96 results about "Online computation" patented technology

The invention discloses a multi-modal KV cache retrieval method and system based on a hybrid architecture, relates to the technical field of data processing, and aims to solve the technical problem that in the prior art, due to the adoption of a full-online computing architecture and lack of a unified cross-modal representation framework, the computing efficiency and the modal fusion effect are difficult to consider at the same time. By constructing a hybrid processing architecture of offline pre-calculation and online processing, multi-modal data is subjected to unified fragmentation coding and KV cache and semantic fingerprints of the multi-modal data are pre-calculated in an offline stage, and cache fragments are intelligently assembled and position codes are dynamically remapped based on a user query intention in an online stage; therefore, online repeated calculation of multi-modal content is fundamentally avoided, coherent alignment and efficient fusion of cross-modal semantics are achieved, the real-time response capability, the cache reuse rate and the multi-modal generation quality of the system are remarkably improved, and high-concurrency scenes can be dynamically adapted.

The invention relates to a reusable rocket landing section fuel optimal conversion time sequence online generation method and device and a medium, and the method mainly comprises the steps: simplifying the complex three-dimensional motion of a rocket vertical landing section into a one-dimensional vertical motion model; establishing a nonlinear equation set which takes the working time of the engine before and after stage conversion as an unknown quantity and enables the rocket to meet the terminal speed and terminal position constraints and the fuel consumption to be optimal; constructing and solving an approximate analytical solution as an initial value of Newton iteration; and based on the obtained initial value, carrying out iterative solution on a double-constraint nonlinear equation set formed by rocket speed and displacement terminal constraint. According to the method, a form based on analytic iteration is adopted, the fuel optimal control in the rocket landing process is realized through real-time online calculation of an engine conversion time sequence and giving of a corresponding thrust adjustment control instruction, and safe and accurate landing is ensured.

The invention discloses an unmanned aerial vehicle wall-attached disturbance suppression method based on strategy optimization, and aims to solve the problems that in the wall-attached flight or operation process of an unmanned aerial vehicle, strategy optimization needs to be updated online, strong constraints such as the minimum safety distance, the attitude angle, the speed and the acceleration are easily violated, airborne computing power is limited, and safety and real-time performance are difficult to consider at the same time. According to the invention, reachability analysis is carried out based on a dynamical model and wall geometry to train a feasible region approximate network so as to construct a safety supervisor, an initial control instruction is generated by adopting a hierarchical strategy, and optimization and correction are carried out under the constraint of a feasible control region to obtain a safety control instruction. Performing event-triggered online updating based on the intervention amount of the safety supervisor, introducing supervision constraint learning which takes a safety control instruction as a supervision target, and replacing a baseline strategy when a candidate strategy meets a switching condition; the technical effects of improving the adherence disturbance suppression capability and reducing the on-line calculation burden on the premise of meeting the safety constraint are achieved.

The invention belongs to the technical field of transformer winding temperature field rapid calculation, and discloses a transformer winding heat flow field calculation method and a related device, working condition data of a transformer winding are obtained, and the working condition data comprise winding current, winding inlet flow velocity and inlet temperature; the working condition data are processed through a trained neural network model, corresponding key physical quantity information of the transformer winding under the working condition data is obtained, and the key physical quantity information comprises temperature distribution of the transformer winding, an oil flow velocity radial component, an oil flow velocity axial component and oil flow internal pressure. The method can meet the requirement for real scene online calculation of the temperature field of the power equipment of the transformer substation, has important application value for state early warning and reliability evaluation of the transformer, and solves the problems that current mainstream multi-physics field numerical calculation software is low in calculation efficiency and difficult to provide technical support for multi-physics field online real scene display of the power equipment.

The application discloses a hand rehabilitation training system and adaptive training method based on brain-muscle coupling and force feedback, comprising a force feedback exoskeleton device, a training task interaction module, an electroencephalogram-electromyogram acquisition module, a synchronous control module, a signal preprocessing and feature extraction module, a brain-muscle coupling analysis module, an adaptive training adjustment module, and a safety constraint and degradation control module. The system is based on a virtual environment, and the user wears the force feedback exoskeleton device to complete the grasping task, and the electroencephalogram and electromyogram signals are synchronously acquired, the brain-muscle coupling indexes of the trial and round are calculated, and the training task difficulty parameters and the force feedback intensity parameters are updated based on the indexes, and the parameter updating process is safety constrained, and the degradation control is executed in abnormal conditions. The application realizes the closed-loop adaptive adjustment of the hand rehabilitation training task difficulty and the force feedback intensity with controllable online calculation, and is helpful to improve the training adaptability and the central-peripheral collaborative training effect.

The invention discloses a judgment feedback method for the selective ore grinding dissociation effect of a stirring mill and a computer system.The judgment feedback method comprises the following steps that S1, a comprehensive evaluation mathematical model Mb of the selective ore grinding dissociation effect is constructed; s2, determining a threshold value T of the mathematical model Mb according to historical field production data or tests; when the Mb value calculated in real time is lower than a threshold value T, the system judges that the current selective ore grinding dissociation effect is poor, and a regulation and control instruction is automatically triggered; and when the Mb value calculated in real time is continuously higher than the threshold value T and is kept stable, the system judges that the current ore grinding effect is good and excessive energy is input, and an energy-saving regulation and control instruction is triggered. According to the method, the comprehensive evaluation value of the ore grinding effect is calculated on line by establishing a comprehensive mathematical model formula Mb, a threshold triggering feedback mechanism is set, key operation parameters of the mill are automatically adjusted, and closed-loop control of evaluation-feedback-regulation is formed.

The invention discloses an anti-parameter disturbance bidirectional converter non-linear control method and system, and the method comprises the following steps: building a three-order non-linear dynamic model of a three-phase voltage source converter, defining the reciprocal 1 / Rc of an AC side resistor R and a DC side load resistor in the non-linear dynamic model as an unknown parameter, and calculating the parameter disturbance of the three-order non-linear dynamic model; obtaining a system dynamic equation with parameter uncertainty of the three-phase voltage source converter; designing a state transformation function, linearizing a system dynamic equation, and mapping a state variable x; s30, a tracking error variable is defined as a difference value between the converted state variable and a reference track of the state variable, and the difference value is used for quantifying the deviation between the actual state and the expected state of the system; defining and designing a parameter updating law of the three estimated values, and calculating and updating unknown parameters on line according to the tracking error variable; and designing and calculating a control input law based on the tracking error variable and the three estimated values, and proving the asymptotic stability of the closed-loop system through a Lyapunov function.

The invention discloses a multi-satellite large-scale MIMO (Multiple Input Multiple Output) downlink precoding method and a multi-satellite large-scale MIMO downlink precoding system considering power amplifier nonlinearity. According to the method, different time delays and Doppler frequency shifts of different satellite signals arriving at a user terminal are considered, and pre-compensation is carried out on each satellite sending signal; the input and output characteristics of the satellite side nonlinear power amplifier are modeled by using the polynomial model; establishing an optimization problem to jointly design a precoding vector and a receiving vector; the signal interference noise distortion ratio in the optimization problem considers interference and distortion from other satellites, interference and distortion in the satellite and a noise item; and converting a sum rate maximization problem into a weighted mean square error minimization problem for solving. And a traditional iteration updating step is further mapped into a learnable network layer structure through a deep expansion algorithm, so that the calculation complexity is reduced. Compared with the traditional method, the method provided by the invention is more in line with the actual transmission condition, can improve the system performance and reduce the online calculation overhead, and is suitable for the precoding design in a multi-satellite communication scene.

This invention discloses a monitoring method and system for the preparation process of ternary cathode materials based on RVAE. It constructs a nonlinear dynamic system model of the sintering process based on a variational autoencoder; assigns different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, derives the loss function of the nonlinear dynamic system model of the sintering process, and trains the model parameters through backpropagation; defines the statistics of the nonlinear dynamic system model of the sintering process based on the cyclic variational autoencoder, and obtains the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation; collects online data as a test set for the nonlinear dynamic system model, calculates the monitoring statistics online and compares them with the control limits to determine whether a fault has occurred. This invention can significantly improve the fault detection rate and false alarm rate, providing a strong guarantee for the stable operation of the sintering process.

The application discloses a large language model inference service key path perception scheduling method for multi-agent workflow, wherein a lightweight DAG registration center is embedded in the inference engine service layer, a standard REST interface is used to accept the task DAG registered by a client, the residual cost (cp_remaining) of each subtask node is calculated online, the scheduling priority of the inference request is converted, the priority queue built in the inference engine is driven to preferentially process the inference request on the key path, the expected cost of a random subtask with uncertain execution rounds is calculated based on a geometric distribution model, and an online cost updating mechanism based on the self-estimation of the large language model is proposed for the subtask with an undetermined cost, so that the online self-adaptation of the scheduling priority is realized through the dynamic updating of the node cost. The application keeps the backward compatibility for the ordinary inference request without carrying the DAG metadata, and can be deployed without modifying the core scheduling logic of the inference engine.

The application discloses a hybrid off-road vehicle energy management method, device, medium and product, and relates to the field of vehicle energy management. The application classifies off-road working conditions based on test data by using a K-means clustering algorithm and relative entropy to obtain a classification result, simplifies variable off-road environments into a transfer probability matrix of electric power demand under different working conditions, greatly reduces the complexity of off-road working condition classification, and realizes an energy management strategy which is adaptively changed in real time according to a driving working condition. Moreover, the application determines an optimal solution corresponding to each type of working condition by using an improved expected SARSA algorithm based on the classification result, reduces online calculation time, and further solves the problems that an energy management strategy based on traditional reinforcement learning is slow in training speed and difficult to realize real-time control of a vehicle.

The invention relates to an online calculation system, method and equipment for a steel pipe rolling tension reducing mill and a medium, and belongs to the technical field of digitization. The method comprises the following steps: inputting production plan information through a superior system or in a manual mode; importing or manually inputting parameters through a template to generate hole pattern data; after a user selects a plan, a system automatically matches corresponding hole pattern data; a user configures and adjusts CEC control strategy parameters and then triggers a calculation process, and the system returns rolling and CEC related result data; the user can choose to store the calculation result in a database for subsequent production reuse or lower-level system call. Hole pattern data can be intelligently matched, and it is ensured that the hole pattern data is highly matched with a current production plan; a hole pattern scheme is flexibly generated, rapid and efficient calculation is achieved, and the operation efficiency is remarkably improved; and historical data storage and reuse functions are provided, the production process is optimized, and repeated input and calculation work is reduced. The whole system effectively improves the production efficiency, the system flexibility and the user interaction experience.

In a method and system for viewer interaction with streaming media, a broadcast device including an operating system level broadcaster interface overlay is utilized to display a scene from a media presentation. The overlay generates image frames from the scene of the media presentation. The image frames are transmitted to a viewer device. Viewer interaction parameters are received from the viewer device and a viewer interaction is displayed on a subsequent scene of the media presentation utilizing the broadcaster interface overlay. On the viewer device, image frames of the media presentation are received from the broadcast device. An operating system level viewer interface overlay is generated on the image frames and the image frames of the media presentation are displayed. Viewer interaction parameters are generated from viewer interaction with the operating system level viewer interface overlay and transmitted to the broadcast device.

The present disclosure relates to the technical field of power system operation management, and discloses a power distribution network line loss online calculation method, comprising the following steps: performing power flow calculation by using power distribution network historical data, establishing a training set and a test set based on the calculation result; constructing a power distribution network line loss calculation model, and performing offline training on the online calculation model using the training set; testing the calculation model training effect by using the test set and outputting an online application model to calculate the power distribution network line loss. The present disclosure can accurately and timely calculate the power distribution network line loss.

1. Name of the product in this design: Electronic device with a graphical user interface for online calculation and display of agricultural algorithms. 2. Purpose of this design: To operate and display the image data results of agricultural algorithm calculations. 3. The key design features of this product are the content and layout of the user interface. 4. The image or photograph that best illustrates the design's key features: the front view. 5. Purpose of the graphical user interface: The main view is the user interface for online calculation of agricultural algorithms. The top is the navigation bar, and the left side is the algorithm list and data list. The button at the top left can switch between the agricultural algorithm list and the data list. The upper right corner displays basic data information. The bottom right side is the input interaction box, where you can select the input image and parameters, and click OK to perform the calculation. The middle section displays the algorithm's computational results in an image.

The application discloses a DCO-OFDM peak-to-average ratio suppression method and system based on multi-stage transfer learning, generates a candidate symbol set through constellation extension; the candidate symbol set is preprocessed into a feature tensor and input into a hybrid deep decision network composed of a local feature extraction module, a global information interaction module and a decision output module in cascade; the network outputs a probabilistic decision weight, and through a fully differentiable channel from the decision weight to PAPR calculation, a composite objective function containing a PAPR performance loss term and a decision entropy regularization term is designed, and a multi-stage progressive transfer learning strategy is adopted for end-to-end unsupervised training. Through the special network architecture and training strategy, the application effectively reduces the PAPR of the DCO-OFDM system, reduces the nonlinear distortion, simultaneously transfers the complex search process to the offline training stage, reduces the online calculation complexity, and improves the convergence speed and the final performance of the model.

A computer-implemented method for predicting coal quality of coal mill based on neural network is provided. The method includes: establishing a prediction model based on a neural network structure; importing cleaned coal mill data for model training; connecting the trained prediction model with a distributed control system (DCS) in real time; calculate online the running coal type of the coal mill; and determining the real-time running coal quality of the coal mill according to the calculation results.

The invention discloses an online computing unloading method and system for edge computing equipment of a graph neural network, and the method comprises the following steps: preprocessing input data, and constructing a node feature matrix of the edge computing equipment; constructing a graph convolution network model, and performing convolution processing on the graph matrix feature data to generate an unloading probability; generating a plurality of candidate adaptive unloading decisions according to the unloading probability, evaluating a weighted calculation rate of each strategy, and selecting an optimal unloading decision; and according to the generated optimal unloading decision, training and updating the model through experience playback. According to the method, the computing task is unloaded to the network edge terminal, the delay and energy consumption problems caused by cloud processing are effectively relieved, and efficient utilization of various edge computing devices in the power grid field is achieved.

The invention belongs to the technical field of computer distributed cache, and provides a dynamic adjustment efficient access cache system and a self-adaptive control method thereof.Operation data of the cache system are collected in real time, and three-dimensional feature vectors including counting heat, vector heat and CPU utilization rate are formed; a QLR-RL decision engine is utilized to perform online calculation to obtain a state S, and the state S determines an action A through linear Q value search; a caching effect generated after the action A is executed is fed back in the form of a reward R, and the TD (0) is driven to be updated online; and executing the operations of hot key splitting, TTL flashing and edge-cloud cache proportion adjustment through a hot key splitter, an effective time flashing device and an edge-cloud proportion adjuster. According to the invention, through millisecond-level dynamic TTL adjustment, the effective time of hot data can be matched with the flow change in real time, and the avalanche probability is reduced to 0; vector hotspots are perceived 30 seconds in advance, edge-cloud proportion optimization is combined, the cache hit rate is increased, and P99 access delay is reduced.

This invention relates to the field of unmanned aerial vehicle (UAV) control technology and includes the following steps: S1, establishing a discrete nonlinear dynamic model with disturbances based on quaternions; S2, designing a lightweight nonlinear model predictive controller based on real-time iteration of sequential quadratic programming; S3, constructing a lightweight MLP disturbance observer to achieve real-time estimation of six-dimensional disturbances; and S4, designing a position-torque dual-layer feedforward compensation strategy. This invention significantly reduces the online computation of NMPC by combining the SQP_RTI real-time iterative framework with a hot-start and matrix pre-storage mechanism, enabling millisecond-level real-time operation on the airborne platform. It employs a 26-dimensional feature fusion MLP network to achieve joint estimation of disturbance forces and torques, eliminating dependence on precise model parameters. The dual-layer separate feedforward design, which corrects the reference trajectory at the position layer and directly compensates for motor speed at the torque layer, avoids coupling interference from a single compensation method, achieving high-precision trajectory tracking even under external disturbances such as wind fields and sudden load changes.

The application discloses an agent control variable acceleration calculation method based on feature embedding, which comprises the following steps: step one, obtaining a low-dimensional feature embedding vector; step two, constructing an embedding-control mapping library; step three, obtaining current state information, importing the feature embedding function, and obtaining a current embedding vector; step four, searching for a candidate control variable, then calculating and outputting the control variable for the current state; and step five, incrementally updating the embedding-control mapping library, and fine-tuning and updating the feature embedding function by using accumulated data. By constructing a task-driven low-dimensional feature embedding space, the high-dimensional state is mapped into a dense vector which retains core information of the control, and the embedding-control mapping library is established. During online calculation, efficient approximate nearest neighbor search and lightweight synthesis are performed in the low-dimensional space, so that the order of magnitude of the control variable calculation speed is improved.

This disclosure provides a method for jointly computing data, an apparatus for participating in jointly computing data, a computer-readable storage medium, and a computer program product. The method includes: a first participant in the joint computing calculating a random verification share value corresponding to itself and obtaining a random verification share value corresponding to a second participant in the joint computing; the first participant obtaining an input share value corresponding to itself and performing corresponding calculations on the input share value to obtain a secret share value corresponding to itself; and the first participant performing verification calculations on the secret share value or the input share value corresponding to itself, based on the random verification share value, to obtain a verification share value corresponding to itself. This disclosure allows for the preparation of large amounts of data in advance when the number of participants is relatively fixed, thus accelerating the speed of online computing.

The invention discloses a DCO-OFDM peak-to-average ratio suppression method and system based on multi-stage transfer learning. The method comprises the following steps: generating a candidate symbol set through constellation extension; preprocessing the candidate symbol set into a feature tensor, and inputting the feature tensor into a hybrid deep decision network formed by cascading a local feature extraction module, a global information interaction module and a decision output module; the network outputs a probabilistic decision weight, designs a composite objective function containing a PAPR performance loss item and a decision entropy regularization item through a full microchannel from the decision weight to PAPR calculation, and performs end-to-end unsupervised training by adopting a multi-stage progressive transfer learning strategy. Through a special network architecture and a training strategy, the PAPR of the DCO-OFDM system is effectively reduced, nonlinear distortion is reduced, meanwhile, a complex search process is transferred to an offline training stage, the online calculation complexity is reduced, and the convergence speed and the final performance of the model are improved.

The invention discloses a vehicle trajectory prediction system and method based on fusion sparse attention. The system comprises a finer-grained TensorCore calling module, a block-based softmax online calculation module, a compressed code generation module and a data preprocessing module. The data preprocessing module is used for carrying out preprocessing analysis of data distribution on sparse node data obtained in a model calculation process; the compressed coding generation module performs compressed coding on sparse node data obtained in a model calculation process; the TensorCore calling module with finer granularity is used for completing a preliminary part of correlation in the fine granularity sparse calculation unit; and the block-based softmax online calculation module analyzes a compression coding result of the compression coding generation module in online calculation, obtains a normalization result of a correlation coefficient of a sparse graph connection relationship, and completes a vehicle trajectory prediction task. The method has the characteristics that the reasoning speed is high, the model precision can be maintained, and the consumption of video memory resources is low.

The invention provides an encoder equiazimuth trigger signal on-line anti-interference processing and verification method, which comprises the following steps: converting encoder signals into high and low level signals in real time, detecting rising edges of Z, A and B signals of an encoder in real time, waiting for XZ, XA and XB seconds after the rising edges are detected, detecting whether the corresponding Z, A and B signals are in high level or not again, and if the signals are true, judging whether the corresponding Z, A and B signals are in high level or not. And if not, determining that the pulse signal is true, otherwise, determining that the rising edge is false triggering of the interference signal, and continuing to detect the rising edges of the Z, A and B signals in real time. The method is simple in calculation logic, facilitates real-time online calculation of high-speed laws such as an FPGA and the like, can effectively filter burrs of encoder signals in real time, prevents false triggering, and greatly improves the real-time anti-interference capability of the encoder signals.

The application discloses an AC solid-state circuit breaker adaptive turn-off method and system, the core of the method is to construct a two-dimensional phase space geometric evolution manifold according to the current amplitude sequence of the fault current and the hardware differential rate signal, realize geometric extrapolation of the short-circuit current peak value containing the non-periodic component through the trajectory curvature characteristics; combined with the cumulative equivalent operation times, the radiator temperature and the standing time and other physical state parameters, input them into the endogenous physical information neural network model, define the safe operation interval with physical certainty by using the hard-coded embedded Cauer thermal network operator; and in the interval, an analytical projection algorithm based on stress cost manifold is adopted to execute constant level optimization of the optimal turn-off time by using the control barrier function. The application can effectively solve the distortion problem of the traditional model under the magnetic saturation working condition and the uncertainty of the soft constraint prediction, while ensuring the breaking reliability, significantly reducing the online calculation complexity.

An online computing system generates and uses usability heatmaps to ensure that AR content is presented in suitable locations within geographic areas. To generate a usability heatmap, the online system accesses geographic data describing a geographic area for which the online system provides AR content services and identifies a set of portions of the geographic area. The online system identifies subsets of the geographic data for each of the portions of the geographic area and computes a usability score for each of the portions. The online system generates a usability heatmap based on the computed usability scores for the portions of the geographic area and stores the heatmap in a heatmap database that stores usability heatmaps for different geographic areas. The online system may use the heatmaps in the database to generate AR content when requested by client devices.

The application discloses a large wind farm wake rapid calculation method and system based on data driving, and creatively studies wind farm wake modeling work through a technical framework of "group division-offline modeling-online calculation". According to historical wind direction and wind speed information, the wake propagation path is mined, and the mean square loss error between the wind speed obtained by model calculation and the actual wind speed is reduced as the target. The wolf swarm algorithm is used to optimize and solve the attenuation parameters of the analytical wake model. The corrected field-level wake model is obtained through offline training. Then, combined with the wind speed and direction prediction results, the unit operation state and the influence of the unit wake, the deployment and application of the online rapid calculation of the wake are realized. The efficiency and precision of real-time wake modeling calculation are greatly improved, and high-reliable data support is provided for subsequent online analysis tasks such as wind turbine load online calculation, output performance evaluation and optimization control strategy of the wind farm.

The application designs an uplink multi-user MIMO precoding transmission method based on clustering, which can adaptively adjust the codebook structure according to the transmission environment by using the statistical characteristics of the channel. The method first needs to sample the channel state information in a period of time, calculates the optimal precoding matrix using the iterative precoding algorithm, and then uses the clustering method to train the precoding codebook suitable for the uplink multi-user MIMO system on the data set, and synchronously deploys the codebook at the base station and the user end. In the transmission, the method is based on the distance between the code word and the optimal precoding matrix to greedily search the codebook. The precoding codebook designed by the method can adapt to the change of channel characteristics, and has better performance than the classic codebook. The greedy strategy for codebook search reduces the complexity of online calculation, and is more suitable for use in actual systems.

This invention discloses an online calculation method for the reliable adjustment potential of flexible loads, belonging to the technical field of flexible load adjustment methods. The online calculation method for the reliable adjustment potential of flexible loads in this invention adopts a data-driven approach, characterizing the operation process of flexible loads as a hidden Markov process, and representing uncertain disturbance events through probabilistic representations of the load state transition process. It achieves a highly reliable adjustment potential calculation by incorporating uncertain disturbance factors faced during the adjustment process, based on the existing maximum adjustment potential. This invention solves the problem of insufficient generalization in model-driven methods and the fusion of disturbance probabilities by establishing a data-driven reliable adjustment potential evaluation model, achieving a unified probabilistic representation of the flexible load operation process. This invention proposes an online evaluation method for the reliable adjustment potential of variable frequency air conditioners. Unlike evaluation methods based on macroscopic influencing factors, this invention combines microscopic probabilistic factors such as room temperature deviation and work-rest patterns to accurately evaluate the reliable adjustment potential of air conditioners under different operating conditions.