64 results about "Offline learning" patented technology

The invention relates to the technical field of intelligent optimization, in particular to a network anti-fiber-breakage protection planning method based on offline learning and improved MCTS, and the method comprises the steps: solving a preset multi-target optimization model of network anti-fiber-breakage protection planning through employing a preset multi-mode multi-target optimization algorithm, obtaining at least one optimal standby path corresponding to each service; learning a topological structure and link weight information of each optimal standby path; and when a link fault exists, dynamically searching a target recovery path of the current damaged service according to the topological structure and link weight information of at least one optimal standby path corresponding to the current damaged service in all services based on a preset improved Monte Carlo tree search algorithm, and outputting the target recovery path. Therefore, the problems of poor timeliness and low service recovery rate in the existing optical network recovery technology are solved, and the performance and efficiency of network fiber breakage resistance protection planning are effectively improved by comprehensively optimizing the shortest recovery time and the maximum service recovery rate.

The invention discloses a risk prediction method and device for an energy storage system, a medium and equipment. According to the method, the adaptive sampling strategy based on model-free offline learning is implemented on the preset edge platform, the monitoring data of the energy storage equipment are dynamically acquired, and then the time service module and the time protocol algorithm are used for performing time-space alignment processing on the data. And then, the processed data are input into a lightweight LSTM-GRU hybrid neural network model, and the model is optimized through sparsification of gating weights of LSTM and GRU units and adjustment of penalty terms of weight absolute values added in a loss function in a training process, so that real-time estimation of the state of the energy storage equipment cell is realized. And finally, in combination with the first monitoring data and the cell state estimation result, according to a preset electrochemical model and a three-dimensional thermal field reconstruction algorithm, predicting the thermal runaway risk possibly occurring in the energy storage system, thereby effectively solving the problem that the risk of the energy storage system cannot be accurately and efficiently predicted in the prior art.

The invention discloses a power distribution system adaptive operation method considering differentiated demands based on joint learning. The method comprises the following steps: 1, establishing a reconfigurable soft switch model; 2, constructing a combined learning agent based on a soft actor-commentator network and a dual deep Q network; 3, performing off-line learning on the reconfigurable soft switch in the established joint learning agent to realize rapid and accurate decision of the reconfigurable soft switch in a time-varying environment; and 4, proposing an adaptive control strategy under the support of the reconfigurable soft switch, including three control modes of loss reduction control, voltage optimization and demand response, and designing an adaptive switching mechanism capable of flexibly switching among multiple modes. Differential regulation and control requirements of the power distribution system are considered, and the flexibility and adaptability of the system are improved by using the reconfigurable soft switch, so that adaptive operation of the power distribution system is realized.

A method and system for anomaly detection from time-series input data. A Gaussian Mixing Model (GMM) learns distribution parameters in an offline learning stage using sample data. The data used for offline learning and for a subsequent online anomaly detection stage are time-series data collected for multiple parameters of a machine operation, such as a robot performing a repetitive set of operations. The method includes aligning the data with a known good reference data file and taking a difference from it before providing the data to the GMM. In an online anomaly detection stage, the GMM calculates a probability that each time-series data point fits the distribution, and then a log-sum calculation is performed on each data file to determine the likelihood that the file contains anomaly data.The log likelihood of the file is compared with previous values, and an alert is issued if there are statistical deviations from the historical data.

This invention discloses a method and system for intelligent fault diagnosis in power plants based on adaptive benchmark matching. The diagnostic method includes the construction of an offline benchmark model library and online adaptive benchmark acquisition. The system consists of a data acquisition and preprocessing module, an offline benchmark model library construction module, an online adaptive benchmark acquisition engine, a parallel fault analysis engine, a fusion decision and knowledge base module, and an alarm and report generation module. Both the diagnostic method and system of this invention completely abandon the traditional method of using static fixed thresholds as the diagnostic basis by combining "offline learning of personalized baselines" with "online multi-level dynamic matching." This allows the diagnostic benchmark to intelligently adjust with environmental conditions, greatly reducing false alarms caused by fluctuations in environmental factors and significantly improving diagnostic accuracy.

Disclosed is a device and method for processing tasks through model-based offline learning. The device includes: a dataset input unit configured to receive an offline dataset for offline reinforcement learning; an initialization unit configured to initialize a world model and a model generation dataset for predicting a state transition and a reward without interacting with a real environment; a model rollout unit configured to expand the offline data of the offline dataset based on the world model to generate an imagined trajectory and generate imaginary data of the model generation dataset; and a learning update unit configured to perform critic update and actor update based on the offline data and the imaginary data.

The invention provides a biologically-inspired robotic fish swimming action learning method and device, and the method comprises the steps: carrying out the extraction of fish key points based on collected biological fish swimming data, and obtaining a biologically-inspired experience database; determining a fish swimming action sequence by using a dynamic trajectory method according to the obtained positions of the joints of the robotic fish; pre-training a pre-constructed agent by using a biological heuristic experience database, and carrying out offline learning based on biological data to obtain a pre-trained agent; training a pre-trained agent based on the fish swimming action sequence, and performing online learning to obtain a robotic fish swimming action learning agent; wherein different states and action vectors are adopted in the online learning stage and the offline learning stage; and controlling the robotic fish to swim based on the robotic fish swim action learning agent. The robot fish swimming is controlled based on the robot fish swimming action learning agent, and the robot fish swimming action can be accurately controlled and optimized.

Offline teaching device (1) comprising at least one processor (2, 3, 4, 5) wherein the processor (2, 3, 4, 5): as a result of a motion program comprising several input teaching points, numerous interpolation points are generated on a motion route of a distal tool endpoint of a robot (100), the motion route being formed under the teaching points according to the motion program; Coordinates of at least one joint point of the robot are calculated when the distal tool endpoint is positioned at each of the interpolation points; and detects whether a disturbance occurs between each of the generated interpolation points and the calculated joint point and a peripheral device.

The invention relates to the technical field of education artificial intelligence, and provides a self-adaptive learning path recommendation method for artificial intelligence general recognition education, which comprises the following steps of: 1, calling a large language model to analyze unstructured course resources of AI general recognition education, mining semantic association and implicit dependency of knowledge points, and constructing a learning path; generating a space-time knowledge graph containing confidence scores, automatically capturing the latest literature and tool update of the field every 72 hours, and iteratively optimizing the topological structure of the graph; 2, deploying data acquisition points to acquire multi-source learning data of a user in real time, and calculating a dynamic cognitive state vector with knowledge points as dimensions by mastering an entropy model; the knowledge graph is automatically iterated through LLM, and the problem of update lag is solved; precise personalized recommendation is realized according to multi-dimensional data and a dynamic algorithm, and cognitive differences are adapted; visually presenting decision logic, and cracking a decision black box; the method can be migrated to multiple fields, supports multi-terminal and offline learning, exceeds an expected adaptive scene, and improves the learning efficiency and credibility.

The invention belongs to the technical field of robot intelligent control, and particularly relates to a mechanical arm tracking control method based on transferable depth incremental reinforcement learning, which effectively solves the nonlinear optimal tracking control problem of a mechanical arm in a model-free mode by constructing a transferable incremental reinforcement learning framework. The method comprises the following steps: firstly, constructing a 1-degree-of-freedom depth increment model by utilizing one-step forward data offline learning, and providing universal dynamic representation for a mechanical arm which is difficult to accurately model; furthermore, an asynchronous depth value network is designed for the one-degree-of-freedom mechanical arm, stable and rapid convergence value function approximation is achieved through a separation base layer and an adaptive layer, and a cross-mechanical-arm migration mechanism is established, so that a pre-training model and a network base layer on the one-degree-of-freedom mechanical arm can be directly migrated to a high-degree-of-freedom mechanical arm subsystem; and the system difference is compensated only by updating the adaptive layer online, so that the repeated training overhead is remarkably reduced, and meanwhile, rapid, robust and adaptive tracking control on the mechanical arms with different degrees of freedom is realized.

The invention provides a data supply acceleration system and method for single-machine to small-scale multi-GPU (Graphic Processing Unit) training scenes, and the system adopts an off-line learning and on-line inference architecture. The system comprises an offline training module, an online portrait generation module, an adaptive strategy selection module, an income inference and replacement execution module, a training cooperative scheduling module and a strategy switching control module. In the off-line stage, the system learns'retention revenue 'of data entries based on historical logs and solidifies a revenue prediction model and a strategy selector. In the online stage, the system generates a system state portrait in real time, an optimal strategy is adaptively selected according to the system state portrait, cache replacement is driven based on profit prediction, and GPU idling is reduced through a collaborative scheduling mechanism. According to the method, the model parameters are updated in an off-line manner, the operation overhead is low, the I / O waiting time and the tail time delay in the deep learning training process can be effectively reduced, and the GPU utilization rate and the overall training throughput rate are improved.

A method and system for anomaly detection from time series input data. A Gaussian mixture model (GMM) uses sample data to learn distribution parameters in an offline learning stage. The data for offline learning and for subsequent online anomaly detection phases is time series data of a plurality of parameters collected, such as machine operations of a robot performing a repeated set of operations. The method includes aligning the data with a known good reference data file and taking a difference from the reference data file prior to providing the data to the GMM. In the online anomaly detection stage, the GMM calculates the probability that each time series data point conforms to the distribution, and performs logarithmic summation calculation on each data file to determine the likelihood that the file contains abnormal data. The file logarithm likelihood is compared to previous values and an alert is issued when there is a statistical difference from historical data.

The invention discloses an unmanned aerial vehicle flight activity full-process task scheduling optimization method, and relates to the technical field of flight activity task scheduling. Historical data is utilized to construct a state vector in an offline stage, and the state vector is composed of task urgency, geographical distribution, current weather, obstacle probability and remaining power; meanwhile, a reward function is constructed according to the task completion time limit, the flight safety cost and the resource consumption cost, when the task is started, field real-time sensor data is introduced, the initial state obtained offline is corrected, a new state vector is formed, the optimal action is output through the deep Q network, and a task allocation scheme reflecting the current environment condition is generated; the method depends on off-line learning results, real-time changes can be reflected in time, and the problem that in traditional scheduling, a static model is actually disjointed from the site is solved; and an offline deep reinforcement learning model is combined with an online real-time data feedback closed loop to realize dynamic coordination of global scheduling and local path re-planning.

The application provides an intelligent vehicle human-machine shared collision avoidance control system and method under a critical working condition, and the method comprises the following steps: collecting and transmitting the current state of the vehicle and the surrounding environment information through a sensing system; using offline learning and large-scale vehicle data to approximately solve the collision avoidance unreachable state set, and using a large amount of pre-collected real vehicle collision avoidance data to perform offline iterative updating on the Hamilton-Jacobi reachability value function and the action-value function; quantifying the collision avoidance ability and the collision avoidance intention of the driver based on the reachability value function and the action-value function, and dynamically allocating the human-machine control weight; generating the machine operation based on the reachability heuristic reinforcement learning algorithm, and combining the human-machine control weight, the driver operation and the machine operation to generate the final operation to be executed and perform the collision avoidance operation. The application can theoretically evaluate the collision risk of the vehicle state and the collision avoidance intention of the driver, and reduce the human-machine conflict to the maximum extent under the premise of ensuring that no collision occurs.

The application provides an agent reinforcement learning method and device based on an iterative policy constraint, comprising: performing policy offline learning on an agent based on a state of the agent in any application scenario; taking an optimized policy obtained through offline learning as an initial policy, and constructing an iterative policy constraint term; introducing the iterative policy constraint term on the basis of online reinforcement learning for maximizing a reward, to generate an optimized target of the agent; and performing policy online reinforcement learning on the agent based on the optimized target. Through iterative updating of the policy constraint, the application can not only avoid a decrease in policy performance in an early online fine-tuning stage of offline-to-online reinforcement learning, but also weaken the policy constraint in a later training stage, to obtain an optimal policy.

This invention discloses a fault mode-driven multi-mode intelligent anomaly diagnosis method and system for power automation systems. The method includes offline learning and online detection phases. In the offline phase, detection knowledge items, including purified historical statistical baselines and ramp-up discrimination models, are periodically generated and stored. In the online phase, the data is preprocessed and adapted, and corresponding dedicated detection operators are activated according to the indicator type, combined with the detection knowledge items for analysis. The dedicated detection operators include ramp-up detection operators for resource-related indicators, pulse detection operators for performance-related indicators, and trend change point detection operators for disk-related indicators. The ramp-up detection operators employ a parallel fusion structure of seasonality analysis and model discriminant analysis. Candidate events output by the detection are converted into alarm records after suppression judgment. This invention achieves accurate and interpretable intelligent fault diagnosis through a closed loop of "offline baseline purification, online mode adaptation, and diagnostic output."

The invention relates to the technical field of laser welding, and discloses a laser welding seam tracking and quality real-time detection system and method based on OCT image guidance. According to the method, through system calibration and offline learning, a coordinate mapping and feature database is established; during welding, the coaxial OCT is used for synchronously scanning a front groove and a rear molten pool area; tracking and correcting a welding path in real time based on the front three-dimensional point cloud; meanwhile, a dynamic feature sequence of the molten pool is extracted and analyzed; inputting the features into a pre-training machine learning model, and judging quality and predicting defects in real time; according to the judgment result, parameters such as laser power and welding speed are adjusted in a self-adaptive mode, and closed-loop control is formed; and full-process data is fed back to optimize the model. According to the invention, synchronous closed-loop control of groove tracking and internal quality detection is realized, and the welding precision and the quality reliability are improved.

The application discloses a general enhancement method for target character recognition in a motion scene and a related device, uses a network structure construction strategy of a double branch, carries out data fusion and parameter exchange through an implicit text statistical branch and an explicit text recognition branch, can control the model size without increasing the model complexity, and reduces unnecessary parameters of the network. Meanwhile, a basic backbone network is provided for subsequent downstream tasks, and a solution can be provided for text analysis in more complex scenes. The offline learning method used in the application can reduce the cost when iterating and strengthening the model. With the increase of data diversity, the model can dynamically adjust the parameters each time, so that the iteration process can adaptively extract and update the parameters of the data, can improve the convergence speed of the iteration process while ensuring the performance of the model without increasing the network complexity.

The application discloses a kind of water turbine blade topological optimization methods combining double-line learning neural network and conformal mapping.First, the water turbine blade model is mapped to two-dimensional plane by conformal mapping algorithm, generates data sample library and corresponding conformal factor.The fluid-structure interaction analysis is used to obtain the external force field and displacement field of the model nodes under specific flow field conditions.The method of the application introduces a global attention U-NET neural network, and uses the conformal factor, structure topology, external force field and displacement field as the input and output of the network, and trains the neural network.The optimization process reduces the dependence on traditional finite element analysis, improves the optimization efficiency, and realizes efficient topological optimization of the water turbine blade through the combination of offline learning and online learning, thereby improving the performance and reliability of the water turbine.

The application discloses a traction system fault detection method based on neighborhood restriction generalized autoencoder and belongs to the technical field of fault diagnosis. In view of defects such as great information loss, insufficient interpretability of potential variables and poor adaptability to nonlinear dynamic systems in existing high-speed train traction system fault detection methods, the method realizes high-precision fault detection through two stages of offline learning and online detection. In the offline learning stage, the Mahalanobis distance is used to determine a sample neighborhood set and weights, a neighborhood restriction generalized autoencoder loss function is constructed by fusing local linear reconstruction error and mutual information regularization terms, and optimal encoders and decoders are trained. Finally, normal state residuals are calculated, and a fault detection threshold is determined based on statistics. In the online detection stage, data are collected in real time and stacked data are constructed, real-time residuals are calculated by using the trained neighborhood restriction generalized autoencoder, and fault detection is realized by comparing statistics and the threshold.

The present application relates to a kind of process industry control method based on offline learning and online constraint fine tuning.Process industry control method includes the historical operation data of the controlled object, constructs the offline data set including state variable, control variable, state transition information and constraint information;Then based on the offline data set, initial control strategy is learned offline to obtain the initial strategy that meets the distribution of historical feasible control behavior;Then the initial strategy is deployed to actual control object, and a small amount of online fine tuning is carried out in combination with real-time operation data;In online control phase, the safety constraint check and correction of strategy output action are carried out, when candidate action does not meet preset safety constraint, action clipping, feasible region projection or back control are executed.The method can reduce the risk of online exploration while improving the adaptability of control system to working condition change and external disturbance, and can balance control accuracy, stability and safety, and is suitable for closed-loop control of water tank liquid level system and other process industry objects.

Example implementations described herein are directed to systems and methods for generation and deployment of automated and autonomous self-learning machine learning models, which can include generating a predictive model and a prescriptive model through an offline learning process at a first system; controlling operations of a second system through deploying the predictive model and the prescriptive model to the second system; and autonomously updating the predictive model and the prescriptive model from feedback from the second system through an online learning process while the prescriptive model and the predictive model are deployed on the second system.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for offline learning using a reward prediction model. One of the methods includes obtaining robot experience data; training a reward prediction model from a first subset of the robot experience data, the reward prediction model receiving a reward input comprising an input observation and generating a reward prediction as output, the reward prediction being a prediction of a task-specific reward corresponding to a particular task assigned to the input observation; processing experiences in the robot experience data using the trained reward prediction model to generate, for each of the processed experiences, a respective reward prediction; and training a policy neural network from (i) the processed experiences and (ii) the respective reward predictions of the processed experiences.

The invention discloses a traction system fault detection method based on a neighborhood limitation generalized auto-encoder, and belongs to the technical field of fault diagnosis. Aiming at the defects of large information loss, insufficient potential variable interpretation, poor adaptability to a nonlinear dynamic system and the like in the existing high-speed train traction system fault detection method, the method provided by the invention realizes high-precision fault detection through two stages of offline learning and online detection. An off-line learning stage: determining a sample neighborhood set and a weight by using a mahalanobis distance, fusing a local linear reconstruction error and a mutual information regular term to construct a neighborhood limitation generalized auto-encoder loss function, and training to obtain an optimal encoder and an optimal decoder; and finally, calculating a normal state residual error, and determining a fault detection threshold value based on the statistical magnitude. And an online detection stage: collecting data in real time and constructing stacked data, calculating a real-time residual error by using the trained neighborhood limitation generalized auto-encoder, and realizing fault detection by comparing a statistical magnitude with a threshold value.

This invention relates to a DRL-based method for optimizing the data upload path of a wirelessly rechargeable drone, comprising: constructing an IoT communication and wireless charging scenario system; establishing a first mathematical model for the onboard battery consumption of the mission drone; establishing a second mathematical model for the data upload channel; establishing a third mathematical model for energy replenishment during the wireless charging process; establishing a fourth mathematical model for the path optimization objective; and determining the state set S, the action set A, and the reward function r. t The optimal path strategy π is obtained through offline learning using an improved Double DQN algorithm. * This invention provides a more convenient charging service for task units that assist base stations in performing data uploads in IoT communication systems. When data real-time requirements are high, the more convenient charging method significantly improves the endurance of the task units, achieving high data upload efficiency in IoT communication systems assisted by drones.

The present utility model relates to a portable stand-alone learning server with local large language model inference, integrated wireless routing infrastructure, and automatic phone-home synchronization for offline educational deployment. The portable server comprises a portable casing, a micro-computer server module, a local storage module, a learning management system, a web server, a local database server, a learning-material repository, a standalone large language model processing block, a local LLM runtime and local API module, an integrated wireless routing and access-point block, an automatic phone-home synchronization module, a local pending synchronization queue, a report generation and upload module, an update package retrieval and application module, and an update-status and auditupload module. Learner devices connect directly to the portable server through the wireless network without requiring an external router, separate access point, cloud AI platform, or internet connection. Educational content delivery, quiz operation, learner-data storage, and local AI-assisted feedback generation occur within the portable server. When no internet connection is available, generated reports, audit records, update-status records, and synchronization data are preserved in the local pending synchronization queue. When an internet connection becomes available, the automatic phone-home synchronization module initiates an outbound connection to a main server, uploads pending reports and status records, retrieves update packages and documents, and causes the update package retrieval and application module to apply updates locally to the learning management system. The arrangement permits offline learning operation with no public IP requirement while enabling later automatic report upload and synchronization with main-server updates without losing locally generated data, and without requiring inbound remote access or continuous internet connectivity.

The present invention discloses a diversified interactive agent construction system based on offline learning. The system includes: an offline loop system and an agent interaction system. The offline loop system includes a process splitting unit, a self-evolution unit, a storage processing unit, and a model calling unit. The agent interaction system includes an information sharing unit and a collaboration unit. The offline loop system uses an offline reinforcement algorithm to process the basic functions of an agent under the condition of a single agent's independence, including thinking function, perception function, and execution function. The agent interaction system uses an interactive adjustment learning algorithm to enhance the information in the offline loop system, and uses a distance contraction adjustment algorithm to improve the collaboration efficiency of the agent interaction system, and conducts information interaction with other single agents to effectively complete complex tasks. The present invention realizes efficient learning of agents in an offline state and improves the data processing efficiency of diversified agents in completing complex tasks.

The invention relates to the field of geotechnical engineering intelligent calculation, and discloses a PINO-based multilayer sludge vacuum loading consolidation prediction method and system, and the method comprises the steps: defining a parameter space, generating a working condition combination, solving and calculating a high-fidelity numerical model in batches, and obtaining and making a data set; determining input and output data structures of a PINO model, taking the PINO as a reference, combining a high-fidelity numerical solution with the PINO, carrying out offline learning pre-training of the PINO based on a data set, and realizing basic operator learning under physical law constraint; pINO instantiation fine tuning is carried out on the basic operator by using a PINN, so that the general solid operator is rapidly adapted to specific engineering parameters, natural transition from operator-level prediction to engineering-level application is realized, and real-time prediction and engineering optimization are carried out. According to the method, millisecond-level consolidation prediction can be realized in a hundred million-level parameter space.

The application discloses a kind of based on spectrum sensing's ultra-wideband radar signal interference detection method, first, the echo signal of ultra-wideband radar and interference signal are modeled to generate a large number of interference data sets for network learning, then the echo matrix after interference is transformed to frequency domain, along the direction of orientation is divided into pulse vector, and the spectrum sensing model of pulse is constructed, then the interference perception network model including offline learning network and online test network is established, the offline training of offline learning network is carried out, the detector with the length L is developed, and the online test network uses detector to perform sliding detection to the echo signal after processing, finally, the interference part in the echo detected is set to 0, complete ultra-wideband radar signal interference detection.The method of the application proposes a new spectrum sensing network, which can better adapt to the complex electromagnetic environment of ultra-wideband radar system, has the advantages of high interference detection probability, fast network learning speed and easy deployment on hardware platform.