An AI model distribution and deployment system and method for cloud edge collaboration

Abstract

The application discloses a kind of AI model distribution and deployment systems and methods for cloud edge coordination, belong to artificial intelligence and edge computing cross technical field, aim at solving the problem of low AI model distribution efficiency and poor deployment adaptability in cloud edge coordination scene;Method includes: collecting multi-dimensional attribute data of edge node and transforming into multi-level capability label, according to the data variation amplitude trigger label differentiation update;Rely on the construction model demand layered description framework of capability label, convert distribution demand into quantitative query condition, through three progressive matching mechanism combined with dynamic weight adjustment to filter target edge node set;Based on the encoding data block of corresponding redundancy generated by target node network state, use distributed cache mechanism to filter adaptive cache node and prestore data block, build cache network transmission to target node;The application realizes the efficient distribution and accurate deployment of AI model in cloud edge coordination environment, improves model transmission reliability and node adaptability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and edge computing, more particularly to an AI model distribution and deployment system and method for cloud-edge collaboration. BACKGROUND

[0002] With the collaborative development of cloud computing and edge computing, the deployment of AI models on the edge side has become an important support for realizing real-time intelligent decision-making. However, the complexity of the edge environment brings many problems to model distribution and deployment. The hardware configurations of edge nodes are diverse, and the computing power and storage capacity of different nodes differ significantly. The traditional unified model distribution method cannot adapt to the hardware characteristics of various nodes, and it is easy to cause the model and the node to be mismatched. At the same time, the edge network state is highly dynamic, and problems such as bandwidth fluctuation and delay change occur frequently, which leads to interruptions and losses during model transmission, affecting the stability and efficiency of transmission. In addition, the resource state of edge nodes changes in real time, and the load fluctuates, so that the static deployment strategy cannot respond to the dynamic adjustment of node resources in time, which may cause unreasonable resource allocation, making it difficult to fully utilize node resources and guarantee the stable operation of the model. These problems jointly affect the efficient landing of AI models in cloud-edge collaborative scenarios. Therefore, in order to overcome these limitations, the present application proposes an AI model distribution and deployment system and method for cloud-edge collaboration. SUMMARY

[0003] In view of the deficiencies in the prior art, the present application aims to provide an AI model distribution and deployment system and method for cloud-edge collaboration to solve the problems of insufficient adaptability of AI model distribution, low transmission reliability and efficiency, and difficulty in realizing precise model deployment in cloud-edge collaborative scenarios.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] An AI model distribution and deployment method for cloud-edge collaboration, comprising:

[0006] Collecting multi-dimensional attribute data of edge nodes and converting it into multi-level capability labels, and triggering differential update of the capability labels according to the change amplitude of the multi-dimensional attribute data;

[0007] Relying on the multi-level capability labels, a model requirement layered description framework is constructed for converting AI model distribution requirements into quantitative query conditions corresponding to the capability labels, and a three-level progressive matching mechanism based on the capability labels is combined with a dynamic weight adjustment strategy to screen a target edge node set from the edge nodes;

[0008] Based on the target edge node network state, the encoding data block corresponding to the redundancy of each target edge node is generated, and a distributed cache mechanism is used to extract the stability characteristics of the edge node according to the capability label of the edge node, which is used to select the cache node suitable for each target edge node and pre-store the encoding data block, build a distributed cache network, and transmit the encoding data block of the AI model to the target edge node.

[0009] Specifically, the specific steps of generating the encoding data block corresponding to the redundancy of each target edge node include:

[0010] The AI model to be transmitted is processed by blocking, and the size of the encoding data block is determined; the network state attribute parameters of the target edge node are extracted according to the capability label of the target edge node, and a network state data set is formed;

[0011] The network state data set is subjected to attribute quantization processing, and the network state attribute parameters are converted into standardized network performance indexes; the redundancy of the encoding data block is determined based on the network performance indexes and a preset redundancy mapping matrix, and the network performance index interval in the redundancy mapping matrix is negatively correlated with the redundancy;

[0012] Based on the calculated redundancy, the blocked AI model is encoded to generate a mixed set of original encoding data blocks and redundant encoding data blocks;

[0013] The integrity of the encoded mixed set is checked, and an encoding parameter dynamic adjustment logic is constructed to automatically trigger the network state data set update.

[0014] Specifically, the specific steps of building a distributed cache network include:

[0015] According to the capability label information of each edge node, the physical distance from the target edge node is calculated, and the network state attribute parameters of the target edge node are extracted and quantitatively scored;

[0016] The upper threshold of the physical distance and the lower threshold of the network link quality score are set to select the cache nodes of the target edge node and construct an initial cache node set;

[0017] According to the functional requirements and performance requirements of the cache nodes, the quantization query conditions corresponding to the capability labels are generated, the initial cache node set is traversed, and a cache node subset is constructed;

[0018] The number range of the cache nodes in the cache node subset is set, and by adjusting the upper threshold of the physical distance and the lower threshold of the network link quality score, the number of the cache node subset is moved to meet the number range of the cache nodes;

[0019] An association mapping between the cache nodes and the target edge nodes is established to form a distributed cache network with the target edge nodes as the center and the cache nodes as the support.

[0020] A capability tag differential update mechanism is used to associate the cache nodes and the target edge nodes, and to automatically trigger a re-screening process to update the cache node subset.

[0021] Specifically, the conversion steps of the multi-level capability tags include:

[0022] Incremental collection strategy is adopted to collect multi-dimensional attribute data of the edge nodes in real time.

[0023] The collected multi-dimensional attribute data is standardized by an attribute standard dictionary, and the attribute standard dictionary matches and converts the multi-dimensional attribute data by a rule engine.

[0024] The standardized multi-dimensional attribute data is divided into hardware inherent attribute data, real-time state attribute data, and environment-related attribute data.

[0025] Based on the tag mapping rules, the hardware inherent attribute data is encoded and converted to generate the capability tags of the basic layer; the real-time state attribute data is calculated and converted to generate the capability tags of the dynamic layer; and the environment-related attribute data is quantized and fused to generate the capability tags of the derived layer.

[0026] The generated capability tags are subjected to consistency check, and a hierarchical storage structure is constructed to store the capability tags of the basic layer, the dynamic layer, and the derived layer.

[0027] Specifically, the tag mapping rules are rules for defining the correspondence between hardware inherent attributes and the capability tags of the basic layer, and the encoding conversion adopts a structured encoding method, and the encoding content includes hardware type identification and attribute feature code.

[0028] Generating the capability tags of the dynamic layer includes: taking the real-time state attribute data within a preset reference period of the edge node as a reference data pool, calculating the relative fluctuation coefficient of the current real-time state attribute and the reference data pool by a sliding window algorithm, and converting the relative fluctuation coefficient and the reference value into a dynamic parameter value in the interval of 0 to 1 as the dynamic layer capability tag.

[0029] Generating the capability tags of the derived layer includes: converting the unstructured environment-related attribute data into a calculable parameter by attribute quantization rules, performing feature fusion on the quantized environment-related attribute parameter and the current values of the capability tags of the basic layer and the dynamic layer, introducing a model deployment scene weight matrix, and generating a multi-dimensional comprehensive score as the derived layer capability tag by weighted calculation.

[0030] Specifically, the specific steps of differential update include:

[0031] The dynamic threshold adjustment mechanism sets the judgment threshold of the change range of the multi-dimensional attribute data by recording the cumulative running time of the edge node from startup to the present, and simultaneously counting the number of changes of the multi-dimensional attribute data within a preset monitoring period, and dynamically matching the judgment threshold corresponding to each multi-dimensional attribute;

[0032] The change difference and change ratio of the multi-dimensional attribute data are calculated, and the change range of each multi-dimensional attribute data is obtained through a preset amplitude calculation rule;

[0033] The multi-dimensional attribute whose change range exceeds the matched judgment threshold is identified and marked as a change attribute; when there is a change attribute, the dynamic parameter values associated with the change attribute in the dynamic layer capability label are updated according to the mapping relationship between the change attribute and the dynamic layer capability label;

[0034] An influence factor list of the change attribute is obtained, and the multi-dimensional comprehensive score associated with the influence factor in the derivative layer capability label is recalculated according to the influence factor list.

[0035] Specifically, the specific steps of converting the AI model distribution requirement into the quantified query condition corresponding to the capability label include:

[0036] Extract the original information of the AI model distribution requirement, and construct a structured model requirement data set;

[0037] Based on the model requirement data set, a three-level structure corresponding to the multi-level capability label is built, including core function requirement, minimum resource requirement and elastic resource requirement;

[0038] The core function requirement is used to match the hardware type identifier and attribute feature code in the basic layer capability label; the minimum resource requirement is used to associate the dynamic parameter value in the dynamic layer capability label; and the elastic resource requirement is used to interface the multi-dimensional comprehensive score of the derivative layer capability label;

[0039] For the core function requirement, the binary matching code corresponding to the hardware type identifier of the basic layer capability label is generated according to the hardware architecture requirement in the model requirement data set, and the compatibility verification bit is determined in the attribute feature code according to the software environment dependency information, forming the quantified query condition of the core function requirement;

[0040] For the minimum resource requirement, the interval threshold corresponding to the dynamic layer capability label is generated according to the resource occupation threshold in the model requirement data set, forming the quantified query condition of the minimum resource requirement;

[0041] For elastic resource demand, a weight coefficient matrix corresponding to each dimension of the multi-dimensional comprehensive score of the derived layer capability label is generated according to the model demand data set, which is used to convert the resource fluctuation tolerance requirement and comprehensive performance expectation in the model demand data set into a quantitative interval based on the multi-dimensional comprehensive score, forming a quantitative query condition of the elastic resource demand;

[0042] A field mapping relationship table of the quantitative query condition and the capability label is established through the distributed mapping rule, which is used to clearly define the association logic of the quantitative query condition and the corresponding fields of the basic layer, the dynamic layer and the derived layer capability label.

[0043] Specifically, the specific steps of screening the target edge node set include:

[0044] According to the quantitative query conditions of the core function demand and the minimum resource demand respectively, the edge nodes are traversed, the capability matching is performed according to the capability label of the edge node, and the candidate node set is generated step by step;

[0045] According to the quantitative query condition of the elastic resource, the multi-dimensional comprehensive score in the derived layer capability label of each edge node is obtained;

[0046] The multi-dimensional comprehensive score is preliminarily weighted and calculated in combination with the weight coefficient matrix, and the basic comprehensive score is obtained;

[0047] According to the attribute change trajectory of each edge node, the capability fluctuation coefficient is calculated by statistically calculating the average change amplitude of the capability label in a preset period, and the update frequency coefficient is generated according to the capability label update frequency;

[0048] The weight value of the derived layer evaluation dimension is corrected according to the fluctuation coefficient and the frequency coefficient; the basic comprehensive score is secondarily weighted and calculated, and the dynamically adjusted comprehensive score is obtained;

[0049] The edge nodes whose dynamically adjusted comprehensive scores are within the quantitative interval set by the elastic resource quantitative query condition are screened out to form the target edge node set.

[0050] Specifically, the transmission of the encoding data block of the AI model to the target edge node includes:

[0051] The generated encoding data block is split according to the shard index, and according to the association relationship table of the distributed cache network, the encoding data block is pre-distributed to each cache node in the cache node subset;

[0052] Based on the role priority sorting of the association relationship table, the main transmission source is determined; the request is sent to the main transmission source cache node;

[0053] After the main transmission source receives the request, the corresponding encoding data block is retrieved according to the shard storage index, transmitted to the temporary buffer area of the target edge node, and the encoding data block receiving state is updated through the verification matching.

[0054] When the main transmission source cannot transmit, the target edge node requests backup encoded data blocks from the sub-priority cache nodes; if all cache nodes are missing, then request for supplementary transmission;

[0055] The target edge node integrates the temporary buffer data blocks according to the shard index, verifies the integrity through overall hash check, and stores them in the persistent storage area of the target edge node.

[0056] An AI model distribution and deployment system for cloud-edge collaboration includes an attribute perception module, a capability label module, an analysis matching module, a data block generation module, a network construction module, and a collaborative transmission module.

[0057] The attribute perception module is used for collecting multi-dimensional attribute data of edge nodes; the capability label module is used for converting multi-dimensional attribute data into multi-level capability labels; the analysis matching module is used for analyzing AI model distribution requirements and screening a target edge node set; the data block generation module generates encoded data blocks based on the network state of the target edge node; the network construction module is used for screening cache nodes and establishing an associated mapping with the target edge node; and the collaborative transmission module is used for receiving, integrating, verifying, and storing encoded data blocks.

[0058] The beneficial effects of the present application are:

[0059] The present application converts multi-dimensional attribute data of edge nodes into multi-level capability labels and realizes differential update, screens target edge nodes in combination with a three-level progressive matching mechanism, generates encoded data blocks with corresponding redundancy based on the network state of the nodes, and transmits data by means of a distributed cache network, which effectively improves the adaptability of AI models and edge nodes, enhances the reliability and efficiency of data transmission, ensures that the model can be accurately and efficiently deployed in a cloud-edge collaborative scenario, and meets the demand of edge environment for real-time intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A flowchart of the AI model distribution and deployment method for cloud-edge collaboration of the present application;

[0061] Figure 2 A flowchart of converting multi-level capability labels of the present application;

[0062] Figure 3 A flowchart of differential update of the present application;

[0063] Figure 4 A flowchart of building a distributed cache network of the present application;

[0064] Figure 5 A schematic diagram of the AI model distribution and deployment system for cloud-edge collaboration of the present application. DETAILED DESCRIPTION

[0065] Referring to Figure 1 The embodiment introduces an AI model distribution and deployment method for cloud-edge collaboration, including:

[0066] Step S1: Continuously collect multi-dimensional attribute data of the edge node through the deployment of the lightweight perception component, convert the multi-dimensional attribute data of the edge node into multi-level capability labels through attribute quantization rules and label mapping rules, and trigger differential update of the capability labels according to the change amplitude of the multi-dimensional attribute data, to provide accurate and real-time capability benchmarks for subsequent node selection. The multi-dimensional attributes include hardware inherent attributes such as architecture type, computing power upper limit, and storage capacity; real-time state attributes such as current load rate, remaining memory, and network bandwidth; and environment-related attributes such as historical stability and software compatibility. The attribute quantization rules are used to convert unstructured environment-related attribute data into calculable parameters, and the label mapping rules are used to divide the labels by levels, with the basic layer corresponding to the hardware inherent attributes, the dynamic layer corresponding to the real-time state attributes, and the derived layer corresponding to the multi-dimensional comprehensive score. When the change amplitude of the multi-dimensional attribute data is lower than the matching judgment threshold, no update operation of the capability labels is performed; when the change amplitude exceeds the matching judgment threshold, i.e., there is a changed attribute, all dynamic parameter values associated with the changed attribute in the dynamic layer capability labels are updated, and the multi-dimensional comprehensive score associated with the influencing factor in the derived layer capability labels is updated, and the change trajectory is recorded for subsequent capability prediction. The lightweight perception component adopts an incremental collection strategy, and only the multi-dimensional attribute data that has changed is collected, and the multi-dimensional attribute data that has not changed directly uses the historical record collected last time, to reduce the resource occupation of the edge node.

[0067] In the embodiment, through the incremental collection strategy of the lightweight perception component, the multi-dimensional attribute data is comprehensively captured while the resource occupation of the edge node is significantly reduced; through the cooperative action of the attribute standard dictionary, the attribute quantization rules, and the label mapping rules, the complex node attributes are converted into structured and hierarchical capability labels, to ensure the standardization and consistency of the attribute description; and the differential update mechanism based on the change amplitude of the multi-dimensional attribute data not only ensures the accurate reflection of the real-time state of the node by the dynamic layer capability labels, but also provides data support for long-term prediction of the node capability through the linkage update of the derived layer capability labels and the change trajectory record by the blockchain, to finally lay a high-quality and high-timeliness capability benchmark for the rapid screening and accurate matching of the subsequent target node.

[0068] Referring to Figure 2 Preferably, the specific steps of converting the multi-dimensional attribute data of the edge node into multi-level capability labels include:

[0069] By deploying a lightweight sensing component at each edge node, multi-dimensional attribute data of the edge node is collected in real time using an incremental collection strategy, wherein the incremental collection strategy refers to collecting only the multi-dimensional attribute data that has changed, and directly using the historical record of the last collection for the multi-dimensional attribute data that has not changed;

[0070] The collected multi-dimensional attribute data is standardized by an attribute standard dictionary, which is a standardized document that includes the standard naming, data format, core dimension list, and extended dimension list of various attributes of the edge node. The standard naming is used to unify the representation of multi-dimensional attributes, such as unifying the upper limit of computing power and maximum computing power as peak computing power. The data format is used to specify the unit and numerical type of multi-dimensional attributes, such as specifying the unit of bandwidth as Mbps. The core dimension list is used to list the attributes required in the AI model distribution process, such as architecture type and real-time load rate. The extended dimension list is used to include optional attributes, such as node geographic location. The attribute standard dictionary matches and converts the multi-dimensional attribute data of the edge node through a rule engine, unifies the multi-dimensional attributes with inconsistent representations into standard naming and format, supplements the missing attributes of the edge node according to the core dimension list, and supplements the missing attributes based on the statistical mean of the same type of edge node. Attributes exceeding the core list are classified into extended dimensions for management.

[0071] The multi-dimensional attribute data that has been standardized is classified and divided into hardware inherent attribute data, real-time state attribute data, and environment-related attribute data based on pre-set feature recognition rules, wherein the feature recognition rules refer to classification logic based on the time stability characteristics and data type characteristics of the attribute data. For example, hardware inherent attribute data has long-term stability characteristics, real-time state attribute data has dynamic change characteristics, and environment-related attribute data is mostly unstructured text type data.

[0072] The hardware inherent attribute data is encoded and converted based on a label mapping rule to generate a basic layer capability label, wherein the label mapping rule refers to a rule that specifies the correspondence between hardware inherent attributes and basic layer capability labels. The encoding conversion uses a structured encoding method, and the encoding content includes hardware type identification and attribute feature code. For example, storage capacity is segmented and encoded in units of 1024 GB, and the generated basic layer capability label is associated and bound with the hardware unique identifier of the edge node to ensure the uniqueness and consistency of the basic layer capability label.

[0073] The real-time state attribute data is calculated and converted to generate the dynamic layer capability label. Specifically, the real-time state attribute data in a preset reference period of the edge node is taken as the reference data pool, wherein the preset reference period refers to a fixed time interval set for calculating the reference value. The relative fluctuation coefficient of the current real-time state attribute and the reference data pool is calculated by a sliding window algorithm. The relative fluctuation coefficient and the reference value are combined and converted into a dynamic parameter value in the interval of 0 to 1, which is taken as the dynamic layer capability label. At the same time, a time decay factor is embedded in the dynamic layer capability label, so that the weight of the recent real-time state attribute data is higher than that of the long-term real-time state attribute data, and the dynamic layer capability label is updated in real time with the change of the real-time state attribute data.

[0074] The environment-related attribute data is quantized and fused to generate the capability label of the derived layer. The unstructured environment-related attribute data is converted into a calculable parameter by attribute quantization rules, for example, the nearly 30 days of fault-free operation is converted into a stability parameter: 1.0. The quantized environment-related attribute parameter is fused with the current value of the basic layer capability label and the current value of the dynamic layer capability label. In the fusion process, a model deployment scene weight matrix is introduced. The scene weight matrix dynamically adjusts the proportion of the evaluation dimensions of the model running efficiency, stability, resource cost, etc. according to different AI model types, and generates a multi-dimensional comprehensive score by weighted calculation, which is taken as the capability label of the derived layer.

[0075] The generated capability labels of the basic layer, the dynamic layer and the derived layer are subjected to consistency check. The check process includes format compliance check and value range verification. The format compliance check is used to ensure that the expression format of the capability labels of each layer is uniform, and the value range verification is used to ensure that the values of the capability labels of each layer are within the preset interval. At the same time, the same type of capability labels of the same type of edge node are compared to ensure the consistency of the description of the capability labels within the edge node cluster. For the capability labels that do not pass the verification, the re-conversion process is triggered.

[0076] A hierarchical storage structure is constructed to store the standardized capability labels of the basic layer, the dynamic layer and the derived layer. The hierarchical storage structure includes the generation time of the capability label, the index of the corresponding multi-dimensional attribute data and the verification information. The capability labels of the basic layer are stored statically, the capability labels of the dynamic layer are stored in time sequence to record the change trajectory of the dynamic layer capability labels, and the capability labels of the derived layer are stored in association to record the influence relationship with the capability labels of the basic layer and the dynamic layer. The storage structure supports quick query and call according to the capability label level and the attribute type.

[0077] Please refer to Figure 3 , preferably, the specific steps of differential update include:

[0078] The dynamic threshold adjustment mechanism is used to dynamically match the judgment threshold of each multi-dimensional attribute. Specifically, a plurality of interval segments of the cumulative running time are pre-configured, and a hierarchical standard of attribute change times in a preset monitoring period is associated with each interval segment. Different interval segments and change time hierarchical combinations correspond to different threshold coefficients. The final judgment threshold is determined by the product of the threshold coefficient and the average amplitude of the historical change of the attribute.

[0079] The incremental comparison method is used to compare each multi-dimensional attribute data with the recently stored historical data, calculate the change difference and change ratio of the multi-dimensional attribute data, and obtain the change amplitude of each multi-dimensional attribute data through a preset amplitude calculation rule. The preset amplitude calculation rule refers to a unified calculation logic set according to the multi-dimensional attribute type. For numerical attributes, the ratio of the change difference to the absolute value of the historical data is taken as the change amplitude. For percentage attributes, the change difference is directly taken as the change amplitude. For enumerated attributes, the attribute change is converted into a corresponding quantitative amplitude value through a preset enumerated value mapping table.

[0080] The multi-dimensional attributes whose change amplitudes exceed the matched judgment threshold are identified and marked as change attributes. When there are change attributes, the update process of the dynamic layer capability label is executed. According to the mapping relationship between the change attributes and the dynamic layer capability label, the values of all dynamic parameters associated with the change attributes in the dynamic layer capability label are updated. Then, the causal analysis algorithm is used for analysis. The causal analysis algorithm takes the field identifier, change amplitude and historical associated data of the change attribute as input, and outputs a list of influence factors, such as hardware field change corresponding to hardware performance factor and network field change corresponding to network environment factor. According to the list of influence factors, the multi-dimensional comprehensive score associated with the output influence factors in the derived layer capability label is recalculated, and the rest of the dimensions remain unchanged.

[0081] When there are no change attributes, i.e. the change amplitudes of all multi-dimensional attribute data are lower than the matched judgment threshold, no update operation of the capability label is performed, and the current values and states of the dynamic layer capability label and the derived layer capability label remain unchanged.

[0082] A deployment attribute change trajectory recording process is adopted to record attribute data changes and capability label updating processes using blockchain storage technology for cases where there are changed attributes. Each record contains a block ID generated based on a node unique identifier and a timestamp, a hash value of multi-dimensional attribute data before the change, a hash value of multi-dimensional attribute data after the change, values before and after dynamic layer capability label adjustment, values before and after derivative layer capability label adjustment, and a list of impact factors output by a causal analysis algorithm. Blocks are chain-connected through the ID of the previous block to form an unalterable trajectory data set. The trajectory data set supports querying and exporting according to node ID, time range, and attribute field, and provides data support for subsequent capability label updating optimization and node capability prediction.

[0083] Step S2: Based on the multi-level capability label, a model requirement hierarchical description framework is constructed, and AI model distribution requirements are converted into quantified query conditions corresponding to the capability label. A three-level progressive matching mechanism based on the capability label hierarchy is designed, and a dynamic weight adjustment strategy is combined to screen a target edge node set from the edge nodes, providing adaptive edge node resources for accurate model distribution. The model requirements are divided into core function requirements, corresponding to the hardware inherent attribute related content such as architecture support and software compatibility in the basic layer capability label; minimum resource requirements, corresponding to the dynamic parameter values related to real-time state attributes such as real-time computing power, storage, and bandwidth in the dynamic layer capability label; and elastic resource requirements, associated with the multi-dimensional comprehensive score of the derivative layer capability label and resource redundancy prediction. The three-level progressive matching mechanism successively calls the basic layer capability label for compatibility filtering, screens edge nodes that meet the architecture requirements by comparing hardware type identifiers and attribute feature codes; calls the dynamic layer capability label to complete resource threshold screening, and judges whether the real-time resource requirements are met based on the dynamic parameter values in the interval of 0 to 1; combines the derivative layer capability label to implement multi-dimensional scoring and sorting, and determines all adaptive nodes that meet the requirements according to the multi-dimensional comprehensive score and scene weight matrix. The dynamic weight adjustment strategy adjusts the weights of the derivative layer evaluation dimensions such as model running efficiency, stability, and resource cost in real time according to the edge node attribute change trajectory and capability label updating frequency. When the edge node capability label is updated, the matching sequence refresh is automatically triggered to ensure that the screening result is synchronized with the latest capability of the edge node. In the matching process, the association analysis of historical matching records and edge node attribute change trajectories is introduced to preferentially select edge nodes with stable capability labels and successful deployment of similar models in the past.

[0084] In the embodiment, step S2 uses the attribute standard dictionary and the specification definition of the label mapping rule to make the docking of the model requirements and the edge node capabilities more targeted and accurate by associating the multi-level capability labels in depth; the three-level progressive matching mechanism makes full use of the structured and hierarchical storage features of the capability labels to improve the screening efficiency through the ordered calling of the capability label levels; the dynamic weight adjustment realized by the edge node attribute change trajectory stored by the blockchain enables the screening results to sensitively reflect the fluctuations in the edge node capabilities; and the linkage with the differentiated update mechanism of the capability labels ensures the real-time selection of the target edge node, and the whole forms a coherent logic from the edge node capability modeling to the target edge node screening, providing precise and dynamically adapted edge node resources for the AI model distribution.

[0085] Preferably, the specific steps of constructing a model requirement hierarchical description framework and converting the AI model distribution requirements into quantified query conditions corresponding to the capability labels include:

[0086] The model configuration file, dependency manifest and running instruction document are parsed through the model metadata analysis interface to extract the original information of the AI model distribution requirements, automatically identify the hardware architecture requirements, software environment dependencies, resource occupancy thresholds and performance expectation indicators involved therein, and integrate these information into a structured model requirement dataset containing field identification, data type and value range;

[0087] Based on the model requirement dataset, a three-level structure corresponding to the multi-level capability labels is built, including core function requirements, minimum resource requirements and elastic resource requirements; the core function requirements refer to the hardware inherent attribute related requirements for the running of the AI model, which are used to match the hardware type identification and attribute characteristic code in the basic layer capability labels; the minimum resource requirements refer to the real-time state attribute related thresholds required for the running of the AI model, which are used to associate the dynamic parameter values in the dynamic layer capability labels; the elastic resource requirements refer to the tolerance requirements of the AI model to the comprehensive performance and resource fluctuations, which are used to interface the multi-dimensional comprehensive score of the derivative layer capability labels.

[0088] For the core function requirements, the hardware type identification and attribute characteristic code matching operation is performed. The binary matching code corresponding to the hardware type identification of the basic layer capability label is generated according to the hardware architecture requirements in the model requirement dataset, the binary matching code is a binary sequence generated by the same coding rule as the hardware type identification, and the compatibility verification bit is determined in the attribute characteristic code according to the software environment dependency information, the compatibility verification bit is a specific binary bit in the attribute characteristic code for marking software compatibility, forming the quantified query condition of the core function requirements based on the combination of the hardware type identification and the attribute characteristic bit, and the rule determines whether it meets the requirements by judging the hardware type identification matching result and the compatibility verification bit state;

[0089] For the minimum resource requirement, the interval threshold corresponding to the dynamic layer capability label is generated according to the resource occupation threshold in the model requirement data set, the interval threshold is a numerical range composed of a lower limit value and an upper limit value, used to define the acceptable dynamic parameter value range, forming the quantitative query condition of the minimum resource requirement; the dynamic layer capability label distribution characteristics of the edge node group are collected in real time, the dynamic layer capability label distribution characteristics are the distribution of the dynamic parameters of a certain number of edge nodes, when the group dynamic parameter mean value fluctuation exceeds the preset proportion, the preset proportion is the mean value variation proportion preset for triggering threshold adjustment, the boundary value of the interval threshold is automatically adjusted, ensuring that the interval threshold setting is adapted to the overall fluctuation trend of the real-time state attribute of the edge node;

[0090] For the minimum resource requirement, the interval threshold corresponding to the dynamic layer capability label is generated according to the resource occupation threshold in the model requirement data set, forming the quantitative query condition of the minimum resource requirement; the dynamic layer capability label distribution characteristics of the edge node group are collected in real time, when the group dynamic parameter mean value fluctuation exceeds the preset proportion, the boundary value of the interval threshold is automatically adjusted, ensuring that the threshold setting is adapted to the overall fluctuation trend of the real-time state attribute of the node;

[0091] For the elastic resource requirement, the weight coefficient matrix corresponding to each dimension of the multi-dimensional comprehensive score of the derived layer capability label is generated according to the performance expectation index in the model requirement data set, the weight coefficient matrix is a matrix composed of weight values of each score dimension, the sum of the weight values is 1, used to reflect the importance of different dimensions; used to convert the resource fluctuation tolerance requirement and the comprehensive performance expectation in the model requirement data set into a quantitative interval based on the multi-dimensional comprehensive score, the resource fluctuation tolerance requirement is the acceptable resource state fluctuation range of the model, the comprehensive performance expectation is the overall expectation of the model running performance, the quantitative interval is a range composed of a lower limit value and an upper limit value, used to define the acceptable comprehensive score range, forming the quantitative query condition of the elastic resource requirement;

[0092] The field mapping relationship table of the quantitative query condition and the capability label is established through the distributed mapping rule, the mapping relationship table is stored in slices according to the capability label level using the distributed storage architecture, and each slice node maintains data consistency through a synchronization mechanism; used to clarify the association logic of each parameter in the quantitative query condition and the corresponding fields of the basic layer, the dynamic layer and the derived layer capability label, realize the automatic matching analysis of the model requirement and the capability label, and ensure that the quantitative query condition can accurately point to the corresponding capability label field.

[0093] Preferably, the specific steps of screening the target edge node set include:

[0094] According to the quantitative query conditions of core function requirements and minimum resource requirements, the edge nodes are traversed, the capability matching is performed according to the capability labels of the edge nodes, and the candidate node set is generated step by step. Specifically, according to the quantitative query conditions of the core function, all edge nodes are traversed, the hardware type identifier and attribute characteristic code of the basic layer capability label of each edge node are compared, the accurate matching of the binary matching code and the hardware type identifier and the state judgment of the compatibility verification bit are performed, and the edge nodes meeting the core function requirements are classified into the first-stage candidate node set. Then, according to the quantitative query conditions of the minimum resource, the first-stage candidate node set is screened, whether the dynamic parameter values in the dynamic layer capability label of each node meet the interval threshold range is checked, and the nodes meeting the requirements are included in the second-stage candidate node set as the final candidate node set.

[0095] According to the quantitative query conditions of the elastic resource, the candidate node set is traversed, and the multi-dimensional comprehensive score in the derived layer capability label of each edge node is obtained. First, the multi-dimensional comprehensive score is preliminarily weighted by combining the weight coefficient matrix to obtain the basic comprehensive score. Then, the dynamic weight adjustment process is started, the capability fluctuation coefficient is calculated according to the attribute change trajectory of each edge node, the mean value of the change amplitude of the capability label in the preset period is determined by statistics, and the update frequency coefficient is generated according to the capability label update frequency. The fluctuation coefficient and the frequency coefficient are substituted into the preset weight adjustment formula to real-time correct the weight values of the derived layer evaluation dimensions such as model running efficiency, stability and resource cost. For example, taking the initial weight value as the reference, the product of the capability fluctuation coefficient and the update frequency coefficient is multiplied by the adjustment factor preset according to the model deployment scene, and then the result is subtracted by 1 and multiplied by the initial weight value. The obtained result is the corrected weight value. The basic comprehensive score is secondarily weighted using the corrected weight value to obtain the dynamically adjusted comprehensive score. Finally, the edge nodes whose dynamically adjusted comprehensive scores are within the quantitative interval set by the elastic resource quantitative query conditions are screened out to form the target edge node set.

[0096] Step S3: Based on the network state of the target edge node, the adaptive network coding technology is used to generate the coding data block corresponding to the redundancy of each target edge node to complete the construction of the basic carrier of data transmission. Then, the distributed cache mechanism is used to extract the stability features of the edge nodes according to the capability labels of the edge nodes to select the cache nodes suitable for each target edge node and pre-store the coding data block to build a distributed cache network. Based on the real-time updated dynamic layer capability label of the target edge node, the transmission parameters are dynamically adjusted to build a collaborative transmission link with the cache nodes as the distribution branch point and the source transmission as the supplement to transmit the coding data block of the AI model to the target edge node to form the full-link technical support from data preprocessing to transmission guarantee, ensuring that the model package reaches the target edge node efficiently and reliably.

[0097] In this embodiment, step S3 dynamically matches the redundancy of the encoded data block with the real-time transmission capability of the node through the precise adaptation of the adaptive network coding technology and the network state of the target edge node, reduces the risk of data loss from the source; the extracted and applied stability characteristics of the edge node by means of the distributed cache mechanism have both efficient data storage and distribution capability and can form an adaptive transmission link with the target edge node to reduce the cross-network transmission delay; the cooperative transmission link based on the dynamic layer capability label adjusts the transmission parameters in real time, and through the organic cooperation of the cache node and the source transmission, it not only plays the efficiency of the nearby distribution in the edge cluster, but also guarantees the data transmission integrity in the extreme network state. The progressive application of the three technical means realizes the efficient and reliable full-process of the AI model from generation to transmission, and lays a solid data transmission foundation for the subsequent rapid deployment of the model in the target edge node.

[0098] Preferably, the specific steps of generating the encoded data block with the corresponding redundancy of each target edge node by using the adaptive network coding technology include:

[0099] The AI model to be transmitted is processed in blocks, the size of the encoded data block is determined according to the network MTU value of the target edge node, and the encoded data block is ensured to adapt to the target edge node.

[0100] According to the capability label of the target edge node, the network state attribute parameters of the target edge node are extracted, including network bandwidth, packet loss rate, transmission delay, etc., to form a network state data set;

[0101] Attribute quantization processing is performed on the network state data set to convert the network state attribute parameters into standardized network performance indexes, and the value range is 0 to 1, and the higher the value, the better the network state;

[0102] The redundancy of the encoded data block is determined based on the network performance index and the preset redundancy mapping matrix, and the network performance index interval in the redundancy mapping matrix is negatively correlated with the redundancy; wherein the redundancy mapping matrix refers to a mapping relationship table that associates the network performance index with the encoding redundancy, which is generated by offline training through statistics of the optimal redundancy corresponding to different network performance indexes in the historical transmission data of the target edge node, combined with the demand threshold of the AI model for transmission stability. Each network performance index interval in the redundancy mapping matrix corresponds to a unique redundancy value.

[0103] Based on the calculated redundancy, the network coding algorithm is used to encode the blocked AI model to generate a mixed set of original encoded data blocks and redundant encoded data blocks, and each encoded data block carries a unique identifier and check information. The unique identifier includes a target node hardware unique identifier and a data block serial number.

[0104] The integrity of the encoded mixed set is checked, an overall data fingerprint is generated through a hash algorithm, and is associated and bound with the capability label feature of the target edge node, to ensure the matching of the data block and the target edge node.

[0105] A dynamic adjustment logic of encoding parameters is constructed, and a differentiated updating mechanism of the capability label of the target edge node is associated: when the change range of the network state attribute parameter in the dynamic layer capability label exceeds the judgment threshold, the network state data set is automatically triggered to update, the network performance index and the redundancy are recalculated, and the redundancy of the subsequent encoded data block is adjusted in real time, to ensure that the encoding strategy is adapted to the latest network state of the node.

[0106] Please refer to Figure 4 , preferably, the cache nodes suitable for each target edge node are screened and the encoded data blocks are pre-stored, and the specific steps of building a distributed cache network include:

[0107] The capability label information of each node is extracted by traversing the edge nodes in the edge cluster except the target edge node; based on the geographic location attribute in the basic layer capability label of the edge node, the physical distance from the target edge node is calculated; the network state attribute parameters of the candidate edge node and the target edge node are extracted in combination with the dynamic layer capability label, the network link quality is quantitatively scored by constructing a network link quality evaluation model; the upper limit threshold of the physical distance and the lower limit threshold of the network link quality score are set, the edge nodes with the physical distance within the set upper limit threshold and the network link quality score higher than the set lower limit threshold are selected as cache nodes and included in the initial cache node set. The network link quality evaluation model refers to a quantitative model for comprehensively evaluating the performance of the data transmission link between the candidate node and the target edge node, the network delay, link bandwidth, packet loss rate and other parameters are standardized, and then dynamic weights are assigned and weighted calculation is performed to obtain the comprehensive score of the link quality, and the dynamic weights are adjusted in real time according to the influence degree of different parameters on the transmission performance.

[0108] According to the functional requirements and performance requirements of the cache nodes, the quantitative query conditions corresponding to the capability labels are generated, the initial cache node set is traversed, the capability labels of the cache nodes are checked, and the cache nodes meeting all the quantitative query conditions are divided into a cache node subset.

[0109] Based on the total amount of the AI model encoded data block and the fault tolerance requirement, the node number range of the cache node subset is set to balance the data storage efficiency and transmission reliability; if the number of nodes in the cache node subset exceeds the upper limit of the node number range, the cache nodes with the closest physical distance and the best network link quality are retained according to the physical distance and the network link quality; if the number of nodes in the cache node subset is less than the lower limit of the node number range, the physical distance upper limit threshold is increased to re-screen, to ensure that the number of cache nodes in the cache node subset meets the requirements.

[0110] An association mapping between the cache nodes and the target edge node is established, an association relationship table is constructed by a node unique identifier, a role priority of each cache node in the subset, a stored encoded data block shard index, and a data synchronization state are recorded, a distributed hash table is used to realize distributed storage and fast query of the association relationship, and a distributed cache network is formed with the target edge node as the center and the cache nodes as the support.

[0111] Before the AI model is transmitted, the encoded data blocks are pre-synchronized to each cache node, an incremental transmission mode is used in the synchronization process, only the encoded data blocks missing from the cache nodes are transmitted, and after the synchronization is completed, a cache check report is generated, including the encoded data block integrity check result, the storage path, and the synchronization timestamp.

[0112] An association cache node and target edge node capability tag differential update mechanism is established, when the capability tag of the cache node no longer meets the quantitative query condition, the re-selection process is automatically triggered to update the cache node subset; the cache data block integrity is checked regularly, if data anomalies are found, data retransmission is preferentially performed through other nodes in the subset, and the continuous availability of the distributed data support layer is ensured.

[0113] Preferably, the specific steps of transmitting the encoded data blocks of the AI model to the target edge node include:

[0114] The generated encoded data blocks are split according to the shard index, and the encoded data blocks are pre-distributed to each cache node in the cache node subset according to the association relationship table of the distributed cache network. Based on the role priority sorting of the association relationship table, the high-priority cache nodes are preferentially activated as the main transmission source, the encoded data blocks are distributed according to the shard importance and the node storage capacity, the high-priority cache nodes store the core shards, the low-priority nodes store the redundant shards and part of the core shard copies, and it is ensured that each encoded data block exists in multiple node backups. After the encoded data blocks are transmitted to the cache nodes, they are written into the shard storage index, and the correspondence between the data block unique identifier and the storage path is recorded.

[0115] The target edge node queries the association relationship table through the distributed hash table, and locates the cache node storing the corresponding shard. The request message carries the unique identifier of the target edge node hardware and the unique identifier of the required encoded data block, and the request is preferentially sent to the main transmission source.

[0116] After the main transmission source receives the request, the corresponding encoded data block is retrieved according to the shard storage index, and the transmission is initiated after the data block integrity is verified by the check information. The encoded data blocks are transmitted according to the path from the cache node to the target edge node, and the target edge node temporarily stores them in the temporary buffer area after receiving them, and updates the receiving state after confirming the correctness through the encoded data block unique identifier check matching.

[0117] If the primary transmission source fails to transmit, the target edge node automatically sends a request to the secondary priority cache node, which transmits the backup encoded data block along the same path. If all cache nodes are missing a certain encoded data block, the target edge node initiates a request to the source, transmits the missing data block along the path from the source to the target edge node, and supplements it to the temporary buffer.

[0118] The target edge node sorts and integrates the data blocks in the temporary buffer according to the shard index, and verifies the integrity through overall hash checking. If there is a shard that fails the check, data recovery is performed based on the redundancy coding rule using the received original blocks and redundant blocks. If the recovery fails, a secondary request for the shard is initiated to the cache node or the source, ensuring that all encoded data blocks are complete and available.

[0119] The integrated complete encoded data block is stored in the persistent storage area of the target edge node, and a transmission completion report is generated, including the receiving path and check result of all encoded data blocks. The report is synchronized to the associated relationship table of the distributed cache network, and the cache nodes update the transmission status of the encoded data blocks accordingly. The target edge node associates its own capability label to record the transmission completion time and data integrity index.

[0120] Please refer to Figure 5 The embodiment introduces an AI model distribution and deployment system for cloud-edge collaboration, including an attribute perception module, a capability label module, an analysis matching module, a data block generation module, a network construction module, and a collaborative transmission module:

[0121] The attribute perception module is responsible for incremental collection and change-triggered re-collection of multi-dimensional attribute data of edge nodes, providing raw data support for capability label generation. The capability label module standardizes the collected data and converts it into multi-level capability labels, realizing hierarchical storage and query management of the labels. The analysis matching module is used to analyze AI model distribution requirements and convert them into quantitative query conditions, and to filter the target edge node set through a three-level matching mechanism. The data block generation module generates encoded data blocks with redundancy based on the network status of the target edge node, completes the construction and effectiveness verification of the data transmission carrier. The network construction module selects cache nodes and establishes an associated mapping with the target edge node, realizing pre-distribution of encoded data blocks and construction of the cache network. The collaborative transmission module schedules the collaborative transmission link of the cache node and the source, and completes the reception, integration, verification, and storage of the encoded data blocks.

[0122] The attribute perception module deploys lightweight perception components, continuously collects multi-dimensional attribute data of edge nodes using incremental collection strategy, and transmits the collected data to the capability label module. At the same time, it associates with the capability label differential update mechanism, and when the change amplitude of multi-dimensional attribute data exceeds the judgment threshold, it triggers the data re-collection process.

[0123] The capability tagging module comprises an attribute standardization unit, a tag conversion unit, and a hierarchical storage unit. The attribute standardization unit standardizes the collected data using an attribute standard dictionary, unifying naming, format, and dimension management. The tag conversion unit, based on tag mapping rules and attribute quantification rules, converts standardized data into capability tags for the basic, dynamic, and derived layers, and performs consistency checks. The hierarchical storage unit uses static storage, time-series storage, and associative storage methods to store capability tags at each level, supporting fast queries by capability tag level and attribute type.

[0124] The parsing and matching module includes a demand layering unit and a three-level matching unit. The demand layering unit extracts AI model distribution requirements through the model metadata parsing interface, constructs a three-level description framework of core functional requirements, minimum resource requirements, and elastic resource requirements, and converts them into quantitative query conditions corresponding to capability tags. The three-level matching unit sequentially calls the basic layer capability tags for compatibility filtering, the dynamic layer capability tags for resource threshold screening, and the derived layer capability tags for multi-dimensional scoring and ranking, and generates a target edge node set by combining a dynamic weight adjustment strategy.

[0125] The data block generation module performs operations such as data segmentation, network performance index quantification, and redundancy calculation based on the network status attribute parameters in the dynamic layer capability label of the target edge node. It uses a network coding algorithm to generate coded data blocks containing unique identifiers and verification information, and verifies the validity of the coded data blocks through an integrity verification unit.

[0126] The network construction module includes a cache node filtering unit, an association mapping unit, and a data pre-distribution unit. The cache node filtering unit selects an initial set of cache nodes based on the geographic location attributes in the edge node's basic layer capability tags and the network parameters in the dynamic layer capability tags, using a network link quality assessment model. It then determines a subset of cache nodes based on quantitative query conditions. The association mapping unit establishes an association table between cache nodes and target edge nodes using a distributed hash table, recording role priority, coded data block sharding index, and data synchronization status. The data pre-distribution unit pre-distributes coded data blocks to cache nodes according to their sharding indexes, synchronizes data using incremental transmission, and generates a cache verification report.

[0127] The cooperative transmission module is composed of a transmission scheduling unit, a data receiving unit and an integration and checking unit. The transmission scheduling unit preferentially activates a high-priority cache node as a main transmission source, switches to a secondary-priority cache node or a source transmission when the main transmission source is unavailable, and constructs a cooperative transmission link. The data receiving unit receives encoded data blocks at a target edge node side and temporarily stores them in a temporary buffer area, checks matching through unique identification of the encoded data blocks. The integration and checking unit integrates the encoded data blocks according to a fragmentation index, verifies integrity through overall hash checking, realizes data recovery based on a redundancy encoding rule, finally writes complete encoded data block sets into a persistent storage area of the target edge node, and generates a transmission completion report.

[0128] Working principle and effects:

[0129] The present application realizes efficient distribution and deployment of AI models in a cloud-edge cooperative scenario by constructing a full-process mechanism of node capability modeling, target node screening and data transmission optimization. The core lies in using multi-level capability labels as a link to dynamically associate edge node attributes and model requirements, and combining adaptive coding and distributed caching technology to improve transmission reliability.

[0130] Specifically, the multi-dimensional attributes of the edge node are first converted into basic layer, dynamic layer and derivative layer capability labels through incremental collection and standardized processing, and differential update is realized based on the attribute change amplitude to provide accurate description of node capabilities. Then, a model requirement hierarchical framework is constructed relying on the capability labels, and the target edge node is screened through three-level progressive matching and dynamic weight adjustment to ensure high adaptability of the model to the node hardware and resource state. Finally, the corresponding redundancy of the encoded data block is generated for the target node network state, the distributed caching network is used to realize efficient transmission and dynamic maintenance of the data block, and the transmission reliability is ensured through cooperative main and backup nodes and integrity checking.

[0131] In summary, the present application realizes accurate matching of nodes and models through dynamic management of capability labels, improves transmission efficiency and stability through adaptive coding and distributed caching, effectively solves the problems of insufficient model distribution adaptability and unreliable transmission in a cloud-edge cooperative scenario, and provides full-link technical support for rapid landing of AI models on the edge side.

[0132] The above only describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the scope of the present application should be considered within the protection scope of the present application. It should be noted that, for ordinary technical personnel in the technical field, some improvements and refinements without departing from the principles of the present application should also be considered within the protection scope of the present application.

Claims

1. A method for distributing and deploying AI models for cloud-edge collaboration, characterized in that, include: Collect multi-dimensional attribute data of edge nodes, transform it into multi-level capability tags, and trigger differentiated updates of capability tags based on the magnitude of changes in multi-dimensional attribute data; Based on the multi-level capability tags, a hierarchical description framework for model requirements is constructed to transform AI model distribution requirements into quantitative query conditions corresponding to capability tags. Based on the three-level progressive matching mechanism of capability tags, combined with a dynamic weight adjustment strategy, a set of target edge nodes is selected from the edge nodes. Based on the network status of the target edge nodes, a coded data block with corresponding redundancy is generated for each target edge node. A distributed caching mechanism is used to extract the stability features of the edge nodes based on their capability tags. This feature is then used to select caching nodes that are compatible with each target edge node and to pre-store the coded data blocks. A distributed caching network is then built to transmit the coded data blocks of the AI ​​model to the target edge nodes. The specific steps for converting AI model distribution requirements into quantitative query conditions corresponding to capability tags include: Extract the raw information of AI model distribution requirements and construct a structured model requirement dataset; Based on the model requirement dataset, a three-level structure corresponding to multi-level capability tags is built, including core functional requirements, minimum resource requirements, and elastic resource requirements. The core functional requirements are used to match the hardware type identifier and attribute feature code in the basic layer capability tags; the minimum resource requirements are used to associate the dynamic parameter values ​​in the dynamic layer capability tags; and the flexible resource requirements are used to connect to the multi-dimensional comprehensive score of the derived layer capability tags. For core functional requirements, binary matching codes corresponding to hardware type identifiers of basic layer capability tags are generated based on hardware architecture requirements in the model requirement dataset. At the same time, compatibility verification bits are determined in attribute feature codes based on software environment dependency information to form quantitative query conditions for core functional requirements. For the minimum resource requirement, the interval threshold corresponding to the dynamic layer capability label is generated based on the resource occupancy threshold in the model requirement dataset, forming a quantitative query condition for the minimum resource requirement. For elastic resource demand, a weight coefficient matrix corresponding to each dimension of the multidimensional comprehensive score of the capability label of the derived layer is generated based on the model demand dataset. This matrix is ​​used to convert the resource fluctuation tolerance requirements and comprehensive performance expectations in the model demand dataset into a quantitative range based on the multidimensional comprehensive score, forming quantitative query conditions for elastic resource demand. A field mapping relationship table between quantitative query conditions and capability tags is established through distributed mapping rules, which is used to clarify the association logic between quantitative query conditions and the corresponding fields of capability tags in the basic layer, dynamic layer, and derivative layer.

2. The AI ​​model distribution and deployment method for cloud-edge collaboration as described in claim 1, characterized in that, The specific steps for generating the coded data block corresponding to the redundancy of each target edge node include: The AI ​​model to be transmitted is divided into blocks, and the size of the encoded data blocks is determined; based on the capability labels of the target edge nodes, the network state attribute parameters of the target edge nodes are extracted to form a network state dataset; The network state dataset is subjected to attribute quantization processing to convert the network state attribute parameters into standardized network performance indices. The redundancy of the encoded data blocks is determined based on the network performance indices and a preset redundancy mapping matrix. The network performance index interval in the redundancy mapping matrix is ​​negatively correlated with the redundancy. Based on the calculated redundancy, the block-based AI model is encoded to generate a mixed set of original encoded data blocks and redundant encoded data blocks. The integrity of the encoded hybrid set is verified, and the logic for dynamically adjusting the encoding parameters is constructed to automatically trigger the update of the network state dataset.

3. The AI ​​model distribution and deployment method for cloud-edge collaboration as described in claim 1, characterized in that, The specific steps for building the distributed caching network include: Based on the capability label information of each edge node, calculate the physical distance to the target edge node, extract the network state attribute parameters of the target edge node, and perform quantitative scoring. Set an upper limit threshold for physical distance and a lower limit threshold for network link quality score, filter cache nodes for target edge nodes, and construct an initial cache node set; Based on the functional and performance requirements of the cache nodes, quantitative query conditions corresponding to the capability tags are generated, and the initial cache node set is traversed to construct a subset of cache nodes; Set the range of the number of cache nodes in the cache node subset. By adjusting the upper limit threshold of physical distance and the lower limit threshold of network link quality score, the number of cache node subsets to be moved meets the range of the number of cache nodes. Establish a mapping between cache nodes and target edge nodes to form a distributed cache network centered on the target edge nodes and supported by cache nodes; A capability tag differentiation update mechanism that associates cache nodes with target edge nodes automatically triggers a re-filtering process to update the subset of cache nodes.

4. The AI ​​model distribution and deployment method for cloud-edge collaboration as described in claim 1, characterized in that, The steps for converting the multi-level capability tags include: An incremental acquisition strategy is used to collect multi-dimensional attribute data of edge nodes in real time; The collected multi-dimensional attribute data is standardized by using an attribute standard dictionary, which in turn matches and transforms the multi-dimensional attribute data using a rule engine. The standardized multi-dimensional attribute data is divided into hardware-inherent attribute data, real-time status attribute data, and environment-related attribute data. Based on the label mapping rules, the inherent attribute data of the hardware is encoded and transformed to generate capability labels for the basic layer; the real-time status attribute data is calculated and transformed to generate capability labels for the dynamic layer; and the environmental related attribute data is quantified and fused to generate capability labels for the derivative layer. The generated capability tags are validated for consistency, and a hierarchical storage structure is constructed to store capability tags for the base layer, dynamic layer, and derived layer.

5. The AI ​​model distribution and deployment method for cloud-edge collaboration as described in claim 4, characterized in that, The label mapping rule is a rule that specifies the correspondence between the inherent attributes of hardware and the capability labels of the basic layer. The encoding conversion adopts a structured encoding method, and the encoding content includes hardware type identifier and attribute feature code. The capability label for generating the dynamic layer includes: using the real-time status attribute data of the edge node within a preset benchmark time period as a benchmark data pool, calculating the relative fluctuation coefficient between the current real-time status attribute and the benchmark data pool through a sliding window algorithm, and combining the relative fluctuation coefficient with the benchmark value to convert it into a dynamic parameter value in the range of 0 to 1 as the dynamic layer capability label. The process of generating capability labels for the derivative layer includes: converting unstructured environmental association attribute data into computable parameters through attribute quantization rules; performing feature fusion between the quantized environmental association attribute parameters and the current values ​​of capability labels for the basic and dynamic layers; introducing a model deployment scenario weight matrix; and generating a multi-dimensional comprehensive score as the capability label for the derivative layer through weighted calculation.

6. The AI ​​model distribution and deployment method for cloud-edge collaboration as described in claim 1, characterized in that, The specific steps of the differential update include: A dynamic threshold adjustment mechanism is adopted to set the judgment threshold for the magnitude of change of multi-dimensional attribute data. The dynamic threshold adjustment mechanism records the cumulative running time of the edge node from startup to the present, and counts the number of changes of multi-dimensional attribute data within the preset monitoring period, and dynamically matches the judgment threshold corresponding to each multi-dimensional attribute. Calculate the difference and percentage of change in multi-dimensional attribute data, and obtain the magnitude of change of each multi-dimensional attribute data through preset magnitude calculation rules; Identify multi-dimensional attributes whose change exceeds the matching threshold and mark them as changed attributes; when changed attributes exist, update the values ​​of all dynamic parameters associated with changed attributes in the dynamic layer capability labels according to the mapping relationship between changed attributes and dynamic layer capability labels. Obtain the list of impact factors for the changed attributes, and recalculate the multidimensional comprehensive score associated with the impact factors in the derived layer capability labels based on the list of impact factors.

7. The AI ​​model distribution and deployment method for cloud-edge collaboration as described in claim 1, characterized in that, The specific steps for filtering the target edge node set include: Based on the quantitative query conditions of core functional requirements and minimum resource requirements, the edge nodes are traversed, and the capability matching is performed according to the capability tags of the edge nodes to generate a candidate node set step by step. Based on the quantitative query conditions of elastic resources, the candidate node set is traversed to obtain the multi-dimensional comprehensive score in the capability tags of each edge node's derived layer; A preliminary weighted calculation of the multidimensional comprehensive score is performed by combining the weight coefficient matrix to obtain the basic comprehensive score; Based on the attribute change trajectory of each edge node, the capability fluctuation coefficient is calculated by statistically analyzing the average change of capability tags within a preset period, and an update frequency coefficient is generated based on the capability tag update frequency. The weight values ​​of the evaluation dimensions of the derivative layer are adjusted based on the volatility coefficient and frequency coefficient; a second weighted calculation is performed on the basic comprehensive score to obtain the dynamically adjusted comprehensive score. The edge nodes whose comprehensive scores after dynamic adjustment fall within the quantitative range set by the quantitative query conditions for elastic resources are selected and formed into a target edge node set.

8. The AI ​​model distribution and deployment method for cloud-edge collaboration as described in claim 1, characterized in that, The process of transmitting the encoded data blocks of the AI ​​model to the target edge node includes: The generated encoded data blocks are split according to the shard index, and the encoded data blocks are pre-distributed to each cache node in the subset of cache nodes according to the association table of the distributed cache network. Based on the role priority sorting in the association table, determine the primary transmission source; send a request to the primary transmission source cache node; After receiving the request, the main transmission source retrieves the corresponding encoded data block according to the fragment storage index, transmits it to the temporary buffer of the target edge node, and updates the received status of the encoded data block by checking the matching. When the primary transmission source fails to transmit, the target edge node requests the transmission of backup encoded data blocks from the secondary priority cache nodes; if all cache nodes are missing, it requests supplementary transmission. The target edge node integrates temporary buffer data blocks according to the shard index, verifies the integrity through overall hash verification, and stores them in the persistent storage area of ​​the target edge node.

9. A cloud-edge collaborative AI model distribution and deployment system, used to implement the cloud-edge collaborative AI model distribution and deployment method according to any one of claims 1-8, characterized in that, It includes an attribute awareness module, a capability tagging module, a parsing and matching module, a data block generation module, a network construction module, and a collaborative transmission module; The attribute awareness module is used to collect multi-dimensional attribute data of edge nodes; the capability tagging module is used to convert multi-dimensional attribute data into multi-level capability tags; the parsing and matching module is used to parse AI model distribution requirements and filter target edge node sets; the data block generation module generates encoded data blocks based on the network status of target edge nodes; the network construction module is used to filter cache nodes and establish association mappings with target edge nodes; and the collaborative transmission module is used to receive, integrate, verify, and store encoded data blocks.

Images