223 results about "Offline learning" patented technology

Disclosed is a method and system for efficiently and accurately tracking three-dimensional (3D) human motion from a two-dimensional (2D) video sequence, even when self-occlusion, motion blur and large limb movements occur. In an offline learning stage, 3D motion capture data is acquired and a prediction model is generated based on the learned motions. A mixture of factor analyzers acts as local dimensionality reducers. Clusters of factor analyzers formed within a globally coordinated low-dimensional space makes it possible to perform multiple hypothesis tracking based on the distribution modes. In the online tracking stage, 3D tracking is performed without requiring any special equipment, clothing, or markers. Instead, motion is tracked in the dimensionality reduced state based on a monocular video sequence.

The invention relates to a traffic flow parameter acquisition method, and discloses a traffic flow parameter acquisition method based on a video and a sensor. The method comprises the steps of video inputting, office learning, target detection, target tracking, online learning, video flow counting, sensor flow counting and multi-data fusing counting. According to the traffic flow parameter acquisition method based on the video and the sensor, a target detector is obtained through offline learning, and traffic video image is detected to rapidly determine the types and quantity of objects in a current video scene; a target track is determined by using target tracking, an online learning mechanism is adopted to intermittently update a target model and the feature vector thereof, and the target detector and a tracker are corrected to ensure that the target can be tracked continuously, and the traffic flow parameter information of a current road segment can be acquired accurately through video detection; moreover, traffic data acquired by the sensor in a road is analyzed through multi-data fusing counting, thereby further increasing the accuracy of the acquired traffic flow parameter data.

The invention discloses a neural network online learning system based on a memristor. A pulse coding mode of a K-bit input vector is improved; the coding pulse corresponding to each bit is expanded into 2m pulses; In this way, the total number of needed coded pulses is K * 2m, each bit of weighted summation calculation is actually carried out for 2m times, finally, summation averaging operation iscarried out at the output end, the influence of accidental factors and noise on the calculation result in the calculation process is reduced through the mode, and therefore the calculation precisionis improved. The memristor array is simultaneously used for forward weighted summation calculation and weight small storage in the neural network; Different from offline learning, The weight in the memristor array needs to be updated once every time a signal is input through online learning, the variable quantity of the weight is mapped into the number of pulses, then the pulses are applied for one-time weight write-in operation, the neural network training speed can be increased, the hardware cost can be reduced, and the power consumption of neural network training is reduced.

The invention discloses an online learning method for optimizing a signalized intersection queuing length. The online learning method comprises the following steps of: 1, selecting states, behaviors and rewards; 2, reinforcing a learning matrix updating formula; 3, establishing a simulation optimizing platform; and 4, carrying out online operation. The online learning method is a signal timing dial optimizing technology which is capable of calculating a globally optimal solution and has the memorability. Compared with the risk neutral reinforcing learning technology, the online learning method has the advantages of no need of advanced offline learning, and better instantaneity and adaptability.

The invention discloses a voice converting method based on deep learning, and belongs to the technical field of voice signal processing. According to the invention, the method includes the following steps: configuring a voice encoding and decoding device AHOcoder as a feature extraction terminal and a voice synthesis terminal, training voice features by using the deep leaning method to separatelyobtain deep features of a source speaker and a target speaker, also obtaining the capability of decoding the deep features to original features, mapping the source speaker and the target speaker by using a BP neural network, thus realizing voice conversion. According to the invention, the method stitches the original features of voice, the combined feature parameters obtained from stitching are deemed to include the dynamic features of the voice features of the speaker, the training of the deep neural network is accelerated by pre-training the deep autoencoders, and by converting the deep features, the method herein obtains quality converted voice even when less voice materials are trained. The method also supports offline learning, and saves computing resources and memory of terminal devices.

A fast variational on-line learning technique for training a transformed hidden Markov model. A simplified general model and an associated estimation algorithm is provided for modeling visual data such as a video sequence. Specifically, once the model has been initialized, an expectation-maximization (“EM”) algorithm is used to learn the one or more object class models, so that the video sequence has high marginal probability under the model. In the expectation step (the “E-Step”), the model parameters are assumed to be correct, and for an input image, probabilistic inference is used to fill in the values of the unobserved or hidden variables, e.g., the object class and appearance. In one embodiment of the invention, a Viterbi algorithm and a latent image is employed for this purpose. In the maximization step (the “M-Step”), the model parameters are adjusted using the values of the unobserved variables calculated in the previous E-step. Instead of using batch processing typically used in EM processing, the system and method according to the invention employs an on-line algorithm that passes through the data only once and which introduces new classes as the new data is observed is proposed. By parameter estimation and inference in the model, visual data is segmented into components which facilitates sophisticated applications in video or image editing, such as, for example, object removal or insertion, tracking and visual surveillance, video browsing, photo organization, video compositing, and meta data creation.

The invention discloses a personalized recommendation system based on deep learning under a social network. The system mainly comprises an offline learning module and an online recommendation module. The offline learning module firstly generates a training sample seat to construct a deep convolutional neural network learning module with an attention mechanism and carries out iterative optimization on parameters in the learning module; and the online recommendation module carries out real-time item recommendation on a newly-registered user based on the learning model obtained through training. Compared with the prior art, the system has the advantages of high accuracy, fast speed and simplicity and easiness in implementation and can be effectively applied to the fields, such as electronic commerce, public opinion monitoring, intelligent transportation and medical treatment and health.

The invention discloses a multipath transmission protocol congestion control method based on reinforcement learning. A Markov decision process is established to express the congestion control in a form. The current network states of sub-streams are expressed by the sizes of congestion windows of the sub-streams and the RTTs of the sub-streams of a sending end, the action of the sending end to adjust the congestion windows and the sending intervals is defined, and a target equation is established to obtain maximum average throughput and a minimum average throughput. A network model is established to simulate and generate a lot of network environments. In different network environments, all actions are executed on the current network environment by continuous trials and errors, and learning and optimization are realized from the feedback provided by the environment. By means of a lot of offline learning, the sending end can make the corresponding action of adjusting the sizes of the congestion windows and the sending intervals in a certain state area, so that the value of the target equation is the maximum.

The invention relates to a rapid behavior identification method and system. According to the method, during the training process, a feature extraction algorithm with low computation complexity is used and dimensionality reduction and normalization are carried out on an obtained feature matrix by using various dimensionality reduction ways, thereby reducing the processing time. During the identification process, a captured video is processed by using an algorithm identical with that applied to the training process; and then the processed video is placed into a model obtained by training to carry out detection and identification. In addition, the system is composed of a video acquisition module, an off-line learning module, a behavior identification module, a multi-thread processing module and a video display module. With a sliding window mechanism, a video captured in real time is processed into a plurality of video clips; and each video clip is processed during behavior identification to guarantee real-time updating of an identification result. Meanwhile, multi-thread processing is carried out on the feature extraction and feature coding process with high computation complexity to guarantee the real-time property of processing identification, so that behaviors occurring in the video can be identified and display visually, rapidly and accurately. The operation is convenient and convenient extension can be realized.

The invention discloses a GIS fault detection system and a GIS fault detection method based on multi-source information fusion and a deep learning network. The GIS fault detection system comprises a multi-source information acquisition and conditioning module, a deep learning module and an information fusion and fault reasoning module, wherein the multi-source information acquisition and conditioning module performs fault state monitoring on a GIS system by adopting a partial discharge time analysis method, a partial discharge phase analysis method and an ultrahigh frequency method separately, extracts corresponding feature vectors separately from obtained current, voltage and electromagnetic information and outputs the feature vectors to the deep learning module; the deep learning module performs online pattern recognition on the three kinds of feature vectors based on the deep learning network obtained through offline learning optimization to acquire corresponding recognition results, and outputs the recognition conclusions to the information fusion and fault reasoning module; and the information fusion and fault reasoning module carries out fusion processing on the three recognition conclusions to obtain a fault feature matrix, and then obtains a fault conclusion by means of a CLIPS reasoning machine. By adopting the GIS fault detection system and the GIS fault detection method, the fault information of the GIS system can be diagnosed quickly, efficiently and precisely.

The invention relates to a deep license plate detection method based on a thermodynamic diagram and key point regression. The method includes an off-line learning phase and a detection phase. The off-line learning phase includes the following four steps of: (1) designing network depth; (2) preparing a training sample set; (3) performing sample labeling; and (4) performing training. According to the step (1), a deep learning network structure is designed, inputted images are adjusted and unified, first-layer convolution and activation function operation are performed on the images, second-layerconvolution and activation function and pooling operation are performed, and third-layer convolution and activation function and pooling operation are performed, and a plurality of task branches areformed, one branch learns the coordinates of plate number plates through one convolutional layer, another branch learns the thermodynamic diagram of the license number plates through another convolutional layer. According to step (2), a batch of vehicle front or rear images is obtained and adopted as a sample set for offline learning, and the larger the total number of the classes of samples is, the better a training effect is, and the sizes of the samples are normalized. According to the method, the offline trained deep network is adopted to characterize a target, and therefore, license platedetection can be performed on the target quickly and steadily.

The invention relates to a facial-recognition-based network teaching method and system. Facial feature identification can be carried out on a student in real time during the whole courseware learning period, so that the network teaching can realize the same effect as the face-to-face teaching. Meanwhile, a certain number of times of failure is allowed, so that a stable and smooth learning environment can be provided for the student in a real-time monitoring mode and the learning of the student can be prevented from being influenced due to the detection failure or other factors. In addition, the system also provides expression management, so that the teacher can know the reflection of the student conveniently during the courseware learning and thus the network teaching environment is close to the face-to-face teaching. On the basis of the off-line learning function provided by the system, the student can carry out learning on any occasion and an objective of real-time monitoring can be achieved. Besides, with the log management module provided by the system, no more network flow is increased because of real-time monitoring when the student uses a mobile terminal for learning, thereby lowering the learning cost.

The invention relates to a man-machine interaction method of a customer service system. The method enables the effect to be continuously optimized by virtue of user feedbacks, and comprises the following steps that a user raises a question and waits the customer service system to answer the question; buttons representing satisfactory and dissatisfactory feedbacks of the user are added on a question answering interface of the customer service system, and the user is invited to give feedback to whether an answer is satisfactory or not; satisfaction and dissatisfaction clicking behaviors of the user are fed back to logs of the customer service system and associated with the question of the user and the answer of a customer service staff at that time; a positive sample and a negative sample of the answer of the customer service staff are generated based on the logs; offline learning is carried out by utilizing the positive sample and the negative sample; and frequently asked question (FAQ) knowledge after the offline learning of the customer service system is updated. According to the method, the user experience is improved and the user does not feel unfriendly; sufficient positive and negative samples are provided for question and answer effects of customer service staffs; and continuously innovated fresh question-asking ways of users can be quickly adapted.

The invention relates to a Q-learning based vehicular ad hoc network routing method and belongs to the technical field of Internet-of-things communication. The method includes that (1) a GPS (global positioning system) is loaded to each vehicle in a network, and the vehicles acquire neighbor node information by passing Hello messages therebetween; (2) a city region is divided into equal grids, the position of each grid represents a different state, and transferring from one grid to the adjacent grid represents an action; (3) a Q-value table is learnt; (4) parameters are set; (5) routing strategies QGrid_G and QGrid_M are selected. Vehicles newly added into the network acquire the Q-value table obtained by offline learning from the neighbor vehicles, and the vehicles can be informed of the optimal next-hop grid of message passing by querying the Q-value table of the message destination grid. The grid sequence that the vehicles mostly frequently travel is taken into consideration from a macroscopic point of view, the vehicle which is mostly likely to arrive at the optimal next-hop grid is selected by considering from a microcosmic point of view, and passing success rate of messages in the urban traffic network is increased effectively by the macroscopic and microcosmic combination mode.

The invention relates to an improved storage battery SOC (State Of Charge) estimation method based on the consistency of a unit cell. In the method, the training data set of a storage battery, SOC predicted data corrected on the basis of consistency, and a self-adoption neural network are use for building a fuzzy rule base with an optimized structure and parameters; and then the fuzzy rule base after off-line learning is implanted into the fuzzy inference machine of an embedded controller of BMS (Battery Management System), so as to carry out on-line correction to storage battery SOC estimation. Compared with the prior art, the invention has the advantage that SOC differences in all unit cells inside the battery can be factually reflected on entire SOC estimation.

The invention discloses a self-learning inverse model control method of an electronic throttle. The control method comprises two parts: 1) taking a radial basic function neural network as a recognizer to create the non-linear model of the electronic throttle; 2) taking another RBF neural network to serve as an inverse model controller to obtain an appropriate control amount. The two RBF neural networks adopt a self-learning strategy which comprises an off-line learning step and an on-line learning step. In off-line learning, the two RBF neural networks adjust and optimize network weighting parameters according to on-site sample data; after the off-line learning is completed, an on-line learning algorithm is adopted to adjust and optimize the network weighting parameters of the two RBF neural networks, and then the control performance can meet the requirements. The invention can overcome the difficulties of nonlinearity, time varying and the like in electronic throttle control and improve the control effect and performance.

Disclosed is a super-resolution face recognition method based on multi-manifold discrimination and analysis. During the training phase, a mapping matrix from a low-high-resolution face image multi-manifold space to a high-resolution face image multi-manifold space is acquired by multi-manifold discrimination and analysis. An intra-class similar graphs and aninter-class similar graph are constructed in an original high-resolution face image multi-manifold space, a discrimination bound term is constructed by utilizing the two neighbor graphs, and a most optimization method is to acquire the mapping matrix by reconstructing a cost function composed of a bound term and the discrimination bound term. During the recognition phase, a low-resolution face image to be recognized is mapped o the high-resolution face image multi-manifold space by the mapping matrix acquired by offline learning, and a high-resolution face image is acquired. Classification and recognition are achieved by a nearest-neighbor classifier according to the Euclidean distance principle in the high-resolution face image multi-manifold space. Compared with a traditional super-resolution method, the super-resolution face recognition method has greatly improved face recognition rate and operation rate.

The invention discloses a frequent subgraph mining based abnormal intrusion detection method, which belongs to an intrusion detection method for network security. In the invention, a frequent subgraph mining theory is introduced; and by using the peculiar generative capacity of a system call sequence transformed into digraph structure, a considerable quantity of effective generative characteristic patterns can be obtained by a small scale of training data, which greatly reduces the dependence of an offline learning process on training data quantity, thereby dealing with various conditions in real-time detection more easily; and meanwhile, an expanded characteristic pattern set has an ability to identify unknown program behaviors, which can effectively reduces the false alarm rate of detecting results.

The invention discloses an adaptive distributed parallel training method for a neural network based on reinforcement learning, and provides an optimal solution for segmentation and scheduling of a large-scale complex neural network. Firstly, the influence of a neural network model structure and calculation attributes on execution performance is analyzed, on this basis, performance factors including calculation cost, communication cost, memory utilization rate and the like are extracted, a multi-dimensional performance evaluation model capable of comprehensively reflecting distributed training performance is constructed, and comprehensive performance of a parallel strategy is improved; secondly, self-adaptive grouping of operators is realized according to attribute characteristics of the operators by utilizing a feed-forward network, the degree of parallelism is determined, and end-to-end strategy search is realized while the search space is reduced; and finally, based on importance sampling, a near-end strategy gradient iteration optimization reinforcement learning model is adopted, an optimal segmentation and scheduling strategy is searched, the strategy network offline learning capability is expanded, and algorithm stability, convergence rate and strategy search performance are improved.

The invention discloses a production method for an injection molding machine workpiece production line. A computer control system integrating a system decision, monitoring and learning functions is established for the characteristics of injection molding machine workpiece manufacturing systems of double production lines so as to realize automatic intelligent optimized production; a system control task is finished by a centralized controller and a stimulation optimizer commonly, monitoring of a production system and running of a control instruction are finished by PLC (programmable logic controller) and a production line state monitor; and a value iteration learning thought is realized by a reinforcement learning device through a support vector machine with a hybrid kernel function, so that the stimulation optimizing capability of a system stimulator can be improved by flexibly utilizing offline learning, the total running time of the control system is shortened, the running quality of the control system is improved, the production cycle of the manufacturing system is shortened, and the equipment utilization rate is increased.

The invention discloses a dangerous goods recognition system and method thereof. According to the system, firstly, goods needing to be recognized is obtained through a camera on an intelligent terminal, and transmits image information to a goods recognition server, a server management module sends the image information to an inference module for image recognition, the inference module performs goods recognition on learned data through an offline training module, when it is read that the goods are dangerous, the inference module transmits the information to an alarm for risk prompts, and when it is read that the goods are not dangerous, the inference module transmits the information to the intelligent terminal for goods information display. The recognition scheme is offline learning, an offline training model is built, a user does not depend on the network, according to the requirement of the user, secondary development can be performed, the period is short, debugging and optimizing are performed according to the application scenarios, and the effect is better.

The invention discloses a real-time driver's braking behavior and state distinguishing method. The method comprises the following steps that 1, original test data of an actual vehicle of a driver is obtained through an actual vehicle testing platform; 2, the test data is preprocessed, wherein the preprocessing includes noise reduction preprocessing and abnormal data processing; 3, a principal component analysis method is utilized to conduct feature extraction and data dimension reduction on preprocessed data; 4, a braking pedal force, a braking pedal stroke and braking pedal stroke change rate data are clustered; 5, offline learning of different braking behaviors and states is completed based on a statistic pattern recognition model, and a behavior training library and a state training library of corresponding quantization representation values are obtained; 6, braking behaviors and states of different drivers under different driving working conditions are distinguished in real time. The method can be used in the technical fields of a brake-by-wire system, a braking energy recovery strategy, vehicle running state monitoring and the like, and can also create conditions for automatic driving decision-making control based on driver's behaviors.

The invention relates to a support vector machine inverse-based bearingless asynchronous motor control system. According to the system, two Park<-1> conversions, two Clark <-1> conversions, two current hysteresis PWM (Pulse Width Modulation) inverters and a bearingless asynchronous motor form a composite controlled object; and the support vector machine inverse of the composite controlled object is created by the support vector machines and integrators through least squares offline learning; the support vector machine inverse and the composite controlled object are connected in series to form a pseudo-linear system which is formed by two linear radial displacement subsystems, a speed subsystem and a magnetic linkage subsystem and is subjected to closed-loop compound control in a multi-internal-model switching manner. The control system provided by the invention can identify real-time equivalent models of the controlled object at improved dynamic response speed and steady state control accuracy, has high robustness against external disturbance and motor parameter variation, and can realize the high-performance dynamic decoupling control of all controlled variables of the bearingless asynchronous motor.

The invention provides a railway locomotive operation control system hybrid task scheduling method based on reinforcement learning, and the method is an offline learning process. According to the method, first hybrid task set data during actual operation or in a simulation experiment of a locomotive operation control system, a hybrid task set is formed and regularization marking is performed on each task in the hybrid task set. Then the task set after regularization marking serves as input of a reinforcement learning system, and a reinforcement learning environment is formed. The reinforcement learning system applies a reinforcement learning algorithm, inspects scheduling objectives of the locomotive operation control system to perform an iterative learning process, a <state-rule> corresponding relation table corresponding to the hybrid task set is generated; and the <state-rule> corresponding relation table is stored in a database. The rule whose frequency of occurrence is the highest is selected from the database as the optimal rule of a current state, and a final <state-rule> corresponding relation table is formed. During operation of the locomotive control system, generation of a real-time scheduling sequence of hybrid tasks is guided according to the <state-rule> corresponding relation table, thereby realizing task scheduling.

The invention relates to a terrain classification and recognition method based on learned vibration information, which aims to solve a problem of low accuracy of the existing classification and recognition method. The terrain classification and recognition method comprises the steps of first, collecting original data of vibration information of a sensor in the x-axis, y-axis and z-axis directionsin a coordinate system; second, performing segmentation processing on the original data to form n vectors with the time length being 1s; third, performing terrain type marking on the n vectors; fourth, converting the segmented n vectors into a frequency domain; fifth, performing off-line learning training on the n vectors converted into the frequency domain by using a multilayer feedforward neuralnetwork to obtain a trained multilayer feedforward neural network; and sixth, obtaining vibration data online in real time, executing the step two to the step four, and performing online classification and recognition by using the trained multilayer feedforward neural network in the step five to obtain the terrain type. The terrain classification and recognition method is applied to the technicalfield of recognition and classification.

The invention provides an intra-frame coding unit rapid selection method based on a logical regression classifier. According to the method, downward partition of one coding unit is modeled into the classification problems of downward partition and non-downward partition, and the two classification problems are solved through the logical regression classifier. Three video sequences are learned offline to obtain an optimal logical regression coefficient of the logical regression classifier, and an optimal decision-making feature is selected through calculation of the effectiveness of a plurality of candidate features. An intra-frame coding unit is selected rapidly through the logical regression classifier obtained by offline learning, so that other unnecessary prediction coding mode calculations are skipped; the intra-frame coding complexity of an HEVC (High Efficiency Video Coding) coder can be lowered effectively; and real-time application of a real-time HEVC coder is facilitated.

The invention discloses a robot parameter identification and contact force monitoring method based on prior kinetic knowledge. The method comprises the following steps: acquiring joint position and speed information and joint acceleration information of a robot through an encoder, and acquiring actual joint torque obtained after current information of an analog quantity monitoring end of a motor driver is processed in real time by a data acquisition card; controlling the robot to move along different tracks as many as possible in a space, acquiring angles, angular speeds, angular accelerationsand actual torque of the robot at different moments as samples, performing offline learning by a gradient descent method, identifying kinetic parameters of the robot, and obtaining an ideal kinetic model of the robot; in the real-time monitoring process, substituting the acquired and processed current information into the ideal kinetic model to obtain theoretical torque of the robot at the moment; and comparing the actual joint torque with the theoretical torque, and if the difference exceeds a threshold range, indicating that collision occurs. According to the method, contact force data canbe accurately acquired at low cost.

The invention discloses an off-line learning system which comprises the following subsystems: a courseware integration subsystem used for realizing analysis processing and document collection processing of a streaming media document in a courseware; a courseware encryption and packaging subsystem used for realizing encryption and packaging of the courseware; an off-line learning package download subsystem used for forming an off-line learning package automatically and realizing automatic detection and download of the off-line learning package; a user identity authentication subsystem used for authenticating user identity; an off-line learning packet loading subsystem used for loading an off-line courseware package and corresponding learning progress information at an off-line learning system client; an off-line learning subsystem used for realizing courseware learning of a user; a learning progress synchronization subsystem used for realizing unification of user learning progress and data synchronization between the off-line learning system and an online learning platform; an off-line learning client software installation updating subsystem; and a USB flash disk automatic detection subsystem. According to the system, the depth and the breadth of network remote training system application can be expanded.

The invention discloses a neighbor searching method and a neighbor searching system. The neighbor searching method includes offline learning and online searching. A learning function of data point number in small regions on two sides of a Hash hyperplane is minimized to enable the Hash hyperplane to penetrate through a data sparse region, and accordingly high accuracy in neighbor searching can be guaranteed. An approximate balance barrel regular term is added to a learning function through approximate balance barrel conditions to enable more balanced data point partition by the Hash hyperplane, and accordingly high searching speed in neighbor searching is guaranteed. Regardless of a small quantity of data or massive data, high-accuracy and high-speed neighbor searching can be realized by the neighbor searching method and the neighbor searching system.