34 results about "Parallel learning" patented technology

The invention discloses a collision avoidance planning method for mobile robots based on deep reinforcement learning in a dynamic environment, and belongs to the technical field of mobile robot navigation. The method of the invention includes the following steps of: collecting raw data through a laser rangefinder, processing the raw data as input of a neural network, and building an LSTM neural network; through an A3C algorithm, outputting corresponding parameters by the neural network, and processing the corresponding parameters to obtain the action of each step of the robot. The scheme of the invention does not need to model the environment, is more suitable for an unknown obstacle environment, adopts an actor-critic framework and a temporal difference algorithm, is more suitable for a continuous motion space while realizing low variance, and realizes the effect of learning while training. The scheme of the invention designs the continuous motion space with a heading angle limitationand uses 4 threads for parallel learning and training, so that compared with general deep reinforcement learning methods, the learning and training time is greatly improved, the sample correlation isreduced, the high utilization of exploration spaces and the diversity of exploration strategies are guaranteed, and thus the algorithm convergence, stability and the success rate of obstacle avoidance can be improved.

An apparatus / method for extracting multiple features from time series data collected from a plurality of sensors and for performing transfer learning on them. There is provided an apparatus including: a multi-feature extraction unit for extracting multiple features from a data stream for each sensor inputted from the plurality of sensors; a transfer-learning model generation unit for extracting useful multi-feature information from a learning model which has finished pre-learning, for the multiple features for forwarding the extracted multi-feature information to a multi-feature learning unit to generate a learning model that performs transfer learning on the multiple features; and the multi-feature learning unit for receiving learning variables from the learning model for each of the multiple features and for performing parallel learning for the multiple features, to calculate and output a loss. In addition, there is provided an apparatus for detecting leaks in plant pipelines.

The invention belongs to the technical field of network security, and particularly relates to a network encryption traffic classification method and system based on multi-feature learning, and the method comprises the steps: carrying out the preprocessing of an original traffic data set, and obtaining a traffic data package vector used for the input of a deep learning model; respectively inputting the traffic data packet vectors into a trained multi-channel CNN model and a trained LSTM model for parallel learning, extracting the data packet space features through the multi-channel CNN model, and extracting traffic time sequence features through the LSTM model; carrying out vector splicing on the data packet space feature and the traffic time sequence feature to obtain an omnibearing traffic feature vector; and inputting the omni-directional traffic feature vector into a neural network full-connection layer, and obtaining an encrypted traffic classification type through a traffic type probability. According to the method, the traffic features can be comprehensively and automatically extracted and utilized from the angles of the spatial features and the time features, the classification capability of the encrypted traffic is improved, and the method has good application value.

The invention discloses a classification forecasting method based on a Bayesian network. The method comprises the following steps: obtaining data to be trained and dividing the data into a plurality of block data; establishing a corresponding Bayesian subnetwork for each block data; and using the Bayesian subnetworks to carry out classification forecasting simultaneously. Through the technical scheme of obtaining a plurality of subnet structures through parallel learning and carrying out classification forecasting simultaneously, time cost of bayesian network learning under the condition of large data quantity is saved, and a better behavior is achieved under a particular data set.

The invention discloses a chemical storage tank abnormity detection algorithm research based on FCM-LSTM, relates to the chemical equipment and neural network fields. According to the method, a multi-layer network architecture model is used as a fault diagnosis method, the advantages of supervised learning and unsupervised learning are combined, and a fault diagnosis mechanism based on data driving is adopted. The method comprises the following steps that firstly, mass data is clustered by using an unsupervised tool class FCM algorithm; according to a specified similarity standard, datasets are divided, the normal data and the fault data belong to different class clusters; a PSO algorithm is used for avoiding random selection of initial values and accelerating the clustering process, a small amount of mark data is obtained to improve the detection performance, then an LSTM is used for carrying out training network on each cluster and offline historical data, finally, multi-subnet parallel learning is carried out, then results are subjected to fitting combination and integrated analysis, and the generalization capacity of the network is improved. The data size processed by the method is larger, more information can be processed, and the application range is wider.

The invention discloses a multi-agent deep reinforcement learning algorithm based on partition experience and multi-thread interaction. Firstly, the algorithm can be used for distinguishing positive experience, negative experience and neutral experience by dividing a reward space by using an experience replay form of a partitioned cache region, and extracting experience data by using a layered random sampling mode during training; and secondly, the algorithm promotes the trial and error process of the intelligent agent and the environment by applying a multi-thread interaction mode, and parameters of a network model are trained through parallel learning of multiple clone bodies of the intelligent agent and integration of learning experience of the clone bodies. The method has the advantages that the multi-agent deep reinforcement learning algorithm based on cache region replay and multi-thread interaction is introduced into the multi-agent deep reinforcement learning algorithm by combining the advantages of a partitioned experience cache region and a multi-thread interaction mode; and the method is superior to an existing model in convergence speed and training efficiency, has higher availability in a multi-agent environment, and can be used for solving the problem of cooperative target tracking of multiple agents.

The invention discloses an event-triggered unmanned surface vehicle cluster distributed cooperative controller, structure and method. Through a communication mode based on event triggering, an unmanned surface vehicle sends own information to a communication network only when a preset event is satisfied, and non-periodic information sending is realized, so the information sending frequency is reduced, and the communication burden of an unmanned surface vehicle communication network is reduced; meanwhile, cooperative action execution of the unmanned surface vehicle is also triggered by an event, and an actuator is aperiodically updated according to the preset performance on demand, so that the updating frequency of the actuator is reduced, continuous action of the actuator is avoided, and the execution burden of the unmanned surface vehicle is relieved; and a parallel learning identification module not only can approach uncertain nonlinear dynamic states in the unmanned surface vehicle, but also can identify unknown control gains, so that cooperative control can still be kept without depending on accurate information of the unmanned surface vehicle, and cooperative control is more flexible and efficient.

The invention provides a network-on-chip routing device and a control method thereof, and the control method comprises the steps: configuring a learning module for each router of the network-on-chip routing device, wherein each router and one learning module form a routing node; by the learning module of each routing node, obtaining the congestion state information of the learning module of the adjacent routing node and then performing parallel learning to obtain the optimal data transmission path of each destination routing node; and performing data transmission according to the optimal datatransmission path. According to the network-on-chip routing device and the control method thereof, the network convergence speed can be increased, the parallel number can be increased, the path planning speed can be greatly increased, meanwhile, a proper path can be found only by obtaining the position information of the destination routing node, resource occupation is reduced, and the path planning time can be shortened.

The invention discloses a construction method and application of a face quality evaluation model. The construction method comprises the following steps: S1, constructing the face quality evaluation model; and S2, inputting a pre-collected training set into the face quality evaluation model, and training the face quality evaluation model by minimizing a weighted sum of differences between a prediction attribute and a real attribute of each face image attribute as a target. The constructed face quality evaluation model comprises a cascaded feature extraction network and a multi-task layer, the multi-task layer comprises a plurality of parallel task branches, and each face image attribute corresponds to one task branch, so that parallel learning is carried out on various related deep learning tasks; relevance among tasks can be considered in the learning process, parameters are shared, and a better generalization effect can be achieved; different face image attributes of the image are perceived through each task branch, detail information of each face image attribute evaluation index can be obtained, and the face quality can be comprehensively and accurately evaluated.

The invention belongs to the technical field of rapid analysis of computer data, and discloses a sub-sampling SVR integration-based short-term power load prediction method; a sub-sampling support vector regression integration method is adopted, and a sub-sampling strategy is executed for realizing parallel learning of a support vector machine, so that it is guaranteed that each single SVR has enough diversity, and the loss amount of information is reduced; and a new group optimization learning algorithm based on single SVR integration is selected, so that it is guaranteed that each SVR integration has enough strength to predict short-term load data. The small-size sub-sampling-based integration method has the classical U-statistical magnitude statistics reasoning property, and by virtue ofintroduction of a small-scale sub-sampling strategy, so that the complexity of an estimator is effectively reduced; and meanwhile, the calculation precision and efficiency are improved, and a relatively high confidence interval is returned. An SSVRE mode is easily transplanted into a parallel computing framework.

The invention relates to a wireless network signal transmission strength calculation method and a computer storage medium, and the method comprises the steps: firstly constructing an input feature variable based on engineering practice analysis during the modeling of a wireless network signal transmission model, and screening an important feature subset from the input feature variable through employing a feature selection theory as an actual input feature vector; generating a virtual input feature vector according to the real input feature vector by adopting an adversarial network, and obtaining an output label corresponding to the virtual input feature vector by adopting a decomposition fuzzy extreme learning machine trained by using real data; then, adopting a multi-layer intuitionisticcondition fuzzy residual neural network as a virtual / real twinning network to extract features of virtual / real input feature vectors, and virtual-real interaction being achieved through a feature sharing laye; and finally, in order to further improve the accuracy and robustness of the model, integrating a plurality of individual parallel learning models by adopting a parallel decomposition fuzzy width neural network on the basis of a staging thought so as to obtain a stronger model generalization ability.

A network device, system, and method are provided. The network device includes a processor. The processor is configured to store a local estimate and a dual variable maintaining an accumulated subgradient for the network device. The processor is further configured to collect values of the dual variable of neighboring network devices. The processor is also configured to form a convex combination with equal weight from the collected dual variable of neighboring network devices. The processor is additionally configured to add a most recent local subgradient for the network device, scaled by a scaling factor, to the convex combination to obtain an updated dual variable. The processor is further configured to update the local estimate by projecting the updated dual variable to a primal space.

The invention discloses a parallel learning soft measurement modeling method for industrial big data. The method comprises the steps of S20, dividing sample data into M training sets, adopting a random configuration network parallel learning strategy combining a point increment algorithm and a block increment algorithm, and synchronously establishing and solving a candidate hidden layer node poolmeeting a supervision mechanism for the M training sets; S30, selecting an optimal candidate node from the candidate hidden layer node pool based on a residual steepest descent principle, and adding the optimal candidate node as a hidden layer growth node to the current network; s40, if the model parameters of the current network reach a stop standard, determining a soft measurement model according to the corresponding model parameters; and S50, if the model parameter of the current network does not reach the stop standard, updating the block number M of the sample data in the next iteration according to the current hidden layer node number, returning to execute the step S20 until the model parameter of the current network reaches the stop standard, and determining the soft measurement model according to the model parameter when the model parameter reaches the stop standard.

The invention provides a service quality evaluation system based on sub-time window deep reinforcement learning. The system comprises a data acquisition module, a model adjustment module, a reward feedback module, a parallel learning module, a Q table updating module, a periodic iteration module, a prediction learning module and a time window adjustment module. According to the method, the problems of low evaluation accuracy, poor timeliness and the like in an existing quality evaluation method are solved by utilizing the interactivity and decision-making ability of reinforcement learning andthe perception ability of deep learning.

The invention discloses a vehicle remote control system based on 5G communication and high-precision positioning and a control device thereof. The control system comprises a laser radar, a high-definition camera, a high-precision positioning antenna, a sound control radar, a millimeter wave radar, a vehicle, a satellite, a time delay detection device, a 5G base station, the control device, a rack and a host. The control device comprises a chassis, a first gear, a second gear, a motor, a signal receiving module, a supporting rod, a first connecting rod, a second connecting rod, a third connecting rod and a first hydraulic cylinder. According to the invention, information of cloud, roads and vehicles is seamlessly connected by using digital and informatization resources, and remote vehicles, a management and control platform and a driving simulator are connected in real time by using development technologies of parallel vision and perception, parallel learning, parallel planning, parallel control and the like, so that the driving of the remote vehicles becomes measurable and controllable. The rapid response of a vehicle system to the environment is improved, the overall system cost is reduced, and vehicle-road interaction, multi-vehicle cooperation, parallel control and safe driving are achieved.

The invention relates to the technical field of image recognition and detection applied to insulators, and provides an insulator image detection method based on unmanned aerial vehicle inspection, which comprises the following steps: 1, acquiring an aerial image set during cruising through shooting equipment carried by an unmanned aerial vehicle; 2, classifying the aerial image set by adopting a manual labeling mode and forming a plurality of related task packets; 3, amplifying a plurality of related task packets in a data amplification mode, and performing parallel learning training on the insulator detection model by utilizing the amplified data packets; 4, in multi-task learning training, introducing a training evaluation formula; if a result value generated by calculation of the training evaluation formula is greater than a set threshold value, immediately stopping continuing to learn the auxiliary task; and 5, verifying the effectiveness of insulator detection through quantitativeanalysis and comparison of experiment results; and the detection performance and the accuracy are improved.

The invention discloses a solar wind speed prediction method based on a multi-task deep learning neural network. A used model comprises a sharing module composed of a one-dimensional convolutional neural network CNN and a long and short term memory neural network LSTM, a main LSTM module and an autoregression layer AR module. The sharing module is used for extracting shallow space and time features shared by a plurality of tasks, the main LSTM module is used for independently extracting features of a main task, and the autoregression module is used for properly correcting a neural network model in a linear mode. A multi-task learning mechanism is adopted to realize multi-step prediction, and a model is constructed. In the field of solar wind speed prediction, introduction of a multi-task learning mechanism is a new attempt. The method has the advantages that a plurality of independent prediction tasks are combined for simultaneous parallel learning, data information and shared public information are fully utilized to help to improve shallow feature extraction, and an independent structure is designed for a main task to improve prediction performance.

The invention discloses a local and global parallel learning-based style transformation method for a ten-million-level pixel digital image, and the method comprises the following steps: S1, constructing a stylized model training sample set, including an original image sample set, a corresponding image retouching sample set obtained by manual processing of a professional image retouching person, and a semantic segmentation image sample set corresponding to the original image sample set; S2, compressing the original image sample set and the corresponding retouching image sample set to obtain a small-size small image training sample set; S3, training to obtain a small graph stylized model; S4, based on the training sample set, cutting the original image sample set to obtain a corresponding slice pair, training and recording coordinate information, and obtaining a slice stylized model; S5, obtaining a fusion model; S6, jointly training the three networks in the steps S3 to S5. The invention further discloses a local and global parallel learning-based style transformation system for the ten-million-level pixel digital image. According to the method, local and global parallel learning is realized, the processing speed is higher, and the effect is better.

The invention discloses a mobile robot collision avoidance planning method based on deep reinforcement learning in a dynamic environment, and belongs to the technical field of mobile robot navigation. The invention collects original data through a laser rangefinder, processes the original data as the input of a neural network, and establishes an LSTM neural network. Through the A3C algorithm, the neural network outputs corresponding parameters, and obtains the action of each step of the robot after processing. The present invention does not need to model the environment, and is more suitable for environments with unknown obstacles. It adopts the actor-critic framework and time difference algorithm, realizes low variance and is more suitable for continuous action spaces, and realizes the effect of learning while training. Design a continuous action space with a heading angle limit, and use 4 threads for parallel learning and training. Compared with the general deep reinforcement learning method, it greatly improves the learning and training time, reduces sample correlation, and ensures the high utilization of the exploration space and exploration strategies. Diversity, thereby improving algorithm convergence, stability and success rate of obstacle avoidance.

The invention discloses a subway passenger flow multi-step prediction method based on a space-time parallel grid neural network, and the method comprises the steps: firstly providing a grid neural network to learn the time relation of the subway passenger flow, and capturing the short-term time correlation of the subway passenger flow; then, proposing an encoder-decoder based on a periodic grid neural network to capture long-term time correlation of subway pedestrian flow; measuring spatial correlation between subway stations through indexes based on transfer flow, and modeling a subway system into a weighted directed graph; combining the propagation graph neural network with an encoder-decoder based on a grid neural network, and learning the dynamic spatial correlation of the subway passenger flow; and executing learning processes of long and short term correlation and dynamic spatial correlation in parallel, and fusing results of the two parties to obtain a final subway passenger flow multi-step prediction result. According to the method, a space-time parallel learning framework is adopted, and the long and short term correlation and the dynamic spatial correlation of the subway passenger flow are effectively learned and applied to multi-step prediction.

The invention discloses a routing method based on hierarchical Q-routing planning, which is to obtain high-efficiency data transmission links through global layered parallel planning by sensing network congestion and interconnection link usage. The algorithm of the present invention is a routing algorithm based on a lookup table. The routing algorithm stores the planned direction in the routing table in the learning module of each router node, and the data packet is obtained by accessing the routing table in the learning module of the router node. path information. The present invention constructs a hierarchical design on the basis of split Q-routing, utilizes multi-layer congestion sensor and multi-layer parallel learning to greatly reduce the convergence time of the algorithm, thereby improving the data transmission efficiency of the on-chip network, and also compressing the routing table , reducing hardware resource consumption.

The invention discloses a fuzzy integrated classifier with high interpretability based on parallel learning. The fuzzy integrated classifier sequentially comprises the following steps of constructing a zero-order TSK fuzzy classifier fastly; constructing an enhanced verification data set, calculating the value of each corresponding output function for each sample of the verification data set, and taking the value of each corresponding output function as an enhanced feature of the original sample so as to form the enhanced verification data set; respectively calling an FCM algorithm for each class on the enhanced verification data set to generate a series of representative center points and corresponding tags thereof, and removing the enhanced features of the representative center points; for any test sample, finding k representative center points closest to the test sample by using the KNN to determine the class label of the test sample. According to the present invention, a parallel learning method is adopted, the operation complexity is low, the structure is simple and easy to understand, the fast learning can be carried out, and the good performance is achieved on the classification problem processing.

A network device, system, and method are provided. The network device includes a processor. The processor is configured to store a local estimate and a dual variable maintaining an accumulated subgradient for the network device. The processor is further configured to collect values of the dual variable of neighboring network devices. The processor is also configured to form a convex combination with equal weight from the collected dual variable of neighboring network devices. The processor is additionally configured to add a most recent local subgradient for the network device, scaled by a scaling factor, to the convex combination to obtain an updated dual variable. The processor is further configured to update the local estimate by projecting the updated dual variable to a primal space.

The invention discloses a Q-learning-based hybrid cloud job scheduling method. A plurality of agents are used for parallel learning, namely, each agent independently performs optimal policy learning; and when one agent obtains a policy meeting a condition that error is less than theta first, knowledge transfer among the agents is carried out. According to the method, a reinforcement learning-based hybrid cloud jog scheduling method is designed by taking minimum user job completion time and waiting time as optimization goals by analyzing an executive process of a user job in a cloud environment, and the convergence of an optimal policy is accelerated by adopting a parallel multi-agent technology, so that the utilization rate of cloud resources is increased and the default rate of a user level protocol is reduced.

The invention discloses a multi-agent deep reinforcement learning algorithm based on partition experience and multi-thread interaction. First, the algorithm uses the experience replay form of the partitioned buffer area to distinguish positive experience, negative experience and neutral experience by dividing the reward space, and uses stratified random sampling to extract these experience data during training. Secondly, the algorithm uses multi-threaded interaction to promote the trial-and-error process between the agent and the environment. Multiple clones of the agent learn in parallel and integrate their learning experience to train the parameters of the network model. The advantages are: the multi-agent deep reinforcement learning algorithm based on buffer replay and multi-thread interaction proposed by the present invention, combined with the advantages of partitioned experience buffer and multi-thread interaction mode, is introduced into the multi-agent deep reinforcement learning algorithm; It is superior to existing models in terms of convergence speed and training efficiency, and has higher usability in a multi-agent environment, and can be used to solve the problem of multi-agent cooperative tracking targets.

The invention provides a subway pedestrian flow prediction method and system based on a space-time parallel grid neural network, and the method comprises the steps: providing a grid neural network to learn the time relation of subway pedestrian flow, and capturing the short-term time correlation of subway pedestrian flow; further capturing the long-term time correlation of the subway pedestrian flow; measuring spatial correlation between subway stations through an index based on transfer flow, and modeling a subway system into a weighted directed graph based on the index; based on construction of a subway weighted directed graph, combining a propagation graph neural network with a grid neural network, and learning dynamic spatial correlation of metro pedestrian flow; and executing the learning processes of the long and short term correlation and the dynamic spatial correlation in parallel, and fusing the results of the two to obtain a final subway pedestrian flow prediction result. According to the method, a space-time parallel learning framework is adopted, the long-term and short-term time correlation and the dynamic space correlation of the subway pedestrian flow can be effectively learned, and the learned knowledge is applied to prediction.

The invention discloses a caenorhabditis elegans detection method based on a multi-task deep neural network, and the method comprises the steps: obtaining an original fluorescence picture of caenorhabditis elegans, carrying out the manual marking of the fluorescence point contour and fluorescence brightness of the original fluorescence picture, carrying out the image expansion and preprocessing, and building a training set; an improved YOLACT network of a multi-task learning mechanism is used for learning, a binary mask and pixel coordinate information of fluorescent dots are obtained according to an instance segmentation result and a fluorescent brightness result of caenorhabditis elegans fluorescent dots of the improved YOLACT network and an instance segmentation result of caenorhabditis elegans polypide and fluorescent dots in an RGB image, and according to pixel coordinates of the binary mask, the fluorescent dots of the caenorhabditis elegans are obtained. And calculating the area of the fluorescent point. According to the detection method, three results are output through parallel learning of caenorhabditis elegans fluorescent dot segmentation and fluorescence degree and size measurement, the accuracy of the output results can be improved through mutual complementation among various loss functions in a multi-task learning mechanism, and the accuracy is improved while the learning efficiency is improved.

The invention discloses a momentum-based parameter identification method for a space robot system, comprising: step 1: establishing a momentum-based parameter identification model for a space robot system; step 2: using a parallel learning method for parameter identification, and every The first step is to update the parameter identification result by using the motion information of the space robot system at the current time and the past time at the same time, and complete the update of the parameter identification result. It realizes that the excitation motion of the system does not need to meet the continuous excitation condition, and it can also ensure that the parameter identification value can converge to the true value at an exponential rate in the global range, which reduces the requirement of the parameter identification method for the system motion, which is beneficial to save fuel consumption and make the system Motion meets other mission requirements.

The invention discloses a routing method based on hierarchical Q-routing planning, which is characterized in that a high-efficiency data transmission link is obtained through global hierarchical parallel planning by sensing the congestion condition of a network and the use condition of an interconnection link. The algorithm provided by the invention is a routing algorithm based on a lookup table, the routing algorithm stores a planned direction in a routing table in a learning module of each router node, and a data packet obtains path information by accessing the routing table in the learning module of the router node where the data packet is located. According to the method, a hierarchical design is constructed on the basis of split Q-routing, and the convergence time of the algorithm is greatly shortened by utilizing a multilayer congestion sensor and multilayer parallel learning, so that the on-chip network data transmission efficiency is improved, the routing table is compressed, and the hardware resource consumption is reduced.