121 results about "Edge model" patented technology

The invention discloses a thermal imaging temperature rise trend early warning system based on time-space sequence prediction, and particularly relates to the technical field of thermal imaging data prediction and early warning. The thermal imaging temperature rise trend early warning system comprises an image conversion module, a fluctuation feature extraction module, an edge prediction module, an anomaly characterization module and a prediction decision module; a temperature dynamic change rate and gradient intensity are calculated, edge model prediction is carried out based on a fluctuation index combination, when the fluctuation index combination does not exceed a stable interval, a lightweight deep network model deployed at a thermal imaging acquisition end is called, and when the fluctuation index combination exceeds the stable interval, a prediction decision module determines whether to switch to a high-order multi-modal model; the space-time information extraction capability is improved by constructing the temperature evolution data body, the prediction path is dynamically controlled based on the fluctuation index combination, and the prediction stability and efficiency are improved; and the abnormal activation index and the prediction offset index are combined to realize adaptive switching of model calling, so that the accuracy and adaptability of the early warning system are enhanced.

The invention relates to the technical field of railways, and discloses a railway passenger train operation state monitoring system based on artificial intelligence, which comprises a multi-modal sensor array, an edge AI judgment platform and a cloud intelligent decision center. According to the railway passenger train operation state monitoring method and system based on artificial intelligence, 28 parameters of a contact network, a running gear and the like are collected in real time through a multi-modal sensor array, an edge AI judgment platform utilizes a cross-modal fusion algorithm to deeply mine multi-source data potential association, the fault diagnosis accuracy is improved, and the fault diagnosis efficiency is improved. The cloud intelligent decision center dynamically aggregates gradient parameters of the edge end model based on federated learning, filters abnormal nodes and updates the abnormal nodes in an hour-level period, and constructs a digital twinborn model in combination with a Spark framework to realize dynamic parameter optimization; the fire-fighting early warning fusion model is linked with the air conditioner pressure and the compartment sealing state correction threshold value through a three-stage mechanism of parameter initial judgment, visual verification and environment verification, it is ensured that the early warning response time is shorter than 200 ms, and cooperation with the overall state of the train is achieved.

The invention discloses an electronic lead seal automatic detection method applied to logistics tracking, and relates to the technical field of Internet of Things safety monitoring. The problems of packaging data loss, tampering risk increase and real-time monitoring failure caused by wireless communication signal interference, transmission delay and state updating lag of an existing electronic lead seal are solved. The method comprises the following steps: fusing multi-physical field sensing data through a space-time correlation sampling algorithm, and constructing an anti-interference characteristic matrix; evaluating communication quality based on the dynamic probability network model, and triggering a multi-path fragmentation concurrent transmission strategy to avoid signal attenuation; a hybrid reasoning model is deployed at an edge end to screen key event data, and the transmission efficiency is optimized in combination with differential coding and an IEEE 1588 clock synchronization mechanism; the cloud end adopts a space-time diagram fusion analysis model to carry out cross-modal abnormal association scoring, corrects misjudgment and updates an edge model; according to the invention, the communication reliability, the real-time transmission efficiency and the anomaly detection precision of the lead sealing state data in a complex environment are obviously improved.

The invention discloses a discrete point-based cutting force prediction method supporting any cutting edge shape, and relates to the technical field of milling machining.The discrete point-based cutting force prediction method comprises the steps that firstly, a workpiece and a tool are both expressed in a discrete model, a milling track is defined as a series of linear segments through NC codes, and a tool and cutting edge model and a Tri-Dexel workpiece model are built; secondly, a tool swept volume is obtained, and material removal is calculated; then subdividing the tool motion into smaller analysis steps to obtain the cutting condition of each turn, determining the cutting meshing area of each analysis step and blade elements participating in cutting, and calculating the nominal instantaneous undeformed chip thickness of each blade element based on the actual feeding direction, the normal vector and the compensation coefficient; then, discretizing the cutting condition into a plurality of orthogonal cutting units, and predicting the cutting force of each participating blade element by adopting a neural network trained based on simulation data; and finally, integrating the cutting forces of all participating blade elements to obtain a synthetic cutting force field of each analysis step. According to the method provided by the invention, the force field prediction capability in a variable geometry processing scene is effectively improved.

The invention discloses a large vision-language model collaborative reasoning method for a cloud side end system, which is characterized by comprising the following steps: deploying a high-performance cloud LVLM model and a cloud archive library at a cloud end; an edge LVLM model and an edge archive library are deployed at an edge end, a retrieval enhancement generation algorithm is introduced into the edge LVLM model, and the edge archive library and a cloud archive library can perform dynamic knowledge interaction; and the task scheduler performs complexity evaluation on the received query task based on a task allocation algorithm, and allocates the query task to an edge end for local processing or allocates the query task to a cloud end for processing according to a complexity evaluation result. According to the method, the reasoning efficiency in the cloud edge end system can be improved, meanwhile, high precision is kept, and the technical problem that the reasoning efficiency is low due to the problems of limited computing resources, communication window intermittency, bandwidth limitation and the like of edge equipment is effectively solved.

The invention relates to a landslide grading early warning method based on multi-modal data change characteristics. The method comprises the following steps: establishing a mountain digital twinborn body; the method comprises the following steps: deploying a multi-node sensor network in a target area, configuring an edge computing unit, collecting geological data in real time, and screening the geological data based on mountain digital twin to form effective local data; constructing a lightweight multi-modal neural network model at each node, dynamically searching hyper-parameters by using an ant colony optimization algorithm, and generating an encryption model weight update quantity packet; the central server dynamically calculates node weights according to disaster feature vectors output by the digital twins, generates a global model through weighted aggregation, and directionally distributes and updates the global model; real-time monitoring data and a model prediction result are fused, millimeter-level disaster evolution simulation is executed through a variable step size physical engine, an advanced early warning signal is triggered when a deduced prediction risk exceeds a threshold value, and the edge model adaptability, federal aggregation precision and early warning advancement are remarkably improved.

The invention belongs to the field of computer artificial intelligence and wireless communication networks, and provides an edge model scheduling method and system fusing graph nerves and a hybrid expert model, and the method comprises the steps: collecting multi-modality, hardware parameters and model performance data, constructing a graph nerve network structure according to the resource condition of edge equipment and task calculation demands, and carrying out the scheduling of the edge model. Therefore, the topological relation and conditional dependence among data, hardware and models are captured; according to the method, feature processing is carried out at the local where data is generated, so that the time delay of data transmission to a cloud or a central server is remarkably reduced, and the resource utilization and model processing efficiency of the whole system is improved; the graph neural network prediction model constructed based on the general features and the new message passing mechanism can effectively deal with the zero sample prediction problem in an unfamiliar hardware equipment environment, does not need to perform special training and optimization for each hardware equipment and generate diversified prediction results, and effectively reduces the risk of prediction result overfitting.

An automatic driving resource arrangement and task unloading method based on model segmentation comprises the steps that in an automatic driving system, each vehicle continuously generates a vehicle-mounted task, and a vehicle task calculation model makes a real-time decision according to collected sensing data; each base station carries out task unloading scheduling on a plurality of vehicles in the coverage range; the MEC server communicates with the base station, and carries out computing resource arrangement and unloading task calculation on the edge side; according to the model segmentation method, a lightweight shallow network is deployed for a vehicle to serve as a vehicle model, and a deep network is deployed for an MEC server to serve as an edge model; adopting a residual network Resnet as a backbone network to realize model segmentation; residual blocks in the Resnet are dynamically segmented by segmentation points; according to resource arrangement and task unloading, collaborative decision-making of resource arrangement and task unloading is modeled as a constraint optimization problem for minimizing long-term task completion time delay, then the constraint optimization problem is decoupled as a resource allocation and task unloading sub-problem, and a double-layer DDPG network is constructed to solve the sub-problem.

The invention discloses a real-time translation recognition system under a cloud service framework, belongs to the technical field of real-time translation, and solves the problems that an existing translation system is insufficient in real-time performance, poor in scene adaptability, weak in privacy protection, slow in model evolution and the like. The dynamic model management engine obtains adaptive slices from a model slice factory according to scenes, equipment states and network quality and distributes the adaptive slices to edges, and the adaptive slices are distributed to a cloud-side collaborative reasoning system; the cloud-side collaborative reasoning system comprises a cloud-side collaborative reasoning system, a cloud-side collaborative reasoning system, a cloud-side collaborative reasoning system and a cloud-side collaborative reasoning system; the multi-modal perception engine fuses audio, images and dialogue history to generate a structured context vector and improve translation context fitting degree, the cloud edge cooperation engine takes an edge model as a core, processes different complexity tasks in combination with a cloud end, and constructs a data closed loop by incremental learning and a federation engine to realize model optimization and privacy protection; according to the system, the real-time performance, accuracy and safety are improved through cloud edge collaboration, dynamic adaptation and continuous learning, and the system is suitable for multi-scene real-time translation.

The invention relates to the technical field of defect detection, and provides a battery film surface defect detection method based on deep learning, and the detection method comprises the following steps: S1, obtaining the image features of a battery film surface image, the image features comprising a color feature and an HSV feature; s2, extracting a battery film surface image according to a V-channel image in the HSV features, and removing an irrelevant background to obtain a first image; s3, performing non-local mean filtering and logarithm transformation on the first image in sequence to obtain a second image; s4, performing connectivity analysis on the second image to screen out defect areas and extract defect features; s5, by analyzing defect areas and features, positioning edge positions of defect features by using an edge model; and S6, detecting the edge position of the defect feature in the step S5 by adopting a sequential similarity detection algorithm, and matching the defect target of the surface image of the primary battery film. The objective of the invention is to improve the accuracy and efficiency of defect detection of the perovskite cell film.

The invention discloses a fault diagnosis method and a fault diagnosis system for an electric valve actuating mechanism. The method comprises the following steps: synchronously acquiring signals through an anti-EMI (Electro-Magnetic Interference) multi-source sensor; adopting complex Morlet wavelet packet decomposition to extract a 1.2-2.4 kHz energy entropy minimum frequency band, and calculating a kurtosis index; separating the third harmonic of the current through variational mode decomposition, and calculating the total distortion rate of the third harmonic; a graph attention network with 12-dimensional features is constructed, and weighted fusion is carried out through a multi-head attention mechanism; the lightweight CNN outputs a fault type, and when the confidence coefficient is less than 0.9, a knowledge graph rule engine is triggered; and updating a threshold value based on a historical diagnosis clustering result, and aggregating edge model parameters by federal learning. The system comprises a wafer-level micro-strain sensing layer, an FPGA accelerated edge computing layer, a cloud platform supporting federated learning, and an AR maintenance guidance and block chain evidence storage module. The early fault detection rate is improved, the false alarm rate under strong EMI is reduced, and the average repair time is shortened.

The invention provides an edge AI model-oriented automatic deployment and updating method and system, and relates to the technical field of artificial intelligence, and the method comprises the steps: carrying out the unified management of the information of edge devices needing to be deployed by a user; transmitting the change parameters of the remote server model to the target edge device based on an automatic pushing mechanism of a time period; through intelligent detection, automatic script generation and remote execution, unattended configuration of the operation environment of the edge device model is realized; by deploying a platform tool on the edge device, a configuration file is intelligently generated, and seamless deployment and optimization of a model reasoning service on the edge device are realized; and automatically generating an API interface and an API document of the edge device model.

The invention discloses a method for determining the demulsification speed of emulsified asphalt, and relates to the technical field of online detection and intelligent monitoring of road engineering materials, and by integrating a lightweight time sequence prediction model and a self-adaptive processing mechanism, the practicability and reliability of the determination of the demulsification speed of the emulsified asphalt are remarkably improved; in the aspect of portability, an edge model based on a gating circulation unit structure is adopted, so that the dependence on computing resources is reduced, the method can be operated on handheld equipment, the problem that traditional laboratory equipment is heavy is avoided, and construction personnel can conveniently and quickly deploy the equipment on site; the real-time performance is enhanced, the adaptive sliding window dynamically adjusts the window length and step length according to the data stability to ensure that the inference frequency is matched with the demulsification process, and the model can immediately respond to environmental changes such as temperature fluctuation and output a demulsification speed estimated value without obvious delay in combination with an online learning mechanism, so that the paving and compacting opportunities are guided; and construction interruption is reduced.

Disclosed is a computer-implemented method and system for training a subject-specific machine learning model to infer inherent subject features from recorded or live video data. The system preprocesses the visual and audio channels, converting audio to text, and employs multiple pre-trained extraction models to generate feature embeddings. Ground truth data is obtained to guide training, where weights are assigned to produce and combine predicted feature values. Model performance is optimized by minimizing error. The trained feature extraction models are deployed on an edge device, while the subject-specific model resides in the cloud. A lightweight edge model, derived via knowledge distillation and model compression, supports local inferencing with reduced reliance on cloud resources. Synchronization ensures iterative updates for sustained accuracy.

The invention discloses an image processing method and device, a medium and a computer program product, and relates to the technical field of semiconductor manufacturing. The image processing method comprises the following steps: acquiring target image samples of a plurality of wafer samples through a scanning electron microscope, and obtaining a first contour image label of each target image sample; based on the target image sample, a first contour image result is generated through a generator of a preset image edge model, iteration training is carried out on the preset image edge model through pixel difference loss, adversarial loss and edge supervision loss until a convergence condition is met, and a target image edge model is obtained; the generator of the target image edge model can output an edge contour image based on the target image of the target wafer. Through multi-dimensional constraints of pixel difference loss, edge supervision loss and adversarial loss, the preset image edge model is trained in the direction of generating an actual edge contour, and the accuracy of the edge contour image is improved.

The invention discloses a low-altitude aircraft intelligent identification system based on an edge network, and belongs to the technical field of intelligent identification. Comprising a data acquisition module used for acquiring time sequence sensor data; the knowledge base construction module is used for constructing a physical rule knowledge base; the penalty value generation module is used for generating a physical violation penalty value; the model generation module is used for obtaining a physical rule embedded model; and the result output module is used for outputting an aircraft intelligent identification result conforming to the physical motion law. According to the method, the physical rule knowledge base is introduced into the edge network, and the lightweight model is trained by using the dual-target collaborative loss function, so that the effect of still outputting a physically consistent aircraft recognition result under the condition of limited computing power is achieved; the problem that in the prior art, a marginal small model generates jumping, over-limit or abnormal tracks due to lack of motion law constraints is solved.

The invention relates to an intelligent dam safety state prediction method and system based on a time sequence large model, and the method comprises the steps: constructing a space-time embedding mechanism comprising dynamic graph position coding and adaptive wavelet position coding through fusing sensor time sequence data, environment factor data and unstructured text data; and cross-time and cross-dam knowledge migration is realized by adopting a multi-scale memory enhancement encoder, and potential causal factors are identified through a causal decoupling decoder, so that the model interpretability is improved. And the prediction result is processed by the physical constraint output layer to ensure that the engineering mechanics law is met. The system supports collaborative deployment of an edge end and a cloud end, the edge end processes data in real time and generates preliminary prediction, and the cloud end operates a complete model and periodically updates edge model parameters. The method is suitable for real-time prediction and grading early warning of the health state of the dam structure, and is widely applied to intelligent water and electricity, infrastructure monitoring and AI-driven predictive maintenance systems.

An edge device provides an agentic framework to manage apps and artificial intelligence (AI) models used by the apps. An app and app metadata are downloaded from a cloud of servers to the device. The app metadata describes requirements of the app for AI models to be used by the app. The agentic framework performs a search in an on-device database that stores the app metadata and model metadata of edge models installed on the device. The search is performed to determine whether one of the edge models satisfies the requirements of the app. Following the search, the agentic framework sets a given edge model already installed on the device as a target model of the app, where the target model satisfies the requirements of the app. The agentic framework then directs the app to use the target model in response to a request for service.

The invention discloses a low-altitude unmanned aerial vehicle autonomous cruise method and system based on cloud edge basic model collaboration, and the method comprises the following steps: S1, collecting an environment RGB image through an airborne monocular camera of an unmanned aerial vehicle, and carrying out the preprocessing of the image, and obtaining a preprocessed gray image; s2, a neural scheduler based on deep reinforcement learning generates a scheduling instruction according to the environment data and the network state reasoned by the navigation model at the previous moment; s3, generating a preliminary flight instruction; s4, generating an optimized flight instruction; s5, generating a structured flight instruction; and S6, the unmanned aerial vehicle executes the preliminary flight instruction or the optimized flight instruction or the structured flight instruction. According to the invention, through dynamic on-demand cooperation and intelligent scheduling of the end-edge-cloud three-level model, resource consumption and delay are substantially reduced, and high-robustness and high-safety autonomous cruise of the unmanned aerial vehicle in a complex open environment is realized.

The invention discloses a ship navigation oil consumption estimation method based on artificial intelligence, and particularly relates to the field of navigation oil consumption estimation, and the method comprises the steps: obtaining a ship navigation state, carrying out manual operation, carrying out the external environment data, carrying out the preprocessing, carrying out the dimension reduction through a feature screening network, generating navigation features, carrying out the distillation training of a lightweight model, and monitoring the actual oil consumption and an estimation error. And circularly updating. According to the ship navigation oil consumption estimation method based on artificial intelligence, through a feature screening network, the data dimension of an input model is reduced, the edge model calculation amount is reduced, the reasoning real-time performance is improved, and the system resource utilization rate and stability are guaranteed; knowledge of the first neural network model is migrated to the lightweight second neural network model, so that the fuel consumption pre-estimation reasoning process responds more quickly, and the real-time requirement is met; by collecting a new data dynamic optimization model, implementing a rollback mechanism and distinguishing short-term fluctuation and long-term degradation, it is guaranteed that the method continuously adapts to ship navigation state changes, and high estimation precision is maintained.

The invention discloses a lightweight image recognition method and system for power plant safety, and relates to the field of image recognition, and the method comprises the steps: constructing a multi-scale adaptive perception image anomaly detection model, and recognizing the violation behaviors of power plant operators in power plant real-time video data through the image anomaly detection model; constructing a multi-stage binocular stereo area distribution model to match the feature points in the binocular image of the power equipment, and calculating the safety distance between the power plant operating personnel and the electrified body based on the matching result; and constructing an edge equipment reasoning model of lightweight knowledge distillation in combination with illegal behaviors and safety distances of power plant operating personnel, and monitoring potential safety hazards of the power plant in real time by using the edge equipment reasoning model. According to the method, the teacher model with high expression ability is deployed on the cloud for training, semantic information and structural knowledge are transmitted to the student model on the edge device through the dynamic adaptation strategy, and the recognition ability of the edge model in a computing power limited scene is remarkably improved.

The invention discloses a power distribution network pseudo measurement modeling generation method, a state estimation method, devices and equipment, and the method comprises the steps: inputting a node feature matrix, an edge feature matrix and an adjacent matrix in a multi-section historical graph data set into a pseudo measurement model of a power distribution network, so as to determine a predicted pseudo measurement value; updating a target parameter of the pseudo-measurement model based on the predicted pseudo-measurement value and the real pseudo-measurement value; repeating the above steps until a trained pseudo measurement model is obtained; the process of determining the predicted pseudo measurement value through the pseudo measurement model comprises the following steps: determining a node model of each node and an edge model of each edge through each hidden layer; determining an initial graph embedding vector based on the node model and the edge model; vector embedding processing is carried out on the initial image through the image pooling layer and the full connection layer so as to determine a predicted pseudo measurement value. According to the method, the prediction pseudo measurement value is determined based on the node model and the edge model, and prediction of power injection pseudo measurement of nodes without power measurement equipment or nodes with measurement data missing is realized.

The invention discloses a multi-modal data processing method and system, computer equipment and a readable storage medium, and the method comprises the steps: carrying out the data processing of an edge-end model and a cloud model, enabling the data processing results of the edge-end model and the cloud model to approach through a semantic and feature weighted total loss function, and carrying out the optimization of the edge-end model; performing adversarial training by taking the optimized edge-end model as a generator and the cloud model as a discriminator, and further improving the edge-end model; when the improved edge-end model is superior to the cloud model, the edge-end model and the cloud model exchange roles, and the edge-end model guides parameter updating of the cloud model; and carrying out data processing on the obtained multi-modal data according to the improved edge end model. Multi-level distillation and adversarial distillation break through traditional distillation bottlenecks from longitudinal knowledge deep extraction and transverse model dynamic optimization, so that the cross-modal feature fusion precision of multi-modal power grid data is guaranteed, and the overall performance of the system is continuously improved through competitive collaborative learning of an edge cloud model.

The invention relates to the crossing field of image processing and artificial intelligence, and discloses a computer image recognition method and system based on machine learning, and the method comprises the steps: fusing a convolutional neural network with a self-adaptive attention mechanism, and constructing a multi-scale feature extraction network; a dynamic adversarial data enhancement strategy is adopted to process the input image to generate an adversarial sample, and the adversarial sample and the style migration image jointly form an enhanced data set; adopting a hierarchical transfer learning framework, initializing source domain pre-training model parameters based on meta learning, and learning a feature mapping relationship between a source domain and a target domain through a domain adaptive network; constructing a probability prediction model based on a variational auto-encoder and a Bayesian Transform, and outputting an image recognition result and a corresponding confidence index; carrying out incremental learning by adopting a federal learning algorithm based on differential privacy, and regularly issuing updated knowledge to the edge model; according to the invention, the image recognition performance and practicability are significantly improved.

The invention belongs to the technical field of image recognition, and provides a remote sensing image building recognition method and system based on multi-neural network integration, and electronic equipment. The method comprises the steps of original edge network construction, two-class standard building data set construction, multi-class model parameter training, positioning information acquisition, remote sensing image acquisition, region related information extraction, main building type determination, airspace rule matching, parameter and rule lowering, model parameter updating and image recognition. According to the invention, the original edge network is constructed, so that the requirements on model parameters and computing power are reduced; through the trained multi-class model parameters, region related information extraction and main building type determination, adaptive matching of the model parameters according to different scene types is realized, the building identification accuracy is improved, and the volume of the edge model is reduced; through airspace rule matching, movement or flight of edge equipment is limited, and potential risks of personnel and buildings are reduced.

Embodiments of the present disclosure relate to a combination of a conditional and marginal model, where the conditional model provides its conditional frame prediction as input to the marginal model. Various embodiments leverage an incremental diffusion process to insert the predicted frame by mixing it with noise and starting the diffusion process part way or at some intermediate level. Some embodiments also minimize the propagation of errors introduced in the process of video generation by recursive prediction of video frames.

The invention provides an incremental cloud learning method and system for vehicle-mounted emotion data, and the method comprises the steps: 1, collecting the multi-modal emotion data of a driver in real time through a vehicle-mounted sensor, and enabling the multi-modal emotion data to comprise a facial image, a voice signal and a physiological signal; 2, preprocessing the multi-modal data, including noise removal, feature extraction and lightweight edge model analysis, and then encrypting the multi-modal data; 3, uploading the encrypted data to a cloud end, and executing incremental learning and federated learning to update the emotion recognition model; 4, verifying the performance of the model based on a preset evaluation strategy, and issuing the updated model to the vehicle-mounted terminal through a privacy protection mechanism; and 5, implementing desensitization, encryption and compliance auditing in the full life cycle of the data to ensure the compatibility of user privacy and laws and regulations. The technical problems that in a vehicle-mounted emotion recognition system, the model precision is low, continuous learning cannot be achieved, and user privacy is difficult to guarantee are solved.

Disclosed is a computer-implemented method and system for training a subject-specific machine learning model to infer inherent subject features from recorded or live video data. The system preprocesses the visual and audio channels, converting audio to text, and employs multiple pre-trained extraction models to generate feature embeddings. Ground truth data is obtained to guide training, where weights are assigned to produce and combine predicted feature values. Model performance is optimized by minimizing error. The trained feature extraction models are deployed on an edge device, while the subject-specific model resides in the cloud. A lightweight edge model, derived via knowledge distillation and model compression, supports local inferencing with reduced reliance on cloud resources. Synchronization ensures iterative updates for sustained accuracy.

An agentic framework is provided on a device that interacts with a cloud and runs apps and artificial intelligence (AI) models. An app and app metadata are downloaded from the cloud to the device. The app metadata describes requirements of the app for AI models to be used by the app. A search is performed in a database on the device that stores the app metadata and model metadata of edge models that are installed on the device to identify a target model among the edge models that satisfies the requirements. The target model is downloaded from a collection of downloadable models in the cloud in response to a determination that the target model is not already installed on the device.

This invention discloses a large-scale visual-language model collaborative inference method for cloud-edge-device systems. Its features include: deploying a high-performance cloud-based LVLM model and a cloud-based archive in the cloud; deploying an edge-based LVLM model and an edge-based archive at the edge, wherein the edge LVLM model incorporates a retrieval-enhanced generation algorithm, and the edge archive and the cloud archive can dynamically interact in terms of knowledge; a task scheduler evaluates the complexity of received query tasks based on a task allocation algorithm, and allocates the query tasks to the edge for local processing or to the cloud for processing based on the complexity evaluation results. This invention can improve the inference efficiency within cloud-edge-device systems while maintaining high accuracy, effectively solving the technical problem of low inference efficiency caused by limited computing resources, intermittent communication windows, and bandwidth limitations of edge devices.