34 results about "Gradient network" patented technology

The invention belongs to the field of additive manufacturing, and discloses a 4D printing method of an intelligent structure with a large deformation function and a product. The method comprises the following steps that (a) a three-dimensional structure of the intelligent structure to be printed is constructed, wherein the intelligent structure to be printed is of a porous gradient network structure, and comprises three layers with the volume fractions of pores sequentially reduced from top to bottom in the vertical direction; (b) a three-dimensional model of the intelligent structure is set up; (c) memory alloy powder is selected as raw materials, and the 3D printing technology is adopted for printing the three-dimensional model; and (d) the printed intelligent structure is subjected to heat preservation, cooling and external field irrigation, the deforming intelligent structure is obtained and subjected to performance testing, and the needed intelligent structure is selected according to performance needs. By means of the 4D printing method, the obtained product responds to the changing external environment immediately, the expectant optimal state is kept all the time, meanwhile,the 4D printing method is not limited by the structure complexity, and the requirements of the mechanical property and the fatigue property are met.

The invention discloses a production method of a proton exchange membrane fuel cell membrane electrode. A screen printing technology and an ultrasonic spraying technology which are two production technologies of membrane electrodes are adopted, an electrocatalyst slurry I undergoes the screen printing technology, an electrocatalyst slurry II undergoes the ultrasonic spraying technology and then is directly or indirectly supported on a proton exchange membrane, and the produced membrane electrode has a catalyst layer with a gradient network structure. The membrane electrode with high performance, long life, multiple layers and gradient network structure is produced by using different contents of an adhesive in different electrocatalyst slurries; and the compact structure of the catalyst layer in a thin layer electrode and an ultrasonic sprayed loose three-dimensional network structure make the interface resistivity of the catalyst layer lower than the that of the three-dimensional network structure, and the area of the introduced thin layer electrode is greater than the spray area when the effective area of the catalyst layer is ensured, so the contact part of a bipolar plate and the thin layer electrode can well facilitate electron conduction in the cell assembling process.

The invention relates to a method for preparing an acrylate polymer emulsion which is used for a pressure sensitive adhesive and is provided with a gradient network structure, which relates to an adhesive based on acrylate polymer. The method comprises the following steps: (1) preparation of a seed emulsion; (2) pre-emulsification of shell layer monomers; and (3) gradient feeding of a shell layer. The acrylate polymer emulsion is prepared through dipping a chain-transferring agent, namely tert dodecyl mercaptan (CTA) into the shell layer to make the concentration discontinuously changed in a phasic mode, wherein poly(butyl acrylate) is taken as a core and a copolymer of acrylic ester and acrylic monomers is taken as the shell layer; and the crosslinking density of the network structure of the shell layer shows gradient change. The acrylate polymer emulsion provided with the gradient network structure can be used as the pressure sensitive adhesive after drying and film forming. The method realizes optimization of the performance of the pressure sensitive adhesive, improves the initial adhesion and simultaneously maintains large peeling strength and continuous viscous force.

A porous membrane of vinylidene fluoride resin, comprising a substantially single layer membrane of vinylidene fluoride resin having two major surfaces sandwiching a certain thickness, including a dense layer that has a small pore size and governs a filtration performance on one major surface side thereof, having an asymmetrical gradient network structure wherein pore sizes continuously increase from the one major surface side to the other opposite major surface side, and satisfying conditions: (a) the dense layer includes a 5 μm-thick portion contiguous to the one major surface showing a porosity A1 of at least 60%, (b) the one major surface shows a pore size P1 of at most 0.30 μm, and (c) the porous membrane shows a ratio Q / P14 of at least 5×104 (m / day·μm4), wherein the ratio Q / P14 denotes a ratio between Q (m / day) which is a value normalized to a whole layer porosity A2=80% of a water permeation rate measured at a test length L=200 mm under the conditions of a pressure difference of 100 kPa and a water temperature of 25° C., and a fourth power P14 of the pore size P1 on the one major surface. The porous membrane is produced through a process including: extruding a melt-kneaded mixture of a vinylidene fluoride resin and a plasticizer through a die into a form of a film, followed by cooling, to form a solidified film; and extracting the plasticizer to recover a porous membrane; wherein the plasticizer is mutually soluble with the vinylidene fluoride resin at a temperature forming the melt-kneaded mixture and further satisfies properties: (i) giving the melt-kneaded mixture with the vinylidene fluoride resin with a crystallization temperature Tc′ (° C.) which is lower by at least 6° C. than a crystallization temperature Tc of the vinylidene fluoride alone, (ii) giving the cooled and solidified product of the melt-kneaded mixture a crystal melting enthalpy ΔH′ (J / g) of at least 53 J / g per weight of the vinylidene fluoride resin as measured by a differential scanning calorimeter (DSC), and (iii) the plasticizer alone showing a viscosity of 200 mPa-s-1000 Pa-s at a temperature of 25° C. as measured according to JIS K7117-2 (using a cone-plate-type rotational viscometer).

The invention discloses a mechanical arm path planning method based on velocity smoothing deterministic policy gradient. The method comprises the steps that a mechanical arm simulation environment with job task feedback is established in a training stage; a previous step mechanical arm action vector is introduced during inputting of a deterministic policy gradient network, and a reinforced learning network framework based on the velocity smoothing deterministic policy gradient is established; network training parameters and the mechanical arm simulation environment are initialized; and samplesare obtained based on the velocity smoothing deterministic policy gradient network and the simulation environment, a training sample database is established, if the training sample quantity reaches the maximum sample quantity, training samples are drawn from the training sample database according to the single time training sample quantity, the velocity smoothing deterministic policy gradient network is trained, and otherwise, next step or the next time of simulation is performed. According to the mechanical arm path planning method provided by the invention, the previous step velocity vectoris added as the network input on the basis of the deterministic policy gradient network, the joint acceleration is effectively decreased, and mechanical arm jitter is reduced.

Disclosed is a porous vinylidene fluoride resin membrane characterized by having two main surfaces, between which a certain thickness is sandwiched, and by having, on the side thereof facing one of the main surfaces, a dense layer that has a small pore diameter and governs the filtration performance, having an asymmetric gradient network structure in which the pore diameter increases continuously from the surface toward the other surface, and satisfying the following requirements (a) to (c): (a) the part of the dense layer which has a thickness of 5 [mu]m and extends continuously from the surface has a porosity (A1) of 60% or higher; (b) the surface has a surface pore diameter (P1) of 0.30 [mu]m or smaller; and (c) the ratio of the amount of water which has passed through in a test length (L) of 200 mm under the conditions of a differential pressure of 100 kPa and a water temperature of 25 DEG C, the water amount being a converted value (Q (m / day)) calculated for an overall porosity (A2) of 80%, to the fourth power of the surface pore diameter (P1) of the surface (P1<4> ([mu]m<4>)), Q / P1<4>, is 5*10<4> (m / day.[mu]m<4>) or higher.

The invention discloses an efficient mechanical arm grabbing deep reinforcement learning reward training method and system, and relates to the technical field of machine learning. A depth camera is used for identifying a target object and a mechanical arm tail end, the sight distance between the center of a mechanical arm tail end actuator and the center of the target object is calculated, and the distance is returned to a computer as a basis for judging whether a mechanical arm can grab the target object or not. Every time the mechanical arm tail end actuator tents, the computer takes the moving distance of the mechanical arm tail end actuator relative to the object, the moving step number of a mechanical arm, the sum of the rotating angles of steering engines of all degrees of freedom of the mechanical arm and the weighted sum of whether the target object is successfully grabbed or not as a reward mechanism of a DDPG depth deterministic strategy gradient network. The end-to-end training process is completed by using a DDPG. The problems that existing mechanical arm control is poor in action coherence and not coordinated are solved.

The invention belongs to the field of target tracking, and particularly relates to a multi-maneuvering target tracking method based on a depth deterministic strategy gradient DDPG. The method includes: establishing a depth deterministic strategy gradient network; generating deep deterministic strategy gradient network input data sk by using an LSTM network; setting actions and returns of the depthdeterministic strategy gradient network to obtain a set depth deterministic strategy gradient network; training the set depth deterministic strategy gradient network according to the depth deterministic strategy gradient network input data sk to obtain a trained depth deterministic strategy gradient network; and performing multi-maneuvering target tracking by using the trained depth deterministicstrategy gradient network. The method is reasonable in resource distribution and high in radar tracking precision, and the problems of no model and Markov decision are solved.

The embodiment of the invention discloses an image processing method and a related device. The image processing method is applied to electronic equipment and comprises the steps that portrait mattingnetwork model application is trained according to sample data in a training set to calculate matting loss of the portrait matting network model application, and the sample data comprises multiple portrait images and matting standard images corresponding to the multiple portrait images; according to the matting loss, adjusting a matting network model included in the portrait matting network model application and various parameters in the edge gradient network model to obtain a trained portrait matting network model application; and performing portrait matting processing by using the trained portrait matting network model application. The embodiment of the invention is beneficial to improving the processing effect of portrait matting.

The invention provides a nuclear operation and maintenance robot shaft hole assembling method based on man-machine cooperation, and belongs to the technical field of industrial robots. A slave end mechanical arm is controlled by a master end mechanical arm to rotate a pin in a gap area of the center of a hole; the pin at the tail end of the slave end mechanical arm makes contact with a plane wherethe hole is located, and adjusting is carried out on the pose of the pin through data returned by a torque sensor in real time; and pose information of the pin and information of the torque sensor are acquired, and the pin is pushed to be inserted into the hole by utilizing a depth deterministic strategy gradient network of continuous action, and then nuclear operation and maintenance robot shafthole assembly based on man-machine cooperation is finished. In order to reduce the radiation dosage borne by operation and maintenance personnel and improve the operation and maintenance efficiency,robot intelligence and human intelligence are fully combined through the method; the reliability of the operation process is enhanced by utilizing experience knowledge of people, and it is guaranteedthat the risk is controllable; and meanwhile, the robot autonomously acts in a local area by utilizing an artificial intelligence algorithm.

The invention relates to a microgrid controller based on an artificial intelligence control method. According to the invention, a historical sample database is used to form an observation set of a deep deterministic policy gradient network to perform action exploration; training is performed according to continuous interactions between the microgrid controller and a microgrid system environment; mapping relationships between microgrid operation conditions and coordination control results are directly constructed; the microgrid controller directly uses a microgrid controller model which is trained offline to carry out rapid online decision making, so that the calculation and response speed is very high; and an offline training sample library is periodically updated to realize rolling optimization of the model and continuous improvement of policies. Therefore, real-time coordination control of the microgrid controller on a power supply, a grid, loads and storage in a system is realized, and the high efficiency and the robustness of the microgrid controller are effectively improved.

The invention discloses an NOMA system resource allocation method based on optimized sample sampling and a storage medium, and belongs to the technical field of mobile communication and wireless networks. The problems that when resources of an NOMA system are allocated through an existing deep reinforcement learning network, samples with important values are not learned possibly, and the learning rate is low are solved. According to the method, the deep reinforcement learning network based on the sample optimization pool is designed, the current channel state information serves as input, the user sum rate serves as the optimization target, each sample TD error serves as the priority, and the optimal user grouping strategy is output through the deep reinforcement learning network; and meanwhile, the optimal distribution power of each user is output by utilizing the depth deterministic strategy gradient network. According to the method, the occurrence probability of valuable samples is improved by introducing the priorities of the samples, the learning rate of the deep reinforcement learning network can be improved, and the convergence speed is increased. The method is mainly used for resource allocation of the NOMA system.

The invention discloses a gradient compression method for distributed DNN training in an edge computing environment. The method comprises the steps: building a selection standard based on a gradient number, and screening a gradient network layer which meets a model compression standard; evaluating gradient importance according to the gradient entropy, adaptively selecting a gradient sparsification threshold value, and performing gradient sparsification compression based on the flexible threshold value; according to gradient residual errors and a momentum correction mechanism, accumulating and optimizing the gradient residual errors, thereby reducing the performance loss of a training model caused by gradient sparsity; quantizing the sparse gradient according to a ternary quantization compression scheme to obtain a sparse ternary tensor; and according to a lossless coding technology, recording a distance of a non-zero gradient in the transmission tensor, performing optimization coding on the distance, and outputting a sparse ternary gradient. According to the sparse ternary gradient compression algorithm based on the gradient number and the gradient entropy, the gradient size of a gradient exchange stage in distributed DNN training can be adaptively compressed; therefore, the communication efficiency of distributed DNN training is effectively improved.

The invention relates to a network frangibility correlation analysis method based on gradient attack hypotheses, which comprises the network environment gradient and frangibility correlation analysis under a gradient network environment. The network environment gradient comprises network environment sub-network recognition and network environment gradient marking. The frangibility correlation analysis under the gradient network environment comprises a pointer initialization device, a transverse direction frangibility correlation analysis device with the same gradient, a longitudinal directionfrangibility correlation analysis device along the gradient direction, and a loop control device. Under the gradient attack hypotheses, facing a large-scale of complex target network, through the network scene gradient, the parallel partitioning of the network frangibility correlation analysis task is turned to be possible, and the foundation is established for the bottleneck of a large-scale application to break through the network frangibility analysis.

The invention provides a mechanical intelligent fault prediction method based on an automatic convolutional neural network, and the method comprises the steps: obtaining an equipment fault signal, carrying out the preprocessing of the equipment fault signal, and obtaining a preprocessed fault signal; constructing an automatic convolutional neural network ACNN fault diagnosis model, wherein the ACNN fault diagnosis model comprises a group of convolutional neural networks CNN and a group of deep deterministic strategy gradient networks DDPG, the convolutional neural network CNN is used for equipment fault prediction, and the deep deterministic strategy gradient network DDPG is used for realizing automatic adjustment of three parameters of learning rate, batch and regularization of the convolutional neural network CNN; training the ACNN fault diagnosis model by using the preprocessed fault signal to obtain a trained ACNN fault diagnosis model; and applying the trained fault diagnosis model to equipment fault diagnosis. The beneficial effects of the invention are that the method achieves the automatic adjustment and optimization of the parameters of the convolutional neural network, and enables the convolutional neural network to be good in fault feature extraction capability.

The invention provides an energy efficiency optimization transmission method based on a complex gradient network. The method comprises the steps of: (A), giving a scalar field to each node on a bottomnetwork; (B), according to the scalar field given to each node on the bottom network, calculating a transmission function value of a current node for performing information transmission to various neighbour nodes; and (C), according to the calculated transmission function value, selecting the best next-hop node of the current node for performing information transmission. By means of the energy efficiency optimization transmission method and system based on the complex gradient network provided by the invention, on the premise that the service quality is ensured, the problems of low network energy efficiency and high energy consumption are solved; and simultaneously, conforming to demands of the time, the method also can save energy and reduce emission.

The invention discloses a power grid regulation and control method based on a layered depth strategy gradient network. A state representation vector and an action representation vector of a power grid are designed for the power grid; the motion space is clustered, so that the motion number of each cluster is equal, a power grid regulation and control model is designed based on a layered strategy gradient network by taking a state representation vector as the input of the network and using a strategy gradient algorithm, the model has two layers, each layer is an independent strategy gradient model, the first layer selects the motion cluster firstly, and the second layer selects the motion cluster; the second layer selects specific actions in the cluster, and continuous decision making is carried out; based on a discrete power grid operation data set of a simulated power grid environment, the model and the simulated power grid operation environment are interacted, a current state is obtained from the simulated power grid operation environment, and a power grid action to be executed is executed by the simulated power grid operation environment, so that the purpose of power grid regulation and control is achieved. The invention provides a feasible means for real-time regulation and control of the power grid.

The invention relates to a self-adaptive parking lot exit barrier gate control method. The method comprises the following steps: receiving downstream and internal queuing information of a parking lot exit; predicting the number of vehicles driving away from each exit in unit time; calculating forbidding time of different exit barrier gate control modes; the method comprises the following steps: taking parking lot exit downstream and internal queuing information and an allowable maximum queuing length as input parameters, and constructing an adaptive parking lot exit barrier gate control model by using a multi-agent deep reinforcement learning framework, establishing a corresponding commentator network and a corresponding performer network, outputting expected benefits, and iteratively training the networks until a reward function converges by utilizing interaction data to obtain the maximum expected benefits; and inputting parking lot exit downstream and internal real-time queuing information and an allowable maximum queuing length, and outputting parking lot exit barrier gate control by using the model. Compared with the prior art, the method has the advantages of consideration of dynamic and static traffic, high flexibility and the like.

The invention provides a mechanical intelligent fault prediction method based on an automatic convolutional neural network, which acquires equipment fault signals and preprocesses them to obtain preprocessed fault signals; constructs an automatic convolutional neural network ACNN fault diagnosis model; The ACNN fault diagnosis model includes a set of convolutional neural network CNN and a set of deep deterministic policy gradient network DDPG; the convolutional neural network CNN is used for equipment fault prediction, and the deep deterministic policy gradient network DDPG is used to realize the The learning rate of the convolutional neural network CNN, batch and three parameters of regularization are automatically adjusted; the ACNN fault diagnosis model is trained using the preprocessed fault signal to obtain the completed ACNN fault diagnosis model; the training The completed fault diagnosis model is applied to equipment fault diagnosis. The beneficial effect provided by the invention is that the automatic adjustment and optimization of the parameters of the convolutional neural network is realized, so that the convolutional neural network has a good ability to extract fault features.

The invention discloses an efficient deep reinforcement learning reward training method and system for grasping a mechanical arm, which relates to the technical field of machine learning. and return this to the computer as a basis for judging whether the robotic arm can grasp the target object. Each time the end effector of the manipulator is tested, the computer calculates the moving distance of the end effector of the manipulator relative to the object, the number of steps moved by the manipulator, the sum of the rotation angles of the steering gear of each degree of freedom of the manipulator, and whether the target object is successfully grasped. And as the reward mechanism of DDPG deep deterministic policy gradient network, the end-to-end training process is completed with DDPG. The present invention improves the problems of poor action continuity and uncoordinated existing in the control of the existing mechanical arm.

The invention discloses a comparison self-supervised learning method based on a multi-network framework. The method comprises the following steps: applying a data augmentation means to each image in a training set to obtain three independent augmented views; respectively inputting the three augmented views into a designed back propagation network, a stop gradient network and a momentum network; respectively calculating loss values of output vectors between the back propagation network and the stop gradient network and between the back propagation network and the momentum network in combination with the negative sample queue, and adding the loss values to obtain a total loss value; performing gradient updating on parameters of the back propagation network through minimizing the total loss value; updating the parameters of the stop gradient network and the momentum network by using the parameters of the back propagation network; and updating the negative sample queue by using the momentum network. On the basis of a classical self-supervised learning method, more positive sample pairs are introduced by using a multi-network framework, and more negative samples are introduced by combining end-to-end and momentum mechanisms, so that a better pre-training effect is achieved.

The invention provides an airspace steganalysis method and system based on gradient network architecture search, which solve the problems of poor performance and long network search time of the existing manual design network. The method comprises the following steps:firstly, defining a steganalysis network search architecture; determining a search architecture and candidate operations of each part of to-be-searched modules in a steganalysis network search architecture, fusing all candidate operations in a search space by using a Softmax function to construct a hyper-parameter network, and then training and optimizing the constructed hyper-parameter network by using a gradient descent method, selecting corresponding operation according to architecture parameters in the trained hyper-parameter network, and completing search and construction of the steganalysis network. Manual intervention in network design is reduced, the defect that the performance of the constructed steganalysis network is limited is overcome, and a network architecture search mode based on gradient updating reduces search time consumption.

The invention discloses a power grid power flow regulation and control decision reasoning method based on a depth deterministic strategy gradient network. The method comprises the following steps: designing state representation vector and an action representation vector of a power network for the power network; designing an inference model based on a depth deterministic strategy gradient network, taking the state representation vectors as the input of an Actor network to obtain a plurality of similar discrete actions, taking the state-action pair vectors as the input of a Critic network, and outputting the value estimation of each state-action pair vector; selecting the action with the highest estimated value as a final action to be executed in the environment in the state; and simulating a power grid operation environment based on the discretized power grid operation data set, interacting the model with the simulated power grid operation environment, obtaining a current state and a final action to be executed from the simulated power grid operation environment, and executing the final action to be executed by the simulated power grid operation environment. The invention provides a feasible means for real-time regulation and control of the power network.

The invention discloses a comparative self-supervised learning method based on a multi-network framework, comprising the steps of: applying a data augmentation method to each image in a training set to obtain three independent augmented views; Input to the designed backpropagation network, stop gradient network and momentum network; calculate the loss value of the output vector between the backpropagation network and the stop gradient network, the backpropagation network and the momentum network, respectively, in combination with the negative sample queue, and add them together. Then, the total loss value is obtained; the parameters of the backpropagation network are gradient updated by minimizing the total loss value; the parameters of the stop gradient network and the momentum network are updated by using the parameters of the backpropagation network; the negative sample queue is updated with the momentum network. The present invention is based on the classic self-supervised learning method, by using a multi-network framework to introduce more positive sample pairs, and combining end-to-end and momentum mechanisms to introduce more negative samples to achieve better pre-training effects.

A NOMA system resource allocation method and storage medium based on optimized sample sampling, belonging to the technical field of mobile communication and wireless network. In order to solve the problem that samples of important value may not be learned when using the existing deep reinforcement learning network to allocate the resources of the NOMA system, and the resulting low learning rate. The present invention designs a sample optimization pool-based deep reinforcement learning network that takes the current channel state information as input, the user sum rate as the optimization target, and the TD error of each sample as the priority, and uses it to output the optimal user grouping strategy , while using a deep deterministic policy gradient network to output the optimal assigned power for each user. The present invention improves the occurrence probability of valuable samples by introducing the priority of samples, can improve the learning rate of the deep reinforcement learning network, and accelerate the convergence speed. It is mainly used for resource allocation of NOMA system.

The invention relates to a network frangibility correlation analysis method based on gradient attack hypotheses, which comprises the network environment gradient and frangibility correlation analysis under a gradient network environment. The network environment gradient comprises network environment sub-network recognition and network environment gradient marking. The frangibility correlation analysis under the gradient network environment comprises a pointer initialization device, a transverse direction frangibility correlation analysis device with the same gradient, a longitudinal directionfrangibility correlation analysis device along the gradient direction, and a loop control device. Under the gradient attack hypotheses, facing a large-scale of complex target network, through the network scene gradient, the parallel partitioning of the network frangibility correlation analysis task is turned to be possible, and the foundation is established for the bottleneck of a large-scale application to break through the network frangibility analysis.

The invention discloses a gradient compression method for distributed DNN training in an edge computing environment, establishes a selection standard based on the number of gradients, and screens the gradient network layer that meets the model compression standard; evaluates the importance of the gradient according to the gradient entropy, and selects it adaptively The threshold of gradient sparsification is based on the flexible threshold to compress the gradient sparsification; according to the gradient residual and momentum correction mechanism, the gradient residual is accumulated and optimized to reduce the performance loss of the training model caused by gradient sparsity; according to the ternary quantization compression scheme, quantization The sparse gradient is obtained to obtain a sparse ternary tensor; according to the lossless coding technology, the distance of the non-zero gradient in the transfer tensor is recorded, and the optimized encoding is performed to output the sparse ternary gradient. The sparse ternary gradient compression algorithm based on the gradient quantity and gradient entropy of the present invention can adaptively compress the gradient size of the gradient exchange stage in the distributed DNN training, and effectively improve the communication efficiency of the distributed DNN training.

The invention discloses a production method of a proton exchange membrane fuel cell membrane electrode. A screen printing technology and an ultrasonic spraying technology which are two production technologies of membrane electrodes are adopted, an electrocatalyst slurry I undergoes the screen printing technology, an electrocatalyst slurry II undergoes the ultrasonic spraying technology and then is directly or indirectly supported on a proton exchange membrane, and the produced membrane electrode has a catalyst layer with a gradient network structure. The membrane electrode with high performance, long life, multiple layers and gradient network structure is produced by using different contents of an adhesive in different electrocatalyst slurries; and the compact structure of the catalyst layer in a thin layer electrode and an ultrasonic sprayed loose three-dimensional network structure make the interface resistivity of the catalyst layer lower than the that of the three-dimensional network structure, and the area of the introduced thin layer electrode is greater than the spray area when the effective area of the catalyst layer is ensured, so the contact part of a bipolar plate and the thin layer electrode can well facilitate electron conduction in the cell assembling process.