72 results about "Gradient projection" patented technology

The invention relates to the technical field of heat management of energy storage systems, and discloses a heat management method of an intelligent distributed liquid cooling energy storage system. The method comprises the following steps: acquiring operation state parameters of a plurality of single batteries and optimizing timestamp synchronization; and processing the temperature data by adopting small-batch time series decomposition, and extracting change trend characteristics and abnormal hot spot positions. And based on the information, optimizing a heat exchange control rule base parameter set through a non-dominated sorting strategy, and generating optimized control parameters and priority rules. And in combination with inlet temperature and flow parameters of the liquid cooling system, a tabu search mechanism is utilized to decide a cooling liquid flow distribution scheme, and a branch control instruction is output. And finally, a flow constraint boundary is dynamically adjusted by adopting an approximate gradient projection technology, and a branch and bound strategy is fused to discretize and optimize a distribution path, so that the branch flow is finely regulated and controlled, and the thermal management precision and reliability of the energy storage system are improved. According to the method, the accuracy, adaptability and stability of thermal management of the liquid cooling energy storage system are improved.

The invention discloses an intelligent archive opening and identifying method based on a large model, particularly relates to the technical field of archive data auditing, and is used for solving the problems of insufficient cross-modal data analysis capability, lagging rule updating and low man-machine cooperation efficiency in the prior art. Fusing cross-modal features of texts, images and metadata through a hybrid expert model to generate multi-modal feature vectors, and dynamically allocating the multi-modal feature vectors to a rule network, a semantic network and a domain network for cooperative processing based on attention weights; the rule network parameters are optimized through gradient projection constraint, and regulation-driven real-time adaptation is achieved; matching sensitive data in combination with a multi-dimensional feature matrix of auditing personnel, and optimizing task allocation accuracy; removing redundant links by utilizing value flow analysis to generate a lightweight process, and recording as a tamper-proof evidence chain through a block chain evidence storage solidification operation; the auditing efficiency and accuracy are improved, and the compliance traceability is guaranteed.

The invention relates to an edge computing network-oriented micro-service and distributed database collaborative deployment system. The system comprises a database copy number dynamic optimization unit which obtains a dynamic optimization strategy of the copy number of a database based on a queuing gradient projection elastic scaling algorithm; a deployment strategy dynamic adjustment unit calculates end-to-end time delay based on the micro-service routing path and the database routing path; calculating a data inconsistency measurement index based on the number of each copy of the database; dynamically adjusting the deployment strategy based on the resource constraint between the end-to-end time delay and the data inconsistency measurement index and the dynamic optimization strategy of the number of copies of the database; the deployment strategy comprises micro-service instance deployment, database copy deployment and optimal routing path selection; according to the MEC scene-oriented distributed database design and fine-grained collaborative deployment optimization method, a message queue, a queuing theory and a Canal database incremental updating mechanism are integrated, and the micro-service application performance and the data query efficiency / reliability are improved.

The invention relates to the technical field of film production control, and discloses a synchronous two-way stretch film production control method and system based on process optimization. The method comprises the following steps: carrying out real-time acquisition and subspace identification on BOPA film production line process parameters, and establishing a state space prediction model; performing multi-objective weighted combination on quality indexes and energy consumption based on the model to form a performance optimization function; setting a process constraint boundary, and constructing a quadratic programming constraint set; and carrying out rolling optimization by applying a gradient projection algorithm, calculating an optimal control sequence and executing the first control instruction. According to the method, the problem of quality fluctuation caused by insufficient adaptive capacity of the model in the BOPA film production process is solved.

The embodiment of the invention provides a large language model optimization method based on variance reduction and momentum acceleration. The method comprises the following steps: a gradient estimation stage of a large language model: initializing a seed list and a projection list; executing multiple independent query iterations, calling a disturbance subprogram, generating a random seed for the large language model, and storing a table; in the disturbance subprogram, the determined gradient projection value is stored in a projection list, and iteration is carried out for next query; after executing multiple independent query iterations, storing a plurality of random seeds and a plurality of gradient projection values corresponding to the random seeds; in the weight updating stage of the large language model, a gradient norm subprogram is called for each layer of the large language model, and a random seed is obtained to reset a random number generator; and determining variance-reduced gradient estimation according to the gradient projection value extracted from the projection list and the reproduced disturbance vector. According to the embodiment of the invention, gradient information queried for multiple times is aggregated to generate low-noise gradient estimation, and fine tuning of a large language model is completed.

The invention relates to the technical field of intelligent medical dispatching, in particular to a multi-terminal collaborative nursing worker resource intelligent allocation method, which comprises the following steps of: firstly, acquiring positioning, road, nursing worker physiology and old people demand data and generating multi-modal standardized data; constructing a three-dimensional digital twin potential field, predicting a corrected potential field by using photons, constructing a matching model by using the corrected potential field, a nursing worker capability vector and an old man demand vector, and performing annealing optimization to obtain initial matching; constructing a nursing worker-old person-time period tripartite graph based on a matching result, and obtaining optimal matching by adopting tension diffusion and gradient projection iteration; double digital signatures are executed on each piece of matching, and a non-homogeneous commitment is cast in the block chain, so that credible performance is realized; the wearing end spiking neural network continuously outputs fatigue probabilities, the fatigue probabilities are mapped into potential energy increments and written back to the potential field, and sub-potential field resolution and zero-knowledge post replacement are triggered and closed-loop updating is carried out. According to the method, the scheduling real-time performance and fairness are improved, the fatigue risk is reduced, and the whole service process is traceable.

The invention relates to the technical field of unmanned aerial vehicle flight control, in particular to a dynamic thrust distribution method for heterogeneous power units of an unmanned aerial vehicle, which comprises the following steps: S1, heterogeneous parameter dynamic modeling: after the unmanned aerial vehicle is unfolded, acquiring the actual response delay time and the maximum thrust threshold value of each power unit through a stepped power loading test; obtaining a blade deformation compensation coefficient through non-contact deformation detection; constructing a heterogeneous parameter matrix; s2, adaptive thrust optimal distribution: a decoupling flight instruction is a thrust component and a torque component under an airframe coordinate system; establishing an optimization model which takes energy consumption and response delay difference minimization as a target and comprises a dynamic balance constraint, a physical limit constraint and a folding position constraint; solving a thrust distribution vector in real time by adopting a gradient projection method; and S3, fault dynamic reconstruction. Through comprehensive consideration of deformation, dynamic balance and physical constraints, the unmanned aerial vehicle can be dynamically adjusted in a complex environment.

The invention discloses a multi-target optimization unit adaptive combustion and control real-time optimization method and system. The method comprises the following steps: acquiring unit operation and decision parameters; constructing a target function with net power; an online Gaussian mixture process regression and uncertainty quantization network (OH-GPR-UQN) is adopted to online identify a dynamic relationship between a decision parameter and a net power objective function, a gradient is estimated, and the uncertainty of the gradient is quantized; on the basis of the gradient and the uncertainty thereof, an adaptive exploration and utilization constraint hill climbing algorithm (BOA-CEHC) based on Bayesian optimization guidance is adopted, and the adjustment step length and direction disturbance of decision parameters are determined by maximizing an acquisition function which considers the utilization gradient and the exploration uncertainty at the same time; processing operation constraints by combining gradient projection and a penalty function to obtain optimized parameters; the optimized parameters are used for closed-loop control and operation data feedback so as to update the OH-GPR-UQN online. According to the method, accurate identification, intelligent optimization and robust processing constraint can be realized, and the comprehensive benefits of the unit are remarkably improved.

The invention relates to the field of artificial intelligence control, in particular to a world model and reinforcement learning collaborative optimization method based on constraint perception gradient projection, which comprises the following steps: constructing a world model based on Transform architecture, performing unified modeling on an environment state, system dynamics and a reward function, and training the model based on reference model data to predict an environment future state; training a reinforcement learning strategy based on a virtual track and a reward signal generated by the world model, and calculating strategy gradient information; an explicit gradient projection operator of constraint perception is designed, a strategy gradient is corrected in real time according to future constraint conditions predicted by the world model, and it is ensured that the gradient updating direction meets the safety constraint and the optimization target at the same time; and synchronously updating world model parameters and reinforcement learning strategy parameters by using the gradient after projection correction to realize collaborative optimization of the world model parameters and the reinforcement learning strategy parameters. According to the method, the key problems of local optimal trap, low convergence speed, insufficient stability and the like in the collaborative optimization process of a traditional method are solved.

The invention discloses a robot motion planning method and system based on a sampling dictionary, and belongs to the robot motion planning technology, the architecture comprises a quantification stage and a prediction stage, in the quantification stage, through a pre-training VQ model, under the condition that re-training or fine tuning is not carried out at the same time, the prediction can be carried out at the same time; the method is directly applied to other constraint planning problems, a gradient projection operator is used for projecting sampling points to a constraint manifold in a prediction stage, and finally, a path from a starting point to a target point is generated by using the improved sampling points and optimized distribution under the condition of considering task space region constraints through a bidirectional planning algorithm. The planning search space is reduced through the pre-trained VQ-MPT model, the sampling area is made to be close to the constraint manifold through the gradient optimization technology, the planning efficiency is improved, and the planning time and path cost are remarkably reduced while it is guaranteed that the trajectory meets the constraint.

The invention provides an attack detection method and system based on Bayesian incremental learning and a storage medium, and the method comprises the steps: 1, collecting a data set, and dividing the data set into a plurality of tasks according to years; 2, using a Bayesian continuous learning framework, taking posterior distribution obtained by learning of a previous task as prior distribution of a current task, and adopting a gradient projection method to project a gradient of the current task to an orthogonal subspace of an old task feature space to obtain projection parameters; step 3, performing label deviation and noise processing on the task; 4, minimizing new task loss to obtain parameters, and adopting a training strategy according to a label deviation and noise processing result; 5, finding an optimal combined solution; and step 6, taking the combined model parameters as initialization parameters of the next task, returning to the step 2, and entering the next round of iteration until training of all tasks is completed. The method has the beneficial effects that the knowledge retention capability can be remarkably improved, and the problem of disastrous forgetting is effectively solved.

The invention discloses a color steel coil production cost optimization method and device and a medium, and the method comprises the steps: determining the production cycle of a single coil according to the time information of a new coil ready signal and a coil unloading completion signal; collecting power data of production line equipment in the production cycle, and determining energy consumption cost according to the energy unit price; according to the preset auxiliary material unit price and the production parameters in the production cycle, the auxiliary material cost is determined; constructing a feature matrix and a unit area cost vector according to a preset batch of single volume production data to train a cost prediction model; according to a preset constraint condition, determining an optimal parameter combination through a gradient projection method so as to minimize the unit area cost; and according to the parameter combination, the coating process parameters and the formula parameters in the next production process are adjusted, so that cost optimization closed-loop control is achieved. The cost accounting error is greatly reduced by defining the single-roll production cycle and accounting the energy consumption and the auxiliary material consumption; and meanwhile, a closed loop is driven through data, so that the production cost of the color steel coil is optimized.

The application belongs to the technical field of load identification of structural health monitoring, and provides a concentrated load positioning method based on principal stress constraint and strain gradient trajectory identification, which comprises the following steps: arranging strain sensors in a monitoring area of a planar structure plate and collecting strain data of measuring points, constructing a strain field inversion function based on a radial basis function and establishing an error optimization function containing a fitting error term and a smooth constraint, introducing a principal stress direction constraint and a gradient projection smooth constraint to form a comprehensive error function and solving to obtain an optimal strain field distribution, then calculating strain gradient vectors of each grid node and tracking a gradient trajectory through a gradient descent method, and finally performing an iterative clustering analysis based on distance statistics on a trajectory end point to output a cluster center as a concentrated load application position identification result. The application can improve the concentrated load positioning accuracy, enhance the anti-noise capability, and realize fast, stable and large-sample-free load position identification.

The invention belongs to the technical field of block chain and artificial intelligence crossing, and particularly relates to a fair federal consensus method and system based on gradient projection. The objective of the invention is to solve the problem that the fairness of a global model is reduced due to unfair client contribution evaluation and improper projection target selection in a gradient conflict resolution mechanism under the condition of non-independent identically distributed data (Non-IID). The method comprises the following steps: initializing a client key and parameters; randomly selecting m clients for local training in each round; calculating an effort value according to the loss change and the accuracy rate, and preferentially assigning a block creator; the server detects gradient conflicts and performs selective projection resolution, and a gradient with a large length is projected to an orthogonal plane with a gradient with a small length; aggregating the conflict-free gradient to update the global model; and circulating to a preset round number and outputting a final model and a creator sequence. According to the scheme, model performance and participation fairness are considered, and consensus efficiency and excitation compatibility of federated learning in a heterogeneous environment are improved.

The application discloses a land-air collaborative composite traffic network design method considering park-and-ride, and plans and designs a land-air collaborative composite traffic network with a vertical take-off and landing field as a center in a future land-air collaborative composite traffic network background, and proposes a corresponding network design model and a solving algorithm, mainly including: (1) a two-level planning model related to site selection, capacity allocation and integration of a landing field of an electric vertical take-off and landing aircraft (eVTOL) and ground parking facilities; (2) for a network equilibrium model in the two-level planning model, a modified improved gradient projection (M-iGP) algorithm is designed in an augmented Lagrange multiplier (ALM) framework; (3) a sensitivity analysis (SAB) algorithm is used to solve the two-level planning model, and the model and the algorithm are verified and analyzed in a Sioux Falls network.

The invention discloses a multi-target radar time resource allocation method based on dynamic priority evaluation, and the method comprises the steps: building a target motion model and a radar observation model, and enabling the target motion model to describe a target motion process through employing a plurality of models containing a possible motion mode; obtaining a radar measurement value according to the target motion model and the radar observation model, and constructing a priority function including target motion characteristics, spatial positions and historical observation information; embedding the priority function into a Bayesian Cramer-Rao lower bound, and establishing a radar resource allocation model which takes a minimum weighted Bayesian Cramer-Rao lower bound sum as a target function, takes a dwell time sum not exceeding a resource scheduling period and takes each target dwell time meeting a constraint range as conditions; and solving the radar resource allocation model through a gradient projection method to obtain a resource allocation result. Reasonable configuration of radar residence time resources can be realized, and the target tracking precision and reliability are effectively improved.

The invention provides an electric vehicle power controller parameter adaptive setting method and system, and relates to the technical field of electric vehicle control, and the method comprises the steps: obtaining a motor torque feedback signal, carrying out the wavelet decomposition, extracting a working condition identifier, determining a phase margin and an amplitude margin offset based on Fourier transform, and obtaining a motor torque feedback signal; parameters are adjusted in combination with a gradient projection algorithm and a primal-dual interior point algorithm, and an optimal parameter vector is obtained through a multi-target particle swarm optimization algorithm. According to the invention, on-line adaptive adjustment of the control parameters is realized, and the power performance and the energy efficiency performance of the electric vehicle are improved.

This invention provides a collaborative optimization method for titanium plate annealing process parameters using multi-agent reinforcement learning, belonging to the field of reinforcement learning technology. The method includes collecting and standardizing annealing process data, encoding process parameters through a multi-layer feature extraction network, and constructing a parameter coupling perception matrix to quantify the collaborative strength. A dual-branch encoding structure is used to extract single-parameter features and interaction features, and the fusion weights are dynamically adjusted based on the coupling perception matrix to complete the state collaborative representation. The collaborative representation is input into a constraint-aware policy network to generate adjustment decisions, which are then executed in a real-world scenario and feedback is collected. Based on the feedback, a reward value is calculated, and training samples are sampled with priority to construct a composite loss function. A gradient projection method is used to update the policy network parameters to within the process feasible region, achieving continuous optimization of the policy network.

According to the electric vehicle charging load prediction method based on the Markov chain, the Markov chain and the improved Kalman filter are fused, and accurate prediction of the electric vehicle charging load is achieved. In the off-line stage, the charge state is discretized into 28 states based on historical data, 96 intra-day time slots are divided, a three-dimensional time-varying state transition matrix is constructed, filtering parameters are set, and state and power mapping is established, and in the on-line stage, charge state priori prediction is obtained through a Markov chain, real-time observation data are fused, and the real-time state of charge is obtained. Lawful optimal charge state distribution is obtained through Kalman filtering recursion and gradient projection method constraint optimization, finally, the distribution serves as a starting point, a transfer matrix is iterated to predict the future charge state, transfer path expected power is accumulated, and the total charging load is output. According to the method, the problems of error accumulation, poor robustness and the like of a traditional model are solved, and the method has high interpretability and engineering practicability.

The invention discloses a large-model-driven charging operation multi-agent collaborative decision-making method and system, and relates to the technical field of electric vehicle charging management. The method comprises the following steps: receiving meteorological data, power grid SCADA data, user electric vehicle data and historical constraint violation cases, and constructing a multi-source heterogeneous data set; and according to the constructed data set, a Stiefel-LoRA fine tuning algorithm is adopted to carry out parameter fine tuning on the pre-trained large language model, an adapter matrix is optimized through Riemannian gradient projection and a QR contraction updating algorithm, a fine-tuned large language model adapted to a charging decision task is obtained, and the fine-tuned large language model outputs an initial charging decision scheme. According to the method, the large model efficiency is improved through Stiefel-LoRA fine tuning, a physical information perception neural network is utilized to ensure that a charging decision strictly meets power grid security constraints, multi-agent collaboration and lexicographical order optimization are realized based on a standardized protocol, the problems of scheme ineffectiveness and decision deadlock of a traditional method are avoided, and the method is suitable for large-scale popularization and application. And safe, reliable and economical intelligent charging operation is realized.

The present disclosure provides computer-implemented method for training a model using a central node and a plurality of client nodes, comprising, at one or more of the plurality of client nodes: receiving a set of model parameters and an initial seed from the central node; determining an update direction based on the initial seed; computing scalar reflecting a sign and a size of an approximation of a projection of a gradient of a loss function on the update direction, using the set of model parameters; sending the scalar of the approximated gradient projection to the central node; receiving a model update scalar from the central node; and updating the set of model parameters in the update direction based on the model update scalar.

The present invention provides a method for adaptively coordinating dynamic droop parameters of heterogeneous battery modules, including: real-time identification of transfer function parameters of each heterogeneous module, using virtual pole supplementation technology to solve the pole alignment problem of different order systems; decomposing the multi-objective optimization problem into three independent sub-problems of low-frequency band SOC balance, mid-frequency band power regulation and high-frequency band transient suppression based on modal decoupling theory; realizing multi-band parallel gradient projection optimization through pre-calculated constrained feasible domain and versioned synchronization mechanism, effectively avoiding thread competition; adaptively adjusting fusion weights according to the convergence quality of each frequency band to generate coordinated droop parameters. The present invention fundamentally solves the problem of mismatched frequency domain responses of heterogeneous battery modules, achieves a technological breakthrough from static pole configuration to dynamic adaptive adjustment, and improves the stability and efficiency of heterogeneous energy storage systems.

A computing system including one or more processing devices configured to receive a weight tensor of a neural network. The one or more processing devices are further configured to execute a gradient descent optimizer that updates the weight tensor over a plurality of projection matrix update intervals. Each of the projection matrix update intervals includes computing a gradient over the weight tensor in each of a plurality of gradient descent iterations. Each of the gradient descent iterations further includes projecting the gradient into a reduced-rank subspace using a projection matrix and updating the weight tensor by performing gradient descent using the projected gradient. Each of the projection matrix update intervals further includes computing a projection matrix error value associated with the projection matrix and updating the projection matrix based at least in part on the projection matrix error value.

The invention provides a hot area identification self-adaption method and system based on density residual errors, and aims to solve the problem that the performance of an existing hot area identification model is reduced due to data distribution differences during cross-domain deployment. The method comprises the following steps: constructing a cross-domain density residual index to quantify the difference of abnormal spatial distribution of a source domain and a target domain; in domain adaptation training, a pseudo-gradient projection mechanism is adopted, and an adversarial weight lambda is dynamically adjusted according to the change trend of the density residual error, so that the model adapts to a target domain; after deployment, macroscopic performance indexes including a false drop rate and an omission rate are monitored through modes such as user feedback, and adaptive updating and incremental training are carried out on the model and hotspot generation parameters based on a monitoring result to form a continuously optimized closed loop. According to the method, a complete closed loop from measurement, adaptation to feedback is established, so that the challenge of hotspot identification in a cross-domain scene is effectively solved, the adaptability and robustness of the model are remarkably improved, and intelligent management of the whole life cycle is realized.

The present application relates to a kind of semantic driven holographic content adaptive transmission method, belong to communication technical field.The method includes the following steps: S1: establish a semantic driven communication model of holographic content transmission of Internet of Things equipment, including semantic model and transmission model;S2: build an optimization problem of joint base station association, semantic compression ratio selection and power allocation, to maximize task utility;S3: establish the sub-problem of power optimization, analyze the feasibility condition of transmission power, and design a kind of power allocation algorithm based on gradient projection method;S4: joint solution base station association, semantic compression ratio selection and power allocation using deep reinforcement learning algorithm.The present application is directed to the non-convexity and NP-hard characteristics of problem, combines numerical optimization and DRL to formulate resource allocation strategy, proposes the reinforcement learning algorithm assisted by gradient projection method, to accelerate the convergence of reinforcement learning model, improves task completion efficiency.

The invention relates to the technical field of wind power prediction, and particularly discloses a gradient synergy and double alignment fused small sample wind power prediction method, which comprises the following steps: designing a Fourier enhanced Transform shared feature extractor to extract general feature representation with periodic perception from wind power time sequence data of a multi-source domain and a target domain; then, constructing a hybrid domain adaptive module, and realizing implicit and explicit dual alignment of feature distribution of a source domain and a target domain through a plurality of adversarial domain classifiers arranged in parallel and multi-core maximum mean difference measurement; and finally, introducing a gradient projection algorithm, carrying out collaborative optimization on conflict gradients of the prediction task and the domain adaptation task in a back propagation process, and eliminating gradient conflicts in multi-task learning. The method effectively improves the feature extraction capability, domain adaptability and optimization stability of the model in a small sample scene, and remarkably improves the prediction precision and robustness in a cross-domain wind power prediction task.

This invention discloses a method and system for constructing a physically guided diffusion model for robust optimization, belonging to the technical field of power system optimization and scheduling. The method includes: constructing a wind farm topology information map and a topology-aware variational autoencoder (VAE). The VAE includes an encoder and a decoder. The encoder encodes the input wind power operation data into latent variables, and the decoder restores the latent variables to the wind power scenario. A diffusion model is constructed. The gradient of the predefined power system operating cost function onto the wind power scenario is projected onto the tangent space of the latent manifold to obtain a gradient-guided operator, which is then introduced into the diffusion model in the next round of solution. This method achieves dimensionality reduction mapping of high-dimensional wind power scenarios by constructing a topology-aware VAE that integrates wind farm topology information and introduces a gradient-guided diffusion process based on the power system operating cost gradient, thus solving the technical problem of low efficiency in solving traditional robust optimization subproblems.

The invention belongs to the technical field of load identification for structural health monitoring, and provides a concentrated load positioning method based on principal stress constraint and strain gradient trajectory identification, which comprises the following steps: arranging a strain sensor in a plane structure plate monitoring area and collecting strain data of a measuring point; and constructing a strain field inversion function based on a radial basis function, establishing an error optimization function containing a fitting error term and a smooth constraint, introducing a principal stress direction constraint and a gradient projection smooth constraint to form a comprehensive error function, and solving to obtain optimal strain field distribution. And further calculating a strain gradient vector of each grid node, tracking a gradient trajectory through a gradient descent method, finally carrying out iterative clustering analysis based on distance statistics on a trajectory end point, and outputting a cluster center as a concentrated load application position identification result. According to the method, the concentrated load positioning precision can be improved, the anti-noise capability is enhanced, and rapid and stable load position identification without a large number of training samples is realized.

The invention relates to the technical field of electric data processing and intelligent algorithms, in particular to an image processing projection scaling optimization method and system based on the symplectic manifold acceleration technology. The method comprises the following steps: for a constraint minimization problem corresponding to an image deblurring or binary classification scene, solving the constraint minimization problem by adopting a zoom gradient projection algorithm to obtain an image deblurring result or a classification result; and a second-order symplectic integrator is introduced to perform position updating and momentum updating in the process of solving the constraint minimization problem by adopting the zoom gradient projection algorithm, so that the convergence efficiency of solving the constraint minimization problem by adopting the zoom gradient projection algorithm is improved.