specific steps: 1, constructing an initialized teacher network; 2, downloading and selecting a plurality of pre-training models and parameters of a visual image task common data set in positive correlation with video behavior identification as an initialized teacher network; 3, establishing a multi-teacher video behavior recognition knowledge base; 4, under the guidance of the multi-teacher network with the weight distributed again, carrying out self-supervised training based on comparative learning on the student network; 5, carrying out performance test on the model video behavior identification on the test data set. The method has the advantages that the image task in positive correlation with the video behavior recognition task serving as the target task is used as the teacher task, the training mode of comparison self-supervised learning is adopted, and the video behavior recognition problem under the condition that high-quality video marking samples are insufficient is solved. Therefore, the accuracy of video behavior identification is effectively improved.

The invention relates to a rolling bearing fault diagnosis method based on VMD and a deep convolutional neural network. The rolling bearing fault diagnosis method comprises the steps of 1, original vibration data of a rolling bearing is collected; 2, variational mode decomposition data processing and neural network training are carried out on the training set vibration data; and step 3, variational mode decomposition is used to carry out data processing on the test vibration data, and a neural network is used to carry out fault diagnosis. For rolling bearing fault detection, a method of combining variational mode decomposition and a deep convolutional neural network is provided, and diagnosis of different fault types and damage degrees of the rolling bearing under the condition of variable working conditions is realized. Vibration data can be decomposed into different limited band eigenmode function components through variational mode decomposition, and a convolutional layer of the deep convolutional neural network can extract local features of each limited band eigenmode function from different angles, so that diversity and comprehensiveness of feature extraction are ensured.

The invention discloses a method and system for training a road anomaly recognition model and a road anomaly recognition method and system. The method for training the road anomaly recognition model comprises the steps: acquiring images of road anomaly conditions and road normal conditions; inputting the images of the road abnormal condition and the road normal condition into a neural network model; and carrying out transfer learning on the preset trainable layer number in the neural network model according to the images of the road abnormal condition and the road normal condition to obtain aneural network model for road abnormal recognition. According to the recognition model, the input full-frame road original image does not need to be positioned and the like, only one operation needs to be carried out on the image, the real-time performance is high, the model training convergence speed is increased by applying the transfer learning technology, and various road abnormal conditions including road ponding, road fire and traffic accidents can be recognized.

The invention discloses a task model acquisition method and device, a storage medium and an electronic device. The method comprises the following steps: obtaining target model parameters obtained by carrying out model training on the first training model, wherein the first training model comprises: a first encoder and a first decoder, the target model parameters comprisea first model parameter corresponding to the first encoder, the input of the first training model is a first image of a first channel of the target optical image, the output of the trained first training model is a prediction image corresponding to a second image of a second channel of the target optical image, and the graph similarity between the prediction image and the second image is greater than or equal to a target similarity threshold; initializing model parameters of a second training model by using the target model parameters to obtain an initial task model, the second training model comprising a second encoderhaving the same structure as the first encoder; and training the initial task model by using the target optical image to obtain a target task model.

The invention relates to the technical field of artificial intelligence, in particular to a three-dimensional human body posture estimation method and device, equipment and a storage medium, and the method comprises the steps: collecting a picture or video frame of a target human body; the picture or the video frame of the target human body is input into a preset three-dimensional human body posture estimation model, a three-dimensional human body posture estimation result of the target human body is obtained, and the preset three-dimensional human body posture estimation model is obtained by training a semantic graph convolutional neural network based on the target key point of the reference human body; and recognizing the actual motion state of the target human body based on the three-dimensional human body posture estimation result. Therefore, the problems of poor real-time performance, poor application effect and the like of three-dimensional human body posture estimation in related technologies are solved.

The invention discloses a visual navigation method based on deep reinforcement learning and direction estimation. The method comprises the following steps: 1) generating an AI2-THOR simulation platform offline data set; 2) extracting image features of the RGB-D image of the offline data set; 3) constructing an A3C deep reinforcement learning model, and taking the features in the step 2) as the input of the model to train the model; 4) migrating the model trained in the step 3) according to the real scene, performing fine tuning on the model to improve the generalization ability, and finally applying the fine-tuned model to the visual navigation of the real scene. According to the method, a three-dimensional geometric method is used, the position and the direction of navigation to the target object are calculated through a direction estimation method, direction features are provided for a deep reinforcement learning model, the model can converge more quickly, the generalization ability is higher, and accurate visual navigation is achieved.

The invention discloses an image segmentation method for a kidney tumor, and the method comprises the following steps: S1, obtaining an abdomen scanning image, and dividing the abdomen scanning imageinto a data set and a training set; S2, performing down-sampling preprocessing on the acquired abdomen scanning image to obtain a scaled image; S3, determining an area of interest of the preprocessedimage in the step S2 by using global position information of the abdominal space, performing image segmentation, and performing training and prediction by a U-shaped kidney tumor segmentation network;S4, outwards expanding the abdomen scanning image in the step S1 for a certain range, segmenting images of the left kidney and the right kidney, interpolating all segmented images, and unifying the interpolated images into the same data distribution to obtain left and right kidney VOI images; S5, performing tumor segmentation prediction on the left and right kidney VOI images by a U-shaped kidneytumor segmentation network. Interference of other organs and tissues is effectively avoided, the accuracy of kidney tumor identification and image segmentation is improved, and efficiency is higher.

The invention discloses a facial expression analysis system for people in a gathering scene. The system comprises a first camera, a second camera, a semicircular sliding rail and a host, wherein the first camera and the second camera are installed at the two ends of the semi-circular sliding rail, and the camera shooting direction of the first camera and the second camera is perpendicular to the tangent line of the semi-circular sliding rail at the cameras. The first camera and the second camera move in front of the measured point in a semi-circular manner along the semi-circular slide rail soas to ensure that a fixed angle is formed between the first camera and the measured point; the semicircular slide rail is mounted on the stage in front of the auditorium; the host comprises a pre-designed analysis algorithm, photos shot by the first camera and the second camera are transmitted to the host for analysis through a local area network, and an analysis result is transmitted to the cloud through the Internet after the host finishes analysis. According to face rotation angle information acquired by the optical sensor, after the host performs judgment, the controller spontaneously drives the camera to move, and manual adjustment is not needed.

The invention discloses a distributed parallel deep neural network performance evaluation method for a super computer, and belongs to the field of high-performance calculation and deep neural networks. The method sequentially comprises the steps of distributed parallel granularity design, deep neural network training framework platform configuration, distributed parallel communication protocol deployment, deep neural network optimization, training test data set slicing and test mode application, and a universal test method is provided for developers. The design of the distributed parallel granularity ensures the comprehensive test of a single node and multiple nodes; the reliability of application layer deployment and bottom layer communication is ensured through coupling of a deep neuralnetwork training framework pytorch and a bottom layer distributed communication framework MPI; the single-node test of different processors and the multi-node test of different processors are helpfulfor realizing the full utilization of super computer computing resources, improving the program computing performance, shortening the neural network training time and improving the neural network training convergence rate.

The invention provides a mixed source positioning method based on a convolutional neural network. According to the technical scheme, the method comprises the following steps: firstly, obtaining a mixed source phase difference matrix by utilizing a radar antenna array; then, inputting the information of the mixed source phase difference matrix into a first convolutional neural network, and calculating the direction of arrival of the mixed source; secondly, removing direction-of-arrival parameters contained in the mixed source phase difference matrix information by utilizing output information of a first convolutional neural network, and inputting the direction-of-arrival parameters into an automatic encoder; and finally, inputting the output of the automatic encoder into the second convolutional neural network, identifying the mixed source and determining the distance of the near-field source. The method can achieve the simultaneous positioning of the near-field source and the far-field source, is high in convergence speed during the training of the provided convolutional neural network, is short in time for calculating the positioning parameters during the use, is high in positioning precision, and is strong in generalization capability.

The invention discloses a satellite component layout temperature field prediction method based on a physical prior neural network, and the method comprises the steps: building a structural model of the satellite component layout according to the layout characteristics of a satellite component; acquiring a plurality of training data, and preprocessing the training data; according to a heat conduction steady-state equation obeyed by a satellite component layout temperature field, constructing a loss function embedded with physical prior; the weight of each prediction point of the temperature field is determined through online data mining, a loss function is updated according to the weights of the prediction points, a regularization item of the loss function is constructed through regional heat flux conservation, and a final loss function is determined; constructing a deep neural network model, and training the deep neural network model by using the preprocessed training data and the final loss function; and utilizing the trained deep neural network model to predict the temperature field of the satellite component layout. According to the method, stable and rapid training of the deep neural network model can be realized by using the training data without labels, and the prediction precision of the model is ensured.

The invention discloses an adaptive feature and data distribution target detection method, and relates to the technical field of computer vision. The invention aims to provide a target detection method for self-adaptive characteristics and data distribution, which can automatically select certain parameters of certain layers as a characteristic pattern to perform target detection, effectively reduce redundant parameters and save the time for optimizing a network. The method comprises the following steps of: 1) inputting an image by adopting a pixel with the width w of 416, a pixel with the height h of 416 and a channel number c of 3; the method comprises the steps of (1) channel weight calculation, (2) size reduction and channel number increase module operation, (3) two-path dense connection module operation, (4) feature map generation, (5) learning, calculation and re-screening of weights of all channels, and (6) regression of target positions and categories.

The invention discloses an end-to-end panoramic image segmentation method based on a query vector. The method comprises the following steps: representing a panoramic image segmentation process and an output result by using the query vector; enabling the query vector (Object Query) to comprise a foreground query vector (Thing Query) and a background query vector (Stuff Query); establishing a panoramic image segmentation model, including respectively establishing an image foreground class segmentation model and an image background class segmentation model; respectively mapping the foreground query vector and the background query vector to an image foreground class segmentation result things and an image background class segmentation result stuff; and carrying out detection training based on the foreground query vector and the background query vector, so that the detection training time of the foreground target is shortened, and end-to-end training and foreground segmentation result and background segmentation result output are realized. The method is simpler in process, lower in calculation complexity and better in performance.

The invention discloses an image generation method based on adaptive gradient cutting under differential privacy protection, and solves the problems of excessive noise and unreasonable cutting in a gradient cutting method based on training data in a conventional privacy protection-based model generation algorithm. Meanwhile, a cutting threshold distribution method based on a model layer gradient cutting method is improved, the efficiency of network learning is improved on the premise that the cutting threshold is reasonably distributed, and the magnitude of gradient noise added after cutting is further reduced under the condition that finer cutting granularity is selected. Therefore, the method can provide a more reasonable gradient cutting mode and lower noise addition. Compared with the prior art, the method has the advantages that the generation model can be more quickly converged to a local optimal solution under the conditions of more appropriate cutting and smaller noise addition, and the purpose of generating a high-quality picture is achieved.