135 results about "Training phase" patented technology

This invention discloses a real-time arrhythmia detection method based on morphological fidelity consistency constraints, comprising the following steps: S1, preprocessing continuous electrocardiogram (ECG) signals and dividing them into overlapping sliding time windows; S2, inputting each time window into an ECG signal encoder obtained through joint training to obtain a latent representation vector for that time window; the joint training is specifically defined as: simultaneously optimizing the encoder parameters using supervised classification signals, contrast consistency signals, and morphological fidelity signals during the training phase; S3, inputting the latent representation vector into a classifier and outputting the class probability of the time window belonging to each arrhythmia category; S4, based on the class probability, outputting real-time detection results using an online decision strategy of threshold hysteresis and cross-window consistency. This invention can simultaneously achieve robust identification under dynamic interference, accurate preservation of clinically critical waveform details, and efficient real-time deployment at the edge. This invention also provides a real-time arrhythmia detection system and device based on morphological fidelity consistency constraints.

This invention discloses a low-light image quality enhancement system and method based on enhanced night scene modeling. The system includes: a sample data construction module, which generates enhanced low-light scene images based on an initial image dataset using a dual-channel hybrid generative network architecture and constructs an optimized training sample library; and a hybrid enhancement model, constructed based on an attention module and a probabilistic graphical model module. During the training phase, the hybrid enhancement model uses the optimized training sample library as input and employs a region-adaptive weighted loss function for parameter optimization. During the application phase, the optimized model receives the low-light image to be processed, estimates the illumination and noise distribution of the low-light image, and generates the enhanced image. This invention significantly improves the coverage of night scene training data and achieves adaptive enhancement of image features under low-light conditions, significantly improving the noise suppression effect in dark areas and the ability to restore image details.

A kind of detection method and device of fake APP, terminal, the method includes: determining training data;Keyword extraction is carried out to the brief introduction of each pair of target APP and the brief introduction of suspected APP, and the data vector of each keyword is determined;Determine the weight value of each keyword, and the data vector of each keyword in the target APP is weighted based on the weight value Operation is carried out to obtain target APP data vector, and obtain suspected APP data vector;The vector comparison result of target APP data vector and suspected APP data vector is calculated, the absolute value of the vector comparison result is determined as first data vector, and the square of the vector comparison result is determined as second data vector;Obtain the training text vector of the pair of target APP and suspected APP;Training fake detection model.The present application can enhance the association between keywords in model training phase, reduce the misjudgment rate.

This application provides a model training method, a speech processing method, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product. The method includes: acquiring a first phoneme sequence sample, a first word sequence sample, and a first alignment relationship between the first phoneme sequence sample and the first word sequence sample; in a first training phase, training an initial prediction model based on the first phoneme sequence sample, the first word sequence sample, the first alignment relationship, the first alignment window, the first phoneme loss weight, and a joint loss function of the initial prediction model; determining a state evaluation index for the initial prediction model; and triggering entry into a second training phase when the state evaluation index meets the phase switching conditions. This application can improve the accuracy of the model's prediction of alignment relationships.

A method and system for training an artificial neural network (ANN) is disclosed, specifically addressing the technical problem of training instability caused by randomly initialized weights. The invention introduces a novel, multi-phase training protocol executed by a processor. During an initial stabilization phase, the processor modifies the backpropagation of a weighted loss function by deprioritizing training on data from classes for which the ANN has a confidence level below a predetermined low threshold (e.g., less than or equal to 25 percent). This initial, focused training prevents erratic gradient updates to the weights of the network's layers, thereby forming stable and rigorous foundational neural pathways. Subsequent to the stabilization phase, a dynamic training phase commences where the network trains on the full dataset, with the training focus adjusted in real-time to reinforce the stable pathways. This protocol results in faster training convergence, a more robust final model, and improved accuracy, particularly for complex datasets, representing a significant technical improvement to the functioning of machine learning systems.

This invention discloses an open-set radar interference pattern recognition method, apparatus, and electronic device. The method includes: matched filtering of the radar received signal to construct a dual-modal input of a one-dimensional range sequence and a two-dimensional time-frequency matrix; feature extraction via a dual-branch network and adaptive fusion based on cosine similarity; introducing cross-modal consistency regularization during the training phase, constructing pseudo-unknown samples by shuffling intra-batch modes, and combining energy constraint loss to compress the energy of known classes and increase the energy of unknown classes to form a clear decision boundary; and adaptively setting a threshold based on the energy distribution of the validation set during the inference phase to achieve accurate identification of known interference and effective rejection of unknown interference. This invention overcomes the performance degradation defects of traditional closed-set methods in the face of unknown interference, and improves the generalization ability, robustness, and engineering practicality of radar interference recognition in complex electromagnetic environments.

This invention provides an adaptive quantization dual-teacher distillation federated learning method for heterogeneous environments, belonging to the fields of adaptive quantization, model compression, knowledge distillation, and federated learning. Its technical solution includes the following steps: S1, system initialization phase; S2, user extraction and model broadcasting phase; S3, local model training phase; S4, adaptive quantization phase; S5, global model update phase; S6, termination judgment phase, when the training rounds reach their maximum value. T If the condition is met, the federated learning task is terminated; otherwise, proceed to S2. This invention is applicable to heterogeneous edge device scenarios with limited computing resources and communication bandwidth, such as smart IoT devices, enabling high-precision, low-communication personalized federated learning while ensuring privacy and security.

This invention provides a 2D convolutional spatiotemporally stimulated dynamic gesture recognition method based on contrastive learning enhancement, comprising: firstly, sampling and preprocessing a dynamic gesture video sequence to obtain multiple frames corresponding to the same gesture action; then, using a shared-parameter 2D convolutional neural network to extract spatial features frame by frame to obtain frame-level feature representations. Based on this, spatiotemporally stimulated enhancement of the frame-level features is performed through multi-scale temporal difference modeling, temporally adaptive weighted aggregation, and spatial attention stimulation to explicitly characterize the dynamic changes of the gesture action. The enhanced features within the same gesture sequence are constructed into a frame-level positive sample set, and a robust positive sample set center vector is calculated. Using this center vector as a positive anchor point, a contrastive learning mechanism based on hard negative sample weighting is introduced to constrain and optimize the feature space. During the training phase, the model parameters are optimized by jointly using classification loss and contrastive learning loss.

The application belongs to the technical field of data center network and distributed computing system optimization, and discloses a physical perception dynamic topology reconstruction method for large model hybrid parallel training, comprising the following steps: collecting training phase signals and network topology physical state variables, and constructing comprehensive effective bandwidth for the current training phase signal task; establishing a quantitative mapping relationship between the network topology physical state variables and the end-to-end delay, obtaining a logical hop number sensitivity parameter and a bandwidth sensitivity parameter, constructing a topology state index based on the comprehensive effective bandwidth, the logical hop number sensitivity parameter and the bandwidth sensitivity parameter, combining the topology state index to perform forward-looking evaluation on candidate actions, performing network topology reconstruction, recording the measured delay and network state after network topology reconstruction, and updating the parameters. The application significantly reduces the end-to-end communication delay and suppresses the long tail delay under the premise of ensuring the hardware physical deployability and control stability, thereby improving the large model training efficiency and stability.

This invention discloses a method and system for locating faults in a wind turbine drivetrain. The method includes: constructing a dual-flow graph neural network model containing two paths: physical flow and latent flow. The input to the model is the features of each node constructed based on multi-channel vibration signals, and the output is an enhanced node representation. The physical flow and latent flow are used to extract explicit and implicit structural correlation features between nodes, respectively. During the training phase, physical consistency constraints are introduced to suppress unreasonable spatial jumps in the implicit correlation structure. During the model operation phase, a historical sample management mechanism based on sample information is combined to enable the model to continuously adapt to new fault categories without forgetting existing knowledge. Finally, the contribution of each sensor node is quantified based on the change in the original discriminant response to generate a fault location result consistent with the drivetrain topology. This invention can achieve refined and interpretable fault location in wind turbine drivetrain scenarios with small sample sizes and incremental learning.

This invention discloses a few-shot industrial anomaly detection method based on dual-guided contrast, relating to the fields of computer vision and industrial defect detection. The method first constructs a few-shot reference set, extracts image features using a pre-trained visual encoder, and then performs dual-guided branch fusion after normal-anomaly guided branching. During the training phase, a global contrast aggregation module and pixel-level contrast loss are introduced, and during the testing phase, the anomaly detection results are output. Compared with existing technologies, this invention can fully utilize the guiding information from a small number of normal and anomaly samples, improving anomaly detection accuracy and cross-class generalization ability, and exhibiting faster inference speed.

This invention relates to the fields of federated learning and cryptography, specifically to a secure three-party aggregation method for federated learning in the Industrial Internet of Things (IIoT). The method includes an initialization phase, a training phase, an online phase, and a verification phase. By setting up an architecture with one honest server and two aggregation servers, each server receives a portion of the secret share of local model gradient parameters sent by various IoT devices, thus protecting the privacy of local data. The honest server assists the two aggregation servers in interactively executing a three-party weight calculation protocol to calculate the secret share of the aggregation weights. They also interactively execute a three-party multi-weight aggregation protocol to calculate the secret share of the gradient parameters of the current aggregation model. Furthermore, a linear homomorphic hashing method is used to verify the correctness of the gradient parameters of the current aggregation model. This method solves the problems of low efficiency, weak security, and low robustness in federated learning during model training and aggregation, and can promote data flow and utilization in IoT scenarios.

This invention discloses a spacecraft pose estimation and uncertainty modeling method based on intrinsic space, belonging to the field of spacecraft pose estimation technology. The method involves directly constructing a hierarchical Bayesian probability model on the rotating manifold SO(3) and translation space, using Fisher and Gaussian distributions respectively, and introducing conjugate priors to achieve the fundamental decomposition and quantification of accidental and cognitive uncertainties. An end-to-end multi-task neural network is used to jointly learn the pose probability model parameters, key points, and segmentation information, and an iterative optimization module is employed to improve estimation accuracy. During the training phase, marginal negative log-likelihood and evidence regularization are jointly optimized; during the inference phase, the probability distribution of pose prediction is obtained through analytical marginalization, and the two types of uncertainty are distinguished. This invention improves pose estimation accuracy while outputting well-calibrated uncertainties, providing a reliable basis for the autonomous and safe operation of spacecraft in orbit.

This application discloses an autonomous driving scene generation method based on a spatiotemporal decoupled world model, belonging to the field of autonomous driving technology. The method includes: firstly, acquiring a multimodal autonomous driving dataset; after preprocessing to construct a state sequence; then, discretizing the state sequence into a pose word sequence and an image word sequence using an equidistant binning strategy and an improved time-aware vector quantization encoder; subsequently, inputting the word sequence into a spatiotemporal multimodal fusion module; and decoupling spatiotemporal information and predicting the potential state features of the next moment by alternately stacking temporal Transformer layers and spatial multimodal Transformer layers; during the training phase, employing a random masking strategy to prevent long-term generation drift, and calculating the cross-entropy loss between predicted words and ground truth words generated by the internal state autoregression module to update the model parameters; finally, generating autonomous driving scene data through binning inverse operation and a time-aware decoder; the method of this invention has the ability to generate long-term, high-fidelity videos.

The DAS seismic noise suppression method based on the conditional latent diffusion model belongs to the fields of machine learning and seismic data processing. This invention utilizes a latent signal representation module encoder to compress high-dimensional seismic data into a low-dimensional latent space, improving computational efficiency while preserving signal characteristics. In the inference phase, noise components are estimated through a noise prediction network, and clean latent variable estimation and state transitions are iteratively performed according to a deterministic sampling strategy. The decoder then reconstructs the denoised seismic data. In the training phase, self-supervised pre-training of the latent signal representation module, supervised training of the noise prediction network, and joint fine-tuning of all network parameters are performed sequentially, using KL divergence loss, reconstruction loss, and mean square error loss to simultaneously optimize network weights. This invention significantly reduces the computational overhead of the diffusion model while suppressing various complex noises.

Systems, methods, apparatuses, and computer program products for consistency between training and inference. A method may include receiving from at least one user equipment during a training phase, a report based on at least one measurement performed based on a first rule. The method may further include training at least one model using the report based on the at least one measurement performed based on the first rule transmitted with a first configuration. The method may further include receiving, from the at least one user equipment during an inference phase, a report based on the at least one measurement performed based on a second rule. In addition, the method may include performing prediction using the at least model and the report based on the at least one measurement performed based on the second rule transmitted with a second configuration during the inference phase.

The present application relates to a kind of training phase perception type SHARP in-network collection communication operator dynamic scheduling system and method, system includes: training framework domain: perception training phase and generate standardization phase descriptor;Host channel adapter domain: descriptor is forwarded to switch SMA by out-band channel, and listens to GPU progress token to realize submission gate;Switch CFU dynamic scheduling domain: according to strategy table, reconfigure parameter in CFU pipeline idle gap, realize no-interruption dynamic scheduling.The present application establishes a light training phase perception collaborative link between training framework in host side and CFU in switch side, realizes the real-time dynamic scheduling of CFU execution strategy with training phase, so that the in-network computing acceleration benefit of SHARP is extended from the current limited All-Reduce scene to the complete communication life cycle of AI large model training.

This application relates to the field of computer technology and provides a method and related equipment for generating intelligent agents for diagnosis. The method generates a virtual object containing temporal pathological feature data based on a preset disease evolution template; controls an initial diagnostic agent to interact with the virtual object in multiple rounds to obtain a temporal diagnostic strategy output by the initial diagnostic agent; generates feedback results for the initial diagnostic agent based on the diagnostic strategy and the target diagnostic scheme corresponding to the virtual object; and optimizes the parameters of the initial diagnostic agent based on the feedback results to generate a target diagnostic agent. By constructing a disease evolution template, this application can generate virtual patient data with continuous temporal features at low cost and on a large scale, enabling the diagnostic agent to be exposed to the complete disease progression logic from the latent stage to the critical stage during the training phase, thereby improving the model's generalization ability.

The application discloses a kind of based on double branch coding and foreground background difference enhancement's focus segmentation method, including the following steps: obtaining GLAS, Kvasir-SEG and BUSI medical segmentation public dataset and the dataset of physician hand marking segmentation result;Data pre-processing, data enhancement and dataset division;DCDB-Net model is constructed, the DCDB-Net is based on double branch coding and foreground background difference enhancement's focus segmentation model;The DCDB-Net model of S3 construction is trained, and parameter adjustment is carried out;Using the trained model is tested in GlaS, Kvasir-SEG and BUSI dataset.The application effectively alleviates the problem that soft boundary between foreground and background is difficult to distinguish, while reducing the interference caused by the coexistence of significant and non-significant objects in the training phase to the model key feature recognition.

Systems and methods are provided for performing anomaly detection. An example method includes, in a training phase, performing time series decomposition on training time series data to extract residuals of the training time series data, the residuals including a plurality of data points of the training time series data, using unsupervised anomaly detection models, identifying and labeling anomalous data points from among the plurality of data points contained in the residuals, based on outputs from the ensemble of unsupervised models including the labeled anomalous data points, obtaining a combined output indicating the labeled anomalous data points, and, using the combined output, training supervised anomaly detection models to detect anomalies in inference time series data In an inference phase, the method includes, using the trained ensemble of supervised anomaly detection models, on real-time, inference time series data.

This invention belongs to the field of natural language processing and dialogue emotion recognition, specifically involving an emotion classification method based on a hybrid expert model and a large-scale model collaboration. The method first encodes the dialogue text to construct basic features. Then, it extracts semantic, contextual, and knowledge-enhanced features through heterogeneous expert networks, and uses a dynamic routing gating network to weightedly fuse the features output by different experts, forming a unified emotion feature representation. Next, this feature representation is mapped to a soft cue vector, concatenated with the word embedding sequence of the original text, and input into a pre-trained large-scale model for emotion inference. During the training phase, cross-entropy loss is used as the target, updating only the parameters of the expert network, routing gating network, and soft cue mapping network, achieving efficient end-to-end parameter training. This invention significantly improves the accuracy of emotion classification by fusing multi-dimensional features and combining the common-sense reasoning ability of a large-scale model, while also achieving higher training efficiency.

This invention discloses a structure- and training-phase-aware adaptive aggregation federated recommendation system and method, belonging to the technical field of federated learning and recommendation systems. The client employs structure-responsive low-rank parameter updates, locally updating the item embedding matrix as a low-rank matrix product, uploading only the lightweight trainable matrix portion. The server calculates the federated gradient norm to determine key learning stages, dynamically adjusting aggregation coefficients based on the determination results, and only enabling performance-based weight adjustments during key stages to perform weighted aggregation of the low-rank matrix. This invention reduces communication overhead by over 90%; the standard deviation decreases from 0.089 to 0.063, significantly improving training stability; the recommendation performance HR@10 remains above 94.2% of the baseline method, and it is fully compatible with homomorphic encrypted secure aggregation protocols. The attached figure is a flowchart of the method of this invention.

This invention discloses a real-time optimization method, system, and device for vehicle-to-grid interaction based on a single-network architecture, belonging to the field of smart grid and electric vehicle charging and discharging optimization. The method includes training and model deployment phases. In the training phase, multi-source data is first collected to construct a state vector. This vector is then combined with a human risk preference model and a quantile mapping function to generate a fused feature vector, constructing an overall optimal action network to output the optimal charging and discharging solution. After constraint correction, the vehicle state and scheduling cost are updated. Then, based on a normalized advantage function, current and target network functions are constructed. Finally, through multiple rounds of training, the optimal action network is output. In the deployment phase, a data storage module collects real-time data, which is then fused with features and fed into the optimal action network to obtain interaction values. An intelligent decision-making module completes the energy interaction, and the interaction data is then fed back to the training module to achieve closed-loop model updates. This method simplifies the computational burden with a single network, achieving multi-objective collaborative optimization of grid load, battery degradation, and vehicle owner benefits.

The application discloses a kind of defense methods for machine learning system poisoning attack, including steps: input image in training phase;Contrast loss, supervision loss and center discrimination loss are calculated;In test phase, input training set and filtering rate, calculate the prediction class label in the image in training set, and classify class center;According to ground truth score and center discrimination score, calculate total score, and rank data according to total score, filter contaminated data;Input actual image, use the model trained by the above steps to identify, and remove the sample of poisoning attack.The application integrates the concept of contrast learning to enhance feature representation, proposes a scoring strategy based on fusion to achieve more effective decision making;The application provides new benchmarks and new indicators to fully study the performance under non-clean settings.

This invention belongs to the field of audio-driven virtual human generation technology, and discloses a real-time streaming audio-driven method and system for generating infinitely long virtual humans. It constructs a causal autoregressive generation mechanism based on frame latent variable blocks, ensuring that the generation of the current frame block depends only on historical frame blocks, a rolling reference frame, audio conditional information, and text conditional information. It also constructs a causal video coding mechanism to perform causal constraint coding on the video data during the training phase, ensuring that the latent variable at any given time is determined only by the current frame and its preceding frames. Furthermore, it constructs a long-term consistency coordination mechanism to suppress generation distribution drift through dynamically updated generation reference frames and to suppress inference pattern drift through rolling relative time position control. Finally, it constructs an autoregressive student model and uses a bidirectional teacher model and a pseudo-score model for distillation constraints, enabling the student model to approximate the generation distribution of the bidirectional diffusion model while maintaining the causal generation structure.

Method and Device for Target Detection in Radar Images The present invention relates to a method (100) for target detection in radar images, comprising the following steps: - creation (102) of a neural network, capable of receiving six distinct data channels as input and optimized for target detection, by combining: - a first U-shaped architecture (A1) of the U-Net type; - a second architecture (A2) comprising three compression stages (54, 56, 58) connected successively, said combination parallelizing the encoding portion of the first architecture (A1) and the second architecture (A2); - a training phase (104) of said neural network; - detection (108) of a target within a current radar image using, during an inference phase (106), said trained neural network. Figure for the abstract: Figure 3

A method is disclosed relating to deep learning-assisted fingerprint-based beam alignment, some embodiments of which may include: obtaining input data including user equipment location, number of user equipment, and desired received signal strength; processing the input data with a neural network having weights determined from a training phase to generate a set of one or more beam pair indices; performing a beam search on at least one subset of the set of beam pair indices; and receiving at least one beam pair index from a vehicle that provides the desired received signal strength.

This invention relates to the field of computer vision technology, providing a method for training a behavior recognition model, a behavior recognition method, and an apparatus. The method and apparatus utilize a frozen pre-trained image-text model (a second video feature extractor and a text encoder) as a general knowledge base to guide a finely tuned student model (a first video feature extractor) with temporal modeling capabilities. A multi-head residual projection network is used for feature-level knowledge transfer and representation alignment. The behavior recognition model trained by this method can accurately understand dynamic behaviors in videos and possesses strong zero-shot reasoning ability (i.e., generalization ability), effectively recognizing behavior categories not seen during the training phase. Furthermore, the entire model framework is clear, training is stable, and the final model exhibits excellent recognition performance for both known and unknown behavior categories under open-vocabulary settings.

The application provides a kind of converter transformer fault prediction method based on deep neural network, it is related to power equipment state monitoring technical field, the steps of the method include obtaining the time series data of the concentration of dissolved gas in oil in a historical period of converter transformer, and pretreatment is carried out;Using fixed-length sliding window mechanism, the data is mapped into a supervised learning sample set, and divided into training set and test set;For the minority class samples containing fault samples and fault precursor samples in the training set, a multivariate dynamic time warping method based on Mahalanobis distance is used for time series sample enhancement, and an enhanced training set is obtained by screening;Based on the enhanced training set, a multi-class deep neural network model is constructed and trained, and the single-window prediction probability of each type of fault is output;After the data preprocessing and window division consistent with the training phase, the real-time time series data of the concentration of dissolved gas in oil is input into the trained multi-class deep neural network model, and the final fault type is output.

Embodiments are associated with unit recommendations for an online establishment backend service. A reinforcement learning apparatus integrates a reinforcement learning algorithm model into the online establishment backend service and plugs streaming data representing user behavior and unit performance into the model. The reinforcement learning algorithm model is used to generate unit recommendations, and periodically collected relevant data is used to update the reinforcement learning algorithm model. In some embodiments, the reinforcement learning apparatus retrieves relevant unit data from a unit historical data store and user behavior data from an archive user behavior data store. The retrieved data can then be fed, during an offline training phase, into the reinforcement learning algorithm model before integrating the reinforcement learning algorithm model into the online establishment backend service (e.g., to improve system performance).