1293 results about "Network layer" patented technology

The invention provides a target recognition model reasoning optimization method and device, and the method comprises the steps: firstly carrying out the structural analysis and sensitivity evaluation of a pre-training model, extracting the structural features of each network layer, activating the distribution features, carrying out the quantitative sensitivity scoring, and constructing a data set reflecting the hierarchical features and fault-tolerant capability; and querying a quantitative configuration knowledge base based on the data set to generate a heterogeneous quantitative strategy. Layered low-bit quantization is executed according to the strategy, and a layered weighted loss function is introduced to carry out quantization perception training, so that precision loss caused by bit width compression is effectively compensated. According to the method, through hierarchical heterogeneous quantification, the model recognition precision is preserved to the maximum extent while high compression ratio and reasoning acceleration are achieved, and particularly, the performance of a high-sensitivity layer is protected. The generated heterogeneous quantitative model remarkably reduces memory occupation and power consumption, is suitable for an edge hardware platform with limited resources, forms a set of complete automatic process from analysis and configuration to training compensation, and has good universality and engineering practical value.

The invention belongs to the technical field of intelligent engineering management, and discloses an automatic monitoring and control method of an engineering management system based on the Internet of Things. According to the method, a four-layer Internet of Things architecture comprising a sensing layer, a network layer, a platform layer and an application layer is constructed, and the four-layer Internet of Things architecture is uploaded to a cloud platform through a 5G / narrowband Internet of Things dual-mode transmission channel. A digital twinborn technology is innovatively adopted to construct a three-dimensional visual engineering model, intelligent identification and prediction of abnormal working conditions are realized through a machine learning algorithm, and a multi-stage linkage control mechanism is established. And when construction deviation or equipment failure is detected, the system automatically generates an optimization control strategy and issues the optimization control strategy to the execution terminal, so that accurate regulation and control of the construction machinery and intelligent pushing of early warning information are realized. According to the method, a manual inspection mode of traditional engineering management is broken through, full-process automatic supervision is realized through data fusion analysis and closed-loop control, the construction quality supervision precision is effectively improved by more than 30%, the safety accident rate is reduced by 50%, and the engineering management efficiency is remarkably improved.

The invention belongs to the technical field of base station equipment management, and discloses a base station equipment management method and system based on big data. The method comprises the following steps: firstly, collecting security situation data of a physical layer, a network layer and an application layer, extracting time sequence behavior characteristics, performing cross-layer correlation analysis, and constructing an equipment security credibility dynamic evaluation model; then, identifying vulnerability indexes of control nodes according to the trust attenuation curve, and constructing a cascade risk conduction model in combination with a topological connection relationship; dynamically dividing a security isolation domain based on the model and generating protection parameters and control rules; further, a differentiated security policy is configured, and a cross-domain collaborative response channel is established; and finally, dynamically adjusting the isolation domain boundary and the security policy by monitoring the security event frequency and the interception success rate in real time. According to the invention, the conversion from passive defense to active prediction is realized, and the safety protection capability and operation stability of the base station equipment are improved.

The invention discloses an unmanned aerial vehicle communication data feature extraction method and an unknown type intrusion detection method. Physical layer information and network layer information of an unmanned aerial vehicle are acquired; wherein the physical layer information comprises the position, attitude, height and command type of the unmanned aerial vehicle, and the network layer information comprises an IP address, a port number, a protocol type and packet metadata; constructing an unmanned aerial vehicle state feature vector according to the physical layer information and the network layer information of the unmanned aerial vehicle; based on the trained CNN branch network, extracting communication data features used for intrusion detection in the unmanned aerial vehicle state feature vectors; a double-branch network structure of the convolutional neural network and the time sequence convolutional network is provided, a loss function is improved, an information entropy regularization item is introduced, efficient detection and distinguishing of unknown attack samples are achieved, the network structure has the advantages of being high in calculation efficiency and light in weight, and the method is suitable for large-scale popularization and application. The method is suitable for being deployed in resource-limited edge computing scenes such as unmanned aerial vehicles.

The invention relates to a computing power resource dynamic scheduling method, device and equipment based on deep reinforcement learning and a medium thereof, and the method comprises the steps: constructing a joint state vector through real-time fusion of a network layer channel state and computing layer node load data, and driving a strategy network to generate transmission parameters and resource allocation actions of cooperative control; the code modulation parameters of the wireless transmission module and the computing resource proportion of the target node are synchronously configured in the execution layer, and dynamic task scheduling in the channel decay environment is achieved; a multi-target reward mechanism is designed to couple transmission bit error rate penalty, resource utilization efficiency and task timeliness evaluation indexes, and a reinforcement learning agent is guided to balance communication stability and computing power demand conflicts; according to the method, strategy network parameters are optimized through time difference errors, closed-loop feedback is formed in combination with channel state prediction and node load updating, the problems of network and calculation layer splitting decision, insufficient dynamic adaptability and multi-target optimization imbalance in the prior art are effectively solved, and the task scheduling success rate in the time-varying wireless environment is improved.

The invention provides an LLM reasoning-oriented heterogeneous core particle architecture simulation and search method and system, and the method comprises the steps: constructing a heterogeneous core particle joint simulation platform which integrates behavior-level modeling aiming at various core particle types, simulates the calculation and data transmission behaviors of a heterogeneous core particle architecture and power consumption area characteristics, forms a simulation architecture, and carries out the simulation of the heterogeneous core particle architecture; carrying out performance evaluation on the simulation architecture; an improved simulated annealing search strategy is adopted to explore the design space of the simulation architecture, and the strategy comprises the following steps: configuring core particles of different types or scales for core particle groups executing tasks of different layers based on calculation and memory access characteristics of different network layers of LLM by adopting packet heterogeneous search, optimizing the search process by adopting a simulated annealing algorithm integrated with Pareto frontier optimization; a hybrid parallel strategy of TP, PP, DP and EP and grouping heterogeneous search are subjected to collaborative optimization, and optimal parallelism combination and task mapping based on core particle grouping are automatically explored.

The invention discloses a three-dimensional point cloud geometric information compression method based on implicit neural representation, and the method comprises the steps: constructing a trunk structure of an implicit neural network through a plurality of sine representation network layers which are connected in series, and introducing a variable-scale position coding mechanism on this basis, the method enables a network to obtain higher geometric reduction precision while keeping a compression ratio, and comprises the following steps: (1) inputting space coordinates of divided voxels into a position coding module with adjustable scale parameters; (2) feeding a coding result into an implicit neural network constructed by a network layer based on sine representation, and outputting the occupancy probability of the voxel through an activation function; (3) in a training stage, the model continuously optimizes parameters, so that the output probability distribution is highly consistent with a real occupied label; (4) after model training is completed, a method of combining an AdaRound second-order quantization optimization strategy and quantization perception training is introduced, network weight is finely adjusted, and quantization errors are reduced; (5) in a reasoning stage, judging whether the voxel is occupied or not according to a preset threshold value, and when the prediction probability exceeds the threshold value, regarding the voxel as occupied; and (6) all voxels judged to be occupied are aggregated, and reconstruction of the geometric structure of the point cloud is completed.

The invention discloses an electric power multi-protocol data conversion method and system based on object model dynamic mapping, and the method comprises the steps: receiving an original protocol message, recognizing a protocol type, extracting a logic name as a unique identifier of equipment, extracting an object identifier as a measurement point identifier, and extracting an electric energy array as a measurement point value; according to the protocol type and the equipment unique identifier, matching a corresponding object model template from a unified object model library; mapping the equipment unique identifier, the measuring point identifier, the measuring point numerical value and the acquisition timestamp obtained by analysis to corresponding fields in a template; and generating an object model data object conforming to a standard format, packaging the object into a message, executing an encryption authentication mechanism of a network layer and a transmission layer through the power Internet of Things security access gateway, and sending the message to an Internet of Things management platform. According to the invention, standardized, automatic and secure access of different protocol terminals is realized through protocol analysis and the unified object model library.

The invention discloses a space-time deep learning-based data-free watershed runoff prediction method and system. The method comprises the following steps: collecting static geographic raster data and dynamic hydro meteorological time series data of multiple watersheds, and performing preprocessing; constructing a double-flow deep learning model fusing space and time features; training a double-flow deep learning model based on the multi-region large-sample watershed data to obtain a general hydrological model; evaluating the adaptability of each network layer in the general hydrological model to a target watershed through a layered unfreezing test, and screening out a key adaptive layer; and under a leave-one-out method cross validation framework, layered progressive unfreezing transfer learning is carried out on the general hydrological model based on the key adaptation layer, and data-free drainage basin runoff prediction is realized. According to the method, multi-basin data driving, spatio-temporal feature fusion and a transfer learning mechanism are organically combined, the conversion of a hydrological modeling norm from local adaptation to global generalization is promoted, and a new path is provided for intelligent prediction of data-free basin runoff.

The invention relates to the crossing field of supply chain management and block chain technologies, discloses a block chain-driven circulating packaging ownership tracing and reuse optimization system, and solves the problems that existing circulating packaging ownership tracing is not credible, the reuse rate is low, data sharing is barrier and settlement is low in efficiency. The system comprises a sensing layer, a network layer, a block chain layer and an application layer, wherein the sensing layer collects and packages full life cycle data; the network layer carries out preprocessing and 5G data transmission; the block chain layer adopts an alliance chain to deploy ownership change, full life cycle records and privacy protection contracts; the application layer is internally provided with a core unit for XGBoost demand prediction and genetic algorithm scheduling, and functions of ownership tracing, multiplexing scheduling and the like are realized. Implementation shows that the ownership tracing credibility is 100%, the reuse rate is improved to 90% or above, the settlement time is shortened from 3 days to 1.5 hours, and the method is suitable for circulating packaging scenes such as cold chains.

The invention relates to the technical field of pantograph-catenary system monitoring, and discloses a nonlinear pantograph-catenary system state prediction method based on a physical information neural network, and the method comprises the steps: obtaining the structural parameters of a catenary and a pantograph in a nonlinear pantograph-catenary system, and the train speed; a multi-expert Mama network based on Fourier transform is constructed; utilizing the first full connection layer to map input features formed by the structural parameters of the contact net and the pantograph and the train speed to a high-dimensional feature space to obtain first high-dimensional features; frequency domain features are extracted from the first high-dimensional features by using a plurality of cascaded multi-expert Mamba network layers, and time sequence features are modeled at the same time; fusing the spliced frequency domain feature and the time sequence feature by using a second full connection layer to obtain a second high-dimensional feature fusing the time-frequency energy and the dynamic evolution information; and performing nonlinear pantograph-catenary system state prediction according to the second high-dimensional features. According to the invention, the state prediction accuracy of the nonlinear pantograph-catenary system can be improved.

The application discloses a Ka dual-circular polarization antenna unit and a flat plate array antenna, which comprises, from top to bottom, a radiation patch layer, a slot coupling layer, a waveguide transmission layer and a feed network layer; a plurality of microstrip line structures are arranged on the radiation patch layer, the microstrip line structure comprises a first microstrip line and a second microstrip line, and the first microstrip line and the second microstrip line form a 90-degree phase difference of electric field orthogonal components; the slot coupling layer comprises a second substrate, the second substrate is provided with a source collapse groove corresponding to the microstrip line structure, and the slot coupling layer and the radiation patch layer are relatively displaced under the action of an external force; the microstrip line structure corresponding to the source collapse groove is switched between a left-handed circular polarization microstrip line and a right-handed circular polarization microstrip line through the relative displacement; and the source collapse groove and the first substrate form a linear polarization feed source. According to the application, each dual-circular polarization feed source is used to form dual-circular polarization, and a polarization partition plate of a feed source waveguide is not relied on, so that the thickness of the antenna structure is reduced.

The invention relates to a fixed-wing unmanned aerial vehicle track anomaly detection method and device and a medium, and the method comprises the steps: building a radar detection noise model based on a radar equation and a signal-to-noise ratio formula; on the basis of the motion equation, simulating the motion process in normal and abnormal states to generate a real track, and adding noise to the real track by using a radar detection noise model to obtain noise-containing simulation track data; constructing a dynamic graph based on the simulated track data; inputting the dynamic graph into a space-time diagram neural network model; the time-space diagram neural network model comprises a diagram attention network layer, a long-short term memory network layer and a linear layer which are connected in sequence; and performing feature extraction and classification on the dynamic graph through a time-space diagram neural network model, and outputting a track anomaly detection result. Therefore, precise detection of tiny track abnormalities of the fixed-wing unmanned aerial vehicle, especially gliding faults and the like, is realized.

The invention relates to the technical field of IT support, and provides a time sequence processing method, device and equipment adopting a quantum pulse neural network, and the method comprises the steps: obtaining time sequence data which comprises network alarm data, network equipment performance index data and network operation and maintenance work order data; encoding the time sequence data into a first quantum state by using a quantum preprocessing layer, inputting the first quantum state into a pulse neural network layer, converting the first quantum state into a time sequence pulse sequence, and processing the time sequence pulse sequence to obtain an output result; by utilizing a quantum attention enhancement mechanism, calculating attention weight of an output result in a quantum state space, and weighting to obtain a second quantum state; and decoding the second quantum state by using the hybrid decoding layer to obtain a final prediction result. The final prediction result is used for realizing fault root cause positioning, abnormal work order identification or network service quality prediction. According to the method, the parallelism of quantum calculation and the superposition characteristic of the quantum state are utilized, and the calculation efficiency can be improved when large-scale time sequence data are processed.

The invention discloses a distributed electric power measurement anomaly detection method and system based on a graph neural network. The method comprises the steps of collecting distributed data; processing the distributed data, and recording a timestamp; converting the nonlinear distributed data into linear data; constructing a graph model by taking each metering device as a node, taking the processed distributed data as a node initial feature and taking a physical connection or communication interaction relationship between the devices as an edge; constructing a graph neural network model; inputting the graph model into a graph neural network model, extracting node space features through a graph convolutional neural network layer, constraining edge weights through a graph attention network layer, and updating node features; and inputting the output of the graph neural network into K-Means clustering, dividing a center node and other nodes, calculating the distance between the center node and the other nodes by using Euclidean distance, setting a threshold value, and if the distance is greater than the threshold value, corresponding metering equipment is abnormal. The method can accurately distinguish the abnormal power fluctuation, shortens the response time, and reduces the omission ratio.

The invention provides a supply and demand matching degree-based multi-target collaborative optimization scheduling system for a light storage building group, and the system comprises a sensing layer which is used for collecting the physical operation state data of the light storage building group in real time; the network layer is used for transmitting data to the control layer; the control layer is used for generating a global scheduling plan by taking maximization of the comprehensive supply and demand matching degree as an optimization target according to the supply and demand matching degree prediction information of the first scale and the time-of-use electricity price information of the power grid; according to the supply and demand matching degree prediction information of the second scale and the physical operation state data, performing rolling correction on the global scheduling plan to generate a correction instruction; according to the supply and demand matching degree prediction information of the third scale and the physical operation state data, performing local autonomous judgment on the correction instruction, and triggering a local response when supply and demand fluctuation is detected; and the application layer is used for displaying the data acquired by the sensing layer and / or the output data of the control layer, so that the self-balancing capability of the system is fundamentally improved.

One or more embodiments of the invention provide a model quantification method and apparatus, a storage medium and a program product. The method comprises the steps of reconstructing a weight of at least one network layer in a to-be-quantized initial model based on an orthogonal rotation matrix; performing quantitative perception training on the reconstructed model to jointly optimize the weight of the at least one network layer and the corresponding rotation matrix; the optimized weight is quantified to generate a target model, and the weight precision of the target model is lower than that of the initial model.

Some embodiments provide a method of executing a distributed application with multiple EPUs that perform computations for the distributed application. The method receives, at a set of one or more scheduling servers, API commands expressed in a first programming layer and related to computations assigned to a set of EPUs and a set of forwarding criteria specifying a manner for forwarding results of the assigned computations among the EPUs in the set. The method generates, from the API commands, a set of network configuration data expressed in a second networking layer that is for configuring forwarding elements of a network to forward the results according to the specified set of forwarding criteria. The generation includes translating programming layer identifiers of EPUs and forwarding operations to networking layer identifiers needed to forward the results to destinations in the network. The method distributes the generated configuration data to multiple forwarding elements that form a network to configure the forwarding elements.

The invention provides a photovoltaic-building group load space-time matching degree prediction system, and the system comprises a sensing layer which is used for collecting photovoltaic side output data and building group side load data; the network layer is used for transmitting the acquired data to the control layer; the control layer is used for constructing a photoelectric-building load coupling model fusing time and space dimensions based on the photovoltaic side output data and the building group side load data, executing multi-time scale collaborative prediction, and calculating a time matching degree, a space matching degree and a comprehensive space-time matching degree based on a prediction result; and the application layer is used for visually displaying the prediction result, the time matching degree, the space matching degree and the comprehensive space-time matching degree. By constructing the photovoltaic-building load coupling model fusing the space-time dimension and executing multi-scale collaborative prediction, accurate prediction of the future state of the photovoltaic output and the building load can be realized, the matching degree of the photovoltaic output and the building load in time and space is quantitatively output, and the defects that prediction is independent and the space-time matching state cannot be evaluated are overcome.

The invention discloses a vehicle-mounted system safety detection and risk assessment system, which belongs to the technical field of vehicle-mounted network safety, and comprises a vehicle-mounted terminal environment sensing module used for collecting multi-level environment information of a vehicle-mounted system, including system layer information, network layer information and safety layer information; the vehicle-mounted configuration file and communication white list detection module is connected to the vehicle-mounted terminal environment sensing module and is used for establishing and maintaining a safety white list library according to the collected system information, comparing the current configuration file, the dynamic library, the port and the certificate state of the system with the safety white list library through periodic scanning, and sending the comparison result to the vehicle-mounted terminal environment sensing module; detecting configuration abnormity and file tampering events; a vehicle-mounted process and control task analysis module; a communication abnormity detection and code identification module; and a safety evaluation and risk feedback module. According to the invention, all-around monitoring and intelligent analysis of vehicle-mounted system configuration, process behaviors and network communication can be realized, and hidden backdoors, abnormal processes and malicious communication can be effectively identified.

The embodiment of the invention provides a model parameter compression method and device for a large language model, equipment and a storage medium. The method comprises the steps of obtaining verification data and inputting the verification data into a large language model to obtain an input activation tensor received by each network layer; obtaining the reasoning confusion degree of the large language model for reasoning the verification data, and taking the minimization of the reasoning confusion degree as an optimization target to carry out iterative cutting decision to obtain a cutting decision vector; performing outlier clipping processing on the input activation tensor of the network layer based on the clipping decision vector to obtain a target activation tensor; for each network layer, calculating an importance score of a model parameter based on the target activation tensor, and pruning the initial model parameter tensor in combination with the importance score to obtain an intermediate model parameter tensor; and performing quantization processing on the intermediate model parameter tensor of the network layer to obtain a target large language model. Therefore, the storage and calculation complexity of the large language model can be reduced while the performance of the large language model is maintained.

The invention discloses a road berth charging and control system for realizing multi-level fault tolerance and a method thereof, and belongs to the technical field of intelligent traffic systems and Internet of Things. The system adopts an edge computing and distributed sensing collaborative architecture, and is composed of an edge controller (ECU) and a sensing and control unit (SCU) deployed in a berth. High availability of the system is ensured through fault-tolerant design of three core dimensions: firstly, perception layer fault tolerance dynamically adjusts data fusion weights of geomagnetism, millimeter wave radar and visual AI according to real-time environment data such as rainfall and electromagnetic interference by integrating an environment perception sub-module; secondly, network layer fault tolerance utilizes an immutable transaction log and a Saga distributed transaction compensation mechanism to realize local charging and digital RMB double offline payment when cloud connection is interrupted, and data consistency is ensured after network recovery; and finally, a hardware layer establishes a neighborhood cooperation protocol based on a signature agent command through fault tolerance, and a healthy node is allowed to act as an agent fault node through safety verification to execute an unlocking instruction. According to the invention, the problems of environmental interference, network interruption, single-point hardware failure and the like in unattended parking management are effectively solved, and the robustness and financial security of the system are remarkably improved.

The invention discloses a multi-fusion intelligent Internet of Things cloud management system, which relates to the technical field of Internet of Things, is constructed based on a digital life entity concept, and comprises an autonomous cell equipment layer, a digital body fluid network layer, a cloud evolution control layer and an application interaction layer. According to the invention, equipment self-discovery and flexible access are realized, manual configuration is not needed, and equipment compatibility and deployment efficiency are improved; routing transmission adapts to multi-dimensional factors, signal transmission delay is reduced, and transmission reliability and resource utilization rate are improved; through global evolutionary optimization and a closed-loop operation mechanism, the system autonomously adapts to a dynamic scene, and the operation performance is continuously improved; a global conflict arbitration mechanism guarantees multi-device cooperation consistency, and operation disorder caused by instruction conflicts is avoided; user target preference dynamic adaptation is supported, personalized optimization is achieved, and the application scene of the system is widened; the distributed architecture and the multi-fusion technology improve the robustness of the system, and avoid the influence of a single-point fault on the overall operation.

The application relates to the field of artificial intelligence, and discloses a model distillation method, which comprises the following steps: at a first computing node of a computing node cluster, distilling a student model by using part of the student model and part of a teacher model, and performing gradient back propagation in the distillation process in the first computing node, without relying on other computing nodes to complete the distillation of a network layer responsible by the first computing node, so that greater computing resource utilization is achieved, and the distillation process is accelerated.

Embodiments of the present disclosure provide an image processing model training method, and a system. The image processing model training method is applied to a first training unit, and comprises: receiving an image processing model sent by a second training unit, and performing model training on the image processing model by using image training data, obtaining update data of each network layer in the image processing model, the update data being used for updating a parameter of each network layer; on the basis of sensitive image data in the image training data, performing data sensitivity detection on each piece of updated data, and determining data sensitivity information corresponding to each piece of updated data; performing data desensitization on the update data by using the data sensitivity information, obtaining desensitized update data of each network layer; and sending the desensitized update data of each network layer to the second training unit for parameter updating, thereby avoiding the problem that sensitive private information may be exposed in a distributed machine learning process.

The invention discloses a food safety traceability system and method based on a block chain technology, and relates to the technical field of supply chain management, and the system comprises a data collection layer, a block chain network layer, an intelligent contract layer and an application layer. The intelligent contract layer comprises a traceability data contract, a business rule contract, a trusted transaction contract, an authority management contract and an early warning trigger contract; the method comprises the steps of supply chain main body registration and authority configuration, business rule intelligent contract deployment, traceability data acquisition and uplink, business rule automatic verification and execution, credible transaction verification and confirmation, traceability query and credible verification, and abnormity early warning and emergency response. According to the invention, automatic execution and credible transaction verification of the business rules are realized through the smart contract, so that the traceability data cannot be tampered, interconnection and intercommunication of data in each link of the supply chain are realized, real-time early warning of abnormity and rapid emergency response are supported, and the credibility, transparency and efficiency of food traceability are improved.

A method in an illustrative embodiment comprises configuring a machine learning system with a multi-layer network architecture comprising at least one neural network and one or more additional network layers, training the neural network at least in part utilizing quantum sampling performed by a quantum computing device, obtaining data characterizing a monitored system, processing at least a portion of the obtained data through at least a portion of the multi-layer network architecture of the machine learning system to generate a prediction of at least one characteristic of the monitored system from the obtained data, and executing at least one automated action relating to the monitored system based at least in part on the generated prediction. The neural network may comprise, for example, a deep belief network (DBN) that includes at least first and second restricted Boltzmann machines (RBMs) of respective first and second different types, or at least one conditional restricted Boltzmann machine (CRBM).

The invention provides a precision target optimization method and system adaptive to a variable precision arithmetic logic unit, a medium, a terminal and a program product. The method comprises the following steps: acquiring an output feature set of each group of an upper layer; the precision generation network layer generates a corresponding precision target according to the output feature set, and the ALU calculation layer generates a prediction result according to the generated precision target; the teacher model generates a reference target and a real label according to the output feature set; constructing a total loss function according to the calculated task loss, precision generation loss and adversarial loss; performing back propagation optimization on the student model based on the constructed total loss function; repeatedly and iteratively training the student model until convergence to obtain a final student model; and deploying the final student model to generate an optimal precision target corresponding to each group. According to the method provided by the invention, the fine precision adjustment of the bit granularity can be realized, the adaptive ability of the model is enhanced, and the matching degree of the precision and the task demand is improved.

The invention discloses a large model reasoning hardware accelerator based on data flow execution, and belongs to the technical field of large model reasoning hardware acceleration. Each node comprises a linear calculation kernel used for sequentially completing query / key / value projection and attention output projection in multi-head attention calculation, dimension raising projection and dimension reduction projection calculation in a feedforward neural network layer, and inverse quantization operation after each projection is output; the multi-head attention calculation kernel is used for completing loading and quantization of query / key / value vectors, multi-head division and cache management, attention score calculation, Softmax normalization, context vector generation and KV cache updating; the RLA kernel is used for completing residual addition, layer normalization, quantization / inverse quantization and nonlinear activation calculation of each layer; the three cores in the same node are located in the same SLR area, and data transmission is carried out through a hardware queue and an HLS blocking mechanism. According to the invention, high hardware utilization rate and low-delay reasoning can be realized.

The invention belongs to the technical field of electric power safety supervision, and discloses an intelligent semantic generation system and method for an electric power operation scene, and the system comprises a sensing layer which is used for collecting multi-modal data in the electric power operation scene; the network layer is used for transmitting the multi-modal data acquired by the sensing layer to the platform layer; the platform layer is used for processing, reasoning and fusing the multi-modal data, generating structured semantic information and sending the structured semantic information to the application layer; and the application layer is used for driving various business applications according to the structured semantic information. According to the invention, a perception technology based on a large model and visual fusion and a natural language processing technology are provided, a visual large model and a multi-mode large model suitable for an electric power operation scene are developed, and through deep fusion of visual perception and intelligent prediction capability, power grid construction is promoted to be upgraded from passive management to prospective intelligent prediction.