Patents
Literature
Hiro is an intelligent assistant for R&D personnel, combined with Patent DNA, to facilitate innovative research.
Hiro

173results about How to "Shorten training time" patented technology

Driving model training method, driver identification method, driving model apparatus, driver identification apparatus, device and medium

The present invention discloses a driving model training method, a driver identification method, a driving model apparatus, a driver identification apparatus, a device, and a medium. The driving modeltraining method includes the following steps that: the training behavior data of a user are acquired, wherein the training behavior data are associated with a user identifier; training driving data associated with the user identifier are obtained on the basis of the training behavior data; positive and negative samples are obtained from the training driving data on the basis of the user identifier, and the positive and negative samples are divided into a training set and a test set; the training set is trained by using a bagging algorithm, so that an original driving model can be obtained; and the test set is adopted to test the original driving model, so that a target driving model can be obtained. With the driving model training method adopted, the generalization of the driving model can be effectively enhanced; the problem of poor recognition results of current driving recognition models can be solved; and the accuracy of identifying the driving of drivers is improved.
Owner:PING AN TECH (SHENZHEN) CO LTD

Neural network equalization method used for indoor visible light communication system

The invention provides a neural network equalization method used for an indoor visible light communication system, and belongs to the visible light wireless communication technology field. The method includes the steps: utilizing a ceiling bounce model to calculate a VLC channel impulse response, carrying out photoelectric conversion for a visible light power signal received by a receiving end, and sending a sequence to a neutral network channel equalizer after amplification sampling; utilizing a heredity algorithm to optimize initialization weights and thresholds among neurons, establishing a neural network for training, and minimizing an error function; and judging an output, restoring the sent sequence, and finally achieving an equalization purpose. According to the scheme, interference among codes is obvious minimized, an error rate is reduced, the communication quality is further improved, a transmission rate that the system can reach is increased, the training time is shortened, and the system complexity is reduced.
Owner:CHONGQING UNIV OF POSTS & TELECOMM

Image super-resolution reconstruction method

The invention relates to an image super-resolution reconstruction method, belongs to the image processing technology field and solves problems that the edge information of an image generated in the prior art is fuzzy, application to multiple magnification times cannot be realized and the reconstruction effect is poor. The method comprises steps that a convolutional neural network for training andlearning is constructed, and the convolutional neural network comprises an LR characteristic extraction layer, a nonlinear mapping layer and an HR reconstruction layer in order from top to bottom; inputted paired LR images and HR images are trained through utilizing the convolutional neural network, training of at least two magnification scales is performed simultaneously, and an optimal parameterset of the convolutional neural network and scale adjustment factors at the corresponding magnification scales are acquired; after the training is completed, the target LR images and the target magnification times are inputted to the convolutional neural network, and the target HR images are acquired. The method is advantaged in that the training speed of the convolutional neural network is fast,after training is completed, and the HR images at any magnification times in the training scale can be acquired in real time.
Owner:CHINA UNIV OF MINING & TECH

A Chinese text sentiment analysis method based on deep learning

The invention discloses a Chinese text sentiment analysis method based on deep learning, and belongs to the technical field of natural language processing. The defects of an unsupervised sentiment analysis method based on English are overcome. The method comprises the following steps: after converting an obtained corpus text into pinyin, pre-training a constructed language model to obtain a pre-trained language model; obtaining a small amount of text data which is in the same field as the corpus text and has emotion categories, converting the text in the text data into pinyin, and training a constructed emotion classification model based on a pre-trained language model to obtain a trained emotion analysis model; and carrying out sentiment classification on the unlabeled text by utilizing the trained sentiment analysis model to obtain a corresponding sentiment category label. The method is used for Chinese text sentiment analysis.
Owner:SICHUAN XW BANK CO LTD

CAN bus-based pig house environmental temperature intelligent monitoring system

The invention relates to a CAN bus-based pig house environmental temperature intelligent monitoring system. The system is characterized in that the intelligent monitoring system is composed of a CAN bus-based pig house environment parameter acquisition and intelligent prediction platform, a pig house environment multi-point temperature fusion model and a pig house environment intelligent prediction model. With the CAN bus-based pig house environmental temperature intelligent monitoring system of the present invention adopted, an existing pig house monitoring system cannot intelligently monitor and predict the temperature of the environment of a pig house according to the non-linearity and large lag of the environmental temperature change of the pig house, the large area of the pig house, complicated temperature change and other characteristics, as a result, control and adjustment of the environmental temperature of the pig house are seriously affected, while, with the CAN bus-based pig house environmental temperature intelligent monitoring system adopted, the above problem can be solved.
Owner:江苏华丽智能科技股份有限公司

Video target detection method based on machine learning

The invention discloses a video target detection method based on machine learning. The method comprises the steps that (1) for an input video, an SSD target detection algorithm is adopted to obtain ato-be-tracked target detection box, and a bounding-box is marked on an image to determine a tracking target; (2) two tracking methods are adopted for each frame of the input video, wherein one tracking method is a light stream tracking algorithm, a tracking point of the next frame is predicted according to a probability, and the tracking point of the next frame is precisely determined through a Euclidean distance and a set threshold value; and the other tracking method is to adopt a full-convolutional neural network and extract high-layer features and low-layer features in the neural network for separate convolution, finally the features are fused into a feature graph through a classifier, and therefore the tracking point of the next frame is precisely determined; and (3) HOG features of the light stream tracking result and the full-convolutional neural network tracking result are extracted, validity discrimination is performed on the two results through a support vector machine (SVM),and the target position of the next frame is determined finally.
Owner:SUN YAT SEN UNIV

Quick evolution method for optimized deep convolution neural network structure

The invention discloses a quick evolution method for an optimized deep convolution neural network structure. The method comprises the following steps that: 1) utilizing an evolution algorithm based onGNP (Gene Network Coding) to effectively construct a nonlinear CNN (Convolutional Neural Network) structure, and carrying out mutation on various hyper-parameters of the CNN structure to search an optimal CNN hyper-parameter combination; 2) in an evolution process, designing a multi-objective network structure evaluation method, simultaneously taking classification accuracy and the complexity degree of the classifier as an optimization objective so as to aim at generating the CNN classifier with high classification accuracy and a simple structure; and 3) putting forwarding an incremental training method, and carrying out filial generation CNN structure training on the basis of a previous generation of CNN structure. By use of the method, the training frequency of the model can be reduced,and the time complexity of an algorithm is lowered.
Owner:ZHEJIANG UNIV OF TECH

Neural network active-disturbance-rejection controller for AC radial magnetic bearing, and construction method thereof

The invention discloses a neural network active-disturbance-rejection controller for an AC radial magnetic bearing, and a construction method thereof. The input of a first tracking differentiator is given radial displacement x*, and the output is a tracking signal x1 and a differential signal x2; the input of a first self-adaptive expanded state observer is controlling quantity u, radial displacement x and three parameters beta01, beta02 and beta03, the other two inputs of a first nonlinear state error feedback control rule is parameters beta1 and beta2, and the output is the controlling quantity u0; the difference of the controlling quantity u0 and an estimation value z3 is the input of a first compensation factor, the output of a second compensation factor is the controlling quantity u,and the controlling quantity u is used as one input of a first self-adaptive active-disturbance-rejection controller. By constructing the self-adaptive extended state observer, the internal disturbance and the externa disturbance of the controlled object are automatically controlled, and the online automatic adjusting of the three parameters beta01, beta02 and beta03 can be realized along the system disturbance change, the estimation and compensation precision on the disturbance by the extended state observer are increased, and the control performance of the active-disturbance-rejection controller is improved.
Owner:JIANGSU UNIV

Image denoising method based on ELM

The invention discloses an image denoising method based on an ELM. The method comprises the following steps: establishing a basic feedforward neural network according to size of a to-be-processed image; aimed at the basic feedforward neural network, establishing a training sample set; using the training sample set to train the basic feedforward neural network based on an ELM method, to obtain a trained neural network; and inputting the to-be-processed image to the trained neural network, and the corresponding output being a de-noised image. Through customizing the training set and using the elm, the method trains connection parameters, so a training process can be completed rapidly, thereby greatly improving training efficiency. The method establishes the network training set according to noise types of an application scene, just the network training set established aimed at the application scene is needed to obtain the trained neural network aimed at different noise types to eliminate noise of an image. The method can be conveniently applied in different noise scenes.
Owner:ZHEJIANG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Motor fault diagnosis method and system based on cavity convolution capsule network

The invention relates to a motor fault diagnosis method and system based on a cavity convolution capsule network, and the method comprises the following steps: (1), obtaining a training sample with alabel, wherein the training sample comprises a motor vibration signal and a corresponding operation state, and the operation state comprises a normal state and a fault type in a fault state; (2) establishing a cavity convolution capsule network, and performing training by using the training sample; and (3) acquiring a to-be-diagnosed motor vibration signal, inputting the to-be-diagnosed motor vibration signal into the trained cavity convolution capsule network, and outputting the operation state of the motor. Compared with the prior art, the method has the advantages that the effective features of the motor signals can be automatically extracted, intelligent fault diagnosis is achieved, the diagnosis accuracy reaches 99% or above, the robustness and generalization ability are high, and theerror recognition rate is remarkably reduced.
Owner:SHANGHAI DIANJI UNIV

Method for constructing convolution neural network in novel network topological structure

The invention relates to the construction of a convolution neural network, and aims at solving problems that a conventional neural network algorithm exerts higher requirements for the processing speed of hardware and cannot be applied in electric household appliances. The method comprises the following steps: determining an interconnection structure among nerve cell nodes, and determining the mutual relation between numerical calculation feedback and forwarding calculation transfer among the nerve cell nodes; building a multilayer perceptron model through employing a reverse propagation neural network, wherein the structure of the reverse propagation neural network comprises a forwarding propagation algorithm; operating a reverse propagation algorithm and the forwarding propagation algorithm at the same time, and carrying out training and calculation of the convolution neural network. The method is suitable for mode recognition of household electric appliances.
Owner:SICHUAN CHANGHONG ELECTRIC CO LTD

Networked control system fault detection method based on neural network prediction

The invention discloses a networked control system fault detection method based on neural network prediction, which comprises four steps of RBF neural network system building, system fault detection function building, system stability judgment and operation and system fault judgment and operation function building. The system building and operation process is simple, the operation efficiency and the operation precision are relatively high, an improved RBF neural network prediction controller is adopted to effectively predict system output data information, and thus, bad influences on the system by packet loss can be effectively cancelled, errors are smaller and training times are reduced through adjusting learning efficiency on the basis of adopting feedback correction on the obtained predicted output value for correction, and better convergence and quicker prediction speed can be obtained. Meanwhile, when fault happens to the system, happening of the fault can be quickly detected according to a designed fault observer and a judgment criterion.
Owner:HENAN POLYTECHNIC UNIV

High-resolution analog beam rapid training method and device

The invention discloses a high-resolution analog beam rapid training method and a high-resolution analog beam rapid training device. The high-resolution analog beam rapid training method comprises the steps of setting relevant parameters of beam training, initializing first iteration, then carrying out many times of iteration while selecting part of code words from a codebook for training in each iteration, comparing a result with a previous iteration result, determining analog beam vectors of each antenna subarray according to an evaluation index, carrying out iteration calculation layer by layer, and completing beam training of one antenna subarray when an iteration terminal condition is satisfied. Compared with the existing beam training method, the high-resolution analog beam rapid training method is low in complexity, only selects part of the code words in the codebook for training through reasonable region search of the codebook, greatly reduces beam training times, and saves training cost. The high-resolution analog beam rapid training method adopts the codebook with large capacity, quantifies all beam directional angles in a circumference at high resolution, has a certain fault-tolerant capability, and ensures effective alignment between transmitted beams and received beams, thereby improving overall performance of the system.
Owner:白盒子(上海)微电子科技有限公司

Electric vehicle power battery SOC intelligent detection device

The invention discloses an electric vehicle power battery SOC intelligent detection device which is characterized in that the intelligent detection device includes a battery parameter acquisition platform and a battery SOC estimation system, the battery parameter acquisition platform collects real-time parameters of voltage, current and temperature of an electric vehicle power battery group, the battery SOC estimation system estimates an SOC value through the collected parameters, and a battery SOC system is a nonlinear, time-delay and multivariable coupling complex real-time system with a high requirement. According to the device, a problem that a conventional detection device can not obtain an ideal effect of the intelligent detection of the electric vehicle power battery SOC.
Owner:安徽惠宏科技有限公司

Millimeter wave MIMO communication multi-subarray cooperative beam alignment method and millimeter wave MIMO communication multi-subarray cooperative beam alignment device

The invention discloses a millimeter wave MIMO communication multi-subarray cooperative beam alignment method and a millimeter wave MIMO communication multi-subarray cooperative beam alignment device. The method comprises the following steps: (1) a receiving end and a sending end analyze a code book corresponding to each subarray and perform space division accordingly, each subarray extracts code words from own code book to form a corresponding subcode book, and a union set of the subcode books formed by extraction can cover the original space; (2) based on the extracted subcode books, the sending end uses a plurality of beam sending signals, and for a sending combination of the sending end, the receiving end simultaneously uses a plurality of beam receiving signals based on the extracted subcode books; (3) by making use of information acquired in a phase of training, a principal direction of channels is calculated, and high-efficiency and low-complexity beam selection is realized further. In the method and the device provided by the invention, by taking full use of a character that a millimeter wave channel has sparsity, a corresponding multi-subarray cooperative training framework and an effective algorithm are provided. Analysis and a simulation experiment both show that the method provided by the invention not only reduces training overhead and calculation complexity greatly, but also has very small corresponding performance loss in comparison with an exhaustive search algorithm.
Owner:白盒子(上海)微电子科技有限公司

Spiking neural network analog circuit based on reinforcement learning

The invention belongs to the technical field of spiking neural networks, and discloses a spiking neural network analog circuit based on reinforcement learning. The spiking neural network analog circuit comprises an input layer nerve cell, a hidden layer nerve cell, an output nerve cell and a synapse; the input layer neurons are connected with the hidden layer neurons through synapses, and the hidden layer neurons are connected with the output neurons through the synapses; and the synapse is used for adjusting the first pulse signal of the pre-stage neuron according to the weight value and thentransmitting the adjusted first pulse signal to the post-stage neuron, and is also used for receiving the second pulse signal output by the post-stage neuron and updating the weight value according to the time difference between the first pulse signal and the second pulse signal and the reward signal. Based on reinforcement learning, a pulse neural network circuit is built, and an XOR classification function is achieved. Compared with a traditional pulse neural network, the method has the advantages of higher training speed and higher accuracy.
Owner:HUAZHONG UNIV OF SCI & TECH

Digitized local area network motor teaching method

The invention discloses a digitized local area network motor teaching method. A digitized motor teaching practically training system is connected with computers for teachers and students in a networking mode, mutual information is transmitted and shared in a local area network by using a motor practically training teaching software, and thereby the teachers on a desk of teacher can know studying and practically training conditions of each student in a macroscopic and microscopic mode and grasp teaching conditions, the digitized local area network motor teaching method has the advantages of improving teaching quality, simplifying practically training teaching procedures, shortening practically training time and enabling practically training teaching to be more flexible, simple and targeted, and the method is particularly suitable for popularizing and using in motor practically training teaching rooms in academies.
Owner:刘坚

Neural network training method, storage medium and equipment

The embodiment of the invention discloses a neural network training method, which comprises the following steps of: constructing a training framework comprising a parameter node and a plurality of training nodes, and updating neural network parameters of the plurality of training nodes and the parameter node; training by each training node, and respectively sending neural network parameters and / orneural network cumulative gradients to the parameter nodes every other preset training steps; fusing the neural network parameters and / or the neural network cumulative gradients of the training nodesby the parameter nodes, and updating the neural network parameters and / or the neural network cumulative gradients of the parameter nodes according to the neural network parameters and / or the neural network cumulative gradients; and each training node performs training again according to the fused neural network parameters and / or the neural network cumulative gradient sent by the parameter node, and the parameter node outputs a neural network model through a preset model training termination condition. According to the neural network training method provided by the embodiment of the invention,the training efficiency of the neural network training method and the performance and training precision of the convergence model can be further improved.
Owner:BEIJING SIMULATION CENT

Power communication network reliability prediction and guarantee method and system based on deep learning

The invention provides a power communication network reliability prediction and guarantee method and a system based on deep learning. According to the method and the system, a deep belief network anda bidirectional LSTM neural network are adopted to carry out feature extraction and prediction on state data in the network and calculated reliability index data respectively, and a network state anda corresponding reliability index in a next effective time period are predicted. Then, the predicted reliability index is evaluated; and if the standard threshold value is not met, network optimization needs to be carried out to improve the reliability of the network, and during optimization, corresponding optical cable optimization, node optimization and service level optimization are selected incombination with the predicted network basic data in the next effective time period, so that the overall reliability of the network is improved. According to the method and the system, the power communication network is optimized by combining the predicted network service state of the next time period, so that the network reliability is improved from the perspective of providing communication service stably for a long time.
Owner:CHINA ELECTRIC POWER RES INST +3

Online collaborative ordering method based on stochastic gradient descent

The invention discloses an online collaborative ordering method. The method comprises the steps that a target function is built through a collaborative ordering method, and a stochastic gradient descent method is used for solving; an online collaborative ordering recommendation system is built for carrying out increment training, a recommendation list is updated in real time, and training and recommendation are achieved at the same time; a data set S is obtained and collaboratively divided into a large part and a small part; rating data of the small-proportion data set is used for building a user-product rating matrix X; a UVT model is obtained through decomposition; the online collaborative ordering method SGDRank and the small-proportion data set are used for off-line updating of a matrix U and a matrix V, and a UVT model is obtained; large-proportion data is used as an online sample to be added into a matrix X; the matrix U and the matrix V are trained on line, the UVT model is updated, the trained X matrix is obtained, and accordingly, online collaborative ordering for data is achieved. Ordering recommendation efficiency can be effectively improved by means of the method.
Owner:PEKING UNIV

Electric automobile power battery SOC (State of Charge) detection system

The invention discloses an electric automobile power battery SOC (State of Charge) detection system. The characteristics lie in that the detection system comprises a battery parameter acquisition platform and a battery SOC estimation system, the battery parameter acquisition platform is responsible for real-time parameter acquisition for the voltage, current and temperature of an automobile power battery pack, and the battery SOC estimation system can accurately estimate a battery SOC value through the acquired parameters; and the battery SOC is a non-linear, delayed, multivariable coupling and complex highly demanding real-time system. The detection system effectively solves a problem that the traditional automobile battery SOC estimation method is difficult to achieve an ideal effect.
Owner:四川欣智造科技有限公司

Target detection model training method and device and electronic equipment

The invention provides a target detection model training method and device and electronic equipment. The method comprises the following steps: acquiring reference information of a reference frame in atraining image; obtaining prediction information obtained through analysis of the target detection model; wherein the prediction information comprises prediction center point information and prediction size information corresponding to all sample points in the center area, the center area is a sub-area of the reference frame, the center point of the center area is an object center point, and theprediction size information comprises sample point information and corresponding distance information; and obtaining and repeatedly training the target detection model according to the first evaluation result and the second evaluation result until the target detection model meets the expected requirements. The second evaluation result comprises the size data corresponding to all the sample points,the data size evaluated by the second evaluation result is large and rapid to obtain, the convergence rate of the target verification model is increased, the training frequency of the target detection model is reduced, and the training speed is increased.
Owner:HANGZHOU FABU TECH CO LTD

Gas outburst prediction early warning method based on big data platform

The invention discloses a gas outburst prediction early warning method based on a big data platform. The method comprises the following steps: A, storing real-time monitoring data in an HDFS (Hadoop Distributed File System) of a Hadoop platform; B, preprocessing the real-time monitoring data by utilizing a linear exponential smoothing method; C, determining a period of the real-time monitoring data and detected outburst prevention data; D, taking the real-time monitoring data in one detection cycle as a data set; E, extracting monitoring data characteristic parameters in each data set; F, combining drilling gas inrush initial velocity qmax and maximum quantity of drilling yields smax with the monitoring data characteristic parameters so as to form gas outburst danger samples; G, performing gas outburst prediction by utilizing a BP neural network so as to obtain predicted values of the drilling gas inrush initial velocity qmax and maximum quantity of drilling yields smax; and H, comparing the predicted values with critical values of tunneling face outburst danger parameters, and judging whether early warning is needed. According to the method disclosed by the invention, the pre-control ability of gas outburst accidents in coal mines is improved.
Owner:HENAN POLYTECHNIC UNIV

Myoelectric gesture recognition method based on RNN-CNN architecture

The invention relates to a myoelectric gesture recognition method based on an RNN-CNN architecture. The method comprises the following steps of performing feature extraction on each channel signal byusing an RNN architecture according to a time sequence characteristic of a myoelectric signal, and further extracting a fused feature map by using a CNN architecture, and mainly comprises the following steps of preprocessing the data, using an RNN module to perform preliminary feature extraction on the preprocessed data, using a fusion module to perform fusion processing on an output result of theRNN; using a CNN module to perform feature extraction and analysis on an output result of the fusion module; and using a classification module to judge the input gesture signal by the model output, namely judging which gesture type the electromyographic signal belongs to according to the currently input electromyographic signal. According to the method, the time sequence relevance and characteristics of the data can be effectively extracted, and meanwhile, the gesture recognition rate is improved; an extreme value point selection and splicing method is introduced at a data preprocessing stage, so that the model training time is reduced, and the mutual interference between the channels is avoided; finally, at the fusion stage, the relevance of the multiple channels is utilized, so that theidentification of the electromyographic signals is facilitated.
Owner:NANJING UNIV OF POSTS & TELECOMM

Railway wagon triangular hole foreign matter detection method

A railway wagon triangular hole foreign matter detection method belongs to the technical field of freight train detection. The objective of the invention is to solve the problems of low efficiency andlow accuracy of existing water leakage hole foreign matter detection. The method comprises the following steps: firstly, collecting images and extracting images containing a triangular hole area, building a sample data set, and wherein the sample data set comprises two sample data sets for triangular hole part positioning and a sample data set for triangular hole foreign matter calibration; respectively training a triangular hole positioning segmentation network and a triangular hole foreign matter segmentation network; in the detection process, collecting a real vehicle passing image, extracting a triangular hole part image, inputting the to-be-detected triangular hole part image into a triangular hole foreign matter segmentation network, and detecting whether foreign matter exists or not; and if the foreign matter exists, inputting the to-be-detected triangular hole part image into the triangular hole positioning segmentation network to carry out triangular hole region positioning,and then judging whether the foreign matter exists in the triangular hole. The method is mainly used for wagon triangular hole foreign matter detection.
Owner:HARBIN KEJIA GENERAL MECHANICAL & ELECTRICAL CO LTD

Intelligent SOC (State of Charge) prediction device for electric vehicle power battery

The invention discloses an intelligent SOC (State of Charge) prediction device for an electric vehicle power battery, which is characterized by comprising a battery parameter acquisition platform and a battery SOC prediction system, wherein the battery parameter acquisition platform is used to acquire real-time parameters of voltage, current, temperature, and ambient temperature of the vehicle power battery pack; and the battery SOC prediction system is used to predict the battery SOC value through the acquired real-time parameters. The battery SOC is a nonlinear, delayed, multivariable-coupling, and complex real-time system with extremely high real-time performance requirements. The problem that the conventional prediction device can not achieve ideal battery SOC prediction precision effects can be effectively solved.
Owner:合肥龙智机电科技有限公司

Subway station air conditioning system energy-saving control method based on deep reinforcement learning

The invention provides a subway station air conditioning system energy-saving control method based on deep reinforcement learning. According to the invention, the method includes collecting data parameters of a subway station air conditioning system; performing moving average filtering processing, normalization and anti-normalization processing on the acquired data, and converting the data into numerical values in a range of 0-1 by using a linear function conversion method; constructing a neural network model of the subway station air conditioning system by using a neural network and the data obtained in the step; determining a state variable, an action variable, a reward signal and a structure of the DDPG agent; and using the multi-step prediction DDPG algorithm for solving the final control strategy. According to the invention, the control method provided by the invention has good temperature tracking performance; compared with a traditional DDPG algorithm, the number of times of agent training is reduced by 86, the system can stably operate under the condition that the system load changes, the station temperature requirement is met, and meanwhile, compared with an operation system in a current practical project, the energy is saved by 17.908%.
Owner:BEIJING UNIV OF CIVIL ENG & ARCHITECTURE

SAR target recognizing algorithm based on guide reconstitution and norm constraint DBN

The invention provides an SAR target recognizing algorithm based on guide reconstitution and norm constraint DBN. In order to solve the problems that an SAR target recognizing algorithm based on DBN is high in network structure complexity, large in training frequency, low in recognizing rate and the like, the guide reconstitution algorithm is put forward to conduct reconstitution preprocessing ontraining samples and testing samples, a one-dimensional image vector is formed through cutting and then extending, low-dimensional features are extracted through a weighted norm constraint deep beliefnetwork (DBN), and targets are classified through Softmax. It is shown through experimental results that the method can reduce the dimension of the image features and the frequency of network training and the network recognizing performance is further improved.
Owner:NORTHWESTERN POLYTECHNICAL UNIV +1

Underwater robot parameter adaptive backstepping control method based on double-BP neural network Q learning technology

ActiveCN111176122AMeet the requirements of real-time online adjustmentShorten training timePosition/course control in three dimensionsAdaptive controlBacksteppingSimulation
The invention discloses an underwater robot parameter adaptive backstepping control method based on a double-BP neural network Q learning technology, belongs to the technical field of underwater robotcontroller parameter adjustment and solves problems that learning efficiency is low when controller parameter adjustment is carried out through a traditional Q learning method and parameters are noteasy to adjust online in real time when controller parameter adjustment is carried out through a traditional backstepping method. According to the method, autonomous on-line adjustment of the parameters of a backstepping method controller is realized by combining a double BP neural network-based Q learning algorithm and a backstepping method, so the requirement that the control parameters can be adjusted on line in real time is met, moreover, due to introduction of the double BP neural networks and an experience playback pool, the Q learning parameter adaptive backstepping control method basedon the double BP neural networks can greatly reduce the number of training times due to the powerful fitting capability, so learning efficiency is improved, and the better control effect is achievedunder the condition that the number of training times is small. The method can be applied to parameter adjustment of the underwater robot controller.
Owner:HARBIN ENG UNIV

Model training method and related equipment

PendingCN113191241ASave storage spaceAchieving resistance to catastrophic forgettingSemantic analysisScene recognitionBatch trainingEngineering
The embodiment of the invention provides a model training method which is applied to the field of artificial intelligence, and the method comprises the steps: obtaining a first neural network model and M batches of batch training samples, M being a positive integer greater than 1; then determining a target incremental training method according to sample distribution characteristics between batches of batch training samples in the M batches of batch training samples, wherein the sample distribution characteristics are related to the degree of catastrophic forgetting generated by the model when incremental training is carried out based on the batches of batch training samples; and using the target incremental training method for realizing catastrophic forgetting resistance when incremental training is performed on the model; and according to the M batches of batch training samples, performing self-supervised training on the first neural network model through a target incremental training method to obtain a second neural network model. According to the method, on the premise that the training time is shortened and the data storage space is saved, the balance between efficiency and performance is realized.
Owner:HUAWEI TECH CO LTD
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Eureka Blog
Learn More
PatSnap group products