21results about How to "Ensure training efficiency" patented technology

The invention relates to a self-processing method based on perception of a real world, and the method comprises the steps: carrying out the symbolization of multi-modal perception data and corresponding processing result data, which are respectively collected, in the steps of optimization training and self-processing, so as to form a word vector for training and a word vector for execution, which have natural semantics; the word vectors for training describe multi-modal sensing data and processing result data corresponding to the collected sensing data, and the word vectors for execution describe the multi-modal sensing data; further, the perception data of the multiple different modes are converted into word vectors with natural semantics, and the information amount contained in the word vectors obtained through symbolization is approximately equal to the sum of the information amount contained in the perception data of the multiple different modes and the information amount contained in the corresponding processing result data. According to the method, the optimization training complexity of the special large language model is reduced, the training efficiency and the response efficiency are improved, and the training cost and the implementation cost are reduced.

The invention belongs to the technical field of lithium battery anomaly detection, and discloses a training method of a lithium battery anomaly sample detection model and a related device. The training method comprises the following steps: based on formation equipment in a lithium battery production process, collecting resource data and calculating to obtain the distribution of the resource data; based on the initially trained lithium battery abnormal sample detection model, obtaining training sample data and calculating to obtain the distribution of the training sample data; and calculating a difference value between the distribution of the resource data and the distribution of the training sample data, and if the difference value is greater than or equal to a preset threshold value, performing incremental training on the initially trained lithium battery abnormal sample detection model based on the resource data and the training sample data. According to the technical scheme of the invention, when the battery data generates domain offset, iterative updating of the model can be realized through incremental training, the generalization ability of the model is enhanced, and the reliability of a model detection result is ensured.

The invention is suitable for the technical field of computer application, and provides an image classification model construction method and device based on a random forest, electronic equipment and a storage medium, and the method comprises the steps: constructing an initial image classification model; generating a feature layer feature matrix through the feature forest layer; generating an enhancement layer feature matrix through the enhancement forest layer; inputting the feature layer feature matrix and the enhancement layer feature matrix into an output layer, and determining the prediction precision of the initial image classification model; when the prediction precision reaches a preset precision condition, determining the initial image classification model as a target image classification model; and otherwise, increasing the number of the first random forest groups and / or the second random forest groups so as to generate a target image classification model. Therefore, the image classification model is constructed by using the integrated structure of the multilayer random forest groups, the model precision is improved by increasing the number of the random forest groups, and the model training efficiency, the generalization ability and the image classification precision of the target image classification model are improved.

The present invention discloses an image super-resolution reconstruction method based on instance space feature modulation and feedback mechanism, which is used to solve the problem of reconstructing low-resolution images into high-resolution images. First, the instance space features of the low-resolution image are extracted by an instance space feature modulator, and the super-resolution reconstruction features of the low-resolution image are characterized by instance space features based on the extracted features. Then, the modulated super-resolution reconstruction features are input into a feedback reconstruction subnet including a projection module for cyclic reconstruction. After multiple reconstruction iterative optimizations, the final super-resolution reconstruction results are obtained. Introducing instance space features into super-resolution reconstruction features and using a recurrent neural network for multiple reconstruction iterative optimizations will not only better reconstruct low-resolution images, but also have significant effects on the texture information of image instances, greatly improving the effect of image reconstruction.

The invention relates to the field of intelligent traffic, deep learning and artificial intelligence, in particular to a deep learning method for urban traffic node mode similarity cognition. Traffic speed data is organized into a graph data format consisting of a feature matrix and an adjacency matrix. The processed speed data are sent into a dynamic multi-graph constructor, an output dynamic adjacency matrix and a space adjacency matrix are sent into a multi-scale graph convolution module together, and the speed data are sent into the multi-scale graph convolution module after being subjected to expansion causal convolution processing in each space-time layer; and splicing from each space-time layer through jump connection to obtain a final result. And finally, analyzing the traffic mode similarity reflected by the dynamic multi-graph. According to the method, the space-time dependent features in the intelligent traffic scene can be extracted more efficiently, and the problem that the perceptual view in the graph convolutional network is limited and the problem that the long-term traffic mode is not considered sufficiently in traffic mode similarity analysis are solved.

The application discloses a kind of reconfigurable intelligent surface auxiliary wireless environment modeling method based on three-dimensional Gaussian splash technology, belong to wireless communication field, this method will three-dimensional Gaussian splash technology from computer graphics field innovatively migrate to RIS auxiliary wireless communication system modeling field, the explicit parameterization thought of 3D-GS is applied to radio frequency electromagnetic field modeling, break through the limitation of traditional deep learning "black box" modeling;Adopt two-stage joint modeling framework: for the characteristics of RIS system, "TX→RIS" and "RIS→RX" cascade modeling strategy is designed, can accurately capture the electromagnetic regulation effect of RIS;With explicit physical parameterization: replace large-scale neural network weight with Gaussian parameter with clear physical meaning, realize light weight;In addition, by constructing complete differentiable rendering process, support efficient gradient descent optimization, ensure training efficiency and convergence.

The invention relates to a self-processing method based on data symbolization and a large model, which comprises the following steps: in the steps of performing model training and responding to input by using the large model to realize self-processing, respectively acquiring various types of sensing data acquired in the two steps and processing result data corresponding to the acquired sensing data acquired in the model training step, the method comprises the following steps of: symbolizing a plurality of different modes of sensing data, namely multi-mode sensing data, to form training data and real-time data, and further converting the plurality of different modes of sensing data into data in a symbolic form, and the information amount contained in the symbolized data is approximately equal to the sum of the information amounts contained in the plurality of different modes of sensing data. On the basis, the multi-modal sensing data can be optimized to a great extent, so that the training complexity of a large model is reduced, the training efficiency and the response efficiency are improved, and the training cost and the implementation cost are reduced.

This invention discloses an implicit reconstruction method for point cloud surfaces based on a piecewise learning strategy. The invention employs a piecewise surface representation and trains an implicit reconstruction network for point cloud surfaces targeting the local symbolic distance field of a 3D shape. The method includes the following steps: using discrete point cloud data as input, a farthest-point sampling strategy is used to generate initial patches; the offset of each sampling point within each patch relative to the patch center is calculated as the relative position of the sampling point. The latent features of the patch are obtained in the neural network encoder, and the symbolic distance values ​​of the relative positions of each sampling point are obtained in the neural network decoder. The relative positions of sampling points located in overlapping areas of different patches are weighted and summed to obtain their corresponding symbolic distance values. A mesh model of the 3D shape is obtained using the Marching Cube algorithm. This invention can reconstruct the overall shape of an object while preserving the fine details of the original shape, and it is robust to point cloud normals and noise.

This invention proposes a structure-condition-driven low-resource multilingual multimodal model training method and system, belonging to the field of artificial intelligence. The method includes: S1: Extracting image structural attributes from multimodal training samples, constructing structure-conditional text sequences, and obtaining text hidden states and visual features; S2: A route selector determines the training path based on the resource and structural attributes of the samples; S3: Freezing the multimodal large-scale model backbone network and dynamically activating a lightweight script adaptation module to update parameters; S4: Aligning visual and textual representations through a structure-conditional cross-attention module to generate a joint representation; S5: Embedding the joint representation, task prefix, target language identifier, and its language into the input decoding and multi-task output modules to obtain the generated result; S6: Constructing a total training objective function to obtain the trained multimodal model. This invention improves the overall performance, stability, and security of multimodal models in low-resource, multilingual, and complex structural scenarios.

One or more embodiments of the invention provide a privacy-protecting machine learning model training method and device. The method comprises the steps of inputting an original sample into a bottom model to obtain a first type of intermediate features; inputting the first type of intermediate features into a feature protection network to transform the first type of intermediate features to obtain transformed features; inputting the transformation features into a trained reverse model so as to restore the transformation features through the reverse model to obtain a reverse sample output by the reverse model; calculating sample loss caused by intermediate feature transformation based on the reverse sample and the original sample, and updating the feature protection network by taking increase of the sample loss as a target; the transformation features are sent to a server side, so that the server side inputs the transformation features into the top model to obtain a prediction label of the original sample; a bottom gradient sent by the server side is received, the bottom model is updated based on the bottom gradient, and the bottom gradient is obtained by the server side through calculation based on the prediction label and the sample label of the original sample.

The invention belongs to the technical field of learning reading, and particularly relates to a training method and device for improving reading ability, a terminal and a storage medium. The training method for improving the reading ability comprises the following steps: S1, acquiring and judging identity information of a user so as to enter a training mode; s2, displaying warm-up training elements, and adjusting display styles of the warm-up training elements to enter visual enhancement training; s3, displaying at least two groups of recognition training elements in different forms for the user to perform potential judgment training; and S4, displaying according to preset formal training content elements of the training mode, so that the user can carry out speed reading training. According to the method and the system, three-step level-by-level reinforced training of warm-up training of vision improvement training, basic training of potential judgment training and reinforced training of speed reading training can be realized, so that the reading ability is improved, and the training efficiency is guaranteed.

This application provides an unsupervised detection method and system for electrical appliance operation anomalies in a virtual power plant, relating to the field of equipment operation status monitoring. The method includes: collecting and preprocessing operating power sequences and timestamp sequences, then concatenating them to form a feature vector set. A single-classification model is constructed, mapping samples to the feature space using a Gaussian radial basis function kernel, and training to obtain a parameterized hypersphere model describing the distribution of normal data. The Lagrange duality problem of the single-classification model is solved to obtain the hypersphere's center vector, radius parameter, and Lagrange multiplier subset. Real-time acquisition of the target electrical appliance's operating data generates a real-time feature vector, and the determination distance between the real-time feature vector and the hypersphere's center is calculated. If the target electrical appliance is determined to be in an abnormal state, the abnormal device is removed from the virtual power plant's resource scheduling pool, and a health warning message is sent. This addresses the technical problems of existing technologies that rely on scarce labeled abnormal data, have high training costs, and poor generalization ability.

The application provides an I / O feature prediction method, a model training method and related devices. The job submission script of a job is converted into a grayscale image, and a deep learning method of image recognition is used to realize the prediction task of the I / O feature. When the I / O feature is predicted, the processing method accumulated in the field of deep learning can be reused, including various classical image recognition network structures, to process the above-mentioned grayscale image to obtain the I / O feature of the job. Compared with the existing method of using a "job name-user name-computing node number" triplet as the I / O feature and using a clustering algorithm to process a large-scale data set for prediction, the I / O feature can be predicted based on the job submission script, the time consumption of the prediction of the I / O feature is reduced, and the I / O feature of the job in the high-performance computing cluster is accurately and efficiently predicted.

The invention discloses a double-layer adaptive training recommendation system and method based on standardized expression increment and a dynamic baseline, and relates to the technical field of artificial intelligence technology, education technology and acknowledgement science, the system has a double-state switching mechanism, provides a clear initialization path for a new user, and ensures the user experience from 0 to N; a standard performance score SPS is introduced into the baseline storage module, and the defect that cross-game types are incomparable is overcome through mean value and variance calibration; the game fatigue GFS model can prevent children from being tired due to repetition; the method comprises the steps of determining an initialization state or a dynamic self-adaptive state by detecting the total number of games of a user; a training load is set, a training course package is generated by using an SPS sequence and NPD data of recent k games, calculating SPSlevel and SPStrend and determining the training load N, finally, difficulty regulation and control in the games are achieved based on DGB, instant excitation of skipping is provided, fatigue switching can be conducted, and training enthusiasm is guaranteed.

The invention discloses a model training video memory optimization method and device, and the method comprises the steps: carrying out the training of a round based on a preset network model, and recording the input time, output time, video memory occupation amount and calculation time of each tensor; according to the transmission speed and the video memory occupation amount, calculating the transfer time overhead of each tensor transferred from the GPU video memory to a main memory; according to the input and output moments of each tensor, obtaining a list capable of hiding the transfer overhead; according to the list capable of hiding the transfer overhead, establishing a large-top heap according to the video memory occupancy of the tensors, and sequentially selecting the tensors at the top of the heap as swap-in and swap-out alternative tensors; when swap-in and swap-out cannot meet video memory requirements under the condition of hiding transfer overhead, video memory occupation is further reduced in combination with a re-calculation means; and carrying out optimization processing according to the video memory optimization strategy, so that the large-scale network model is trained on the target equipment. According to the invention, the model training speed can be improved while the video memory occupation is effectively reduced.

The invention discloses a seasonal trend decomposition-based short-term photovoltaic power generation power interval combined prediction method, and belongs to the field of solar power generation. The problems that in the prior art, prediction accuracy is low, and balance between precision and coverage rate is difficult to achieve are solved. According to the technical scheme, the method comprises the following steps: decomposing power generation power data into a season term, a trend term and a remainder term by adopting season trend decomposition; keeping the seasonal item unchanged, and performing addition operation in a final point prediction result; a trend term is predicted by adopting an ARIMA model; the rest items are predicted through a BiTCN-BiGRU-Attention (BiTCN-BiGRU-Attention) model; adding the obtained prediction results to obtain a point prediction result; and obtaining a final interval prediction result through the interval combination measurement model. According to the scheme of the invention, high prediction accuracy can be realized, and meanwhile, certain training efficiency can be ensured.

The invention discloses a self-organizing network-oriented decentralized device-level federated forgetting task scheduling method, which comprises the following steps of: performing decentralized device-level federated forgetting processing on the basis of task scheduling information through a device-level federated forgetting task scheduling model to obtain a task scheduling scheme, and realizing task scheduling of a self-organizing network through the task scheduling scheme; the equipment-level federal forgetting task scheduling model comprises the following steps: partitioning clients in a self-organizing network by adopting a heuristic client dynamic partitioning strategy based on a DDFU-SON algorithm; and dynamically evaluating the data heterogeneity of the self-organizing network to perform dynamic aggregation so as to realize task scheduling. The invention discloses a self-organizing network-oriented decentralized device-level federation forgetting task scheduling method and system, a storage medium, an edge device forgetting method focused on the dynamic access condition of a client, a partition selection and topology construction strategy of a newly accessed client, and a training stability guarantee mechanism.

The embodiment of the invention discloses a detection method, device and equipment and a computer readable storage medium. The method comprises the following steps: acquiring index weight information of each entity; calculating distance measurement information between the index data sequences of every two entities in a specified time period based on the index weight information of each entity to obtain multiple pieces of distance measurement information; determining a reference entity from the plurality of entities based on the plurality of pieces of distance measurement information, and performing phase alignment on an index data sequence of each entity in a specified time period and an index data sequence of the reference entity in the specified time period based on the index weight information of each entity to obtain an index data sequence after phase alignment; and performing model training based on the index data sequence after phase alignment, and detecting an entity to be detected based on the trained model. Through the method, the detection cost can be reduced.

The present application relates to a kind of main hoist remaining life prediction method and device, wherein, method includes: S1, the vibration time series data and working condition parameter time series data of the main hoist to be measured are acquired, and the measured vibration signal and measured working condition parameter signal of single operation are segmented, detection data group is constructed and time information is generated;S2, using detection data group to identify working condition, input the first self-encoding model corresponding to working condition, obtain the vibration signal after reconstruction and physical constraint residual, and the model is based on physical constraint dynamics equation to build;S3, total deviation is calculated according to reconstructed signal and physical constraint residual, and deviation sequence is obtained according to time sequence;S4, failure curve is generated based on deviation sequence, and then the remaining life is obtained.The present application can effectively separate working condition influence, fusion physical priori knowledge, realize accurate and reliable life prediction under the condition of variable working condition and variable load.

The present application provides a method and device for synthesizing speech for a specific speaker, which encodes a phoneme sequence according to the position of a target type text in an annotated text to obtain a position-encoded text with positional features, and uses the phoneme sequence and the position-encoded text to train a speech synthesis model. Correspondingly, the speech synthesis model designed by the present application splices the vector representation of the phoneme sequence and the vector representation of the position-encoded text through a splicing module and inputs them into an encoder, so that the encoder can learn the rhythmic latent features and the positional latent features of the target type text in parallel without increasing the complexity and input them into a variance adapter. The variance adapter strengthens the fusion of the speech features and the above features, and while ensuring that the model complexity is small, the trained model synthesizes the speech of a specific speaker with not only a sense of frustration, but also a sense of emphasis on the position of the target type text.

The embodiment of the invention discloses an image processing model training method and device, electronic equipment and a computer readable storage medium, and the method comprises the steps: inputting an image into an image processing model, and obtaining a feature value of the image; performing smooth quantization on the feature value to obtain a hash-like code of the image; determining a loss function based on the Hash-like code and the feature value of the image; deriving the loss function to obtain a gradient value; when the sign of the characteristic value is inconsistent with the sign of the gradient value, updating the gradient value; and training the image processing model based on the updated gradient value to obtain a trained image processing model. According to the scheme, the problem of dead bits in the model training process can be solved.