74results about How to "Achieve adaptive" patented technology

The invention discloses a CFD calculation method and system based on a neural network and adaptive parameter optimization, and belongs to the technical field of computational fluid dynamics. Initializing a flow field by reading calculation parameters and a grid file; solving the N-S equation, recording solving parameters and outputting an intermediate flow field; inputting the intermediate flow field into a pre-trained physical constraint neural network, sequentially applying boundary condition hard constraint and soft constraint based on control equation residual error, and outputting a corrected flow field; based on the convergence dynamic characteristics and the solving parameters, outputting a parameter adjustment amount through a strategy network of reinforcement learning training, and dynamically optimizing the solving parameters; and taking the corrected flow field as an initial flow field of the next iteration step, updating solving parameters, and circularly executing until convergence. According to the method, the physical prior is embedded into the neural network, intelligent self-adaptive regulation and control of solving parameters are realized through reinforcement learning, and the prediction precision and convergence efficiency of complex flow simulation are remarkably improved.

The invention discloses a voice interaction method, a user terminal, a server side, equipment, a medium and a product. The method comprises the steps that the user terminal acquires voice input of a user; the voice input is compressed and coded, first voice input is obtained, and the mode of compressed coding is dynamically selected according to the content of the voice input; performing voice recognition, text reasoning and voice synthesis on the first voice input to obtain a target synthesized voice; or, the first voice input is transmitted to the server side, the target synthetic voice obtained by the server side is received, in the transmission process, the first voice input is divided into a voice segment and a silent segment, and transmission of the voice segment is completed preferentially; and locally playing the target synthesized voice. Through the dynamic environment perception and self-adaptive noise reduction method, transmission delay is reduced, context understanding is enhanced through text reasoning, low-delay and high-naturalness real-time voice interaction is achieved, and user experience is improved.

This invention relates to a rocker arm nozzle for irregularly shaped spray areas, comprising a mounting assembly. The mounting assembly has an elevation adjustment component on its surface, and a nozzle assembly is mounted on top of the elevation adjustment component. The mounting assembly includes a lower tube, with a mounting base fixedly mounted at its bottom. The elevation adjustment component includes an angle adjustment ball fixedly mounted on the surface of the lower tube, and a sliding block movably mounted on the outer surface of the angle adjustment ball. In this rocker arm nozzle for irregularly shaped spray areas, a sliding track is fitted onto the track balls. Furthermore, an upper tube is rotatably connected via a connecting shaft and a docking rod, allowing the upper tube to change its elevation angle based on the lower tube via the connecting shaft. Utilizing the striking force of the rocker arm, the sliding track slides laterally across several track balls. By pre-setting the positions of each slider, the elevation angle of the nozzle assembly can be changed, thus achieving self-adaptation to irregular spray areas.

The invention provides an image mirror surface highlight removal method and device based on a gradient mask and electronic equipment. The method comprises the following steps: acquiring a mirror surface highlight image and prior semantic information thereof; based on the prior semantic information, performing mirror surface highlight detection on the mirror surface highlight image to obtain a highlight detection result; according to the highlight detection result and the gray scale and illumination information of the mirror surface highlight image, constructing a gradient mask used for representing highlight intensity change; and based on the gradient mask and the prior semantic information, performing image restoration on a highlight area of the mirror surface highlight image, and performing adaptive fusion on a restoration result and a non-highlight area in the mirror surface highlight image to obtain an output image after mirror surface highlight is removed. According to the technical scheme, by introducing the prior semantic information and the gradient mask, accurate detection and smooth restoration of the mirror surface highlight area are realized, and the visual quality of the image can be improved.

This invention relates to the field of casting technology, and more particularly to a high-strength drive wheel casting process. The steps include: Step 1, alloy material preparation; Step 2, melting and micro-alloying treatment; Step 3, mold preparation and pouring; Step 4, solidification and cooling staged control; and Step 5, heat treatment process. By optimizing the alloy composition, improving the mold structure, precisely controlling melting and pouring, and adapting the heat treatment process, this invention solves the technical problems of insufficient strength, high defect rate, and poor forming accuracy in existing drive wheels. The invention also improves the structure and adapts the cooling chamber to the process, combining a controlled atmosphere cooling chamber with a staged cooling process. It incorporates partition isolation components, waste heat recovery components, and a temperature control unit, achieving independent partitioned cooling of the casting, precise temperature control, and waste heat reuse. This solves the problems of "airflow interference, uneven cooling, and energy waste" in existing cooling chambers.

The application discloses a data-driven optimal control method for a grid-connected inverter. The method comprises the following steps: constructing a DeePC optimization controller of the grid-connected inverter on the grid side; acquiring a data matrix of the grid-connected inverter in a preset historical period and an initial trajectory after a preset adjacent period, inputting the controller, setting parameters for different control targets and types of the grid-connected inverter, outputting an optimal input data sequence after processing, generating a sinusoidal pulse width modulation wave signal and acting on the grid-connected inverter, and realizing data-driven optimal control. The method can automatically perceive and adapt to the dynamic characteristics of the actual power grid through real-time data-driven predictive control, can present or enhance the control behavior of the grid-constructing or grid-following inverter by modifying the cost parameters, can realize multi-mode flexible, optimal and coordinated control, has excellent dynamic response and steady-state performance, is beneficial to safe and stable operation, and improves the ability to adapt to complex and variable operating conditions and power grid strength.

The utility model provides a welding tool for a lengthened ladder stand. The welding tool comprises a supporting piece, a welding piece and a welding piece, wherein the supporting piece is used for bearing vertical pipes on the two sides of the ladder stand; the positioning piece is used for bearing a transverse plate in the middle of the ladder; the mounting parts are used for fixing ladder legs on two sides of the ladder; the supporting piece, the positioning piece and the mounting piece are all mounted on the bottom plate, and a working position for assembling the crawling ladder is formed among the supporting piece, the positioning piece and the mounting piece; a plurality of bottom plates are arranged, and the plurality of bottom plates with the supporting pieces, the positioning pieces and the mounting pieces are connected with one another to form the welding tool matched with the length of the crawling ladder. The beneficial effect of the utility model is that (this is the beneficial effect corresponding to the above subject).

The invention discloses a wireless network positioning method, server and system, and relates to the technical field of wireless communication and positioning navigation. The method comprises the following steps: receiving a signal sent by a to-be-positioned target node through a plurality of positioning reference nodes and extracting an original positioning parameter sequence; then, improved Z-score exception elimination based on median and median absolute deviation and adaptive attenuation coefficient smoothing based on average inter-frame hopping are carried out on the sequence, and a stable estimated value and a measurement variance are output; and finally, according to the number of the effective positioning reference nodes, adaptively switching a CHAN algorithm or a weighted least square method to carry out position calculation. The improved Z-score filtering and dynamic weight fusion technology is adopted, the defects in the prior art are overcome, the problem that a fixed algorithm is poor in robustness when the reference node topology change is positioned is solved through an algorithm self-adaptive switching mechanism, the precision and adaptability are remarkably improved in a complex scene, and meanwhile the deployment and maintenance cost is reduced.

The application discloses a kind of water meter multi-network adaptive switching method, system, terminal and storage medium.The method is applied to the intelligent water meter integrated at least two networks;Method includes: periodically detecting the performance data of each network of intelligent water meter;Performance data includes: one or more data in signal strength, packet loss rate and communication delay;According to the performance data of each network, it is judged whether to switch the original network, and after confirming the switching of the original network, each standby network is switched priority ordering;Before switching the original network, the data that is not transmitted successfully is cached to local, and breakpoint information is recorded;Based on the switching priority ordering result, attempt to connect standby network, and based on breakpoint information, carry out breakpoint resume through the standby network connected successfully.The application realizes the self-adaptation, zero loss, low-power consumption communication of water meter in weak signal environment.It is especially suitable for water meter data acquisition in complex environment such as remote mountainous area, signal blind area.

The invention provides an intelligent temperature control method, system and equipment for a high-temperature purification furnace, and relates to the technical field of intelligent temperature control, and the method comprises the steps: carrying out the multi-modal state tracking of the high-temperature purification furnace through a three-dimensional perception network, and obtaining multi-modal time sequence perception data; performing time sequence change feature extraction of the multi-modal time sequence perception data based on a sliding time window to obtain K impurity dynamic feature vectors; after K single-dimensional temperature control strategies are matched in the feature strategy mapping space, conflict resolution fusion is carried out, and a combined temperature control strategy is output; and in the dynamic temperature control execution process, on-line self-adaptive updating of the combined temperature control strategy is executed based on the purification evolution deviation until the material purity reaches the semiconductor grade standard. The technical problems that in the prior art, due to the fact that a single-point temperature sensor or a single modal parameter is generally adopted for temperature control over the high-temperature purification furnace, the sensing capacity for multi-factor changes is insufficient, temperature control adjustment lags or fails, and finally the material purity is affected are solved.

This application relates to an in-line capping machine suitable for bottle caps of different sizes, belonging to the technical field of packaging machinery. It includes a connecting plate, with multiple first supports fixedly connected to the bottom end of the connecting plate, and multiple second supports fixedly connected to the top end of the connecting plate. A cap-sorting compartment is fixedly connected to the top end of each of the second supports, and an operating table is fixedly connected to the bottom end of each first support. An adjustment mechanism is fixedly connected to the front exterior of the cap-sorting compartment. A sliding compartment is fixedly connected to the bottom end of the connecting plate, and a moving block is slidably connected to the inner wall of the sliding compartment. Multiple connecting mechanisms are fixedly connected to the bottom end of the moving block. This application features a lower pressure plate that moves up and down via a threaded structure between a nut and the lower pressure plate. The lower pressure plate's height adjustment adapts to bottle caps of different heights, thereby achieving adaptation to bottle caps of different diameters and heights, and improving the in-line capping machine's ability to handle diverse bottle caps.

This invention relates to a super-resolution reconstruction method for UAV images of kiwifruit sunburn. Compared with existing technologies, it solves the problems of poor model generalization due to the lack of accurate modeling of the actual physical degradation of UAVs, and the difficulty of feature extraction networks in simultaneously considering global semantics and local high-frequency lesion details, resulting in blurred reconstruction. The invention includes the following steps: acquisition and preprocessing of UAV images of kiwifruit sunburn; generation of multi-scale paired datasets; construction of a super-resolution reconstruction model for kiwifruit sunburn images; training of the super-resolution reconstruction model for kiwifruit sunburn images; acquisition and preprocessing of UAV images; and super-resolution reconstruction of UAV images. This invention significantly eliminates boundary artifacts in image processing by constructing a multi-scale attention and spatial-frequency fusion mechanism, accurately restoring the high-frequency edge texture details of small lesions at different scales, and reducing data acquisition time.

The present application provides a kind of geologic multi-source constraint adaptive full waveform inversion method and system, constructs multiple complementary constraint norms, organically fuses multiple modal prior information, forms unified multi-source constraint framework, compared with traditional single type prior constraint, the present application can more comprehensively utilize available prior information, guide inversion to the direction that both conforms to geophysical data and geological understanding converges;While introducing adaptive regularization parameter strategy based on Bayes theory, realize the adaptive update of weight factor, avoid the subjectivity and randomness caused by the dependence of regularization parameter on experience selection in traditional method, make the inversion result more stable, and do not need tedious manual parameter adjustment, can dynamically adjust the constraint strength according to the data noise level, enhance the guiding role of prior constraint in low signal-to-noise ratio, more rely on observation data in high signal-to-noise ratio;At the same time, through the combination of model compression and forward simulation calculation, the calculation efficiency is greatly improved.

The application relates to the technical field of medical image processing, in particular to a small kidney stone automatic segmentation and three-dimensional positioning method based on a CT image. The method determines local saliency according to the gray level difference between the inner layer analysis area and the outer layer background area of pixels and the gray level fluctuation degree of the outer layer, filters out candidate points, obtains the structure regularity of the candidate points according to the radial monotonicity and direction consistency of the gray level distribution in the candidate point window, divides a plurality of concentric annular zones with the candidate points as the center, calculates the annular zone gray levels of the annular zones, determines the scale matching degree according to the annular zone gray level variation trend and the relative concentration degree of the annular zone gray level difference, determines the comprehensive confidence degree according to the saliency, the regularity and the matching degree, adjusts the contrast parameter of the level set edge indication function according to the comprehensive confidence degree, drives the level set evolution by using the adjusted function, obtains the segmentation contour and the three-dimensional space positioning of the target, and thus the weak edge missing detection and the false positive rate of the target are reduced.

The application provides a tracking control method, device and equipment of an adaptive hybrid path, the tracking control method inputs system state information into an unmanned vehicle plane kinematics model, obtains deviation information of the unmanned vehicle relative to a reference path and curvature information of a reference path point; for a lateral control layer, inputs a plurality of candidate front wheel steering angles and the deviation information into a steering angle optimization model, obtains a candidate front wheel steering angle with minimum cost, and inputs the candidate front wheel steering angle into a front wheel steering angle correction model, outputs a steering angle residual correction amount, and superimposes the candidate front wheel steering angle and the residual correction amount to generate a target front steering angle; for a longitudinal control layer, inputs the curvature information and the target front steering angle into a target vehicle speed planning model, obtains a target vehicle speed; inputs a current vehicle speed and the target vehicle speed into an adaptive speed regulation relationship model, obtains a longitudinal acceleration, and the safety of path tracking control of the unmanned vehicle is improved in the aspects of tracking accuracy and steering smoothness under complex curvature road conditions.

This application discloses a power module control method and device based on artificial intelligence, relating to the field of power control technology. The method includes: collecting sample operating parameters and corresponding operating states of the power module to construct deep learning samples; training a power module operation monitoring model using an improved deep learning algorithm; collecting real-time operating parameters and using the monitoring model to identify the target operating state; mapping the real-time parameters and target state into a state space input to a reinforcement learning agent, and outputting control actions; controlling the power module's operation based on the actions and updating the reward function value to construct historical experience; when the online optimization cycle arrives, extracting historical experience to perform online optimization of the reinforcement learning agent; otherwise, proceeding to the next control cycle. This application accurately identifies the power supply state through deep learning, combines it with reinforcement learning for dynamic control, and utilizes a specific improved search algorithm to optimize model parameters, achieving adaptive and high-efficiency control of the power module and improving the power supply's stability and response speed.

The present application relates to the technical field of liquid crystal display backlight module detection, and discloses a backlight module light leakage determination method and system based on machine learning. The backlight module light leakage determination method based on machine learning comprises the following steps: collecting backlight module optical characteristic matrices; screening suspicious light leakage areas according to adaptive threshold segmentation of ambient light; implementing local contrast enhancement on the suspicious light leakage areas, combining multi-scale pyramid analysis to strengthen weak light leakage signals; quantifying micro-area light leakage structure complexity by means of box dimension calculation and multi-fractal spectrum analysis, and constructing a fractal feature vector; establishing a light leakage and defect mapping based on a machine learning correlation model, and outputting light leakage details and defect types after scoring. The present application combines fractal geometry theory, image enhancement technology and machine learning algorithms, realizes accurate quantification of light leakage morphology and automatic recognition of defect types, improves detection accuracy, reduces false positive rate, reduces missed detection rate, and improves detection sensitivity.

The invention belongs to the technical field of path planning, and more specifically relates to a limited space unmanned aerial vehicle bridge simulation inspection path planning method. The method comprises the following steps: sensing and subscribing point cloud information of an environment by an unmanned aerial vehicle, wherein the point cloud information comes from a laser radar, a depth camera or a fusion point cloud; filtering: carrying out point cloud noisy point self-adaptive multi-level filtering on the point cloud information, executing nine-step self-adaptive noisy point filtering, and outputting filtered point cloud and a free space F; planning and bridge-imitating inspection path generation are carried out, an initial path is generated through a narrow channel RRT *, a safe flight inspection corridor SFIC is constructed, and a middle bridge-imitating inspection path is generated in the SFIC in an optimized mode. The problems of insufficient environmental perception precision, low safety and efficiency of path planning, lack of dynamic adaptability of path planning and low path quality during the inspection process of an existing unmanned aerial vehicle in a limited space in a bridge are solved.

This application provides a user mining method and apparatus, electronic device, program product, and storage medium. The method includes: acquiring user data through public API interfaces of multiple social networking platforms and storing the acquired user data in an intelligent computing performance-enhanced file storage service deployed on a cloud computing platform; using a data processing cluster on the cloud computing platform to read the user data from the intelligent computing performance-enhanced file storage service, reconstructing features based on the user data, and storing the reconstructed feature data back to the intelligent computing performance-enhanced file storage service; constructing and training a reinforcement learning model on the cloud computing platform based on the reconstructed feature data; storing historical experience data of the agent in the intelligent computing performance-enhanced file storage service during the training process; deploying the trained reinforcement learning model to a production environment and using the reinforcement learning model to perform user mining based on real-time acquired user data.

The application discloses a kind of personalized federated learning method and system based on multi-dimensional adaptive aggregation and local dynamic representation alignment, the method in federated learning process, first by multi-dimensional adaptive aggregation mechanism, comprehensive historical local characteristics, round attenuation and model structure characteristics, to build personalized training starting point for client;Subsequently in the local training process, introduce local dynamic representation alignment mechanism, dynamically update class prototype and build contrast loss, to enhance feature discrimination ability;Further, the importance of each layer parameter of the model is evaluated by parameter importance fair quantization mechanism, and a layer importance mask is generated to guide the next round of refined aggregation.Based on the above technical solutions, the application effectively alleviates the personalized and generalization imbalance problem under the condition of data heterogeneity, while improving the model accuracy, effectively reducing the communication round and communication overhead.

The application discloses a coal furnace mixed coal parameter determination method, device and equipment, wherein the method comprises the following steps: obtaining historical power plant operation data and historical coal type data; based on the historical power plant operation data and the historical coal type data, a coal power coupling model for representing the mapping relationship between the coal quality parameters and the coal furnace performance indexes is trained; current power grid operation data and current coal furnace operation data are obtained, and based on the current power grid operation data and the current coal furnace operation data, a constraint condition is constructed; a target function is constructed based on the coal furnace performance indexes, a mixed coal parameter determination model is constructed based on the target function, the coal power coupling model and the constraint condition, the mixed coal parameter determination model is solved, and the target mixed coal parameter corresponding to the coal furnace on the power plant side is determined; and the target mixed coal parameter is used for representing the mixed coal proportion of different coal types. Through the above method, the mixed coal parameter can be quickly and accurately determined, and the power grid peak regulation demand can be adapted.

The invention discloses a method for controlling a stage device to move through an electric push rod, and the method comprises the following steps: S1, constructing a plurality of groups of device control channels, and carrying out the adaptive matching and safety binding of each group of device control channels and a target stage device based on the dynamic evaluation of device fingerprints and channel capabilities; s2, generating a visual interface for each group of distributed equipment control channels, and integrally displaying static equipment information, dynamic operation parameters and graded alarm information; s3, switching between a manual operation mode and an automatic operation mode in response to the switching instruction subjected to authority authentication; s4, configuring an equipment operation rule based on the operation requirement of the stage equipment, marshalling multiple groups of equipment control channels which need to cooperatively operate, and configuring a synchronous marshalling strategy; and S5, based on the received control instruction, driving an electric push rod with touch unlocking and zero point simulation functions, and executing accurate position motion control.

The utility model discloses a kind of pipeline connecting piece welding tool, it is related to pipeline connecting piece welding technical field, specifically a kind of pipeline connecting piece welding tool, including tool table, the inner side wall of tool table and inner bottom are all installed with clamp jaw, the inner wall of clamp jaw is installed with locking mechanism, locking mechanism includes arc plate, push block, sliding block and spring, arc plate is divided into two layers, through hole is opened in the arc plate of inner layer, push block is distributed in through hole and can move towards the direction of center, sliding slot is opened in the arc plate of inner layer, sliding slot is communicated through hole, sliding block is in sliding slot, the slope of both ends is consistent with the slope of the both sides of push block, one end of spring is fixed in the inner surface of outer layer arc plate, the other end is fixed in the surface of push block close to outer layer arc plate, by setting locking mechanism including push block, sliding block, spring and arc plate with spring pre-tightening force, self-adapting coating and flexible locking to three-way pipe outer surface during the process of clamp jaw clamping are realized.

This invention relates to the field of data security technology and provides an artificial intelligence-based data security tracing method and system. The method collects multi-dimensional operational data of industrial robots during continuous production, including continuous physical parameter sequences and discrete action command sequences; establishes a dynamic sliding window baseline model to track slow parameter drift and constructs a normal action pattern library to record high-frequency action sub-sequences; monitors baseline deviation events and rare pattern events in real time, cross-validates these two types of events on the timeline, and marks overlapping time windows that point to the same robot as suspicious contamination periods, tracing the data injection source in reverse; excludes data within suspicious contamination periods from the baseline update queue, and stores rare action sub-sequences in a blacklist pattern library. This invention can adapt to changes in equipment status, integrate multi-dimensional data cross-validation, accurately identify and trace data contamination attacks, and ensure data security in continuous industrial production.

This invention discloses a method and system for engine and automatic transmission self-adaptation. The method includes: acquiring a signal indicating that the engine electronic control unit (ECU) is ready; acquiring the current gear ratio and current gear of the transmission; setting the current gear ratio as the minimum gear ratio of the transmission and setting the current gear as the maximum gear of the transmission; adjusting the transmission operating conditions multiple times; sequentially acquiring the adjusted gear ratio and adjusted gear of the transmission; determining whether there is an adjusted gear ratio less than the minimum gear ratio of the transmission; if there is an adjusted gear ratio less than the minimum gear ratio of the transmission, setting the adjusted gear ratio as the minimum gear ratio of the transmission and setting the adjusted gear corresponding to the adjusted gear ratio as the maximum gear of the transmission, until there is no adjusted gear ratio less than the minimum gear ratio of the transmission. The engine ECU acquires the transmission gear and gear ratio to complete self-learning of the transmission, achieving engine self-adaptation to the transmission and realizing the universality of engine strategy and calibration.

The application provides a method, device, medium and program product for implementing multi-level cloud resource management, which comprises the following steps: receiving a request of a senior user for requesting a junior user to execute an action through a user request agent; obtaining a trust score of the senior user at the current time through a dynamic trust engine, and determining a policy related to the action; modeling real-time credibility of the senior user when proxying the junior user resources as a Bayesian learning process by the dynamic trust engine; the trust score is subject to Beta distribution; according to a pre-constructed objective function and constraint condition, calculating an activation vector of an optimal policy related to the operation to obtain an optimal solution; when the optimal solution is greater than or equal to a predetermined utility threshold, determining whether the permission range of the senior user contains the operation according to the activation vector, and if yes, calculating a token survival time of the operation b and issuing a corresponding dynamic proxy token. By using the above technical scheme, fine-grained and dynamic multi-level cloud resource management can be realized for the proxy behavior.

This invention discloses an air-cushion massage smart sofa, comprising: a backrest frame and an airbag structure surrounding the backrest frame. The backrest frame includes a backrest body and a massage mechanism, which is installed on the human-contact side of the backrest frame. The backrest frame has a mounting cavity corresponding to the massage mechanism, and the massage mechanism is installed within the mounting cavity. The airbag structure covers the backrest frame and avoids obstructing the massage mechanism. The smart massage sofa also includes a control unit, which is communicatively connected to both the massage mechanism and the airbag structure to coordinate the control of the massage mechanism's movement mode and the airbag structure's inflation / deflation state. Through the coordinated control of the control unit, this invention organically integrates the airbag structure and the massage mechanism. When the mechanical massage head performs deep therapeutic actions such as tapping and kneading, the surrounding airbags can simultaneously inflate to provide support.

The invention discloses a self-adaptive variable sample length mineral resource reserve estimation method, device and equipment, and relates to the cross technical field of geostatistics and intelligent mineral resource reserve estimation. The technical problems of low mineral resource reserve estimation precision and poor adaptive ability caused by equal-length preprocessing and single fixed power exponent in a traditional distance power inverse proportional method can be solved. The method comprises the following steps: determining an effective sample corresponding to each to-be-estimated block of a to-be-estimated mining area; according to the adaptive coupling coefficient, the distance attenuation coefficient, the corresponding relative distance of each effective sample, the sample length influence coefficient and the corresponding relative original sample length of each effective sample, constructing a corresponding weight item of each effective sample, and based on a particle swarm algorithm, determining a weight item parameter optimal value; and calculating target grade values of the to-be-estimated blocks according to the parameter optimal values and the weight items, and calculating the total mineral resource reserves of the to-be-estimated mining area according to the target grade values of all the to-be-estimated blocks.

The application discloses an elastic file preheating and distribution method for a multi-machine room large model, which comprises the following steps: collecting basic resource information and model file resource information of each machine room through a global resource sensing scheduler; obtaining preheating task instructions corresponding to each machine room based on the model file resource information and the running condition of the corresponding machine room through a model-resource matching strategy; based on the preheating task instructions, scanning a model cache state, preferentially selecting a same-machine-room node to transmit a model file through a P2P network, and performing storage cleaning according to a least recently used algorithm or a use frequency strategy; intercepting an access request of an inference instance to a model file through a unified file access module, and based on the position information of the model file, redirecting the access request to a P2P cache node or a backend storage device of the machine room through a routing strategy, so that transparent access to the underlying storage difference is realized.