38 results about "Neural network learning" patented technology

This invention discloses a method for predicting and protecting the overheating risk of asynchronous motors, specifically relating to the field of motor control and protection technology. Based on an intelligent fusion model, it estimates temperature and thermal stress field, calculates the rate of change of thermal stress, non-uniformity, and hotspot trends, and calculates a dynamic health index using historical data. These parameters are then input into a multi-objective reinforcement learning controller to optimize long-term health and short-term performance, generating a thermal shaping control vector to regulate the motor. This invention combines physical mechanisms with data-driven approaches through an intelligent fusion model, utilizing a graph neural network to learn the structure of the heat conduction graph and verify physical laws, thereby improving the accuracy of thermal state estimation. It achieves a multi-dimensional risk characterization combining transient impact and cumulative effects through the rate of change of thermal stress, non-uniformity, hotspot trends, and dynamic health index. By optimizing long-term health and short-term performance losses, a thermal shaping control vector is generated to achieve regulation from passive protection to active prevention, extending the motor's service life.

The application provides a heterogeneous graph representation learning method based on attribute completion and category balance integration, comprising: representing minority class nodes in a heterogeneous graph by using a high-order enhanced category balance mechanism to obtain a category-balanced heterogeneous graph; based on a self-supervised heterogeneous attribute completion mechanism, introducing a mask autoencoder to learn more optimal attribute embedding for attribute observable nodes, and introducing an attention mechanism to complete attributes for attribute missing nodes; inputting the completed heterogeneous graph into a heterogeneous graph neural network to learn the final node embedding and perform a node classification task or a node clustering task. The method overcomes the defect of insufficient processing of cross-type category imbalance in the heterogeneous graph, solves the representation deviation and classification accuracy decline problem caused by cross-type imbalance propagation effect, overcomes the defect of ignoring the attribute missing problem in the category imbalance heterogeneous graph, breaks the coupling negative influence of attribute missing and category imbalance, and improves the representation quality of minority class nodes.

This invention discloses a grayscale image prediction method based on a lightweight neural network, comprising: 1) training an LNN network; and 2) grayscale image prediction. This method uses a neural network to learn and model the complex relationships between image pixels, achieving high-precision image prediction.

The application discloses a kind of silicon steel sheet visual quality detection method and experimental system based on microservice, including the following steps;Step 1, build remote service management platform based on microservice, meet the user's custom storage needs through database interaction API, namely design corresponding data storage structure and convenient and fast data interaction API;Step 2, offline acquisition industrial area camera's checkerboard image and defect silicon steel sheet sample image, calculate the row and column scale factor between pixel and actual size;Step 3, offline training model and deployment model;Using the method of deep neural network learning, build PyTorch deep learning framework, use YOLOV7 model to carry out visual quality detection;Step 4, online silicon steel sheet visual quality detection experiment;Step 5, visual quality detection result display.The application carries out visual quality detection experiment by machine vision software, with high degree of automation, wide measurement range, fast, high precision, the characteristics of accurate detection.

The application provides a vehicle fault diagnosis method and device based on multi-source data fusion and a storage medium. The method comprises: collecting real-time state data and real-time alarm data of a target part, obtaining external environment data and map navigation data; after data preprocessing, the multi-source data set target features are obtained by fusion, the target features extracted from the multi-source data set are input into the adaptive neural fuzzy inference system, the target features are mapped to multiple fuzzy sets, and the member function is used for fuzzification; the rule weight of each target feature is calculated to reflect the fuzzy strength; the neural network learning algorithm is used to train the inference system, the member function and the rule weight are adjusted, the fuzzy output is converted into a specific numerical value or a fault level in the defuzzification process by using the output member function, so that the health state of the part is judged. The application provides a fault diagnosis output with high interpretability, improves the accuracy, and can perform fault diagnosis in a wide range of vehicle operating environments.

The present disclosure relates to deep learning based channel buffer compression, and in particular, a method and system of neural network learning with compression is provided. The method includes performing channel estimation on a reference signal (RS), compressing the channel estimation of the RS with a neural network, decompressing the compressed channel estimation with a neural network, and interpolating the decompressed channel estimation.

This invention discloses a low-light image enhancement method based on frequency domain decoupling and implicit neural representation, comprising: acquiring an input low-light image. I ; to image I Input the bi-branch multi-scale denoising module to obtain a clean, denoised image. I Den Perform a Fourier transform on the denoised image to decompose it into amplitude components. A and phase components Φ The pixel coordinates are positionally encoded and jointly modeled with amplitude features. The resulting residual is then input into an implicit neural network to learn the amplitude residual, thus obtaining the corrected amplitude. A′ Phase components Φ The enhanced phase component is obtained by using a feature enhancement network and an attention mechanism to enhance edge and texture details in the phase. Φ′ Finally, the corrected amplitude A′ and enhanced phase Φ′ The enhanced image is obtained through inverse Fourier transform. I Fixed This invention solves the problems of noise amplification, severe coupling between illumination and structure, insufficient utilization of frequency domain information, and spatial discontinuity caused by discrete modeling in existing methods under low light conditions.

This invention provides a 3D reconstruction method based on structured Gaussian representation and neural kernel fields. Taking multi-view images, camera intrinsic and extrinsic parameters, and point clouds as input, it first constructs and optimizes a structured Gaussian representation, then constructs a neural kernel field for the scene, where sparse voxel meshes and anchor points serve as the basic spatial structure of the kernel function. Features related to the kernel function are learned and predicted through a neural network. This invention introduces a mutual guidance mechanism between the structured Gaussian representation and the neural kernel field to collaboratively optimize the two representations. Finally, the structured Gaussian representation is used for efficient view rendering, outputting a high-precision 3D surface model. This method leverages the compactness and efficiency of structured Gaussian representation in handling large-scale scenes, combined with the scalability and robustness of neural kernel fields in implicit surface learning, to achieve high-quality rendering and reconstruction output, suitable for understanding and applying complex and large-scale 3D scenes.

To improve the confidentiality of data to be anonymized without increasing the amount of data in a neural network model in anonymizing data using a neural network.SOLUTION: Encryption means is provided that generates encrypted data by encrypting data to be encrypted using encryption key data using a neural network trained in advance.SELECTED DRAWING: Figure 1

This invention belongs to the field of rotor optimization technology and discloses a rotor aerodynamic shape optimization method and system based on airfoil geometric feasibility constraints. It transforms the high-dimensional, complex three-dimensional rotor shape into low-dimensional mathematical vectors using CST and SVD parameterization methods; constructs a large-scale aerodynamic database using the low-fidelity but extremely fast XROTOR tool; utilizes a deep neural network MLP to learn the nonlinear mapping relationship between design variables and aerodynamic performance, constructing a high-precision surrogate model to replace expensive CFD simulations; and finally introduces airfoil geometric feasibility constraints based on modal space Euclidean distance into the SLSQP gradient optimization algorithm, rapidly searching for the global optimal solution within the surrogate model space while ensuring the geometric rationality and manufacturability of the optimization results.

This paper discloses a system and method for calibrating broadcast video sources. The computational system retrieves multiple broadcast video sources comprising multiple video frames. The computational system generates a trained neural network by generating multiple training datasets based on the broadcast video sources and learning to generate a homography matrix for each of the multiple frames through a neural network. The computational system receives a target broadcast video source containing a target motion event. The computational system divides the target broadcast video source into multiple target frames. The computational system generates a target homography matrix for each of the multiple target frames via a neural network. The computational system calibrates the target broadcast video source by warping each target frame with the corresponding target homography matrix.

The present invention relates to a method and system for learning a spiking neural network based on transformation recognition learning. According to the present invention, the method comprises an ANN generation step of generating an analog artificial neural network (ANN) model and inputting variable data, a transformation recognition learning step of simulating a spiking neural network (SNN) model using one or more activation functions in the ANN model, and an SNN generation step of generating an SNN model by correcting the parameters and weights of a layer based on the simulation results.

This invention provides a method and apparatus for predicting protein binding sites based on dynamic masked graph convolution, belonging to the field of neural network learning technology. The method includes generating an initial embedding representation of residues based on amino acid sequences; inputting the initial embedding representation of residues into a pre-trained embedding enhancement module to generate enhanced embedding representations of residues; constructing a residue graph based on the three-dimensional structural information of the protein to be predicted, and using the enhanced embedding representations of residues as initial features of each residue node in the residue graph; inputting the residue graph into a trained protein binding site prediction model to obtain final features after multiple iterations; and generating the predicted probability of each residue being a binding site based on the final features. This invention optimizes the information propagation efficiency of the graph structure through a multi-round dynamic masked graph convolution mechanism, thereby improving the overall accuracy and robustness of binding site prediction.

The embodiment discloses an unmanned ship encircling surrounding layered game control method under unreliable communication, considers that target information is not completely known and unreliable communication conditions of intermittent communication and false data injection (FDI) exist simultaneously, adopts a prescribed time flexible Nash equilibrium target state estimator converging under intermittent communication, ensures that the target estimation pose can converge to a saddle point according to the prescribed time and performance under intermittent communication, and based on this, a prescribed performance guidance law under the prescribed time is designed, longitudinal reference speed of the unmanned ship and bow reference speed of the unmanned ship are acquired, a cost function is established, then Hamilton-Jacobi-Isaacs equation is established, ideal optimal dynamic control law and optimal false data are acquired, and a critic neural network learning structure is used for approximation, and ideal optimal dynamic control law after approximation and optimal false data after approximation are acquired.

This application applies to the field of integrated circuit technology, providing a solid-state storage I / O scheduling system, method, solid-state drive (SSD), and electronic device. The system includes an SSD and a processor. The SSD integrates a main control chip and a neural network processing unit, which are vertically interconnected. The processor is installed in the electronic device, and the SSD is connected to or installed inside the electronic device. The processor is configured to: extract feature data of read / write requests when generating read / write requests for the SSD; the neural network processing unit is configured to: read the feature data; read the memory state of the SSD from the main control chip; run a neural network learning model based on the memory state and feature data to obtain scheduling information for read / write requests; and the processor is configured to: read the scheduling information and schedule read / write requests based on the scheduling information. This system can improve the I / O performance of the SSD.

The subject of the invention is a method for developing a computer tool that utilises learning in a feedforward neural network FNN for the remote diagnosis of chemical element content in plants. The method includes collecting a training dataset, consisting of hyperspectral data from cultivated fields and measurement data on element content in plant leaves. The hyperspectral data is normalised and converted into input vectors, after which the dataset is divided into a training set and a test set. The training of the FNN involves an input layer with 64 neurons, a hidden layer with 32 neurons, and an output layer with a single neuron performing a regression function. The network is trained to match the predicted element content values to actual measurement data. The trained network is validated using the test dataset to ensure diagnostic accuracy. The method is applicable in precision agriculture for analysing the chemical element content in plants. The invention also concerns a computer-implemented method for the remote diagnosis of chemical element content in plants using neural network learning.

A weight transfer apparatus for a neuromorphic device includes a memory storing a weight transfer program for the neuromorphic device, and a processor configured to execute the weight transfer program. The weight transfer program builds an artificial neural network learning model, transfers a weight of the built artificial neural network learning model to the neuromorphic device, determines whether a synaptic cell included in the neuromorphic device to which the weight is transferred is defective, rebuilds the artificial neural network learning model after the artificial neural network learning model sets a weight corresponding to a defective synaptic cell to 0, and transfers a weight of the rebuilt artificial neural network learning model to the neuromorphic device.

This invention discloses a deep learning-based power allocation method for downlink in a cellular-free system, comprising: S1, S2 constructing a fully connected neural network structure model, transforming the problem between the power allocation variable set and the solution of traditional algorithms into a problem between the weights W and bias term b of the trained neural network, with the aim of minimizing the loss function; S3. This invention uses M-MMSE precoding to allocate power in the downlink of a large-scale MIMO network. First, a feedforward neural network with fully connected layers, consisting of an input layer, hidden layers, and an output layer, is employed to generate an estimate of the optimal power allocation vector. The neural network learns the estimates from all users to satisfy power constraints and improve estimation accuracy, significantly increasing convergence speed and reducing computational complexity.

This invention discloses a block-adaptive gradient optimization method, system, and storage medium for large model pre-training, relating to the field of neural network learning technology. The method includes: dividing the trainable parameters of the neural network to be trained into multiple parameter blocks according to layer type and functional role; acquiring the current gradient of each parameter block during training, and fusing the squared element of the current gradient with the existing second-order moment statistics of the corresponding parameter block using an exponential moving average to maintain the block-level second-order moment statistics; calculating the block-adaptive step size based on the block-level second-order moment statistics, and performing parameter updates according to the block-adaptive step size; and recalculating the block-adaptive step size based on the updated block-level second-order moment statistics in subsequent training steps. Through the technical solution of this invention, different parameter blocks can obtain update step sizes adapted to their gradient change characteristics, reducing the oscillation risk of high-curvature parameter blocks and improving the convergence stability and training efficiency of large model pre-training.

The invention discloses a spinosad culture medium optimization method and device based on machine learning, and belongs to the technical field of microbial fermentation. According to the method, deep learning and Bayesian optimization technologies are combined to optimize the ratio of the spinosad fermentation culture medium, firstly, the characteristics of the culture medium are learned through a deep neural network, and then model parameters are adjusted by using a Bayesian optimization algorithm, so that intelligent control and optimization of the ratio of the microbial culture medium are realized; the microorganism production efficiency and the product quality are improved. According to the method, technical support is provided for the fields of medicine production, food processing, environmental protection and the like through intelligent analysis and accurate prediction, and the method has remarkable market prospects and application value.

A PID parameter tuning method, in response to the fulfillment of a preset trigger condition, employs a spiking neural network (SNN) to learn the error and adjust the control parameters of the PID algorithm to achieve a preset accuracy. Specifically, when the preset trigger condition is met, the SNN learns the error generated by the PID algorithm and adjusts the control parameters accordingly. This ensures the PID algorithm meets the preset accuracy, thereby improving its precision. Furthermore, the SNN is lightweight and can be stored in a microcontroller, allowing for online parameter tuning, improving the intelligence of the control system and reducing parameter tuning costs.

This invention discloses an intelligent management system and method for parking coupons based on multi-source data analysis in the field of smart parking. To address the problems of incorrect grouping and distribution due to miscorrelation of multi-source data, resource waste caused by inefficient distribution, and difficulty in controlling entry pressure during peak hours, this invention cleanses, standardizes, and aligns multi-source data such as parking entry / exit data, vehicle identification, payment orders, coupon redemption and verification, user accounts, and parking lot operation status. It constructs a weighted heterogeneous graph and learns user representations based on a heterogeneous graph neural network with edge weights. User grouping is performed under both mandatory and non-mandatory grouping constraints. Distribution instruction sets are generated through two-level distribution optimization at the group and user levels, achieving the technical effects of improving the refinement of grouping and distribution, increasing budget utilization, and suppressing operational pressure during high-load periods.

Method for segmenting objects (1) with their own motion using an ego-vehicle (2) for execution in a neural network, wherein the ego-vehicle (2) comprises at least the following components: - at least one computer (3) with a processor for processing digital data and with a data storage for holding digital data for the processor; and - at least one measuring sensor (4) for sensing the current vehicle environment (5) of the ego-vehicle (2), wherein the method comprises at least the following continuously repeating steps: a. using the measuring sensor (4), generating a plurality of measurement data (6, 7) by sensing objects (1) in a current vehicle environment (5) of the ego-vehicle (2) and providing the measurement data (6, 7) from the sensed vehicle environment (5) to the computer (3); and b.By means of the processor, to generate segmented motion data (8), the current measurement data (6) are enriched with a number of measurement data (7) generated in previous sequences and position-corrected to the current measurement data (6) in a position group (9), wherein the number of sequences can be adjusted as required for measurement data that are known to be highly noisy, wherein the number is increased for measurement data that are known to be highly noisy compared to less noisy measurement data, wherein the following are used for enrichment: - previous measurement data (7) which are used for the current task in step b.relevant, and- previous measurement data (7) which are learned as relevant by the neural network, wherein the current measurement data (6) and a number of previous measurement data (7) are superimposed by adding a presumed and / or measured change in position of the ego vehicle (2) to the previous data and / or subtracting it from the current measurement data (6), wherein only those measurement data (7) of the previous sequences are used for enrichment which have the identical and / or same speed and / or, in the case of a radar sensor, the same radar cross-section.

Monitoring self-discharge in operating battery cells (i.e., actively charging or discharging) is disclosed. Cell voltage data over time is used to detect cell-to-cell charge voltage imbalance in a battery while battery cells are at their peak charge voltage. Resistors across each cell are switched on or off according to the average charge voltage imbalance. Actual resistor on-times are used as input to train a neural network in real time. The neural network learns the resistor duty cycle needed by each battery cell to precisely balance its self-discharge losses. If the self-discharge changes, the neural network retrains itself as the new data arrives. The neural network may include a Long Short-Term Memory (LSTM) network for each cell in the battery, which can learn the relative self-discharge rate of the respective cell.

A method, computer system, and computer program product for enhancing feature engineering based on unsupervised learning of associated external knowledge embedding are provided. The embodiment may include receiving, by a processor, input data as a table and a name of a column. The embodiment may also include analyzing the column to identify multisets of concepts or sequences of concepts. The embodiment may further include automatically expanding the column by linking the identified multisets or the sequences of the concepts with corresponding concepts in an external knowledge graph. The embodiment may also include training a neural network to learn embedding vectors of concept multi-sets in the expanded column of the tables, wherein the training is unsupervised without provision of labels of data when the neural network learns an embedding of the multisets of concepts with an objective to minimize a reconstruction error of the identified multisets of concepts.

A method for sintering dimension compensation in binder jet additive manufacturing is disclosed. This method constructs a thermoelastic-viscoplastic constitutive model of the green body sintering process using binder jet additive manufacturing, employs the finite element method (FEM) for numerical simulation, and predicts deformation data. The deformed node coordinates output from the FEM simulation are extracted to form a deformed coordinate point cloud. Simultaneously, the original design model is discretized into an original design coordinate point cloud, achieving matching between the original and deformed coordinate point clouds. A GA-BP neural network is used to learn the nonlinear mapping relationship between the original design coordinate point cloud and the compensation model coordinate point cloud, generating a compensated geometric model. Three indicators—curvature change energy, Fourier high-frequency energy ratio, and straightness linearity error—are introduced to evaluate the contour quality of the compensation model. The compensation model is then used for binder jet additive manufacturing printing, and the green body undergoes curing and sintering to obtain the finished part. This invention improves the manufacturing precision of parts and belongs to the field of metal 3D printing technology.

A computer based method for an electronic marketing campaign from a customer to a contact receives a campaign having an email message body and a historic profile of a previous campaign by the customer. The email message body includes text and image data. The email message body is preprocessed based upon the campaign and the historic profile to produce campaign training data. A neural network learning model is trained with the campaign training data. The neural network provides a subject line recommendation inference, and named entity recognition is performed on the subject line recommendation.

In general, systems and methods are provided for training a neural network based on sample screen recordings wherein the neural network learns to minimize a reconstruction error between an original image input to the VAE, and an image reconstructed by the VAE. The systems and methods can provide determining and displaying significant points in time of the frame based on a similarity between compressed frames of a screen recording based on the trained VAE.