135results about How to "Increase training time" patented technology

The invention discloses a wind power forecasting method based on a genetic algorithm optimization BP neural network, comprising the steps: acquiring forecasting reference data from a data processing module of a wind power forecasting system; establishing a forecasting model of the BP neural network to the reference data, adopting a plurality of population codes corresponding to different structures of the BP neural network, encoding the weight number and threshold of the neural network by every population to generate individuals with different lengths, evolving and optimizing every population by using selection, intersection and variation operations of the genetic algorithm, and finally judging convergence conditions and selecting optimal individual; then initiating the neural network, further training the network by using momentum BP algorithm with variable learning rate till up to convergence, forecasting wind power by using the network; and finally, repeatedly using a forecasted valve to carry out a plurality of times of forecasting in a circle of forecast for realizing multi-step forecasting with spacing time interval. In the invention, the forecasting precision is improved, the calculation time is decreased, and the stability is enhanced.

The invention designs a target detection method based on multi-feature fusion of a full convolutional network. The main technical features are that constructing a full convolutional neural network having six convolutional layer groups; using the first five groups of convolutional layers of the convolutional neural network to extract image features, and outputting the image features to be fused toform a fusion feature chart; performing convolutional processing on the fused feature map to directly generate a fixed number of target frames of different sizes; and calculating a classification error and positioning error between the target frames generated by the convolutional neural network and real frames, utilizing a random gradient descent method to reduce a training error to obtain parameters of a final training model, and finally carrying out a test to obtain a target detection result. The target detection method based on multi-feature fusion of the full convolutional network utilizesa powerful capability of representing a target of a deep convolutional network, constructs the full convolutional neural network used for target detection, proposes a new fusion feature method, improves detection speed and precision of an algorithm, and obtains a good target detection result.

The invention discloses an extreme learning machine-based hyperspectral remote sensing image ground object classification method. An original extreme learning machine network is expanded into a hierarchical multi-channel fusion network. In terms of network training, the method is different from the least squares algorithm-based output weight solving strategy of the original ELM (extreme learning machine) and the global iterative optimization strategy of a deep learning network; according to the method of the invention, a greedy layer-by-layer training mode is adopted to train a hierarchical network layer by layer, and therefore, the training time of the network is greatly shortened; and in the layer-by-layer training process, a l1 regular optimization item is added into the training solving model of each layer of the network separately, so that parameter solving results are sparser, and the risk of over-fitting can be lowered. In terms of network functions, A single-hidden layer ELM network focus on solving the fitting and classification problems of simple data, while the different levels of the network model provided by the invention achieve target data feature learning or feature fusion, the network model of the invention integrates the advantages of high training speed and strong generalization capacity of the single-hidden layer ELM network, and therefore, the in-orbit realization of the model is facilitated, and the requirements of emergency response tasks can be satisfied.

The invention relates to a method, a device and a storage medium for improving the prediction performance of a depth learning network, belonging to the field of artificial intelligence and computer technology. The method comprises the following steps of: iteratively training a first neural network model by m cycles using a preset training set; verifying the trained neural network model by the preset verification set; determining the corresponding performance indexes based on the scenarios of the neural network mode; selecting multiple neural network models whose performance indexes reach the fusion standard from the neural network models trained by m iterative training; acquiring network parameters of a plurality of selected neural network models, and fusing the network parameters of the plurality of neural network models; assigning the fused network parameters to the second neural network model to obtain a neural network model with the fused network parameters; not only does not increase the training time of the model, but also can improve the network prediction efficiency, and can meet the different needs of the application scenario.

The invention discloses a method and a system for defending DDoS attack detection facing 5G network slicing, and aims to solve the problem that the later defense efficiency of DDoS attack is low caused by the fact that sub-equipment attacked by DDoS under nodes cannot be judged purposefully in the prior art. Accurate feature extraction is carried for flow table information, a Hash table using equipment as a bond is established for recording abnormal data flows of the equipment, and discarding and shielding treatment is carried out in later period, so that purposeful DDoS attack judgment is carried out on the sub-equipment under the nodes is realized, accurate and efficient DDoS detection defense is realized, the attack detection efficiency is improved, the attack relief efficiency after attack detection is improved simultaneously, and security assurance is provided for a 5G network with a great deal of SDN technology. The method and the system are applicable for related fields of network detection defense.

The invention discloses a substation pointer instrument identification method based on an improved YOLOV3 model, and the method comprises the following steps: firstly collecting an instrument image, making a data set, and carrying out the calibration; then, clustering the bounding box through a Mini Batch Kmeans algorithm to find an optimal clustering coordinate; modifying a framework network DarkNet-53 of the basic YOLOV3 into a lightweight network MobileNet, and accelerating a training process by virtue of a better activation function; modifying a loss function of coordinate prediction to enable the model to better fit instrument data; and finally, enabling the trained model to be better applied to the detection and identification task of the substation inspection robot, and a small-target instrument panel and a multi-target instrument panel can be quickly and accurately obtained for subsequent processing in the detection process. On the premise that the accuracy is guaranteed, the detection speed is increased, the real-time performance is enhanced, and the detection effect on small targets and multiple instrument panels is greatly improved.

The invention discloses a method for detecting human body parts by performing parallel statistical learning based on three-dimensional depth image information. Specific to the problems of complex deformation, difficulty to describe, and the like, of human body parts (head, hands and feet), a novel feature, namely universal feature, for reflecting diversity of the human body parts, is constructed. A parallel statistical learning method is utilized to select effective and sufficient novel features and form a parallel cascaded classifier, so as to perform real-time efficient detection on the human body parts.

The invention discloses a human face identification method and apparatus, and belongs to the field of human face identification. The method comprises: performing feature extraction on a to-be-identified human face image by using a plurality of pre-trained convolutional neural networks to obtain a plurality of sub-feature vectors of the to-be-identified human face image, wherein the sub-feature vectors of the to-be-identified human face image are same in number of dimensions; normalizing the sub-feature vectors of the to-be-identified human face image; performing addition on the normalized sub-feature vectors of the to-be-identified human face image, and multiplying the sum of the normalized sub-feature vectors by a coefficient to obtain a union feature vector of the to-be-identified human face image; and performing human face identification by using the union feature vector of the to-be-identified human face image or / and the sub-feature vectors of the to-be-identified human face image. According to the human face identification method and apparatus, the training time of the convolutional neural networks is shortened, the over-fitting of the convolutional neural networks is avoided, and the operation is simple and convenient; and identification modes are more diversified and the accuracy is higher.

The invention discloses an anti-attack and defense method based on an LSTM (Long Short Term Memory) detector. The method comprises the following steps that: 1) generating a candidate detector of an LSTM-FNN (Fuzzy Neural Network) structure, and asking the detector to detect abnormal samples as more as possible on a premise that the detector does not misjudge a normal sample; 2) storing the candidate detector into a register queue, detecting a training dataset by an abnormality detector taken out of the register queue, and deleting the detected abnormal sample to enable different abnormality detectors to cover different abnormal areas; and 3) in a detection stage, detecting a detected sample by all abnormality detectors, and comprehensively judging whether the sample is abnormal or not. Byuse of the method, a neural network is combined with a negative selection algorithm, each LSTM detector guarantees that the normal sample can not be misjudged, the abnormality detector set guaranteesthat the abnormal situation may be covered, and an algorithm detection effect is improved.

The invention discloses an eye fundus image cup-disc segmentation method based on a generative adversarial mechanism. The method includes the steps that data enhancement is carried out on a single-channel or multi-channel color eye fundus image, the fundus image is segmented through a U-Net network, the predictive segmentation image will be sent to the discriminator network to identify the true and false, the true and false judgment loss returns back to adjust a model generated by the U-Net network, after many times of running of a generative adversarial network, and finally the optimal opticdisc segmentation model and optic cup segmentation model are obtained. The method achieves the optic disc and optic cup segmentation detection.

The invention discloses a multi-functional basketball dribbling sparring device comprising a simulation human body, an arm movable mechanism, a height adjustment mechanism, a human body movement mechanism and a control mechanism. The simulation human body comprises an upper body, a lower body and simulation arms; the upper body and the lower body are capable of being telescopically connected; the simulation arms are hinged with human body shoulders through the arm movable mechanism; a leg brace is arranged at the bottom of the lower body; a chassis is disposed at the lower end of the leg brace; the arm movable mechanism located on the upper portion of the simulation human body comprises a two-arm spacing adjustment mechanism and an arm stretching adjustment mechanism; the height adjustment mechanism is located on the central axis portion of the simulation human body; the human body movement mechanism positioned at the bottom of the chassis comprises a mobile walking mechanism; a receiver is mounted on the lower body. The multi-functional basketball dribbling sparring device has the advantages of effectively reducing the burden on a coach, simulating actions of opponents below a variety of circumstances; providing rich diversity of basketball training, being beneficial to training and deployment of basketball tactics, and improving dribbling level of basketball players.

The invention belongs to the technical field of named entity recognition, and particularly relates to a Chinese named entity recognition method and device based on a RoBERTa-BiGRU-LAN model, and the method comprises the steps: converting a to-be-processed Chinese corpus into a word vector sequence; inputting the obtained word vector sequence into a first layer of BiGRU-LAN of a RoBERTa-BiGRU-LAN model, and obtaining a coding sequence fused with local information; inputting the obtained coding sequence into a second layer of BiGRU-LAN of the RoBERTa-BiGRU-LAN model, and obtaining attention distribution fused with global information; and obtaining a named entity identification result according to the obtained attention distribution. According to the improved word embedding method disclosed by the invention, Chinese representation is better carried out, and meanwhile, BiLSTM-CRF is improved into BiGRU-LAN, so that the parameters of the model are reduced, the complexity of the model is reduced, and the training time is saved.

The invention discloses a batch A3C reinforcement learning method for an agent to explore a 3D maze. In order to achieve the goal of relatively short training time and small memory loss, the inventionuses a batch-based reinforcement learning method to train a neural network. The neural network is divided into two parts. The first part mainly consists of several convolution layers and MLP, and thelow-dimensional representation of the original screen pixels is obtained. The second part is an LSTM (Long-Short-Term Memory) model. The input of the LSTM is the output of the MLP of the first part,and the cell output of the last time step of the LSTM is circumscribed with two MLPs, which are respectively used to predict the probability distribution of the action a in the current state and the prediction of the state value v in the current state. Combining the efficient reinforcement learning algorithm and depth learning method, the agent can explore 3D maze independently, and the agent cansuccessfully explore 3D maze environment with relatively short training time and small memory consumption.

The invention discloses a remote sensing image building extraction method and system based on a U-Net network, and electronic equipment. A multi-scale module is added to a decoding layer of a U-Net network, and the hole convolution network is introduced, the receptive field can be expanded under the condition that the resolution is not lost through hole convolution, so that the semantic information mining capacity of the network can be improved while detail information is reserved, and meanwhile, the multi-scale feature obtaining capacity of the network is enhanced through the multi-scale module; according to the invention, the convolution mode of the convolution layer is set as filling; that is, after convolution, the size of the feature map is completely unchanged; the original feature map is actually shrunk by 2; in this way, each time the feature map passes through a convolution layer , the size of the feature map is reduced by two times; by the adoption of the convolution model, the size of the feature map output through the four coding layers and the last coding layer is shrunk to be one sixteenth of the size of the input picture after the feature map passes through 4 encoding layers, the image resolution is recovered through deconvolution operation, the size of the feature map begins to be enlarged at the moment, and the training time is effectively shortened.

The invention discloses a width learning system based on multi-feature extraction, and the system comprises four sub-width learning systems, wherein each sub-width learning system comprises a featurenode, an enhancement node and a sub-node; each sub-width learning system firstly extracts an image feature from the image data set, combines the image features extracted from the image data set to obtain respective feature nodes, and then enhances the respective feature nodes through an enhanced mapping function to form corresponding enhanced nodes; after each sub-width learning system forms an enhancement node, the feature node of each sub-width learning system is combined with the corresponding enhancement node and then is connected to the sub-node of the sub-width learning system, and thenthe output of the sub-node of each sub-width learning system is normalized and then is connected to the final output layer. The method has the advantages of short model training time and high classification accuracy in the aspect of complex data set classification.

The invention discloses a student comprehensive quality evaluation method based on a genetic algorithm optimization BP neural network. The method comprises following steps: S1. constructing sample data in a manner of acquiring sample data of all quality indexes of a student, wherein the sample data servers as training samples of the BP neural network; S2. determining a network topology structure in a manner of determining the number of BP neural network hidden layers and the number of nerve cells in each layer and initializing weight thresholds of the neural network; S3. optimizing weight thresholds in a manner of optimizing the weight thresholds of the BP neural network through a genetic algorithm; S4. training and testing in a manner of training the BP neural network and performing tests by the use of data which is not trained; and S5. evaluating the comprehensive quality of the student in a manner of inputting the data of all the quality indexes of the student to the trained BP neural network so as to evaluate the comprehensive quality of the student. By employing the method, the efficiency and accuracy of the student comprehensive quality evaluation can be improved.

The invention discloses a neurological lower limb recovery training device which comprises a base. A guide plate is fixedly connected with the upper surface of the base, a protective plate is fixedlyconnected with the upper surface of the guide plate, a guide groove is formed in a surface of the guide plate, first rotary shafts are slidably nested in the inner wall of the guide groove, two ends of the two first rotary shafts are fixedly connected with first hinge rods by bearings, a power device is fixedly mounted on the upper surface of the base and comprises threaded pipes, the upper surfaces of the threaded pipes are fixedly connected with the surfaces of the first rotary shafts by reinforcing ribs, a transverse slide groove and a longitudinal slide groove are formed in a surface of the protective plate, and second rotary shafts are slidably nested in the inner wall of the transverse slide groove. The neurological lower limb recovery training device has the advantages that effectsof automatically controlling the first hinge rods in a reciprocating manner can be realized by the neurological lower limb recovery training device, threaded rods can be driven by a motor to be connected with the threaded pipes in a threaded manner, the motor can clockwise and anticlockwise rotate under the control when in contact with a first travel switch and a second travel switch, and accordingly the neurological lower limb recovery training device is safe and is high in efficiency.

The invention discloses a vehicle ranging method based on deep neural network. The vehicle ranging method comprises the following steps: S1, acquiring target vehicle images and extracting image coordinates of the target vehicles; S2, building predication network, loading training sample including the image coordinates of the target vehicles, and training the samples through a deep neural network model to calculate network model parameters of the predication network; and S3, using the images of the target vehicles as input of the trained predication network, and predicating a target vehicle distance through a forward propagation algorithm. The vehicle ranging method disclosed by the invention has the advantages that the ground height and pitch data of a camera are not needed to be known inadvance, so that the total identification correction ratio is increased, and the training time is lengthened; a ranging geometrical model is not needed to be built in advance, so that the problem of lower fitting degree of manual modeling is solved, the trouble caused by conventional geometrical ranging is solved, and the problem of low ranging accuracy obtained by firstly modeling and subsequently predicting is solved.

The invention discloses a boiler combustion efficiency predicting method based on a support vector machine incremental algorithm. The boiler combustion efficiency predicting method based on the support vector machine incremental algorithm is characterized by including the following steps: (1) a kernel function is selected; (2) an initial data set is formed; (3) the initial data is pre-treated; (4) a training sample is taken out and tested; (5) a sensitivity coefficient Epsilon is 0.0001, a training precision is 0.00001 and the default values of a penalty coefficient C and a width coefficient sigmate Sigma are respectively 10 and 0.0001; (6) generalization is determined; (7) the optimum coefficient pair is selected; (8) an initial classifier Omega 0, a support vector set and a non-support vector set are obtained through training; (9) sample points which are not in line with a generalized karush-kuhn-tucker (KKT) condition, namely yif (xi)>1 are found out in a newly added sample set X1; (10) a new set is formed; (11) in terms of X, a classifier Omega and a support vector SV are determined; (12) a support vector machine predicting model on boiler combustion efficiency is established. Less input coefficients are input so as to facilitate measuring, a complicated calculation process is removed, training time of working conditions of boiler combustion is shortened, a requirement for online calculation of a distributed control system (DCS) is met and prediction precision is high.

The invention relates to a monitoring video multi-dimensional pedestrian behavior intelligent analysis system and method. The system comprises an individual behavior analysis module, a small group behavior analysis module, a group behavior analysis module, a scene analysis module, a pedestrian behavior dimension evaluation module, an intelligent gateway, an application logic integration module anda video source management module. The method comprises the following steps: acquiring a monitoring video data stream, extracting video frames from the monitoring video data stream, and performing scene classification and pedestrian behavior dimension selection by using key video frame data; combining the scene and pedestrian behavior dimension information; performing decisions selected by the pedestrian behavior analysis module and transmission of management data; and performing intelligent behavior recognition through a pedestrian behavior analysis model of a scene integrated in the pedestrian behavior analysis module under the selected behavior dimension, and outputting a final pedestrian behavior recognition result. Pedestrian behavior recognition in various complex monitoring scenes and pedestrian behavior dimensions can be supported.

The invention discloses an online comment automatic reply method based on deep semantic matching, which is used for finding out online comments closest to input comment semantics in a database by combining sentence vector cosine similarity and multi-dimensional emotion matching degree. The method specifically comprises the steps that feature words of different topics are obtained through Canopy +Kmeans clustering, and on this basis, topic feature word expansion is conducted through a topic model CorEx based on priori knowledge. Meanwhile, a BERT-BiLSTM emotion analysis model is constructed, and multi-dimensional emotion analysis is carried out on online comments according to theme feature words obtained through clustering and by means of dependency syntactic analysis. Online comments arematched with the most similar semantics in the database by combining sentence vector cosine similarity and a multi-dimensional sentiment analysis result, data enhancement EDA operation is performed onmerchant replies of the comments, and sentences with the highest sentence vector cosine similarity are selected as automatic reply contents. According to the invention, automatic reply online comments can be provided for merchants conveniently, efficiently and accurately.

The invention discloses a radar jamming equipment simulation training system, and belongs to the technical field of simulated training. The system comprises a radar jamming device and a simulated training system controlling the radar jamming device to work normally. The radar jamming device comprises a charge station and a hamming station. A supporting platform is installed at the side wall of theradar jamming device. A first seat, a second seat and a jamming seat are installed at the surface of the supporting platform in parallel; the simulated training system comprises an equipment simulated training system and an electronic blue force and evaluating system. An electronic blue force module inside the electronic blue force and evaluating system is used for generating the complex electromagnetic environment, the charge station and the jamming station simulates real equipment operation training, the equipment simulated training system enables training workers to perform actual operation training in the simulated environment, and after training is finished, the training performance is evaluated through the evaluating system, so that the training cost is lowered, the training time isprolonged, and the fighting capacity of a radar jamming team is improved.

The invention belongs to the virtual reality technology field and discloses a meteorological radar training control system and method based on virtual reality and a computer. The system comprises VR hardware equipment, wherein the VR hardware equipment is composed of VR external members. The method comprises steps that (1), system management is carried out, including database and user management;(2), VR modeling is carried out in combination with model type classification of meteorological radars, and key models partially have electrical characteristics; (3), common virtual instruments required for meteorological radar testing and training are provided; and (4), a secondary development interface and functions are provided. The system and the method are advantaged in that the virtual reality (VR) technology is integrated into meteorological radar training, hardware and travel cost for training is reduced, safety of training personnel is enhanced, an online education scene is made to berealer through an immersive scene provided by the VR, and the training effect is improved. Remote VR implementation is employed, time, personnel and fund constraints of traditional centralized training are avoided, and training timeliness is higher.

The invention discloses a crane simulated operation training system, belonging to the simulation system. The system comprises a simulated operation part, sensing systems or switches, a programmable controller and an image display system, wherein all the sensing systems or switches are connected with a first-level signal processor singlechip microcomputer by data lines, the singlechip microcomputer is connected with the programmable controller by a data interface, and the programmable controller is connected with the display system by an image data line. After the sensing systems or switches collect action signals of operating parts of an operating platform, and send the signals to the singlechip microcomputer by the signal data line; after being processed by the singlechip, the signals are input into the programmable controller by the data interface; and all operating actions of the crane are synchronously displayed on a screen by the display system after signal processing of the programmable controller. The system of the invention has the advantages of being capable of simulating all operating tasks of the crane, displaying the lifelike working environment, field and assignment, prolonging the training time of students, lowering the training risk and lowering training cost while improving the teaching quality of the training institution.

The invention relates to a robot target grabbing detection method based on a continuous path. The continuous path is obtained on a grabbed objected, geometric center points of overlapped grabbing areas on the grabbed object are connected, and a path set is obtained; a redundancy path is removed from the path set, confidence evaluation is carried out, and the effective continuous path is obtained;on the basis of the YOLO V3 model, a multiscale grabbing detection model of the grabbed object is built; and under a Darknet frame, the multiscale grabbing detection model is trained, an image containing a grabbed object is input to the trained multiscale grabbing detection model, and an output grabbing parameter is obtained. The method has the beneficial effects that firstly, all feasible grabbing area distributions can be described, the grabbing probability of the grabbing area can be more accurately evaluated and predicated, rapid convergence of the grabbing detection model is facilitated,the model training time is shortened, multiple grabbing areas with different positions and different scales can be predicated at the same time, and multiple grabbing chooses are provided for actual grabbing operation.

The invention relates to an improved binocular stereo matching method based on PSMNet, and the method comprises the steps: obtaining a binocular image, and constructing a backbone network based on PSMNet; wherein the network comprises a deep convolutional network used for extracting left and right feature maps of a binocular image; a pyramid pooling structure used for extracting multi-scale targetfeatures of the left and right feature maps; a matching cost volume used for performing cost aggregation on the multi-scale target features to obtain a 3D feature module; a 3D convolution structure used for carrying out subsequent cost calculation on the 3D feature module; giving different weights to different feature points by introducing a channel attention mechanism to improve the structure ofa matching cost volume; designing a network structure based on an encoding process and a decoding process to improve a 3D convolution structure to obtain an improved PSMNet-based backbone network; and carrying out stereo matching on the binocular image. The stereo matching method can enable the network structure to obtain faster training time and higher parallax precision, and has better practicability.

The invention provides an analogue simulation system for a tower crane. The analogue simulation system for the tower crane comprises a control board, a master controller, a sensor, a single-chip microcomputer and a main unit, wherein the control board is used for receiving a tower crane operating gesture of a student and sending the tower crane operating gesture of the student to the master controller, the master controller is used for converting the received operating gesture into the corresponding magnetic induction intensity and sending the magnetic induction intensity to the sensor, the sensor is used for converting the received magnetic induction intensity into a corresponding digital voltage signal and sending the digital voltage signal to the single-chip microcomputer, the single-chip microcomputer is used for conducting data processing on the received digital voltage signal and sending the processed data to the main unit, and the main unit is used for displaying the scene of simulating operation conducted on the tower crane according to the received processed data and calculating and displaying an operation result. By the adoption of the analogue simulation system for the tower crane, flexible training can be achieved, inconvenient teaching caused by the problems such as the environment or the site is avoided fundamentally, the student is taught to master various operation skills before operating the real tower crane, teaching time and teaching cycle can be shortened, and resource consumption and capital cost are reduced greatly.

The invention belongs to the technical field of model optimization, and discloses a parameter selection optimization method, equipment and terminal in random forest model training, and the parameter selection optimization method in random forest model training comprises the steps of determining the parameter influence of a random forest; building a parameter optimization algorithm based on QGA-RF; and performing random forest optimization based on the quantum genetic algorithm. Experiments prove that through QGA optimization, the classification performance of the random forest algorithm is improved, and the training time of the model is within an acceptable range; compared with the GA, the QGA has better global search capability and is not easy to fall into a local optimal solution. Meanwhile, an improved QGA is used for optimizing the random forest classification model, the influence of two parameters in the random forest on the model classification performance is given, a pair of optimal parameter solutions are searched through the QGA, and finally the effectiveness of the method is proved through experiments.

The invention discloses a ball-throwing device used for practicing tennis serving action, which comprises a base plate, a support 1 and a support 2. The top of the base plate is sequentially installed with a support 2, a bearing frame and a support 1 from left to right. The bearing The inner cavity at the top of the frame is movably equipped with a rotating bearing, and the middle part of the rotating bearing is fixed with a parabolic power wheel. One end of the rotating bearing on the left side of the parabolic power wheel is connected with the rotating shaft of the parabolic motor. There is a parabolic guide cylinder, and the inner cavity of the parabolic guide cylinder is provided with a push rod. The present invention is a ball throwing device used for tennis serving practice. The device has a simple structure and is easy to use. The device also effectively reduces the physical energy consumption of athletes and improves the training efficiency of athletes.