An AI model full life cycle collaborative management method and system

Abstract

The application relates to the technical field of artificial intelligence, in particular to an AI model full life cycle collaborative management method and system. The method is characterized in that: a plurality of source heterogeneous data are acquired, dynamic data management is performed on the plurality of source heterogeneous data, a standardized data set is formed, a metadata graph is constructed, and unified correlation management of data, models and training processes is realized. AI model construction and configuration are completed according to a preset development mode, and model training is performed based on the configuration to schedule computing resources, so that a target AI model is generated. After the model completes training, deployment and inference service publishing are performed, and the model running performance is continuously monitored, and data management updating or model retraining is automatically triggered when the performance changes, so that closed-loop collaborative management of the AI model full life cycle is realized. The application effectively guarantees the stability, accuracy and business adaptation capability of the long-term operation of the AI model, and improves the automation level and overall efficiency of the AI model full life cycle management.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a collaborative management method and system for the entire lifecycle of an AI model. Background Technology

[0002] In related technologies, with the deepening of enterprise digital transformation and intelligent upgrading, artificial intelligence models have become an important technical means to support data analysis, business forecasting, intelligent decision-making, and automated control. AI models are typically built and trained using multi-source heterogeneous data, encompassing various data formats such as relational databases, log data, data lakes, file systems, and interface data. They undergo multiple stages, including data collection, data governance, model development, model training, deployment and operation, and performance evaluation, forming a complete model lifecycle management process. During this process, factors such as data quality, model configuration, computing power scheduling, and operational feedback all significantly impact the final model performance.

[0003] However, existing AI model management methods still have significant technical shortcomings in terms of full lifecycle collaboration. On the one hand, the governance process of multi-source heterogeneous data is often disconnected from the AI ​​model development process. Data integration and cleaning rely heavily on independent data governance tools, while model development and training take place in a separate environment. The lack of a unified metadata management and traceability mechanism makes it difficult to establish a clear correlation between data sources, processing procedures, and model results. When data structure or quality changes, the model struggles to detect and respond promptly. On the other hand, existing AI model training and deployment processes typically lack dynamic adjustment capabilities linked to data status and model performance. Once training is complete, the model enters the deployment and operation phase, making it difficult to promptly detect changes in business scenarios, data distribution drift, or model performance degradation. Manual intervention for data updates or model retraining is often necessary, resulting in long response cycles and high maintenance costs. Furthermore, existing methods generally lack a closed-loop feedback mechanism that spans data governance, model training, deployment and operation, and performance monitoring. This prevents the automatic driving of data governance and model iteration by changes in model performance, thus hindering the stability, adaptability, and continuous optimization capabilities of AI models in complex and long-term operating environments.

[0004] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0005] The main objective of this application is to propose a collaborative management method and system for the entire lifecycle of AI models, which can effectively ensure the stability, accuracy, and business adaptability of AI models in long-term operation, and improve the automation level and overall efficiency of AI model lifecycle management.

[0006] To achieve the above objectives, one aspect of this application proposes a collaborative management method for the entire lifecycle of an AI model, the method comprising the following steps: Acquire multi-source heterogeneous data; Based on the multi-source heterogeneous data, perform dynamic data governance operations to generate a standardized dataset; Based on the standardized dataset, a metadata graph is constructed; Based on the metadata graph, AI model construction is performed according to a preset development mode to generate an AI model configuration to be trained; Based on the AI ​​model configuration, computing resources are scheduled to perform AI model training to generate the trained target AI model; Based on the target AI model, perform model deployment and inference service release processing; The performance metrics of the target AI model are continuously monitored during its operation. When a performance change is detected, a data governance update or model retraining process associated with the metadata graph is triggered to form a closed-loop collaborative management of the entire lifecycle of the AI ​​model.

[0007] In some embodiments, the dynamic data governance operations include data cleaning, structure standardization, and automatic metadata collection.

[0008] In some embodiments, the step of performing dynamic data governance operations based on the multi-source heterogeneous data to generate a standardized dataset includes: Data structure parsing is performed on the multi-source heterogeneous data to identify the data format, field type, and field semantics of different data sources, and to generate data structure parsing results. Based on the data structure parsing results, data cleaning processing is performed on the multi-source heterogeneous data to eliminate abnormal data and form valid data; Based on the cleaned and valid data, structural standardization processing is performed to map data from different data sources into a unified data structure. During the data cleaning and structural standardization process, metadata information corresponding to the multi-source heterogeneous data is collected simultaneously. A standardized dataset is generated based on the data that has undergone structural standardization and is associated with the metadata information.

[0009] In some embodiments, constructing a metadata graph based on the standardized dataset includes: Metadata is extracted from the data tables, fields, and their source information in the standardized dataset to obtain a metadata set; the metadata set includes data structure information, field semantic information, data source identifiers, and update feature information; Based on the metadata set, a metadata association model is constructed that includes data entity nodes and relation edges. The data entity nodes include data source nodes, data table nodes, and field nodes. The relation edges are used to represent the inclusion relationship, source relationship, and dependency relationship between the data entity nodes. Perform unified identification and semantic mapping processing on the metadata association model; A metadata graph is generated based on a metadata association model that completes semantic mapping.

[0010] In some embodiments, the development mode includes canvas mode and expert mode; The canvas mode is used to perform AI model building operations based on a visual workflow orchestration method; The expert mode is used to perform AI model building operations based on code.

[0011] In some embodiments, the canvas mode specifically includes: Based on the metadata graph, the data structure information and field semantic information corresponding to the standardized dataset are automatically loaded into the visualization canvas interface to generate configurable data nodes; Based on the data nodes, according to the preset process constraint rules, data preprocessing nodes, model building nodes and model evaluation nodes are configured in sequence, and data dependencies between each node are established to form a candidate model building process. Based on the metadata graph, consistency and dependency integrity checks are performed on the input and output fields of each node in the candidate model construction process to obtain the model construction process that passes the checks. Based on the verified model building process, a structured model configuration description file is generated, and the model configuration description file is output as the configuration of the AI ​​model to be trained.

[0012] In some embodiments, the expert mode specifically includes: Based on the metadata graph, determine the data tables and field semantic information corresponding to the data to be modeled in the standardized dataset; In the code development environment, based on the data table and field semantic information, model training code is written to clarify the feature input items, model calculation structure and training parameters at the code level; Execute the model training code to generate model parameters and performance metrics corresponding to the feature input items and model computation structure; Write the model parameters and performance metrics into and associate them with the metadata graph; Based on the model parameters and performance metrics written into the metadata graph, structured model configuration description information is generated; The model configuration description information is output as the configuration of the AI ​​model to be trained.

[0013] In some embodiments, scheduling computing resources to perform AI model training based on the AI ​​model configuration to generate a trained target AI model includes: Based on the AI ​​model configuration, the model structure information, algorithm framework, feature input constraints and training parameters are determined, and a model training task is generated. Based on the model training task, and considering the required computational scale and data scale of the task, the corresponding computing resources are scheduled and allocated. Based on the computing resources, the model is trained on the standardized dataset according to the model training task, the model parameters are iteratively updated, and the trained model result is generated. After the model training is completed, the model results and corresponding training metrics are written into and associated with the metadata graph to generate the target AI model.

[0014] To achieve the above objectives, another aspect of this application proposes an AI model full lifecycle collaborative management system, the system comprising: The data acquisition module is used to acquire heterogeneous data from multiple sources. The dynamic data governance module is used to perform dynamic data governance operations based on the multi-source heterogeneous data to generate a standardized dataset; The metadata management module is used to construct a metadata graph based on the standardized dataset. The AI ​​model building module is used to build AI models according to the metadata graph and a preset development mode, and generate AI model configurations to be trained. The training scheduling module is used to schedule computing resources to perform AI model training based on the AI ​​model configuration, so as to generate a target AI model that has been trained. The deployment service module is used to perform model deployment and inference service release processing based on the target AI model; The feedback module is used to continuously monitor the performance indicators of the target AI model during its operation. When a performance change is detected, it triggers a data governance update or model retraining process associated with the metadata graph, so as to form a closed-loop collaborative management of the entire life cycle of the AI ​​model.

[0015] To achieve the above objectives, another aspect of this application proposes an AI model full lifecycle collaborative management system. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0016] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0017] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0018] The embodiments of this application include at least the following beneficial effects: This application provides a collaborative management method and system for the entire lifecycle of AI models. This solution achieves collaborative processing of data cleaning, standardization, and consistency constraints through unified acquisition and dynamic data governance of multi-source heterogeneous data, effectively reducing the impact of data quality differences on model training and inference performance, and improving the reusability and reliability of data assets. Based on a metadata graph constructed from standardized datasets, data sources, feature structures, processing rules, and model configurations are associated and modeled, making the model building, training, deployment, and operation and maintenance processes traceable and interpretable, avoiding information fragmentation and configuration mismatch problems during model development. By introducing a preset development mode to generate model configurations and combining it with a unified scheduling mechanism for computing resources, collaborative optimization of the model training process and resource usage strategies is achieved, improving model training efficiency and resource utilization. Furthermore, after model deployment and inference service release, model performance is continuously monitored, and when performance changes are detected, data governance updates or model retraining processes associated with the metadata graph are automatically triggered, thus forming a closed-loop feedback mechanism between data, model, and performance. This invention can effectively address issues such as data distribution drift, changes in business scenarios, and model performance degradation, significantly improving the stability, adaptability, and long-term performance of AI models in complex application environments, and enabling collaborative management of the entire lifecycle of AI models from development to operation and maintenance. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a collaborative management method for the entire lifecycle of an AI model, as provided in an embodiment of this application. Figure 2 yes Figure 1 A flowchart illustrating step S5 in the process; Figure 3 This is a schematic diagram of a collaborative management system for the entire lifecycle of an AI model provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0021] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0022] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0024] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0025] OpenMetadata is an open-source metadata management platform for unified management of data assets and metadata. It is used to collect, store, and manage data sources, data tables, fields, data lineage, and related information to support traceable associations between data and models.

[0026] Elasticsearch is a distributed search and analysis engine used to index structured or semi-structured data and perform efficient retrieval. In this embodiment, it is used to index and quickly query metadata graph-related information.

[0027] Neo4j refers to a graph database system based on a graph data model, used to store and manage nodes and the relationships between nodes. In the embodiments of this application, it is used to perform graph-structured storage and analysis of entity nodes and their associated relationships in the metadata graph.

[0028] The OpenMetadata API refers to the application programming interface provided by the OpenMetadata platform, which is used to automatically collect, query, update, and manage metadata objects to support the construction and dynamic maintenance of the metadata graph.

[0029] The MLflow API refers to the application programming interface provided by the MLflow platform, which is used to record, manage, and query parameters, metrics, model versions, and running results during the AI ​​model training process to support traceable management of the model training process.

[0030] TensorFlow is a deep learning framework for building and training machine learning models. It supports defining model structures in a computational graph manner and is used to perform model training and inference computations.

[0031] PyTorch is a deep learning framework based on dynamic graph mechanisms, used to build, train, and optimize neural network models, supporting flexible model structure definitions and gradient calculations.

[0032] DolphinScheduler is a distributed workflow scheduling system used for the orchestration, scheduling, and execution management of data processing tasks, model training tasks, or related computational tasks.

[0033] MLflow is a platform for AI model lifecycle management. It is used to uniformly manage and record model experiments, model parameters, model versions, and model deployment information to support collaborative management of model development, training, and operation and maintenance processes.

[0034] This application provides a method and system for collaborative management of the entire lifecycle of an AI model. This solution achieves collaborative processing of data cleaning, standardization, and consistency constraints through unified acquisition and dynamic data governance of multi-source heterogeneous data. This effectively reduces the impact of data quality differences on model training and inference performance, and improves the reusability and reliability of data assets. Based on a metadata graph constructed from standardized datasets, data sources, feature structures, processing rules, and model configurations are linked through modeling, making the model building, training, deployment, and operation processes traceable and explainable, avoiding information fragmentation and configuration mismatch issues during model development. By introducing a preset development mode to generate model configurations and combining it with a unified scheduling mechanism for computing resources, collaborative optimization of the model training process and resource usage strategies is achieved, improving model training efficiency and resource utilization. Furthermore, after model deployment and inference service release, model performance is continuously monitored, and when performance changes are detected, data governance updates or model retraining processes associated with the metadata graph are automatically triggered, thus forming a closed-loop feedback mechanism between data, model, and performance. This invention can effectively address issues such as data distribution drift, changes in business scenarios, and model performance degradation, significantly improving the stability, adaptability, and long-term performance of AI models in complex application environments, and enabling collaborative management of the entire lifecycle of AI models from development to operation and maintenance.

[0035] This application provides a collaborative management method for the entire lifecycle of an AI model, relating to the field of artificial intelligence technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the collaborative management method for the entire lifecycle of an AI model, but is not limited to the above forms.

[0036] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0037] Figure 1 This is an optional flowchart of an AI model full lifecycle collaborative management method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S1 to S7: S1: Acquire multi-source heterogeneous data; In this embodiment, the step of acquiring multi-source heterogeneous data is performed. This multi-source heterogeneous data originates from data carriers with different business scenarios and data formats, including structured data, semi-structured data, and unstructured data. Since the data sources differ in data organization, field structure, data update frequency, and access methods, data from different sources is collected uniformly during the data acquisition phase to ensure that subsequent processing can be performed under the same methodological flow.

[0038] In practice, data acquisition operations are performed using corresponding methods based on the different types of data sources. For structured data, data content is acquired according to preset acquisition rules; for semi-structured or unstructured data, data content is acquired through file monitoring, API calls, or log parsing. During the data acquisition process, the source identifier, acquisition time, and acquisition order information are recorded for each batch of data to distinguish data sets from different sources and batches.

[0039] S2: Based on multi-source heterogeneous data, perform dynamic data governance operations to generate standardized datasets; among which, dynamic data governance operations include data cleaning, structure standardization, and automatic metadata collection.

[0040] This includes performing dynamic data governance operations based on multi-source heterogeneous data to generate standardized datasets, including: S21: Perform data structure parsing on multi-source heterogeneous data, identify the data format, field type and field semantics of different data sources, and generate data structure parsing results; S22: Based on the data structure parsing results, perform data cleaning processing on multi-source heterogeneous data to eliminate abnormal data and form valid data; S23: Based on the cleaned and valid data, perform structural standardization processing to map data from different data sources into a unified data structure; S24: During the data cleaning and structural standardization process, metadata information corresponding to multi-source heterogeneous data is collected simultaneously. S25: Generate a standardized dataset based on the data that has been structurally standardized and associated with metadata information.

[0041] In this embodiment, for the acquired multi-source heterogeneous data, data structure parsing processing is first performed on the multi-source heterogeneous data. Specifically, according to the different data source types, the data format, field structure, and field type information of database data, file data, data lake data, and interface data are parsed respectively, and the business semantic meaning corresponding to each field is identified, forming a data structure parsing result containing data source identifier, field set, field type, and field semantic description, thereby providing a unified structured input basis for subsequent data governance processing.

[0042] After completing the data structure parsing, data cleaning is performed on the multi-source heterogeneous data based on the parsing results. In practice, the validity of data records is verified according to field types and semantic constraints. Abnormal values, missing values, duplicate records, and data with non-compliant formats are identified and processed. At the same time, issues such as encoding differences, time format differences, and invalid characters in data from different sources are uniformly corrected to ensure that the cleaned data meets the requirements of usability and consistency, thus forming a cleaned and valid data set.

[0043] After the cleaned and valid data is generated, structural standardization processing is further performed based on the valid data. Specifically, fields from different data sources but with the same business meaning are uniformly mapped, and field types, field names, and data units are standardized and converted according to preset rules. For data with inconsistent structural levels, field expansion, merging, or aggregation processing is performed to ensure that data from different sources are consistent at the structural level, thereby obtaining a standardized data representation with a unified structure and clear semantics.

[0044] During the data cleaning and structural standardization process, metadata information related to the data processing is collected simultaneously. This collected metadata includes data source information, data structure information, field meaning descriptions, cleaning and transformation rules, and data processing time. The metadata information is then linked to the corresponding data entities to facilitate subsequent tracing of the data processing process and data lineage.

[0045] After structural standardization and association with metadata information, a standardized dataset is generated based on the data. Specifically, the standardized data is stored in a unified format, and a corresponding dataset identifier and version information are generated for each dataset. At the same time, the associated metadata information is also saved, enabling the standardized dataset to be directly referenced by subsequent steps and providing a stable and reliable data foundation for metadata graph construction and AI model development and training.

[0046] S3: Construct a metadata graph based on a standardized dataset; Among these, a metadata graph is constructed based on a standardized dataset, including: S31: Extract metadata from the data tables, fields, and their source information in the standardized dataset to obtain a metadata set; the metadata set includes data structure information, field semantic information, data source identifiers, and update feature information; S32: Based on the metadata set, construct a metadata association model containing data entity nodes and relation edges. The data entity nodes include data source nodes, data table nodes, and field nodes. Relationship edges are used to represent the inclusion relationship, source relationship, and dependency relationship between data entity nodes. S33: Perform unified identification and semantic mapping processing on the metadata association model; S34: Generate a metadata graph based on the metadata association model that completes semantic mapping.

[0047] In this embodiment, metadata extraction is first performed based on a standardized dataset. Specifically, after pre-configuring data source credentials and collection rules, the metadata collection process calls OpenMetadata to automatically collect data tables, fields, and their source information involved in the standardized dataset. This process obtains information such as data table structure, field names, field types, field descriptions, data source identifiers, and update times, thereby forming a metadata set containing data structure information, field semantic information, data source identifiers, and update characteristic information.

[0048] After obtaining the metadata set, a metadata association model is constructed based on the metadata set. In specific implementation, different objects in the metadata set are abstracted into data entity nodes, including data source nodes that represent the source of data, data table nodes that represent standardized datasets, and field nodes that represent field-level information. At the same time, according to the data generation and processing process, relational edges are constructed through the acquisition interface (API) provided by OpenMetadata to describe the inclusion relationship between data tables and fields, the source relationship between data sources and data tables, and the dependency relationship between data tables formed due to cleaning, transformation, or processing, thereby forming a complete data association model.

[0049] After constructing the data entity nodes and relationship edges, a unified identification process is performed on the metadata association model. Specifically, through OpenMetadata's built-in entity identification mechanism, a unique identifier is generated for each data source node, data table node, and field node, and the identifier is bound to the corresponding data source identifier, collection task identifier, and update time feature. When the data table structure or field changes, a new version identifier is automatically generated, enabling different versions of metadata objects to be accurately distinguished in the graph, thereby ensuring the consistency and traceability of metadata relationships during continuous evolution.

[0050] After completing the unified identifier processing, semantic mapping processing is performed on the metadata association model. In specific implementation, the metadata management interface provided by OpenMetadata is used to standardize the semantic descriptions, business tags, and ownership information of fields in field nodes, and to semantically align fields with the same or similar business meanings in different data tables. At the same time, semantic associations are established between data processing tasks, dependencies, and corresponding data entity nodes, so that the metadata association model can simultaneously express data structure relationships, processing dependencies, and business semantic relationships.

[0051] After semantic mapping, a metadata graph is generated based on the metadata association model. Specifically, data entity nodes and relationship edges are written to metadata storage and indexed using Elasticsearch to support fast retrieval and querying. Simultaneously, the metadata association results can be exported to Neo4j for graph relationship analysis and visualization. When the data structure or update characteristics change, incremental data collection and update processing is triggered, ensuring that the generated metadata graph continuously reflects the latest state of the standardized dataset. This provides a traceable metadata foundation for subsequent AI model construction, training, deployment, and closed-loop optimization.

[0052] S4: Based on the metadata graph, execute AI model construction according to the preset development mode to generate the AI ​​model configuration to be trained; The development modes include canvas mode and expert mode; Canvas mode is used to perform AI model building operations based on a visual workflow orchestration approach; Expert mode is used to perform AI model building operations based on code.

[0053] Specifically, the canvas mode includes: Based on the metadata graph, the data structure information and field semantic information corresponding to the standardized dataset are automatically loaded in the visualization canvas interface to generate configurable data nodes; Based on data nodes, and in accordance with preset process constraints, data preprocessing nodes, model building nodes, and model evaluation nodes are configured sequentially, and data dependencies between nodes are established to form a candidate model building process. Based on the metadata graph, consistency and dependency integrity checks are performed on the input and output fields of each node in the candidate model construction process to obtain the model construction process that passes the checks. Based on the validated model building process, a structured model configuration description file is generated, and the model configuration description file is used as the output configuration of the AI ​​model to be trained.

[0054] Specifically, the expert model includes: Based on the metadata graph, determine the data tables and field semantic information corresponding to the data to be modeled in the standardized dataset; In the code development environment, based on the data table and field semantic information, write model training code to clarify the feature input items, model calculation structure and training parameters at the code level; Execute the model training code to generate model parameters and performance metrics corresponding to the feature inputs and model computation structure; Write and associate model parameters and performance metrics with the metadata graph; Based on the model parameters and performance metrics written into the metadata graph, structured model configuration description information is generated. The model configuration description information is output as the configuration of the AI ​​model to be trained.

[0055] In this embodiment, AI model building is performed according to a preset development mode based on the metadata graph to generate an AI model configuration to be trained. The development modes include a canvas mode and an expert mode, which are used to meet the model building needs under different development methods. The canvas mode focuses on model building based on visual workflow orchestration, while the expert mode focuses on model building based on code. Both use the metadata graph as a unified basis for data semantics and structure.

[0056] Specifically, in canvas mode, based on the metadata graph, the data structure information and field semantic information corresponding to the standardized dataset are automatically loaded into the visualization canvas interface. Specifically, the OpenMetadata API is called to read the field names, data types, business descriptions, and lineage information recorded in the data table nodes and field nodes, and this information is mapped to configurable data nodes in the canvas, ensuring that each data node clearly corresponds to a specific data table, field, and version identifier in OpenMetadata.

[0057] After loading the data nodes, based on these nodes, data preprocessing nodes, model building nodes, and model evaluation nodes are sequentially configured in the canvas according to preset process constraints. Data dependencies between these nodes are established to form the candidate model building process. In practice, data preprocessing operators, model algorithm components (based on TensorFlow or PyTorch), and evaluation metric components are selected through the Canvas interface. Field-level input-output mapping relationships are clearly defined between nodes, ensuring that the candidate process structurally meets the requirements of complete data dependencies and a clear execution order.

[0058] After the candidate model building process is established, consistency and dependency integrity checks are performed on the process based on the metadata graph. Specifically, in conjunction with the field type definitions, semantic descriptions, and lineage relationships in OpenMetadata, type matching checks and dependency closure checks are performed on the input and output fields of each node. When the checks pass, a structured model configuration description file is generated based on the model building process. The model configuration description file describes the data source, feature fields, model structure, and parameter configuration in JSON or YAML format, and serves as the configuration output for the AI ​​model to be trained.

[0059] Specifically, in expert mode, the first step is to determine the data tables and field semantic information corresponding to the data to be modeled in the standardized dataset based on the metadata graph. Specifically, by querying the data table entities and field entities in OpenMetadata, the field names, field types, field semantic labels, and data lineage corresponding to the data to be modeled are obtained, ensuring that the data fields used in the code development phase are consistent with the standardized dataset in both semantics and structure.

[0060] In the code development environment, model training code is written based on the data table and field semantic information. In specific implementation, Python is used to write the model training logic, explicitly specifying the feature input fields, feature types, and model calculation structure in the code, and defining the training parameters; at the same time, during the model training process, the MLflow API is called to record model parameters, training metrics, and running information, so that the model training process is traceable at both the parameter and metric levels.

[0061] After executing the model training code, the model parameters and performance metrics corresponding to the feature input items and model computation structure are generated, and the model parameters and performance metrics are registered through the MLflow register_model() interface. Subsequently, the model identifier, model version, associated data table and field information are synchronously written to OpenMetadata, and a structured model configuration description is generated based on the information. The model configuration description is used as the configuration output of the AI ​​model to be trained.

[0062] S5: Based on the AI ​​model configuration, schedule computing resources to perform AI model training to generate the trained target AI model; Among them, reference Figure 2 As shown, based on the AI ​​model configuration, computing resources are scheduled to perform AI model training to generate the trained target AI model, including: S51: Based on the AI ​​model configuration, determine the model structure information, algorithm framework, feature input constraints and training parameters, and generate the model training task; S52: Based on the model training task, and considering the required computational scale and data scale, schedule and allocate the corresponding computing resources. S53: Based on computing resources, perform model training on the standardized dataset according to the model training task, iteratively update the model parameters, and generate the trained model results; S54: After the model training is completed, write the model results and corresponding training metrics into the metadata graph and associate them to generate the target AI model.

[0063] In this embodiment, a model training task is first generated based on the AI ​​model configuration. Specifically, when the AI ​​model is configured in canvas mode using a generated JSON or YAML configuration file, the model structure information, feature input field constraints, training parameters, and the algorithm framework used are parsed from it. When the AI ​​model is configured in expert mode using a structured configuration description, the Python training entry point, model computation structure, feature input order, and training parameter set are parsed from it. Based on the parsing results, the algorithm framework used for model training is determined to be either TensorFlow or PyTorch, and the version identifier of the standardized dataset used and the semantic constraints of the fields are written into the training task description, thereby forming an executable and traceable model training task.

[0064] After generating the model training task, computing resources are scheduled and allocated based on the computational and data scale of the training task. Specifically, the required CPU, GPU, memory, and storage resources for the training task are assessed according to the model parameter size, training epochs, batch size, and the data volume of the standardized dataset. The model training task is then triggered via DolphinScheduler using a Directed Acyclic Graph (DAG) approach, ensuring that the training task executes while meeting the completion conditions of the upstream data governance task. Simultaneously, task dependencies, parallelism control, and retry strategies are set during the scheduling process to guarantee the consistency of the model training process and the data preparation process in terms of time and state.

[0065] After allocating computing resources, model training is performed on the standardized dataset according to the model training task. Specifically, based on the resolved feature input constraints, the corresponding data version is loaded from the standardized dataset, and model training is performed within the selected algorithm framework. The model parameters are iteratively updated through forward computation, backpropagation, and parameter update processes until the preset training epochs or convergence conditions are reached. During the training process, MLflow continuously records training parameters, running status, and training metrics, ensuring the reproducibility and comparability of the model training process.

[0066] After model training is complete, the trained model results and corresponding training metrics are written to and associated with the metadata graph to generate the trained target AI model. Specifically, the trained model is registered using the MLflow register_model() interface, generating a model identifier and model version information. The model version is then associated with its corresponding training data version and feature field semantic information. Simultaneously, the model identifier, model version, and training metrics are written to OpenMetadata, enabling the metadata graph to fully express the relationship between "data—training task—model—version," thus forming a traceable target AI model and providing a foundation for subsequent model deployment and operational monitoring.

[0067] S6: Based on the target AI model, perform model deployment and inference service release processing; In this embodiment, a pre-deployment preparation process is performed on the target AI model. Specifically, the input data structure, output format, model running parameters, and dependent environment information of the target AI model are uniformly sorted out, and a clear association is established between the target AI model and its corresponding training data version, feature field set, and training evaluation results. This ensures that the model can perform inference processing under consistent data semantics and operating conditions in subsequent running stages, avoiding distortion of inference results due to environmental differences or data inconsistencies.

[0068] Subsequently, based on a pre-defined deployment strategy, the target AI model is deployed. The deployment strategy includes at least the selection of the model execution mode, the determination of resource configuration methods, and rules for loading execution parameters. In practice, based on the concurrency scale of model inference requests, real-time requirements, and resource constraints, an appropriate deployment method is selected to load the target AI model, placing it in a callable running state. During deployment, the model deployment time, deployment configuration parameters, and the corresponding model version identifier are recorded for subsequent tracking and comparative analysis of the model's running status.

[0069] After model deployment, the target AI model undergoes further inference service deployment. Specifically, based on the model's input and output specifications, a unified inference call interface description is generated, clearly defining the supported input parameter types, value ranges, and output result structures. This enables business processes to submit inference requests to the target AI model and receive inference results through the interface. Simultaneously, the inference service deployment information is associated with and stored with the model version information. This ensures accurate location of the currently running model instance and its corresponding data and parameter status during subsequent model upgrades or rollbacks, thereby guaranteeing the stability and controllability of the model's runtime.

[0070] S7: Continuously monitor the performance metrics of the target AI model during its operation. When performance changes are detected, trigger data governance updates or model retraining processes associated with the metadata graph to form a closed-loop collaborative management of the entire lifecycle of the AI ​​model.

[0071] In this embodiment, after the target AI model is deployed and begins providing inference services, continuous performance monitoring is performed on its operation. Specifically, according to a preset monitoring cycle, the operational status of the AI ​​model in actual inference scenarios is collected. This operational status includes at least changes in inference accuracy, fluctuations in inference response latency, and resource consumption levels during model operation. The collected performance metrics are associated with the corresponding model version identifier and inference timestamp to form a continuous performance data sequence during model operation.

[0072] During performance monitoring, when changes in model performance metrics are detected, analysis and judgment are performed on these changes. Specifically, the currently collected performance metrics are compared with preset performance thresholds or historical benchmark performance metrics. When it is determined that the model performance is continuously declining, fluctuating abnormally, or failing to meet expected requirements, based on the established data-model correlation, the training data version, data governance rules, and historical training records on which the target AI model relies are traced and analyzed to identify possible changes in data quality or model capability degradation corresponding to the performance changes.

[0073] After completing the analysis and judgment, the corresponding closed-loop collaborative processing flow is triggered based on the analysis results. Specifically, when it is determined that the model performance change is related to the quality of the input data, data governance update processing associated with the target AI model is triggered. This involves re-cleaning, structural standardization, or repairing the relevant data and generating an updated standardized dataset. When it is determined that the model performance change is related to a decline in the model's generalization ability, model retraining is triggered, and a new model version is generated based on the latest standardized dataset. Through these methods, a continuously linked closed-loop processing flow is formed between model operation, data governance, and model training, thereby achieving collaborative management throughout the entire lifecycle of the AI ​​model.

[0074] Please see Figure 3 This application also provides an AI model full lifecycle collaborative management system, the system including: The data acquisition module is used to acquire heterogeneous data from multiple sources. The dynamic data governance module is used to perform dynamic data governance operations based on multi-source heterogeneous data to generate standardized datasets; The metadata management module is used to build a metadata graph based on a standardized dataset; The AI ​​model building module is used to build AI models according to the metadata graph and a preset development mode, and generate the configuration of the AI ​​model to be trained. The training scheduling module is used to schedule computing resources to perform AI model training based on AI model configuration, so as to generate the target AI model after training. The deployment service module is used to perform model deployment and inference service release processing based on the target AI model; The feedback module is used to continuously monitor the performance indicators of the target AI model during its operation. When a performance change is detected, it triggers a data governance update or model retraining process associated with the metadata graph, so as to form a closed-loop collaborative management of the entire life cycle of the AI ​​model.

[0075] This application also provides an AI model lifecycle collaborative management system. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This system can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0076] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0077] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0078] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0079] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0080] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0081] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0082] This application provides a collaborative management method and system for the entire lifecycle of AI models. This solution achieves collaborative processing of data cleaning, standardization, and consistency constraints through unified acquisition and dynamic data governance of multi-source heterogeneous data. This effectively reduces the impact of data quality differences on model training and inference performance, improving the reusability and reliability of data assets. Based on a metadata graph constructed from standardized datasets, data sources, feature structures, processing rules, and model configurations are correlated and modeled, making the model building, training, deployment, and operation processes traceable and interpretable, avoiding information fragmentation and configuration mismatch issues during model development. By introducing a preset development mode to generate model configurations and combining it with a unified scheduling mechanism for computing resources, collaborative optimization of the model training process and resource usage strategies is achieved, improving model training efficiency and resource utilization. Furthermore, after model deployment and inference service release, model performance is continuously monitored, and when performance changes are detected, data governance updates or model retraining processes associated with the metadata graph are automatically triggered, thus forming a closed-loop feedback mechanism between data, model, and performance. This invention can effectively address issues such as data distribution drift, changes in business scenarios, and model performance degradation, significantly improving the stability, adaptability, and long-term performance of AI models in complex application environments, and enabling collaborative management of the entire lifecycle of AI models from development to operation and maintenance.

[0083] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0084] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0085] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0086] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0087] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0088] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A collaborative management method for the entire lifecycle of an AI model, characterized in that, The method includes the following steps: Acquire multi-source heterogeneous data; Based on the multi-source heterogeneous data, perform dynamic data governance operations to generate a standardized dataset; Based on the standardized dataset, a metadata graph is constructed; Based on the metadata graph, AI model construction is performed according to a preset development mode to generate an AI model configuration to be trained; Based on the AI ​​model configuration, computing resources are scheduled to perform AI model training to generate the trained target AI model; Based on the target AI model, perform model deployment and inference service release processing; The performance metrics of the target AI model are continuously monitored during its operation. When a performance change is detected, a data governance update or model retraining process associated with the metadata graph is triggered to form a closed-loop collaborative management of the entire lifecycle of the AI ​​model.

2. The method according to claim 1, characterized in that, The dynamic data governance operations include data cleaning, structure standardization, and automatic metadata collection.

3. The method according to claim 2, characterized in that, The step of performing dynamic data governance operations based on the multi-source heterogeneous data to generate a standardized dataset includes: Data structure parsing is performed on the multi-source heterogeneous data to identify the data format, field type, and field semantics of different data sources, and to generate data structure parsing results. Based on the data structure parsing results, data cleaning processing is performed on the multi-source heterogeneous data to eliminate abnormal data and form valid data; Based on the cleaned and valid data, structural standardization processing is performed to map data from different data sources into a unified data structure. During the data cleaning and structural standardization process, metadata information corresponding to the multi-source heterogeneous data is collected simultaneously. A standardized dataset is generated based on the data that has undergone structural standardization and is associated with the metadata information.

4. The method according to claim 1, characterized in that, The construction of the metadata graph based on the standardized dataset includes: Metadata is extracted from the data tables, fields, and their source information in the standardized dataset to obtain a metadata set; the metadata set includes data structure information, field semantic information, data source identifiers, and update feature information; Based on the metadata set, a metadata association model is constructed that includes data entity nodes and relation edges. The data entity nodes include data source nodes, data table nodes, and field nodes. The relation edges are used to represent the inclusion relationship, source relationship, and dependency relationship between the data entity nodes. Perform unified identification and semantic mapping processing on the metadata association model; A metadata graph is generated based on a metadata association model that completes semantic mapping.

5. The method according to claim 1, characterized in that, The development modes include canvas mode and expert mode; The canvas mode is used to perform AI model building operations based on a visual workflow orchestration method; The expert mode is used to perform AI model building operations based on code.

6. The method according to claim 5, characterized in that, The canvas modes specifically include: Based on the metadata graph, the data structure information and field semantic information corresponding to the standardized dataset are automatically loaded into the visualization canvas interface to generate configurable data nodes; Based on the data nodes, according to the preset process constraint rules, data preprocessing nodes, model building nodes and model evaluation nodes are configured in sequence, and data dependencies between each node are established to form a candidate model building process. Based on the metadata graph, consistency and dependency integrity checks are performed on the input and output fields of each node in the candidate model construction process to obtain the model construction process that passes the checks. Based on the verified model building process, a structured model configuration description file is generated, and the model configuration description file is output as the configuration of the AI ​​model to be trained.

7. The method according to claim 5, characterized in that, The expert mode specifically includes: Based on the metadata graph, determine the data tables and field semantic information corresponding to the data to be modeled in the standardized dataset; In the code development environment, based on the data table and field semantic information, model training code is written to clarify the feature input items, model calculation structure and training parameters at the code level; Execute the model training code to generate model parameters and performance metrics corresponding to the feature input items and model computation structure; Write the model parameters and performance metrics into and associate them with the metadata graph; Based on the model parameters and performance metrics written into the metadata graph, structured model configuration description information is generated; The model configuration description information is output as the configuration of the AI ​​model to be trained.

8. The method according to claim 1, characterized in that, The step of scheduling computing resources to perform AI model training based on the AI ​​model configuration to generate a trained target AI model includes: Based on the AI ​​model configuration, the model structure information, algorithm framework, feature input constraints and training parameters are determined, and a model training task is generated. Based on the model training task, and considering the required computational scale and data scale of the task, the corresponding computing resources are scheduled and allocated. Based on the computing resources, the model is trained on the standardized dataset according to the model training task, the model parameters are iteratively updated, and the trained model result is generated. After the model training is completed, the model results and corresponding training metrics are written into and associated with the metadata graph to generate the target AI model.

9. A collaborative management system for the entire lifecycle of an AI model, characterized in that, The system includes: The data acquisition module is used to acquire heterogeneous data from multiple sources. The dynamic data governance module is used to perform dynamic data governance operations based on the multi-source heterogeneous data to generate a standardized dataset; The metadata management module is used to construct a metadata graph based on the standardized dataset. The AI ​​model building module is used to build AI models according to the metadata graph and a preset development mode, and generate AI model configurations to be trained. The training scheduling module is used to schedule computing resources to perform AI model training based on the AI ​​model configuration, so as to generate a target AI model that has been trained. The deployment service module is used to perform model deployment and inference service release processing based on the target AI model; The feedback module is used to continuously monitor the performance indicators of the target AI model during its operation. When a performance change is detected, it triggers a data governance update or model retraining process associated with the metadata graph, so as to form a closed-loop collaborative management of the entire life cycle of the AI ​​model.

10. A collaborative management system for the entire lifecycle of an AI model, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1-8.

Images