13results about How to "Fully learn" patented technology

The invention discloses a multi-layer semantic perception, distillation and semi-supervised cooperative training target detection method, and the method specifically comprises the following steps: constructing a data set: extracting a first part of images from image data, marking the first part of images to construct a supervised target detection data set, and taking the remaining images as an unmarked image data set, the data volume of the unlabeled image data set is greater than that of the supervised target detection data set; teacher model optimization: performing supervision training on the teacher model on the supervised target detection data set; constructing a teacher model, and executing self-distillation training on the teacher model; pseudo labels are generated for the unlabeled images through a teacher model, and adaptive screening is carried out based on confidence distribution; mapping a pseudo label to a strong enhanced sample through enhanced geometric transformation, carrying out semi-supervised training by using the strong enhanced sample and the pseudo label, and introducing a feature layer distillation constraint at the same time; optimizing a student model; and outputting the target detection model obtained through training.

This disclosure relates to a training method for a multi-label recognition model for speech recognition and a smart home device. The multi-label recognition model includes a shared feature extraction network and multiple branch recognition networks. The shared feature extraction network outputs shared features based on audio data, and the multiple branch recognition networks output multiple recognition labels based on the shared features. The training method includes: acquiring training data, which includes audio data and the ground truth value of at least one of multiple recognition labels associated with the audio data; using the multi-label recognition model to output a predicted value for each of the multiple recognition labels based on the audio data; calculating a branch loss function value between the ground truth value and the predicted value for each of the at least one recognition label; calculating a total loss function value based on the branch loss function value; and updating the parameters of the multi-label recognition model based on the total loss function value.

The present application relates to the technical field of aerospace artificial intelligence, and provides a space debris collision risk probability prediction method based on deep learning, which takes ESA Kelvins real rendezvous data as a benchmark, and uses Kessler to generate a synthetic CDM sequence to expand the training sample to improve the coverage of high risk and complex situations; at the same time, a plurality of update records and their associated related events are expressed as a dynamic heterogeneous graph, a dynamic graph neural network is used to learn the law of risk evolution over time and the similarity between events can be optionally used for knowledge transfer, so that the continuous regression prediction result of the final collision risk probability of the rendezvous event can still be stably output under the condition of missing and noise, and the prediction is quickly refreshed when new CDM arrives through the local graph incremental update mechanism, which significantly improves the accuracy, robustness and online availability of risk prediction.

The invention discloses a train transmission system fault diagnosis method based on parameter alignment, and the method comprises the steps: S1, constructing a personalized federated learning system which comprises a server and a plurality of clients; s2, the server issues the initialization model to each client; s3, carrying out local training by the clients by utilizing respective data, and calculating an aggregation weight; s4, the server generates a new global model and issues the new global model to each client; s5, executing a cross-round parameter semantic alignment strategy by the client, and updating the local model; repeating the steps S3 to S5 until a preset maximum training round is reached; and S6, inputting the state data of the train transmission system into the client local model after the maximum training round is ended, and outputting a fault diagnosis result of the train transmission system. According to the invention, on the premise of ensuring data privacy of each client, efficient cooperative training of the global model and the local model is realized, and the overall accuracy of fault diagnosis of the train transmission system is improved.

The invention relates to the technical field of signal processing, and discloses a simultaneous same-frequency full-duplex self-interference elimination method and system, and the method comprises the steps: carrying out the data preprocessing of a generated self-interference signal, generating a self-interference signal containing a nonlinear factor, and constructing a training data set according to the self-interference signal; establishing a plurality of CNN-LSTM attention models for extracting nonlinear features and time sequence features of the self-interference signals; an auxiliary sub-module is introduced in the training stage of the model, so that the model has internal knowledge migration, and knowledge distillation of a feature level is realized; training the model, and performing parameter optimization by adopting a loss function which considers the amplitude and the phase of the complex signal at the same time; based on the trained model, linear and nonlinear two-stage self-interference elimination is carried out on the received signal, self-interference suppression in a full duplex system is realized, and the problems that the existing full duplex nonlinear self-interference elimination method based on deep learning is insufficient in feature extraction capability, limited in supervision information and low in training speed are effectively solved through the method.

The invention discloses a short-term offshore wind power prediction method, and the method comprises the steps: S1, collecting wind power time sequence data, and carrying out the preprocessing of the wind power time sequence data; s2, realizing fine division of the wind power plant group based on multi-source fusion spatial-temporal clustering analysis, and obtaining spatial-temporal correlation among the wind power plants; s3, constructing an offshore wind power prediction model based on a bidirectional space-time convolution module, a space-time attention module and a residual feature fusion module, and performing preliminary prediction on the offshore wind power; s4, performing maximum overlapping double-domain cooperative decomposition on the prediction error of the preliminary prediction of the offshore wind power, and extracting error information of different frequency bands; s5, introducing an attention mechanism to dynamically distribute weights, focusing key information to drive model feedback learning, improving offshore wind power prediction precision, and outputting a prediction result; and S6, performing comparative analysis, error analysis and economical efficiency analysis on the output prediction result. The problem that an existing offshore wind power prediction model is poor in prediction precision is solved.

The invention belongs to the field of FTU fault identification, and particularly relates to a zero sequence voltage abrupt change characteristic grounding fault identification method and system for FTUs, and the method comprises the steps: firstly obtaining a power distribution network bus zero sequence voltage discrete sequence, calculating a continuous sampling point difference absolute value sequence, and constructing an abrupt change characteristic quantity sequence through sliding time window integration; the time window features are input into a pre-trained capsule neural network, and the primary capsule layer encodes the time window features into primary capsule vectors; and updating the connection weight between the primary capsule and the high-level capsule to be converged through iterative computation, and judging a fault according to the converged high-level capsule vector: judging the fault and outputting a starting signal when the model length of the capsule in a grounding fault state exceeds a first threshold value and the included angle between the capsule and the fault feature base vector is smaller than a second threshold value, otherwise, judging that the system is normal. The FTU identification precision and efficiency can be improved.

The application provides a prompt learning method based on unsupervised knowledge distillation, comprising: a supervised fine-tuning stage, taking a first visual language model as a teacher model, freezing first pre-training parameters of the teacher model, and performing supervised fine-tuning on the teacher model through labeled samples to optimize first learnable prompt parameters of the teacher model; an unsupervised distillation stage, taking a second visual language model as a student model, freezing second pre-training parameters of the student model, aligning inference results of the student model and the teacher model on unlabeled samples, and migrating discriminative knowledge of the teacher model to second learnable prompt parameters of the student model. The application also provides a prompt learning device based on unsupervised knowledge distillation, a storage medium and an electronic device. Therefore, the application significantly improves the adaptation effect of the visual language model on downstream tasks, improves the generalization performance of the visual language model, and has low training and inference costs.

This application relates to an AI model image generation method, system, device, and medium, belonging to the technical field of image processing. It involves collecting multimodal data of the target garment, including mannequin images, flat static images, and material close-up images; selecting a reference image from a target pose library whose pose matches the desired pose, and extracting the corresponding pose skeleton image as the target pose skeleton image; inputting the flat static image and the material close-up image into a visual language model to generate a text description; and inputting the target pose skeleton image, mannequin image, and text description into a pre-trained LoRA diffusion model to generate a preliminary dressed image. The LoRA diffusion model's input layer includes an independent pose condition injection channel, through which the target pose skeleton image is input into the LoRA diffusion model. This application has the beneficial technical effect of avoiding stiff and unnatural AI model poses, thereby improving the aesthetics of the generated images.

The invention provides a mode perception prediction model for building load imbalance data, relates to the field of power load prediction, and recognizes typical operation modes in building load data through a K-means clustering algorithm, including startup impact, morning, afternoon and shutdown dormancy. A time window feature enhancement strategy is introduced to extend an original one-dimensional feature space to a three-dimensional feature space, a mode-aware oversampling method PP-SMOTE is constructed, an optimal balance factor is obtained through a balance factor optimization strategy, and the quality of data generated by the PP-SMOTE is verified; the PP-SMOTE is applied to a LightGBM prediction model, and the optimal balance between prediction precision and data distribution characteristics is realized through a method for determining the optimal data balance degree through a multi-balance-factor comparison experiment.

The application discloses a photovoltaic and energy storage power grid power generation power prediction method based on a model-agnostic meta-learning method, solves a photovoltaic and energy storage power grid power generation power prediction problem through a meta-learning technology, designs a generation model, combines task characteristics, splits prediction problems of all devices into power generation device power prediction problems in different power station ranges, and through a meta-learning method, meta-knowledge learned from one task can be transferred between different tasks, and the conversion of the meta-knowledge between different tasks plays a key role in power generation power prediction. The application jointly models dynamic time sequence characteristics and static characteristics related to power generation efficiency, learns a representative model initialization parameter value, and significantly improves fitting speed and prediction accuracy of the model in a new scene.

The invention provides a driving road defect detection platform and method for a large transport vehicle. The driving road defect detection platform comprises a road defect processing device, a road surface information display and a road surface information acquisition device, wherein the road surface information display and the road surface information acquisition device are respectively connected with the road defect processing device; wherein the road defect processing device is internally provided with a circuit board, the middle position of the circuit board is provided with a central processing unit and a static random access memory located on the side of the central processing unit, and the central processing unit is internally provided with an improved YOLOv8n model; compared with the prior art, the road defect processing device has the following beneficial effects that the flexible characteristic of the road defect processing device provides an effective means for intelligentization of road defect recognition, and in the driving process, when the road surface information acquisition device detects and recognizes the road defects, the road defect processing device can detect and recognize the road defects; and an improved YOLOv8n model arranged in the central processing unit can carry out pavement defect marking and give an alarm through an alarm on the road defect processing device to remind a driver to make a corresponding defect avoidance means.

The invention relates to the technical field of lithium ion batteries, in particular to an energy storage lithium ion battery health state prediction method, which comprises the following steps that sparse test point data are acquired through a battery aging experiment, and each test point comprises a cycle number, internal resistance, an increment health state characteristic and a health state label; systematic slicing and resampling are carried out on a time axis by utilizing an aging mining method, training samples containing different circulation history paths are generated, and a mapping relation between circulation history and health state attenuation is learned by adopting a deep learning model, so that high-precision health state prediction is realized; the method has the beneficial effects that the problem of scarcity of battery aging data is solved, and the health state prediction precision is remarkably improved; the method has good generalization ability and has flexibility and expandability.