Hosting and collaborative optimization method for AI model and data version mapping
By constructing a two-way version mapping relationship and adaptability assessment, the problem of insufficient mapping in AI model and data version management is solved, enabling accurate traceability and optimization, and improving iteration efficiency and traceability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-03-17
AI Technical Summary
The existing AI model and data version management lacks accurate mapping, poor hosting collaboration, and insufficient optimization focus, resulting in difficulties in tracing version after model iteration, low efficiency, and blind optimization direction.
A two-way version mapping relationship is constructed. By generating unique identifiers for model and data versions, mapping contracts and blockchain records, accurate traceability and dynamic collaborative hosting of model and data versions are achieved, and adaptation evaluation and optimization are carried out based on the mapping relationship.
It enables precise tracking of model and data versions, improves iteration efficiency, solves version mismatch problems, ensures accurate optimization direction, forms a closed loop of hosting and optimization, and improves the iteration efficiency and traceability of AI models.
Smart Images
Figure CN121684155A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the intersection of AI model management, data version control, and intelligent hosting, specifically a hosting and collaborative optimization method for mapping AI models to data versions. Background Technology
[0002] In the existing AI-related technologies, model management and data management are mostly independent systems, with core issues concentrated in three aspects.
[0003] First, version association is missing: existing solutions can only record the model version or data version separately, without establishing a precise mapping relationship between the model version and the corresponding training / inference data version. As a result, after model iteration, it is impossible to trace the specific data version that the model version depends on. Once the model has performance problems, it is difficult to determine whether it is caused by the model architecture or the data quality.
[0004] Secondly, there is insufficient hosting collaboration: the hosting of models and data is mostly "static storage" and no collaborative mechanism is designed for the needs of data version updates and replacements during model iteration. When the data version is upgraded, it is necessary to manually match and retrain the model, which is inefficient and prone to version mismatch.
[0005] Third, the optimization lacks specificity: model optimization relies heavily on manual experience to adjust parameters or replace data, without analyzing the impact of different data versions on model performance based on the mapping relationship between model version and data version. This leads to a blind optimization direction and makes it impossible to achieve precise optimization that "matches the data version to the model version".
[0006] While existing technologies attempt to achieve separate hosting of models or data, or explore automated training of models, they have not designed a dedicated mapping mechanism and hosting-optimization collaborative architecture to address the strong correlation between model versions and data versions. Furthermore, they have not established a dynamic optimization model based on the mapping relationship. This fails to meet the needs of "version traceability, collaborative hosting, and precise optimization" during the iteration of AI models. It is fundamentally different from existing inventions related to data assetization and securitization (which focus on the circulation of data value) and fails to address the gap in collaborative management of model and data versions. Summary of the Invention
[0007] The purpose of this invention is to provide a hosting and collaborative optimization method for mapping AI model and data versions, aiming to solve the problems of lack of accurate mapping between AI model and data versions, poor hosting collaboration, and insufficient optimization targeting in the prior art. This invention provides a hosting and collaborative optimization method for mapping AI model and data versions, realizing bidirectional traceability of model and data versions, dynamic collaborative hosting, and accurate optimization based on mapping relationships, thereby improving the iteration efficiency and traceability of AI models.
[0008] To address the aforementioned technical problems, this invention provides a managed and collaborative optimization method for mapping AI models to data versions, with the specific steps as follows: 1. Version Mapping Layer: Constructs a two-way version mapping relationship. Version identifier generation: On the model side, the model version management module generates a unique version identifier for each iteration of the model (format: “Model-Domain-Iteration Sequence Number-Timestamp”, such as “Model-Med-005-202X0915”); on the data side, the data version management module generates a unique version identifier for each version of data (including original data, labeled data, and derived data) (format: “Data-Domain-Update Sequence Number-Timestamp”, such as “Data-Med-012-202X0915”).
[0009] Two-way mapping construction: The version mapping contract connects with the model and data version management module, captures the training data list (including data type and data volume) of the model version and the application model list (including model name and iteration stage) of the data version, and generates the "Version Mapping Table" with "Model Version Identifier - Data Version Identifier - Mapping Weight (data's contribution to the model)" as the core fields, and writes it to the blockchain to achieve tamper-proof record.
[0010] Mapping update: When the model iterates (a new model version is generated) or the data is updated (a new data version is generated), the version mapping contract automatically triggers the mapping update, calculates the fit between the new model version and the existing data version, or the fit between the new data version and the existing model version, updates the "Version Mapping Table" and synchronizes it to the blockchain.
[0011] 2. Collaborative Hosting Layer: Enables collaborative hosting under mapping relationships. Related storage: The collaborative hosting contract connects to the distributed storage system, and stores the model version file and the corresponding data version file in a directory structure of "model version identifier - data version identifier" according to the relationship of the "version mapping table". At the same time, the storage address and encryption key (bound to the version identifier) are written into the blockchain, and only authorized nodes can call the corresponding model and data through the version identifier.
[0012] Version linkage scheduling: When a user calls a model version through the terminal, the collaborative hosting contract automatically queries the "Version Mapping Table" to match all data versions corresponding to the model version, and pushes the data version list and mapping weight; if the user selects a data version, the contract automatically calls the data version and the associated model version to avoid version mismatch.
[0013] Version consistency verification: The collaborative hosting contract is set to a 24-hour verification cycle, which regularly compares the relationship between the "Version Mapping Table" in the blockchain with the actual storage relationship between the model and the data in the storage system. If it is found that a model version is stored separately from the corresponding data version (such as the data version being accidentally deleted), an alert is immediately triggered and the backup data version is called to restore the relationship.
[0014] 3. Dynamic Optimization Layer: Precise Optimization Based on Mapping Relationships Adaptability assessment: The dynamic optimization module connects to the model performance monitoring system to obtain performance indicators (such as accuracy, recall, and inference speed) for each model version. Combined with the mapping weights in the "Version Mapping Table", a three-dimensional evaluation model of "model version-data version-performance indicators" is constructed to calculate the adaptability score (out of 100) between different data versions and model versions.
[0015] Optimization Direction Generation: Based on the fit evaluation results, the dynamic optimization module automatically generates optimization schemes: If the fit of a certain model version with the existing data version is less than 70 points, it outputs "data version update suggestions" (such as supplementing a certain type of labeled data); if the fit of a certain data version with multiple model versions differs greatly, it outputs "model parameter adjustment suggestions" (such as adjusting the number of convolution kernels); if the fit meets the standard but the model performance is insufficient, it outputs "version combination optimization suggestions" (such as replacing with a data version with a higher fit).
[0016] Optimize closed-loop execution: After the user confirms the optimization plan, the dynamic optimization module triggers the collaborative hosting contract, calls the corresponding model version and target data version (or the updated data version), and starts automated training and verification; after training is completed, a new model version is generated, the data version management module synchronously records the derived data versions generated during the optimization process, and the version mapping contract automatically updates the mapping relationship between the new model version and the data version, forming a closed loop of "mapping-evaluation-optimization-remapping".
[0017] In summary, due to the adoption of the above-mentioned technologies, the beneficial effects of this invention are: 1. Two-way version mapping mechanism: Unlike the existing technology that manages model and data versions separately, this invention constructs a two-way association relationship of "model version-data version-mapping weight" through exclusive identifiers and mapping contracts, so as to achieve accurate traceability of the two and solve the problem of version mismatch.
[0018] 2. Hosting-Optimization Collaboration Closed Loop: Breaking through the limitations of the existing "static hosting", version mapping relationship is integrated into the hosting process to achieve linkage scheduling of model and data versions; at the same time, adaptability assessment is carried out based on the mapping relationship to generate targeted optimization solutions, forming a closed loop of "optimization during hosting and updating mapping after optimization" to improve iteration efficiency.
[0019] 3. Multi-dimensional fit evaluation model: The innovative design combines mapping weights and model performance indicators to evaluate fit, which can not only determine the degree of matching between the model and the data version, but also accurately locate the optimization direction (data update / parameter adjustment / version change), solving the problem of blind optimization in traditional optimization. Attached Figure Description
[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention, making other features, objects, and advantages of the invention more apparent. The illustrative embodiments of the invention illustrated in the drawings and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a system framework diagram of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] In the description of this invention, it should be understood that the terms indicating orientation or positional relationship are based on the orientation or positional relationship shown in the drawings and are only for the convenience of describing the invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.
[0023] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific context of the specification.
[0024] This invention provides a managed and collaborative optimization method for mapping AI models to data versions. (a) Setting up the implementation environment The system upon which this method relies includes a model version management module, a data version management module, blockchain nodes (using the Hyperledger Fabric consortium blockchain), a distributed storage system (IPFS), a dynamic optimization module, and user terminals (model developers and data managers). The smart contract module contains version mapping contracts and collaborative escrow contracts, deployed on the consortium blockchain nodes; the dynamic optimization module integrates a fit evaluation algorithm and an automated training interface.
[0025] (II) Implementation Steps (Taking the AI Diagnostic Model for Lung Nodules as an Example) Version mapping layer operations A medical AI company (the developer) iterated its lung nodule diagnosis model to version 8. The model version management module generated the identifier "Model-Med-Lung-008-202X0920". The corresponding training data is the updated lung CT annotation data (including 3,000 new data entries), and the data version management module generated the identifier "Data-Med-Lung-020-202X0920".
[0026] The version mapping contract retrieves the training data list of model 008 (including Data-Med-Lung-020 and historical data 019), calculates the contribution of data 020 to model 008 (mapping weight 65%) and the contribution of data 019 (35%), generates the "Version Mapping Table", and writes it to the blockchain.
[0027] Collaborative managed layer operations The collaborative escrow contract encrypts and stores the Model 008 file and Data 020 and 019 files in IPFS according to the directory structure "Model-Med-Lung-008-Data-Med-Lung-020 / 019", obtains the storage address " / ipfs / QmAbc... / Model-Med-Lung-008", and writes the address and encryption key into the blockchain.
[0028] Developers can access model 008 via the terminal, and the collaborative hosting contract will automatically query the mapping table and push data versions 020 (weight 65%) and 019 (weight 35%). Developers can then select data version 020, and the contract will synchronously call this data version and model 008 for subsequent optimization.
[0029] Dynamic optimization layer operations The dynamic optimization module detected that the accuracy of model 008 version in the test set was 89% (lower than the target of 92%). Based on the mapping table, the fit was calculated as follows: the fit between data 020 version and model 008 version was 82 points, and the fit between data 019 version and model 008 version was 75 points.
[0030] Module analysis revealed that the fit was satisfactory but the performance was insufficient, so the following "version combination optimization suggestion" was output: replace the low-quality labeled data (about 500 records) in data version 020, generate a new data version 021, and retrain model version 008.
[0031] After confirmation by the R&D personnel, the dynamic optimization module calls the Model 008 version and Data 021 version (newly generated identifier "Data-Med-Lung-021-202X0922") in the collaborative hosting contract, starts automated training, and generates Model 009 version (identifier "Model-Med-Lung-009-202X0922"); the version mapping contract automatically updates the mapping relationship between Model 009 version and Data 021 version (weight 72%), writes it to the blockchain, and completes the optimization loop.
[0032] Example: This example uses an AI predictive maintenance model for CNC machine tools in the industrial field as an application scenario to verify the applicability of this invention in the iterative management of industrial AI models. The FISCOBCOS consortium blockchain is used, the smart contract is written in Go, and the distributed storage system is an industrial-grade distributed cloud storage system compatible with the S3 protocol. The data owner is a heavy machinery manufacturing plant (hereinafter referred to as "the plant"), the model developer is an industrial AI technology company (hereinafter referred to as "the AI company"), and the core assets are the CNC machine tool vibration predictive maintenance model (used to predict equipment failures) and the corresponding equipment operating data (including parameters such as vibration, oil temperature, and current).
[0033] I. Setting up the implementation environment Blockchain node configuration: Six consortium blockchain nodes are set up, held by the factory, AI company, equipment manufacturer, third-party industrial testing agency, cloud storage service provider and safety production supervision department respectively. The consensus mechanism adopts RBFT (high-efficiency fault-tolerant consensus algorithm) to ensure the security and traceability of industrial data and model version information.
[0034] Core module adaptation: The model version management module adds an "Industrial Model Parameter Annotation Submodule" (records the equipment model and fault type for which the model is applied); the data version management module adds an "Equipment Data Classification Submodule" (classifies data by parameter type and equipment number); and the smart contract adds an "Industrial Data Version Verification Submodule" (adapts to the real-time requirements of equipment data) on the basis of the original version mapping contract and collaborative hosting contract.
[0035] Terminal and storage setup: The factory is equipped with equipment data acquisition terminals (connecting to sensors of 50 CNC machine tools to collect operating parameters in real time), the AI company is equipped with model development terminals (integrating TensorFlow training interface), and the industrial-grade distributed cloud storage system is deployed with a redundant backup mechanism. It is connected to the collaborative hosting contract through an industrial-grade SDK to support real-time uploading of equipment data and rapid access to model files. Specific implementation steps
[0036] (a) Version mapping layer operations Version identifier generation: The AI company iterated its predictive maintenance model for CNC machine tools to version 6. For the fault prediction scenario of the "CK6150 CNC machine tool" in the factory, the model version management module generated a unique identifier "Model-Industrial-CK6150-006-202X1010" (format: Model-Domain-Equipment Model-Iteration Number-Timestamp); Monthly updates of the operating data of the factory's 50 CNC machine tools were completed (adding 100,000 vibration parameter data points). The data version management module categorized the data by "Equipment Number + Parameter Type". The data version identifiers are as follows: vibration parameter data version identifier is “Data-Industrial-CK6150-Vib-015-202X1010” (format: Data-domain-equipment model-parameter type-update number-timestamp), and oil temperature and current parameter data version identifiers are “Data-Industrial-CK6150-Oil-015-202X1010” and “Data-Industrial-CK6150-Cur-015-202X1010”, respectively.
[0037] Bidirectional mapping construction: The version mapping contract captures the training data list of model 006 (including vibration data version 015, oil temperature data version 015, and current data version 015, with data volumes of 500,000, 450,000, and 400,000 respectively). The mapping weight is calculated based on the proportion of data gradients during model training: vibration data has the highest contribution to model fault prediction (mapping weight 55%), followed by oil temperature data (30%), and current data has the lowest contribution (15%). The contract uses “Model-Industrial-CK6150-006-202X1010”, “Data-Industrial-CK6150-Vib-015-202X1010 (55%)”, “Data-Industrial-CK6150-Oil-015-202X1010 (30%)”, and “Data-Industrial-CK6150-Cur-015-202X1010 (15%)” as core fields to generate an “Industrial Model-Data Version Mapping Table”, which is then written to the blockchain node.
[0038] Mapping Update: Ten days later, the factory adds two CK6150 CNC machine tools, collecting an additional 20,000 vibration parameter data points. The data version management module generates vibration data version 016 (identified as "Data-Industrial-CK6150-Vib-016-202X1020"). The version mapping contract automatically triggers an update, calculating the compatibility between vibration data version 016 and model version 006 (88% compatibility based on historical data contribution), adding the new data version and mapping weight (temporarily set at 52%) to the "Version Mapping Table," and synchronizing it to all consortium blockchain nodes.
[0039] (ii) Collaborative Hosting Layer Operation Related Storage: After receiving the "Version Mapping Table" from the Version Mapping Contract, the Collaborative Hosting Contract encrypts the Model 006 file (including model structure and weight parameters) and the corresponding Data 015 and 016 files using the SM4 national cryptographic algorithm according to the directory structure of "Model Version Identifier - Equipment Model - Data Version Identifier" (e.g., "Model-Industrial-CK6150-006-202X1010 / CK6150 / Data-Industrial-CK6150-Vib-015-016). The encrypted files are then uploaded to the industrial-grade distributed cloud storage system, obtaining the storage address "oss: / / industrial / CK6150 / Model-006 / Data-015-016". The contract writes the storage address and encryption key (binding the model and data version identifiers) into the blockchain. Only factory equipment management terminals, AI company R&D terminals, and monitoring nodes can decrypt and access the data using the version identifier.
[0040] Version-linked scheduling: The factory equipment management department calls model version 006 via the terminal to predict the failure risk of CNC machine tool number "CK6150-003". The collaborative hosting contract queries the "Version Mapping Table" in the blockchain and automatically pushes the list of three data versions corresponding to model version 006 (vibration version 015 / 016, oil temperature version 015, and current version 015) and their mapping weights; the equipment management personnel select the vibration version 016 containing the newly added equipment data, and the contract immediately synchronously calls the vibration version 016, oil temperature version 015, and current version 015 data, running in conjunction with model version 006, and outputs the failure prediction result for the equipment (failure probability of 12% in the next 72 hours).
[0041] Version consistency verification: The collaborative hosting contract is set with a 12-hour verification cycle (to adapt to the real-time data requirements of industrial equipment). At 00:00 and 12:00 daily, the blockchain's "Version Mapping Table" is automatically compared with the actual association between the data and the storage system. In one verification, it was discovered that the oil temperature data version 015 file was lost in the storage system due to a storage node failure. The contract immediately triggered an alert, sending a "data version missing" notification to the factory and AI company terminals, and calling upon redundant backups in cloud storage to restore the associated storage of oil temperature data version 015 and model version 006, ensuring version consistency.
[0042] (III) Dynamic Optimization Layer Operation Adaptability Assessment: The dynamic optimization module connects to the factory's equipment fault record system to obtain the actual prediction performance of Model 006: The fault prediction accuracy for 50 pieces of equipment over the past 30 days was 85%, with one instance of "prediction miss" for equipment number "CK6150-012" (the model predicted a fault probability of 8%, but a minor fault actually occurred within 48 hours). The module, combined with the mapping weights of the "Version Mapping Table," constructs a three-dimensional evaluation model: "Model 006 - Data Version (Vibration 015 / 016, Oil Temperature 015, Current 015) - Prediction Accuracy," calculating the adaptability score: Vibration 016 version has an adaptability score of 86 with Model 006, Oil Temperature 015 version 82, Current 015 version 78, and the overall adaptability score is 82 (lower than the target of 88).
[0043] Optimization direction generation: The dynamic optimization module analyzes the adaptation results and finds that the current data version (version 015) has the lowest adaptation, and the prediction weight of the current parameter of the corresponding device is only 15%. Combined with the fault record (the current fluctuation was not effectively captured during the "CK6150-012" fault), the optimization suggestion is: "Update the current data version to version 016 (supplement the current fluctuation data before the fault), adjust the feature extraction weight of the current parameter in the model, and retrain the model to version 007".
[0044] Optimized closed-loop execution: After the AI company's R&D personnel confirm the optimization plan, the dynamic optimization module triggers the collaborative hosting contract, calling model version 006, vibration version 016, oil temperature version 015, and the newly added current version 016 (identified as "Data-Industrial-CK6150-Cur-016-202X1025"). Automated training is initiated through the TensorFlow training interface: the number of convolution kernels for the current parameters is adjusted (increased from 16 to 32), with a training cycle of 50 rounds. After training, model version 007 (identified as "Model-Industrial-CK6150-007-202X1026") is generated. The data version management module records the derived data generated during training (such as current parameter feature maps), generating the derived data version identifier "Data-Industrial-CK6150-Cur-Der-001-202X1026". The version mapping contract automatically updates the "Version Mapping Table," adding mapping relationships between Model 007 and various data versions (Vibration 016 version 53%, Oil Temperature 015 version 29%, Current 016 version 18%), and writes this information to the blockchain. Subsequent testing shows that the prediction accuracy of Model 007 has improved to 91%, completing the optimization loop.
Claims
1. A method for hosting and collaborative optimization of AI model and data version mapping, characterized in that: Comprise the following steps: (1) Version mapping layer operation: Version identification generation: the model version management module generates a unique model version identification for the iterative model, and the data version management module generates a unique data version identification for the updated data; Bidirectional mapping construction: the version mapping contract grabs the training data list of the model version and the application model list of the data version, generates a《version mapping table》with "model version identification-data version identification-mapping weight" as the core field, and writes it into the blockchain; Mapping relationship update: when the model is iterated or the data is updated, the version mapping contract automatically updates the《version mapping table》and synchronizes it to the blockchain; (2) Collaborative hosting layer operation: Associated storage: the collaborative hosting contract encrypts and stores the model version file and the corresponding data version file according to the《version mapping table》association relationship in the distributed system, and writes the storage address and encryption key into the blockchain; Version linkage scheduling: when calling the model version, the collaborative hosting contract queries the《version mapping table》, pushes the corresponding data version list, and synchronously calls the selected data version and model version; Version consistency check: regularly compare the mapping relationship in the blockchain with the actual association in the storage system, and trigger an alarm and restore the association if an anomaly is found; (3) Dynamic optimization layer operation: Adaptation evaluation: the dynamic optimization module combines the mapping weight of the《version mapping table》with the model performance indicators to build a three-dimensional evaluation model and calculate the adaptation score of the model version and the data version; Optimization direction generation: according to the adaptation score, output the suggestion of data version update, model parameter adjustment or version combination optimization; Optimization closed loop execution: after confirming the optimization scheme, call the corresponding model and data version to start automated training, generate a new model version, and synchronize the version mapping relationship to form an optimization closed loop.
2. The AI model and data version mapping hosting and collaborative optimization method according to claim 1, characterized in that, The model version identification format is "Model-field-iteration number-timestamp", the data version identification format is "Data-field-update number-timestamp", and the field includes medical, industrial, and autonomous driving.
3. The AI model and data version mapping hosting and collaborative optimization method of claim 1, wherein, The mapping weight is the performance contribution of the data version to the corresponding model version, which is calculated by the data gradient proportion in the model training process.
4. The AI model and data version mapping hosting and collaborative optimization method according to claim 1, characterized in that, The model performance indicators include accuracy, recall rate, precision, inference speed, and generalization ability.
5. The AI model and data version mapping hosting and collaborative optimization method according to claim 1, characterized in that, The distributed storage system is an IPFS system or a distributed cloud storage system compatible with S3 protocol, and the encryption method uses AES-256 or SM4 algorithm.
6. The AI model and data version mapping hosting and collaborative optimization method according to claim 1, characterized in that, The blockchain platform is Hyperledger Fabric consortium chain or FISCOBCOS consortium chain, and the smart contract is written in Go language or Solidity language.
7. The AI model and data version mapping hosting and collaborative optimization method according to claim 1, characterized in that, The dynamic optimization module realizes automated training by calling the training interface of TensorFlow or PyTorch.