129results about How to "Improve convergence rate" patented technology

The invention discloses a mobile robot path planning method based on an improved RRT* algorithm. The method introduces a target biasing strategy into a standard RRT* algorithm so as to reduce the randomness of sampling points; provides an avoidance step length extension method in order that a random tree can reasonably stay away from an obstacle area and avoids falling into a local minimum; and smoothes a path obtained by the improved RRT* algorithm by using a reverse sequence connection method smoothing strategy, so as to reduce the direction-changing operations of the robot and achieve the stable movement of the robot. Compared with an original standard RRT* algorithm, the improved RRT* algorithm has a better planned path and takes less time.

The invention discloses an MB-RRT-based unmanned aerial vehicle two-dimensional track planning method, comprising the steps of initializing a tree and environmental information; importing obstacle information, and setting a number of iterations; judging whether the number of iterations is arrived, if so, performing down-sampling on the generated path point and optimizing the generated path line by adopting an interpolation algorithm; otherwise, generating a random sampling point, looking for a point nearest to the random sampling point in the tree, generating an adaptive step length according to the point, generating a final interpolation point according to the step length, judging whether the distance between the interpolation point and the root is greater than the current optimal path length, if not, performing collision detection on the path, adding the interpolation point to the tree and optimizing adjacent nodes around the interpolation point; if not, performing connection detection and connection on the tree. The method is high in convergence rate and small in memory occupation space, solves the problem of limitation of growth nearby an obstacle, and can be directly applied to unmanned aerial vehicle control.

The invention discloses a federated learning method and system based on batch size and gradient compression ratio adjustment, which are used for improving model training performance. The method comprise the following steps: in a federated learning scene, enabling a plurality of terminals to share uplink wireless channel resources; completing the training of a neural network model together with anedge server based on training data of a local terminal; in the model training process, enabling the terminal to calculate the gradient by adopting a batch method in local calculation, and in the uplink transmission process, compressing the gradient before transmission; adjusting the batch size and the gradient compression rate according to the computing power of each terminal and the channel stateof each terminal, so as to improve the convergence rate of model training while ensuring the training time and not reducing the accuracy of the model.

The method discloses a robust feature tracking and stereoscopic vision positioning technology based on image processing and machine vision. The technology can integrate inertial information and visual information and achieve reliable stereoscopic vision positioning under camera waggling conditions and outdoor light conditions. Images are collected through a binocular video camera in real time, and rotation information of the camera is collected with an inertial measurement unit. Feature points in the images are extracted with a feature extraction algorithm, and the feature points of the left image and the feature points of the right images are matched stereoscopically. The inertial information is combined and the inertia and the KLT algorithm are integrated to track the feature points, so that the reliability of the feature tracking is promoted. Three-dimensional information of the feature points is restored according to the double vision geometric principle. Motion parameters of the camera are obtained through position information of the feature points with the Gaussian and Newton iteration method. The accuracy of visual positioning is further promoted with the RANSIC algorithm. The whole process is iterated continuously, and thus real-time calculation of the posture and the position of the camera is achieved.

The invention discloses a monocular light field image unsupervised depth estimation method based on a convolutional neural network. According to the method, the disclosed large-scale light field imagedata set is firstly used as a training set, and samples of the training set tend to be balanced through data enhancement and data expansion; an improved ResNet50 network model is constructed; an encoder and a decoder are used for extracting high-level and low-level features of a model respectively, results of the encoder and the decoder are fused through a dense difference structure, meanwhile, asuper-resolution shielding detection network is additionally constructed, and the shielding problem between all visual angles can be accurately predicted through deep learning; the objective functionbased on the light field image depth estimation task is a multi-loss function, the preprocessed image is trained through a pre-defined network model, and finally generalization evaluation is carriedout on the network model on a test set. According to the method, the preprocessing effect on the light field image of the complex scene is obvious, and the effect of more accurate light field image unsupervised depth estimation is achieved.

The invention relates to a rapid diagnosis and scoring method for a full-scale pathological section based on deep learning. Preprocessing is carried out on a full-scale pathological section staining map; the node number of a full-connection layer and an output layer of a traditional AlexNet neural network is changed to meet the needs of practical problems, a marked training sample set is selectedto train two AlexNet neural network models for diagnosis and scoring, and high-dimensional feature information of a lesion area is extracted; with the two improved AlexNet neural network models aftertraining, the full-scale pathological section staining map is diagnosed and scored; and according to a diagnosis predicted probability, a probability heat map is drawn and the lesion area is identified visually, statistics of proportions of small sampling block numbers with different lesion degrees is carried out, and the lesion degree of a tissue is scored. Therefore, the diagnosis and Gleason scoring of the full-scale pprostate tissue pathological section are realized automatically; and the accuracy rate and the calculation rate exceed the average level of the artificial diagnosis substantially.

The invention discloses a humanoid robot gait control method based on model correlated reinforcement learning. The method comprises steps of 1) defining a reinforcement learning framework for a stable control task in forward and backward movements of a humanoid robot; 2) carrying out gait control of the humanoid robot with a model correlated reinforcement learning method based on the sparse online Gaussian process; and 3) improving a motion selection method of a reinforcement learning humanoid robot controller by a PID controller, and taking the improved operation as an optimizing initial point for the PID controller obtaining the motion selection operation of the reinforcement learning controller. The invention utilizes reinforcement learning to control gaits of the humanoid robot in movement, and thus the movement control of the humanoid robot can be automatically adjusted via interaction with the environment, a better control effect is achieved, and the humanoid robot is enabled to be stable in forward and backward directions.

The invention discloses an ancient font classification method based on a convolutional neural network. According to the method, firstly, an ancient font category image data set is crawled by using a crawler technology; through data expansion, training set samples tend to be balanced; graying processing is carried out on the balanced training set sample and setting an image size to a target image size; histogram equalization processing is performed on the sample set, isolated noise points are removed in the image through an N8 connected noise reduction algorithm, and finally binarization processing is performed on the image based on a fuzzy set theory and by using a Shannon entropy function, so that detail features of the image are well reserved; based on the objective function of the classification task. The center loss function and the traditional cross entropy loss function are matched for use. The inter-class distance is increased. The intra-class distance is reduced. The distinguishing capability of features is improved to a certain extent, preprocessed images are trained through a pre-defined network model, and the accuracy of a classification result is evaluated through a confusion matrix. According to the method. The preprocessing effect on the degraded ancient font image is remarkable, and a more accurate ancient font classification effect is achieved by optimizing parameter setting and utilizing appropriate training skills to train the convolutional neural network model.

The invention discloses a bottom material sonar image classification method based on a convolution neural network, belonging to the technical field of image classification. The method comprises obtaining the sonar image of seafloor bottom, preprocessing the image with denoising and enhancement, extracting the edge shape based on a canny algorithm, and generating a gray scale-basic element co-occurrence matrix, constructing a convolution neural network classifier structure and a sample set, training a neural network, obtaining classification model and realizing bottom material sonar image classification. The present invention is directed to the graphical characteristics of bottom material sonar images of seafloor, the disadvantage of using single method is solved, and the learning strategyof the convolution neural network classifier is used to learn and train different types of seabed sediment, and finally the classification model with classification function is obtained, and the purpose of fast and accurate classification of seabed sediment sonar images is achieved.

The invention provides a high-performance computing cluster dynamic node operation method. The method comprises the following steps: providing a unified system file system space and diskless booting based on an Infiniband network, providing cluster operation scheduling, power on / off control, system monitoring and like functions based on the Ethernet. Through above mentioned way, the deployment efficiency of the high-performance computing cluster can be improved, a fault rate is lowered, the job submitting is simplified, the energy source use rate and the capital use rate are improved, and the use efficiency and performance are promoted.

The invention discloses an active noise control system and method based on an improved FxLMS algorithm. The active noise control system mainly comprises five modules as follows: (1) the FxLMS algorithm, (2) a secondary channel, (3) a performance monitor, (4) a variable-power white noise generator, and (5) a main channel path. The invention aims at improving noise reduction performance of an ANC system as well as modeling precision and convergence rate of the secondary channel; in accordance with innovation points, a training signal (auxiliary random white noise) of the secondary channel undergoes power scheduling, and then performance in (2) is observed, wherein injection of the auxiliary random white noise in the (4) is stopped when a following relation is satisfied: [mu]Smax-[mu]S<[alpha](1*10<-5><[alpha]<1*10<-3>), and injection of the auxiliary random white noise in the (4) is started once again when a following relation is satisfied: 20log10|f(n)|<0. Therefore, mutual transformation of the secondary channel between online modeling and offline modeling is achieved, and the finally obtained ANC system is relatively high in modeling precision and noise reduction performance. Theactive noise control system has the characteristics of simple structure, easy adaption to environmental change and big-variance auxiliary white noise and the like.

The invention provides a self-adaptive device and a method for analyzing global power flow of generation, transmission and distribution, which belongs to the technical field of power transmission and distribution of power systems. The device comprises an upper computer and a lower computer, wherein the lower computer comprises a power-flow feedback module; a global power system is divided into a master system and a slave system; according to the characteristics of each part of the power system, the master system adopts a Newton-Raphson power-flow analysis method, while the slave system adopts a forward-backward substitution power-flow analysis method; and the master system and the slave system are organically linked together through master-slave-system association nodes in the middle. The invention has the advantages that the device adopts the power-flow feedback module, uses slave-system node-voltage signals obtained from last power-flow calculation of the slave system for next power-flow calculation of the slave system, and realizes the optimization of the prior forward-backward substitution algorithms, and as the method adopts the forward-backward substitution method which reduces iterations compared with the prior forward-backward substitution method, the device for analyzing global power flow not only guarantees accurate power-flow analysis, but also raises convergence rate and saves storage space.

The invention discloses a mixed-experience multi-agent reinforcement learning motion planning method, namely an ME-MADDPG algorithm. According to the method, through MADDPG algorithm training, when a sample is generated, experience is generated through exploration and learning, high-quality experience of successfully planning multiple unmanned aerial vehicles to a target through an artificial potential field method is added, and the two kinds of experience are stored in different experience pools. During training, a neural network collects samples from the two experience pools through dynamic sampling according to the changing probability, state information and environment information of the agents serve as input of the neural network, and the speeds of the multiple agents serve as output. Meanwhile, the neural network is slowly updated in the training process, the training of a multi-agent motion planning strategy is stably completed, and finally, the agents autonomously avoid obstacles in a complex environment and can smoothly reach the respective target positions. According to the method, a motion planning strategy with better stability and adaptability can be efficiently trained in a complex dynamic environment.

The invention discloses a WSAN actuator task distribution method based on BA-BPNN data fusion, and the method employs a BA optimization BP neural network to build a data fusion model. The method specifically comprises the steps: employing a bat algorithm to optimize the weight value and threshold value of the BP neural network, building a data fusion model, carrying out the data fusion of the sensor node information, and obtaining the task distribution information of an actuator node. The bat algorithm is a meta heuristic type group intelligent optimization algorithm, employs an echo positioning method of a miniature bat under the condition of different transmitting speeds and responses, can achieve a precise capturing and obstacle avoidance random search algorithm. The BP neural network is a multilayer feedforward neural network which can search a global optimal value in a training process, and can increase the convergence rate of the network. The method searches the optimal parameter of the BP neural network through the positioning updating of bats, is more precise in data fusion, and is more reasonable in task distribution of an actuator.

The invention discloses an all-PMU (phase measurement unit)-orientated robust state estimation method comprising the following steps of: initializing a network topology model and network parameters of a power system, and updating a switch measurement value and a PMU measurement value; performing topological contraction and measurement matching; identifying the range of observable nodes; performing equivalent measurement transformation; and establishing a state estimation model and solving. According to the all-PMU-orientated robust state estimation method disclosed by the invention, the convergence rate, the qualification rate and the calculation frequency of a state estimation system are remarkably improved, the maintenance workload of the state estimation system is greatly reduced, and the foundation is laid for rapid sensing, comprehensive sensing and accurate sensing for a power grid. The robust state estimation method disclosed by the invention has remarkable improvement on the aspects of calculation accuracy, calculation speed and convergence rate compared with the existing methods.

The invention discloses a method for eliminating an underwater acoustic channel interference in underwater acoustic communication. The method comprises: step one, establishing an impulse response timedomain model and an impulse response frequency domain model of an underwater acoustic channel; step two, carrying out training by using a training sequence according to the impulse response time domain model of the underwater acoustic channel to obtain a channel impulse response matrix; step three, eliminating a channel interference of an information sequence by the training sequence by using thechannel impulse response matrix, thereby obtaining an information sequence after the interface elimination; step four, carrying out balancing between channels on the information sequence after the interface elimination to obtain estimation of the information sequence after channel balancing; and step five, carrying out soft-decision decoding on the estimation of the information sequence to obtaina final hard decision bit. According to the method disclosed by the invention, an accurate feedback symbol can be obtained; the inter-symbol interference is eliminated well; and the error propagationeffect is improved.

The invention discloses an efficient neural network training scheduling method based on a heterogeneous distributed system. The method comprises the following steps: firstly, detecting and analyzing resource dynamic changes in a distributed system through a resource detection system; a training process is decomposed into internal iteration and external iteration as important subsets of a task scheduling system, and then the task scheduling system adaptively modifies environmental parameters and schedules and calculates according to distributed system node state information provided by a resource detection system. Related experiments performed under a public data set show that the method has better robustness and expandability on the premise of ensuring high accuracy and convergence rate.

The invention relates to a high-dimensional multi-target set evolutionary optimization method based on preference of a decision maker. According to the method, the objective function of an original optimization problem is converted into an expectation function according to the preferential area of each target given by the decision maker; the expectation function optimization problem is converted into a two-target optimization problem with a set formed by multiple solutions of the original optimization problem as a new decision variable and the hypervolume and the satisfaction degree of the preference of the decision maker as a new objective function; an internal self-adaptive crossing strategy of individuals of the set is designed according to the hypervolume contribution degree of the solutions of the original optimization problem in the set and the satisfaction degree of the preference of the decision maker; furthermore, an individual variation strategy of the set is designed by means of the updating of particles in the PSO algorithm and the idea of a globally optimal solution and a locally optimal solution, so that a Pareto optimal solution set satisfying the preference of the decision maker and meeting the requirement for convergence and distributivity balance is obtained.

The invention relates to a robust adaptive autopilot control algorithm for a navigation simulator. Aiming at various ship models in navigation simulator training, a nonlinear mathematic model for ship motion is obtained based on half theoretical calculation and half experimental result, and the nonlinear mathematic model considers the interferences of wind, wave, current, and the like and relates to a low-speed shallow-water model. The designed autopilot algorithm is based on a closed-loop gain shaping simple robust control algorithm and is a new ship autopilot robust adaptive control algorithm having robustness and adaptability by combining a neurotic network direct controlling training method. The robust adaptive autopilot control algorithm has adaptability by considering the robustness of a ship motion control algorithm as well as the complexity of ship navigation and solves the problem that the prior algorithm has either robustness or adaptability and can not have both.

The invention provides a sensor array steady adaptive beamforming method comprising the following steps: using array observation data covariance matrix to calculate orientation vector; calculating array observation data space power spectrum according a sample covariance matrix and the orientation vector; determining an orientation section of desired signals and interference signals according to a spectrum peak position of the space power spectrum; averaging space power spectrums outside the orientation section, and calculating an average noise power; using a convex optimization method to obtain reconstructed a signal covariance matrix and an interference plus noise covariance matrix; calculating a product matrix of the said two matrixes, and using the feature vectors corresponding to the maximum feature value to obtain beamforming device factors. The method can effectively solve the performance saturation problems in the adaptive beamforming process, so the adaptive beamforming is high in stability, and easy to be applied.

The invention discloses a multi-unmanned aerial vehicle motion planning method based on an artificial potential field method and MADDPG. According to the method, high-quality experience of successfully planning a plurality of unmanned aerial vehicles to a target through the artificial potential field method is added on the basis of the original multi-unmanned aerial vehicle exploration environment experience, through the MADDPG algorithm training, samples are collected from exploration environment experience and high-quality experience at a certain probability, state information and environment information of each unmanned aerial vehicle serve as input of a neural network, the speeds of the multiple unmanned aerial vehicles serve as output, training of a motion planning strategy is completed, the multi-unmanned aerial vehicle autonomous obstacle avoidance in a complex environment is realized, and the target position is safely and quickly reached. According to the method, the Q values of the multiple unmanned aerial vehicles in different states and different actions are fully learned, the robustness of the strategy is improved, an excellent strategy with higher adaptability and higher flexibility is trained, and the method has a good application prospect in a scene of cooperative motion planning of the multiple unmanned aerial vehicles.

The invention discloses a multi-user detection method for an SCMA (Sparse code multiple access) communication system for dynamic message scheduling and belongs to the field of single detection of the wireless communication system. The method adopts a dynamic scheduling strategy and utilizes a residual (the difference degree before and after updating of one message) as a measurement criteria. During each of the second to tmaxth iteration processes, the method comprises the following steps of firstly, selecting a user node ujmax and a resource node ckmax which are corresponding to a maximum residual value according to the residual values from all user nodes to resource nodes, which are computed in the last iteration; updating messages from the resource node ckmax to all corresponding user nodes except for the user node ua of the ujmax; and updating messages from the user node ua to all corresponding resource nodes except for the user node cb of the ckmax. According to the method, the dynamic scheduling method updates the non-convergent node messages in the iteration preferentially, so that the iterative decoding convergence rate of the whole codon can be accelerated; the updated node messages are utilized in real time in each iteration, so that the utilization rate of the updated nodes can be effectively improved; and the computational complexity is low and the BER (Bit Error Rate) performance is excellent.

The invention discloses a preventive maintenance decision-making optimization model for key components of a train bogie based on maximum and minimum ant colony algorithms. The preventive maintenance decision-making optimization model for key components of the train bogie and the constraint conditions there of are built, and then the opportunity maintenance threshold and maintenance cost are obtained according to the maximum and minimum ant colony algorithms. A method employed by the model comprises the following steps: setting a node set Ci; setting relevant parameters; placing m ants on the nodes to start visit, and then counting the path length of each ant, and recording as a current best solution; updating the path information according to an information update principle until all nodesare visited; evaluating the solution of the access path of each ant according to an optimization target of a model E, and selecting a solution with the shorter path as an updating value, wherein thenumber of ants visiting the path will be gradually increased with the shorter path and stronger pheromones; finally, judging the relationship between the number of current iterations and the total number of times, and finding out an optimal solution. The model is high in calculation speed, is high on solving relation, and can effectively reduce the maintenance cost of the key parts.

The invention discloses a personalized federated meta-learning method for data isomerism. The method comprises the following steps: determining an automatic encoder structure in an initialization stage of each client and a meta-model structure in a personalized stage; initializing parameters of a federation training stage; grouping the clients according to the local data distribution vectors uploaded by the clients; aggregating the client models in each group, and issuing the aggregated client models to the clients in the group to carry out the next round of iteration; and after federation training is finished, the client performs fine tuning on the meta-model in the group and local data thereof to generate a personalized model. According to the method, when the clients participate in federation training, the clients with approximate data distribution are dynamically divided into the same group according to the local data distribution vectors uploaded in each round, and the corresponding meta-model is set for each group, so that the problems of slow model convergence and low accuracy caused in a highly heterogeneous data environment are solved.

The invention provides a parameter and full-interval SOC joint estimation method of a power battery. Power battery parameter identification is performed by adopting an adaptive recursive square root algorithm, a parameter updating rate can be adaptively changed based on a working condition and a system state, meanwhile, a problem of numerical instability caused by rounding errors of an embedded system is avoided, and a convergence rate of parameter identification and stability and reliability of a result are improved. Controllable correction of SOC is achieved through a correction starting judgment module, a SOC correction amount is restrained under conditions of an open-circuit voltage platform period, a low SOC interval, a low temperature and the like, and stability of SOC estimation isguaranteed. A SOC correction function is started in a non-platform period, and a SOC error is rapidly corrected so that the method has relatively high applicability and relatively high estimation precision for a whole system battery in a whole SOC interval.

The invention discloses an interference management method under partial channel state information in an uplink. The method mainly solves the problem in the prior art that a resource distribution method is great in information interaction cost and an association control method is low in convergence speed. The method comprises following steps that: 1) a macro base station carries out association control to all users by a history subgradient descent method; 2), a home base station distributes resources to every user associated with the home base station according to a constructed home user utility function; 3), the macro base station firstly selects a reference home base station for every user associated with the macro base station and then distributes resources to every user associated with the macro base station according to a constructed macro user utility function. In application of the method, the cross layer interference and the information interaction cost are reduced; meanwhile, the system throughput and the convergence speed of the association control are improved; and the method is applicable to a macro base station and home base station coexisting heterogeneous wireless network.

The invention provides a route synchronization method, a device and a communication system. The method comprises obtaining routing information which includes route prefix information and route forwarding information; and synchronizing the route prefix information and the route forwarding information with a receiving device separately according to a preset transmission order. Thus, by using an approach of separately synchronizing the route prefix information and the route forwarding information, the number of synchronized routes between devices in unit time is increased so as to achieve an effect of increasing the convergence rate of the routes.

The invention belongs to the technical field of communication, and relates to a joint detection method for signals of a massive MIMO uplink system. According to the scheme of the invention, the methodincludes the following steps that: firstly, an iterative matrix is constructed, and a low-complexity iterative scheme is proposed based on the constructed matrix; and then, a steepest descent methodand the proposed iterative method are mixed to accelerate the proposed iterative method; in addition, the computational complexity is reduced by skillfully utilizing the properties of inversion of partitioned matrices and a matrix-vector multiplication; and then, the detailed convergence proof and complexity analysis are performed; and finally, simulations prove that the BER performance of the proposed algorithm is better than that of most existing iterative algorithms and the near-optimal performance of an MMSE algorithm can be reached in a small number of iterations.

The invention discloses a fast underwater robot three-dimensional path planning method with target-oriented centralized optimization. A target-oriented Gaussian sampling strategy is introduced in an RRT* path planning method, thereby reducing the randomness of sampling points when underwater obstacle distribution is scarce; random disturbance is combined for sampling, a random tree can be reasonably away from an obstacle area, falling into the local minimum is avoided, and search is quickly orientated to the target; and a centralized optimization search strategy is used to perform path optimization processing on the initial path acquired in the invention, the optimization convergence rate and the path quality are improved, and progressive optimization is realized. Compared with the traditional RRT* method, the improved method disclosed in the invention plans a better initial path, the path optimization speed is quicker, and consumption of time and memory is greatly reduced.

The invention discloses a weld joint radiographic inspection negative film image enhancement method, a storage medium and equipment. The method comprises the following steps: acquiring an original weld joint radiographic inspection negative film image by weld joint radiographic inspection negative film scanning equipment; constructing a training set and a test set of weld joint radiographic inspection negative film images, wherein the training set and the test set comprise image pairs of all weld joint radiographic inspection negative film image data subjected to enhancement processing and original data; creating a convolutional neural network model, wherein the convolutional neural network model comprises an encoding and decoding structure of a lower sampling layer and an upper sampling layer; and training the constructed network model by using the image data of the training set of the weld joint radiographic inspection negative image to generate a network model, and testing the modeleffect by using the image data of the test set of the weld joint radiographic inspection negative image. The weld joint radiographic inspection negative film image enhancement method is learned by using a deep learning technology, and has a good effect of processing the weld joint radiographic inspection negative film image.