20results about How to "Speed ​​up the convergence process" patented technology

The present application relates to a kind of 5G / 5G-A network slice resource allocation methods supporting high reliability low latency service and enhanced mobile bandwidth service coexistence, belong to mobile communication field.It includes: obtaining URLLC and eMBB service coexistence 5G / 5G-A network parameter information;Optimization target and constraint condition of minimum physical resource block usage are constructed;According to system model and optimization target, establish the hierarchical PPO model assisted by transfer learning, and design state space, action space, reward function and action space truncation mechanism;Pre-training and the parameters of the pre-trained resource allocation model are migrated to hierarchical PPO model;Hierarchical PPO model is trained, and network parameters are updated according to loss function until model converges, obtain the physical resource block allocation method for URLLC and eMBB service coexistence.The present application minimizes physical resource block usage while guaranteeing slice service quality, improves resource utilization.

The invention discloses a three-level T-type inverter grid-connected control method based on a sliding-mode observer, and relates to the technical field of grid-connected control. According to the three-level T-type inverter grid-connected control method based on the sliding-mode observer, a preset time sliding-mode control scheme based on the sliding-mode observer is provided, a continuous Sigmoid function is introduced to replace a nonlinear sign function to construct a sliding-mode observation reaching law, signals can be effectively and smoothly controlled, and the control precision of the three-level T-type inverter grid-connected control method based on the sliding-mode observer is improved. The inherent high-frequency buffeting phenomenon of traditional sliding mode control due to a sign function is remarkably suppressed, and meanwhile, the convergence speed and tracking precision of the system are improved. On the basis of observing power grid voltage serving as external disturbance through a sliding-mode observer, feedforward compensation is performed on the external disturbance, and the feedforward compensation is applied to a sliding-mode control reaching law of preset time so as to suppress buffeting and accelerate the convergence process of output current. And the timeliness and the stability of grid-connected control of the three-level T-type inverter based on the sliding-mode observer are improved.

This invention provides a multi-UAV sensor data acquisition trajectory planning method based on deep reinforcement learning, comprising the following steps: S1, establishing the basic framework of a UAV-assisted sensor data acquisition model in a land scenario; S2, constructing a persistent trajectory planning strategy based on an improved deep reinforcement learning algorithm according to the basic framework, and generating a UAV flight trajectory map that satisfies energy balance constraints and maximizes data acquisition value based on velocity vectors. This invention introduces a feature extraction unit module containing self-attention weight calculation to extract spatiotemporal differential features of sensor observations based on traditional multi-agent deep reinforcement learning algorithms, and combines a packet retransmission mechanism and an energy balance coefficient to optimize UAV flight energy consumption while improving the value and reliability of data acquisition. This system improves the processing efficiency of UAVs for dynamic monitoring tasks.

The present application relates to the technical field of robots, and in particular to a motion generation method, system, device and storage medium for a humanoid robot, comprising: physical simulation rendering generates a visual video containing visual representation of joint motion state of the robot; analyzing the visual video to obtain frame images, natural language instructions and motion sequences, decoding output trajectory data; and outputting control instructions through a hierarchical architecture of high-level motion planning and bottom-level dynamics tracking to drive the robot to perform actions. The present application implicitly encodes the joint state of the humanoid robot in the visual features, improves the inference efficiency and generalization ability of the model, and the hierarchical decoupling architecture of the large model high-level trajectory planning and the reinforcement learning bottom-level dynamics tracking controller pays a small performance cost, which can compress the inference time delay to within 50ms and the parameter quantity to below 2 billion, and balances the high generalization of the stateless VLA and the small model, few-step inference, edge low latency, and low power requirement.

The invention discloses an emergency evacuation planning method based on improved ant colony algorithm and genetic fusion, and belongs to the technical field of emergency management and public safety. The method comprises the following steps: constructing a multi-objective optimization model taking minimization of total evacuation time T, total road section risk R and total personnel panic index P as objectives, and solving the multi-objective optimization model based on fusion of an improved ant colony algorithm and a genetic algorithm to obtain an optimal evacuation path scheme. The evacuation efficiency, safety and personnel psychological factors are comprehensively considered, the state transition probability and the pheromone updating strategy of the ant colony algorithm are improved, the global search mechanism of the genetic algorithm is introduced for deep fusion, and finally, the evacuation efficiency can be improved in a complex evacuation scene. And a global optimal evacuation path giving consideration to timeliness, safety and personnel psychological factors is quickly planned, and the evacuation efficiency is effectively improved.

The application discloses a sea surface small target detection method based on multi-domain multi-dimensional feature combination, and adaptively extracts deep features of sea clutter and target echo signals by using a stack sparse auto-encoder, so that the feature dimension is improved to ensure the feature feasibility. Since the feature extraction is a complete adaptive process, the complexity of the model is reduced. Meanwhile, aiming at the problem that the feature distinction degree of the sea clutter and target echo data in a single domain is low, a time-frequency domain feature combination method is provided to improve the feature difference and ensure the stable and efficient detection performance of the detector. Through an adaptive genetic algorithm, the convergence process of the super parameter group optimization is accelerated, and the local optimization of the super parameter is prevented to a certain extent, so that the final detection probability is effectively improved. The experimental results show that the detector provided by the application has better detection effect on high sea state data, the detection probability is improved by 27.6%, and the sea surface small target detection under high sea state can be coped with.

The invention discloses a deep tight reservoir crustal stress field boundary condition inversion method integrated with machine learning, and belongs to the field of petroleum and natural gas engineering. The method comprises the following steps: firstly, constructing a fine three-dimensional heterogeneous modeling model through high-precision scanning and image processing; then, multi-source data such as hydraulic fracturing, acoustic emission and logging are fused, and the real ground stress of a target point is determined; an integrated machine learning model trained by an optimization algorithm is utilized to intelligently invert and generate a high-precision boundary load required by the crustal stress site; and finally realizing reliable prediction of the global crustal stress field. The method is used for providing key and reliable geomechanical basis and boundary condition input for well location trajectory optimization, hydraulic fracturing scheme design and development scheme adjustment of a deep tight reservoir, and effectively supporting safe and efficient drilling and effective utilization of reserves.

This invention discloses a SAR image classification method and system based on scattering cues and optimal transmission, belonging to the field of synthetic aperture radar image classification technology. The main steps include: thresholding and connected component analysis of the SAR image to extract the coordinates of the image scattering center; generating scattering cues text and cues vector sequences based on the normalized coordinates of the scattering center, obtaining scattering perception text feature vectors and image feature vectors to achieve cross-modal alignment; introducing uniform edge probability to supplement global category constraints, dynamically separating clean and noisy samples; and optimizing model parameters through joint training to achieve robust SAR image classification. This invention, by introducing prior knowledge of the scattering structure unique to SAR images and combining it with the cross-modal capability of a visual-language model, effectively improves the classification accuracy and robustness under noisy label conditions, making it suitable for complex remote sensing scenarios such as disaster monitoring, military reconnaissance, and marine surveillance.

The invention discloses an unmanned aerial vehicle cluster collaborative task planning method and system based on a large language model and an environment feedback mechanism. The method comprises the steps of performing knowledge extraction and text segmentation based on an unmanned aerial vehicle combat expert knowledge base to obtain an expert knowledge text fragment set; performing semantic coding and similarity screening according to the expert knowledge text fragment set to construct strategy cue words; inputting the strategy cue word into a large language model to generate an initial unmanned aerial vehicle cluster collaborative task planning strategy; performing unmanned aerial vehicle simulation deduction based on the initial unmanned aerial vehicle cluster collaborative task planning strategy to obtain multiple sections of trajectory data; constructing an external trajectory knowledge base based on the trajectory data, and screening a target trajectory text based on trajectory freshness and text similarity; and based on the target trajectory text, the historical dialogue information and a preset cue word, constructing a strategy evolution cue word and inputting the strategy evolution cue word into the large language model to generate an optimized unmanned aerial vehicle cluster collaborative task planning strategy. According to the invention, the adaptability, rationality and reliability of task planning can be improved.

This invention relates to the fields of cheminformatics and high-value utilization of solid waste, and particularly to a method for optimizing process parameters in the preparation of geopolymers from copper tailings and fly ash based on cheminformatics. The method involves constructing a heterogeneous reaction map of the geopolymer precursor and generating a silicon-aluminum-oxygen network topological chemical fingerprint vector. A dynamic graph convolution thermodynamic constraint collaborative prediction model is established, embedding physicochemical hard constraints to output the mapping relationship between process parameters and target performance. A chemical fingerprint similarity-weighted Pareto front adaptive exploration algorithm is used for global optimization, combined with in-situ vibrational spectroscopy to achieve online updating of the heterogeneous reaction map and iterative calibration of the model. This invention can effectively shorten the optimization cycle, comprehensively improve the target performance, and produce geopolymers with excellent mechanical properties and heavy metal solidification effects, providing efficient and precise technical support for the large-scale high-value utilization of copper tailings and fly ash.

The application discloses a surgery room infection management system based on big data, which comprises a data acquisition module, a preoperative surgery room infection risk assessment module, an intraoperative surgery room infection dynamic monitoring module, a postoperative surgery room infection terminal assessment module and a surgery room infection intelligent management module. The application relates to the technical field of data processing, in particular to a surgery room infection management system based on big data. The application innovatively combines preoperative, intraoperative and postoperative surgery room infections to realize dynamic monitoring and whole-process control. The application introduces a segmented convergence factor and an iterative level reverse learning optimization population strategy to improve the optimization algorithm, thereby improving the model hyperparameter optimization precision and the accuracy of model output results. The application designs a bidirectional convolution path, combines a gated recurrent unit and simultaneously adopts a model semi-supervised loss function with a supervised loss, a consistency regularization loss and a pseudo label loss, thereby significantly improving the precision and real-time performance of the monitoring model.

The invention discloses a system and a method for adjusting a dynamic learning rate of TPM key agreement of a tree parity machine based on synchronous history. Aiming at the technical problems of low efficiency and insufficient stability caused by the fact that a fixed learning rate mechanism in traditional TPM key negotiation is difficult to adapt to dynamic change of a synchronization process, the system records a recent synchronization result by introducing an annular buffer area, and calculates and quantifies a synchronization progress index according to the recent synchronization result. And a negative correlation mapping function between the synchronization progress and the learning rate is constructed, and adaptive dynamic adjustment of the learning rate is realized. According to the method, convergence can be accelerated by adopting a large learning rate in the initial stage of synchronization, and stable and accurate synchronization can be ensured by adopting a small learning rate in the later stage of synchronization, so that the efficiency of key negotiation and the robustness of the system are remarkably improved on the premise of ensuring the safety.

The invention provides a transient voltage dynamic prediction system and method for a direct current sending end system based on an LSTM network, and belongs to the technical field of transient voltage prediction.The method comprises the steps that electrical quantity data of the direct current sending end system are collected in real time through a synchronous phasor measurement unit, and data preprocessing is executed; extracting time domain and frequency domain features, and screening out key feature subsets having significant influence on transient voltage prediction; constructing an LSTM deep learning model, and setting model hyper-parameters; training the model by using the training set, adjusting parameters to minimize a prediction error, and evaluating and optimizing the model by using the test set; and preprocessing electrical quantity data acquired in real time, inputting the preprocessed electrical quantity data into the trained LSTM deep learning model, predicting a transient voltage value at a future moment, and outputting the transient voltage value to an electrical control system. By adopting the LSTM network-based transient voltage dynamic prediction system and method of the direct current sending end system, the problems of low prediction precision and slow response speed in a traditional prediction method are solved.

The application belongs to the field of home energy management, and relates to an energy management method and system based on sparse federated reinforcement learning, comprising: collecting local energy consumption information at each time step, and receiving real-time electricity price information from a power company; constructing an initial energy management model, and training the initial energy management model to obtain a final energy management model; based on the energy consumption information and the real-time electricity price information, executing an intelligent energy management algorithm through the energy management model to obtain an energy management strategy; generating control instructions for installed household devices through the energy management strategy, and managing the installed household devices through the control instructions; fundamentally solving the memory bottleneck problem of edge devices, realizing the localization deployment of complex algorithms, significantly reducing the communication overhead of federated learning, and improving the system training efficiency and stability.

The invention discloses a radar-oriented order constraint Jacobi eigenvalue self-sorting method and system, solves the problem of resource consumption caused by an additional sorting module in a Jacobi iteration eigenvalue decomposition post-processing flow in the prior art, realizes direct output of ordered eigenvalues and eigenvectors, and omits an independent sorting module. The method comprises the steps that a real symmetric matrix A obtained through radar baseband signal sampling and covariance estimation is processed to obtain a compressed array, and a feature vector matrix V is initialized; in the parallel iteration process, index pairs are generated according to a scheduling sequence, sub-matrixes are extracted, two types of asymmetric rotation operators are generated through a condition driving mechanism, and a feature vector matrix is synchronously updated to track feature vectors; global data replacement is carried out, and non-diagonal element energy judgment convergence is calculated; after iteration is ended, feature values arranged in a descending order can be directly extracted from the main diagonal of the final compressed array, and corresponding feature vectors are output and used for direction of arrival estimation.

The application provides a speed spectrum automatic picking method based on MeanShift clustering analysis, comprising the following steps: step 1, pre-processing artificial picking speed control points; step 2, training an initial constraint model of a work area speed; step 3, cleaning a speed spectrum according to the speed constraint model; step 4, picking a time-speed pair based on the MeanShift method; step 5, updating and optimizing the work area speed constraint model according to the speed picking result in step 4; step 6, repeating steps 3 to 5 until an iteration condition is reached; and step 7, outputting a speed spectrum automatic picking result. The speed spectrum automatic picking method based on MeanShift clustering analysis can effectively reduce the labor and time cost consumed in manual picking, and further improve the precision of speed automatic picking by introducing a three-dimensional speed field nonlinear multivariate regression model based on DNN as a constraint model for pre-processing the speed spectrum.

The present disclosure provides an address resolution method, an electronic device, a storage medium and a computer program product, relating to the technical field of data processing. The address resolution method comprises: obtaining an address sequence to be processed; performing vectorization processing on the address sequence to be processed to obtain a target embedding vector; and using an address resolution model to predict the target embedding vector to obtain an address label sequence, wherein the address resolution model comprises an encoder, a decoder and a fully connected linear layer, the decoder comprises a neural network, the dimension of the output of the hidden layer in the neural network is equal to the length of the address sequence, and the number of neurons of the fully connected linear layer is equal to the dimension of the address label space, so as to improve the accuracy of cross-border logistics international address resolution.

This invention proposes a method for detecting fake reviews based on big data, comprising: acquiring a review dataset, wherein the data in the review dataset includes review texts from multiple users and corresponding user behavior data, and performing data preprocessing; extracting review text features from the review dataset, and extracting user behavior features and review behavior features from the user behavior data; defining relationship categories based on the review text and user behavior features, and constructing a multidimensional relationship enhancement graph (MREGC) based on multiple relationships; using a dynamic adaptive feature enhancement graph neural network to learn features on the MREGC relationship graph, obtaining the embedding representation of each node in the multidimensional relationship enhancement graph, and aggregating the embedding representations; inputting the aggregated features into a classifier to determine whether a review is a fake review, and outputting a fake review label if the review is classified as a fake review; this invention, by fusing residual networks and multi-relationship review graphs, fully considers the multidimensional relationships and deep features between reviews, enabling the model to more accurately identify fake reviews, and significantly improving the accuracy and robustness of fake review detection.

The invention provides a feature-driven retired power battery health state estimation method and system, and relates to the field of battery health state estimation, and the method comprises the steps: constructing an algorithm library, and presetting the capability vector of each algorithm in the algorithm library; acquiring the voltage and current of the retired power battery at the current moment, and acquiring a feature vector at the current moment in combination with a preset battery equivalent circuit model; based on the feature vector, calculating a matching score of each algorithm in an algorithm library through a dynamic weighted scoring mechanism, and determining a current execution algorithm according to the matching score; performing parameter identification on the battery equivalent circuit model by adopting a current execution algorithm to obtain an optimal model parameter, and calculating the current health state of the retired power battery according to the optimal model parameter; and calculating the convergence performance of the currently executed algorithm to update the historical reputation of the algorithm in the algorithm library for matching score calculation at the next moment. The method can adapt to various working conditions, and the estimation precision and the convergence speed are improved.

The invention discloses a large model accelerated training method based on staged learning, and relates to the technical field of large model training. The method comprises the following steps: step a, dividing a training process of a model into a continuous core structure learning period, a detail feature rich period and a final fine tuning period; b, in the core structure learning period, a first information bottleneck constraint is injected behind a first middle layer of the network, and a loss function of the first information bottleneck constraint is divergence obtained through calculation based on output features of the first middle layer and variational prior distribution. According to the method, a three-stage progressive training framework from core feature learning to detail feature enrichment to final fine adjustment is constructed, differentiated information bottleneck constraints are applied to each stage, the model is guided to follow a feature learning rule from macroscopic to microscopic, calculation waste of redundant features at the initial stage of training is effectively avoided, and the training efficiency is improved. The overall convergence process of the model is accelerated, and the accuracy is improved.