256 results about "Pooling" patented technology

The invention discloses a data processing method for hardware acceleration of a convolutional neural network. By analyzing the parallel characteristics of the convolutional neural network, and in combination with the parallel processing capability of hardware, the hardware acceleration is carried out on the convolutional neural network. The acceleration scheme is used for performing acceleration improvement on the Tiny-yolo network from three aspects that (1) the processing speed of the Tiny-yolo network is increased through multi-channel parallel input; (2) the convolution computing speed ofthe Tiny-yolo network is increased through parallel computing; and (3) the pooling process time of the Tiny-yolo network is shortened through pooling embedding. The method greatly increases the detection speed of the convolutional neural network.

Disclosed embodiments relate to virtual pooling of a local resource in a balloon network. In an example, a balloon network has geographic zones, including at least a first and a second geographic zone. Further, each balloon has a sustainable utilization rate for a local resource. Each balloon utilizes the local resource according to a respective first utilization rate when located in the first geographic zone and utilizes the local resource according to a second utilization rate when located in the second geographic zone. For one or more of the balloons, the sustainable utilization rate is less than the respective first utilization rate and greater than the respective second utilization rate. However, the balloons are operable to move between the geographic zones such that the first utilization rate is substantially continuous in first geographic zone.

The invention relates to a lightweight remote sensing target detection method based on SEYOLOv3, which belongs to the technical field of target detection, and comprises the following steps: 1, takinga YOLOv3 algorithm as a basic model framework, and in order to reduce network parameters and improve network reasoning speed, designing a lightweight trunk feature extraction network; 2, in order to improve the scale invariance of the features and reduce the over-fitting risk, a spatial pyramid pooling (SPP) algorithm is provided, and pooling of three scales is carried out to obtain an output feature vector with a fixed length; a spatial attention model SE module is introduced, useless information is further compressed, and useful information is enhanced; and 3, updating parameters through iterative training to obtain a final network model, adopting multi-scale prediction by utilizing the model, and predicting a final result through detection heads of three scales. According to the method,while the reasoning speed of the network is effectively improved, the precision is ensured, the feature expression capability of the network is enhanced, and the scale invariance is improved.

The invention discloses a welding simulator virtual weld defect detection method based on deep learning. A convolutional neural network CNN is used to realize virtual weld defect detection, the convolutional neural network CNN comprises an input layer, a convolutional layer, an excitation layer, a pooling layer and a full connection layer from top to bottom, an output value of an upper layer between two adjacent layers is used as an input value of a lower layer, and the invention belongs to the technical field of weld joint detection, and provides the welding simulator virtual weld joint defect detection method based on deep learning. The method comprises the following steps: firstly, on the basis of analyzing and collecting data characteristics, designing an identification algorithm of awelding seam image, performing graying processing and segmentation processing on the image, and remarkably separating a welding seam region from a non-welding seam region; secondly, constructing a deep learning network and expanding the collected data set; and finally, performing recognition training on the defect features by using a training learning framework.

The invention discloses a garbage classification method based on a hybrid convolutional neural network, and belongs to the technical field of garbage classification and recovery. The method solves theproblems that an existing method is low in garbage classification precision and long in required training time. According to a hybrid convolutional neural network model, a convolutional layer, batchstandardization, a maximum pooling layer and a full connection layer are flexibly applied, and BN batch standardization is applied to each convolutional layer and each full connection layer, so that the feature extraction capability of the model is further enhanced, the effect of each layer is brought into full play, and a relatively good classification result is obtained. By utilizing the regularization effect of the BN layer, the maximum pooling layer is properly added to perform statistics on the features, the feature dimension is reduced, the representation capability is improved, fittingcan be well performed, the convergence speed is high, the parameter quantity is small, the calculation complexity is low, and the method has obvious advantages compared with a traditional convolutional neural network. Meanwhile, an optimizer of SGDM + Nesterov is adopted in the model, and finally the classification accuracy of the model on the image reaches 92.6%. The method can be applied to household garbage classification.

The invention provides a rolling bearing fault diagnosis method based on a space pooling network. The method comprises the steps of: collecting vibration signals of a rolling bearing under a fault state and a normal state, cutting collected vibration signals of the rolling bearing to form samples, and dividing the samples into a training set, a verification set and a test set; inputting samples inthe training set and the verification set into a convolutional neural network to train the convolutional neural network and adjusting the structure of the convolutional neural network; adding spatialpooling attention after determining the last "convolution+pooling" unit of the convolutional neural network so as to achieve feature weighting, adding two spatial pooling layers and a softmax classifier, thereby completing the construction of a spatial pooling model; inputting the samples of the training set and the verification set into the space pooling network to perform parameter updating, inputting the samples of the test set into the trained space pooling network to obtain a bearing state type, comparing the bearing state type with a label, and calculating to obtain diagnosis precision.

The invention discloses a lightweight CNN model calculation accelerator based on an FPGA, relates to the technical field of hardware acceleration, and aims to solve the problem of low operation speedof an accelerator in the prior art, and the lightweight CNN model calculation accelerator comprises a weight cache region, a normalization layer, a convolution layer, a pooling layer, a full connection layer and a Softmax classifier. According to the invention, the characteristics of fast parallel computing, low power consumption and high flexibility of FPGA are utilized; the CNN accelerator design for the lightweight network using the depth separable convolution structure is carried out, the neural network can be be deployed in a resource-limited occasion, the calculation efficiency of an algorithm is greatly improved, and the operation speed of the algorithm is increased.

The application discloses a method for knowledge extraction based on TextCNN, comprising: S10, collecting first training data, and constructing a character vector dictionary and a word vector dictionary; S20, constructing a first convolutional neural network, and training the first convolutional neural network based on a first optimization algorithm, the first convolutional neural network comprises a first embedding layer, a first multilayer convolution, and a first softmax function connected in turn; S30, constructing a second convolutional neural network, and training the second convolutional neural network based on a second optimization algorithm, the second convolutional neural network comprises a second embedding layer, a second multilayer convolution, a pooling layer, two fully-connected layers and a second softmax function, the second embedding layer connected in turn; S40, extracting a knowledge graph triple of the to-be-predicted data according to an entity tagging prediction output by the first trained convolutional neural network and an entity relationship prediction output by the second trained convolutional neural network.

The present invention discloses a human action classification method based on a video local feature dictionary. The method comprises a step of extracting a local feature from a training video with a category label, wherein a feature package is formed by the feature vector set in each segment of video, a step of grouping feature packages, using a multi-instance learning method to learn a local feature classifier, using a mode of cross-validation in the multi-instance learning, and marking multiple instances with largest ranking in each package as positive instances in updating the positive instances, a step of taking a learned classifier as the dictionary of feature encoding, and using a maximum pooling method to pool the local feature response to obtain the global vector expression of a video, and a step of using a global feature vector to learn, obtaining the classifier of each action category, and using the classifier to classify the action in a new video. According to the invention, the improvement of the accuracy of estimation is facilitated, the memory of an initial value by classification is avoided, and the accuracy of estimated positive samples is ensured at the same time.

The invention discloses a garbage classification method based on a lightweight convolutional neural network, and belongs to the technical field of garbage classification. The method solves the problemthat an existing method cannot have low model complexity and high classification precision at the same time. According to the method, a feature extraction layer is divided into nine parts, wherein convolution of each part adopts a method of combining depth separable convolution and common convolution, convolution kernels alternately adopt 1 * 1 and 3 * 3 in size, and batch normalization processing is carried out on a convolution result of each time. Different from a common ReLU activation function and a Flatten connection layer, the model provided by the invention adopts Leaky ReLU as the activation function and a global average pooling layer as the connection layer. Experimental results show that after the network is trained and tested on a TrashNet data set, the accuracy of 93.02% is obtained, the classification precision is high, the complexity of the model is low, and the classification precision and the complexity of the model can be considered at the same time. The method can beapplied to intelligent garbage classification.

The invention discloses a statement recognition method and device, and relates to the technical field of human-computer interaction. According to the specific implementation scheme, the method comprises the steps of obtaining a to-be-recognized statement and corresponding feature information, wherein the feature information comprises a word segmentation result, a part-of-speech recognition resultand an entity recognition result of the statement; obtaining a trained dialogue understanding model, wherein the dialogue understanding model comprises a backbone neural network, a slot branch connected with the backbone neural network, a pooling layer connected with the backbone neural network, and an intention branch and an intention slot relationship branch which are respectively connected withthe pooling layer; inputting the feature information into a trained dialogue understanding model; obtaining the intention and the slot position of the statement, so that the intention slot positionrelationship branch is added into the dialogue understanding model, the relationship between the intention and the slot position can be considered when the dialogue understanding model is trained or the statement is recognized, the matching degree between the output intention and the slot position is improved, and the statement recognition efficiency of the dialogue understanding model is improved.

The invention discloses a knowledge mining method for an intelligent IETM fault maintenance record text. The knowledge mining method comprises the steps of 1, collecting equipment fault maintenance record text data; 2, establishing an equipment fault word bank; 3, obtaining a lexical item document matrix; 4, generating a theme document matrix; 5, training a label feature matrix; 6, constructing aneural network; and 7, classifying the equipment fault maintenance record texts. According to the method, the LDA topic model is used for extracting lexical item document matrix features; obtaining label feature matrix, providing a classification label of the fault maintenance record text; equipment fault maintenance record texts can be quickly classified; in addition, a pooling layer in the convolutional neural network is replaced with a circulating layer, the convolutional layer in the convolutional neural network has the advantage of weight sharing, the circulating layer has the advantage of solving the problem of long-term dependency of texts, and sufficient knowledge data is provided for a fault maintenance auxiliary system in an intelligent IETM platform.

The invention relates to an integrated control system for power grid dispatching. Compared with the prior art, the system achieves the integration of the main and standby systems mainly through threelayers, and achieves the integration of resource layers through active-active storage. The system provides active-active access, zero data loss, and sharing and storage, achieves the unified management of various types of magnetic disk arrays, and achieves the resource pooling of different types of disk arrays at a data center. A network layer executes the high-performance and high-reliability data communication of main and standby equipment through a bare optical fiber and a wavelength division interconnection link. The system achieves the integration of a computer layer through a database, achieves the high usability, load balance and migration dispatching through a database module, guarantees the real-time synchronization of data between two stations, and achieves the business integrityin an operation process.

The invention discloses a data transmission method and device for transverse federation learning, terminal equipment and a computer readable storage medium. The method comprises the following steps: receiving a residual pooling model uploaded by each client in transverse federation learning, wherein each client determines a to-be-trained model in transverse federated learning according to an aggregation result received from the server and an initial to-be-trained model, performs local model training on the to-be-trained model by using local training data, calculates a residual error, and performs pooling compression on the residual error to obtain a residual error pooling model; performing preset aggregation operation on the received residual pooling model to obtain an aggregation result;and issuing the aggregation result to each client, so that each client performs a new round of model training according to the aggregation result. According to the invention, the data transmission quantity between the client and the server is reduced, and the problems of complex implementation process and loss of model precision caused by intermediate transmission through model cutting are avoided.

The embodiment of the invention relates to the technical field of cloud computing, and discloses a cloud game resource scheduling method, which is applied to a server and comprises the following stepsof: obtaining a target pooling region and a pooling number of a cloud game; applying instance resources with the same pooling number for the target pooling region, wherein the pooling instance resources are used as pooling resources of the target pooling region; installing and starting a cloud game on the pooling resources of the target pooling region; and in response to a cloud game request of auser terminal belonging to the target pooling region, allocating instance resources of the pooling resources to the user terminal. According to the cloud game resource scheduling method, the server and the storage medium provided by the embodiment of the invention, the waiting time of the user is shortened, and the game experience of the user is improved.

Embodiments select demand forecast parameters for a demand model for a first item. Embodiments receive historical sales data for a plurality of items on a per store basis and receive a plurality of seasonality curves for the first item of the plurality of items, each seasonality curve corresponding to a different pooling level for the first item. Embodiments determine a correlation for each of the seasonality curves at each pooling level and determine a root mean squared error (“RMSE”) for each determined correlation. Embodiments determine a score for each pooling level, the score based on the corresponding correlation, RMSE and a penalty and select one of the seasonality curves based on the determined scores. Embodiments use the demand model and the selected seasonality curve to determine a demand forecast for the first item, the demand forecast including a prediction of future sales data for the first item.

The invention discloses a vectorization implementation method for pooling of a multi-sample multi-channel convolutional neural network. The method comprises the following steps: step 1, storing inputfeature data set data of a convolutional neural network pooling layer according to a sample dimension priority mode; 2, dividing an input feature data set data matrix into a plurality of matrix blocksby the vector processor according to columns; 3, sequentially extracting matrix blocks with specified sizes by the vector processor according to rows, and transmitting the matrix blocks to a data buffer area of an array memory of the vector processor; 4, each core of the vector processor performs pooling vectorization calculation on the matrix blocks in the respective data buffer area in parallel, and calculation results are transmitted to an off-chip memory in sequence; and step 5, repeating the step 3 to the step 4 until all pooling layer calculation is completed. The method can give full play to the calculation performance of the vector processor, and has the advantages of simple implementation method, high implementation efficiency, low power consumption, good effect and the like.

The invention provides a three-dimensional point cloud processing method, device and equipment. The method comprises the following steps: acquiring point cloud data comprising a plurality of points; inputting the point cloud data into a pre-trained convolutional neural network; the convolutional neural network comprises a convolution module for geometric feature information sharing; for each pointin the point cloud data, obtaining a neighbor point of the point in the Euclidean space based on the convolution module, and determining a neighbor point of the point in the feature value space basedon the neighbor point of the point in the Euclidean space; aggregating the neighbor points of the points in the Euclidean space and the neighbor points of the points in the feature value space to obtain aggregated features; and performing feature learning on the aggregated features by using a multi-layer perceptron, and performing maximum pooling operation on the dimensions of the neighbor pointsto obtain output features. A convolution structure is directly constructed to process three-dimensional point cloud data, so that the problems of data missing and data redundancy are effectively relieved.

The invention discloses a convolutional neural network processor for edge calculation. A simple control instruction for the convolutional neural network is provided, convolutional neural network basicoperations such as a convolutional layer, a pooling layer, a ReLU activation function and a full connection layer can be realized, and the applicability of the accelerator for the convolutional neural network is realized through combination sorting of the instructions. The accelerator realizes a high-efficiency pulsation calculation array, so that the reusability of data can be increased to the maximum extent in a data reading process, and the access frequency of the pulsation calculation array to a data cache unit is reduced. The accelerator supports the calculation of two data precisions of16-bit fixed point number and 8-bit fixed point number at the same time, and can realize the calculation of the mixing precision of different layers of the same network, thereby greatly reducing thepower consumption of the accelerator. According to the accelerator, the minimum scheduling frequency of data in a cache part is guaranteed, power consumption is reduced. Meanwhile, convolution layer calculation and full connection layer calculation share the same pulsation calculation array, and the resource utilization rate is greatly increased.

The invention provides a method for predicting the service life of a bearing of a wind driven generator. Themethod includes the following steps: obtaining full life cycle vibration signal data of a bearing, and generating an original data set; performing normalization processing on the original data set; building an improved multi-scale neural network model; inputting normalized data into an input layer of the model, setting dilated convolution layers with different convolution kernel scales, and obtaining abstract features of input signals layer by layer; setting a global average pooling layer, and inputting the extracted abstract features into the global average pooling layer to obtain output features of the improved multi-scale 1DCNN model; inputting the output features of the global average pooling layer into an LSTM model, and extracting bearing performance degradation information implied in the output features through a multi-layer LSTM memory unit; and predicting the residual life of the bearing according to the extracted bearing performance degradation information to obtain a prediction result. According to the method, the residual life of the wind driven generator bearing can be efficiently and accurately predicted.

The invention discloses a neural network compression method, device and equipment and a computer readable storage medium, and the method comprises the steps: carrying out the pooling operation of a convolution kernel in a to-be-compressed neural network, and obtaining a pooling value corresponding to the convolution kernel; performing integer quantization operation on the pooling value to obtain an integer value; and cutting the convolution kernel with the corresponding integer value as the target value in the convolution kernels so as to compress the neural network to be compressed. Accordingto the invention, the burden of the equipment in operating the neural network is reduced, so that the edge equipment with limited computing power can also operate the neural network.

The embodiment of the invention relates to the technical field of cloud games, and discloses a cloud game installation method and device, electronic equipment and a storage medium. Game instances aredivided into pooling groups by creating the pooling groups and utilizing a multicast technology, and based on multicast message transmission characteristics, in a one-to-many scene, a server side onlyneeds to copy a game package body once. Therefore, the game package body can be sent to all instances in the multicast group, and then a game can be installed. When a user requests to start a cloud game, a server side can distribute a game instance to the user, so that the game requested by the user can be installed and operated in the game instance, a pooling group is created for the cloud gamein advance, the game instance in the pooling group is installed, the instance distribution process is shortened, and the game starting speed is increased. Therefore, the time for the user to wait forthe game to pull up is greatly reduced and the user game experience is improved.

A computer based on a spiking neural network, includes at least one maximum pooling layer. In response to an input spike received by a neuron of the maximum pooling layer, the device is configured so as to receive the address of the activated synapse. The device comprises an address comparator configured so as to compare the address of the activated synapse with a set of reference addresses. Each reference address is associated with a hardness value and with a pooling neuron. The device activates a neuron of the maximum pooling layer if the address of the activated synapse is equal to one of the reference addresses and the hardness value associated with this reference address has the highest value from among the hardness values associated with the other reference addresses of the set.