Data management system based on model training

Abstract

The embodiment of the invention provides a data management system based on model training. The system comprises a control unit, a data processing unit, a model training unit and a storage unit, the control unit is used for sending an instruction to each unit and receiving information fed back by each unit; the data processing unit is used for receiving the data processing instruction of the control unit, completing data processing according to the data processing instruction, and sending the processed data to the model training unit; the model training unit is used for receiving the training instruction of the control unit and the processed data sent by the data processing unit, and completing model training on a target model according to the training instruction and the processed data; and the storage unit is used for receiving the process data sent by the data processing unit, receiving the related information of the model trained by the model training unit and the process data of a training instruction, and storing the received information. The system is used for achieving the technical effect of improving the data management efficiency in model training.

Description

Technical Field

[0001] The present application relates to the technical field of information systems, and in particular to a data management system based on model training. Background Art

[0002] In today's era of rapid information development, data has become one of the most important production factors. Big data management and application majors are facing unprecedented opportunities and challenges. With the rapid development of artificial intelligence technology, the application prospects of model training and data management systems in various industries are becoming more and more broad.

[0003] In the prior art, the data management system for model training mainly implements the model training process through a data acquisition system, a preprocessing system and a model training system.

[0004] However, in the prior art, the resource consumption of data collection and processing is large, the accuracy of the trained model is not high, and at the same time, as the data and model are updated, the management of the data cannot meet the needs of users. Therefore, the data management system based on model training in the prior art has the technical problem of low management efficiency. Summary of the invention

[0005] The embodiment of the present application provides a data management system based on model training, which is used to achieve the technical effect of improving the efficiency of data management during model training.

[0006] In a first aspect, an embodiment of the present application provides a data management system based on model training, the data management system comprising: a control unit, a data processing unit, a model training unit and a storage unit;

[0007] A control unit, used to send instructions to each unit and receive feedback information from each unit;

[0008] A data processing unit is used to receive a data processing instruction from the control unit, complete data processing according to the data processing instruction, send progress information and completion results of the data processing to the control unit, send the processed data to the model training unit, and send process data of the data processing instruction to the storage unit;

[0009] A model training unit is used to receive the training instructions of the control unit and the processed data sent by the data processing unit, complete the model training for the target model according to the training instructions and the processed data, send the progress information and performance indicator information of the model training to the control unit, and send the information related to the trained model and the process data of the training instructions to the storage unit;

[0010] The storage unit is used to receive process data sent by the data processing unit, and is also used to receive model-related information trained by the model training unit and process data of training instructions, and is also used to store the received information.

[0011] Optionally, the data processing unit includes: a data acquisition module, a data cleaning and annotation module, and a feature engineering module;

[0012] A data acquisition module is used to receive a data acquisition instruction from the control unit, complete data acquisition according to the data acquisition instruction, send progress information and acquisition status of the acquired data to the control unit, send the acquired data to the data cleaning and annotation module, and send the acquired process data to the storage unit;

[0013] A data cleaning and labeling module is used to receive data cleaning and labeling instructions from the control unit, and is also used to receive collected data sent by the data acquisition module, perform data cleaning and data labeling on the collected data according to the data cleaning and labeling instructions, obtain data after data cleaning and data labeling, send progress information and result information of data cleaning and data labeling to the control unit, send the data after data cleaning and data labeling to the feature engineering module, and send process data of data cleaning and data labeling to the storage unit;

[0014] The feature engineering module is used to receive the data feature parameter setting instruction of the control unit, and is also used to receive the data after data cleaning and data annotation sent by the data cleaning and annotation module, set the feature parameters of the data after data cleaning and data annotation according to the data feature parameter setting instruction, obtain the data after the feature parameter setting, send the progress information and result information of the feature parameter setting of the data after data cleaning and data annotation to the control unit, send the data after the feature parameter setting to the model training unit, and send the process data of the feature parameter setting of the data after data cleaning and data annotation to the storage unit.

[0015] Optionally, the data acquisition module includes: a data acquisition scheduler and a data screening submodule;

[0016] A data collection scheduler is used to receive data collection instructions from the control unit, determine priority information of the data collection instructions, and obtain resource status information of the model training and data management system, collect data from multiple databases according to the resource status information of the model training and data management system and the priority information of the data collection instructions, obtain the collected data, send progress information and collection status of the collected data to the control unit, and send the collected data to the storage unit;

[0017] The data filtering submodule is used to receive the filtering instructions sent by the control unit, and is also used to receive the collected data sent by the data acquisition scheduler, filter the collected data according to the filtering instructions, preset data type, preset data scale, and preset data quality standard, obtain the filtered data, and send the filtered data to the data cleaning and annotation module.

[0018] Optionally, the data cleaning and labeling module includes: a data cleaning submodule and a data labeling submodule;

[0019] A data cleaning submodule is used to receive a data cleaning instruction from the control unit, and is also used to receive the collected data sent by the data acquisition module, clean the collected data according to the data cleaning instruction and the preset cleaning rules, obtain cleaned data, send the progress information and result information of the data cleaning to the control unit, send the cleaned data to the data annotation submodule, and send the process data of the data cleaning to the storage unit;

[0020] The data labeling submodule is used to receive the data labeling instructions from the control unit, and is also used to receive the cleaned data sent by the data cleaning submodule. According to the data labeling instructions, an active learning algorithm is used to label the cleaned data to obtain the labeled data, and the progress information and result information of the labeling are sent to the control unit, the labeled data is sent to the feature engineering module, and the labeled process data is sent to the storage unit.

[0021] Optionally, the feature engineering module includes: a feature selection submodule and a feature transformation submodule;

[0022] A feature selection submodule is used to receive a feature selection instruction from the control unit, and is also used to receive the data after data cleaning and data labeling sent by the data cleaning and labeling module, screen the data after data cleaning and data labeling according to the feature selection instruction, information gain and chi-square test algorithm, determine multiple data features, select multiple data features according to a method combining deep learning and traditional feature selection, determine multiple selected data features, send progress information and result information of feature selection to the control unit, send the multiple selected data features to the feature transformation submodule, and send process data of feature selection to the storage unit;

[0023] The feature transformation submodule is used to receive the feature transformation instruction of the control unit, and is also used to receive the multiple selected data features sent by the feature selection submodule, and according to the feature transformation instruction, perform data standardization operation on the multiple selected data features to obtain standardized data, determine the data type of the standardized data, perform feature transformation on data of different data types to obtain transformed data, send the progress information and result information of the feature transformation to the control unit, send the transformed data to the model training unit, and send the process data of the feature transformation to the storage unit; wherein the data type includes numerical data and non-numerical data, the numerical data is normalized to obtain normalized data, and the non-numerical data is discretized to obtain discretized data.

[0024] Optionally, the model training unit includes: a model selection module, a model parameter initialization module and a model training module;

[0025] A model selection module is used to receive a model selection instruction from the control unit and also to receive processed data sent by the data processing unit. The model selection module includes multiple models. The multiple models are selected according to the model selection instruction and the data features of the processed data to determine the target model, and the result information of the model selection is sent to the control unit. The target model and the processed data are sent to the model parameter initialization module;

[0026] A model parameter initialization module is used to receive a parameter initialization instruction from the control unit, and is also used to receive a target model and processed data sent by the model selection module, initialize the parameters of the target model according to the parameter initialization instruction and the data features of the processed data, obtain the model after parameter initialization, send the model after parameter initialization and the processed data to the model training module, and send the process data of parameter initialization to the storage unit;

[0027] The model training module is used to receive the model training instructions from the control unit, and is also used to receive the parameter-initialized model and processed data sent by the model parameter initialization module. According to the model training instructions, the processed data is input into the parameter-initialized model, and the parameter-initialized model is trained according to multiple optimization algorithms. The speed of the training process is improved through distributed computing technology, and the trained model is obtained. The progress information and result information of the training are sent to the control unit, and the process data in training is sent to the storage unit.

[0028] Optionally, the model training unit further includes: a model verification module, a model evaluation module and a visualization display module;

[0029] A model verification module is used to receive a model verification instruction from the control unit, and is also used to receive a trained model sent by the model training module, and verify the trained model through a cross-validation method according to the model verification instruction to obtain multiple verification results and a verified model, and send the verified model to the model evaluation module, send the process data of the model verification to the storage unit, and send the progress information of the model verification and multiple verification results to the control unit; wherein the verification results include accuracy, recall rate, mean square error information and robustness index;

[0030] A model evaluation module, used to receive a model evaluation instruction from a control unit, and also used to receive a verified model sent by a model verification module, determine a model type of the verified model according to the model evaluation instruction, determine a corresponding evaluation index according to the model type of the verified model, evaluate the verified model using multiple evaluation indexes, obtain multiple evaluation results, send the multiple evaluation results to a visualization display module, send progress information and result information of the evaluation to the control unit, and send process data of the evaluation to a storage unit; wherein the model types of the verified model include classification models and regression models, the evaluation indexes corresponding to the classification model include accuracy, recall rate and F1 value, and the evaluation indexes corresponding to the regression model include mean square error information and mean absolute error information;

[0031] The visualization display module is used to receive the display instructions of the control unit and also to receive multiple evaluation results sent by the model evaluation module. According to the display instructions, it determines the display form of each evaluation indicator and sends the display results corresponding to each evaluation result to the control unit and the storage unit respectively.

[0032] Optionally, the storage unit includes: a model storage and management module, and a data storage and indexing module;

[0033] The model storage and management module is used to receive the model storage management instruction of the control unit, and is also used to receive the trained model sent by the model training unit, store and manage the trained model according to the model storage management instruction, and send the storage status of the trained model to the control unit;

[0034] The data storage and indexing module is used to receive the data storage indexing instructions of the control unit, and is also used to receive the process data sent by the data processing unit and the process data sent by the model training unit, and manage multiple process data according to the data storage indexing instructions.

[0035] Optionally, the model storage and management module includes: a model storage submodule, a model index submodule and a model update submodule;

[0036] The model storage submodule is used to receive the model storage instruction of the control unit, and is also used to receive the trained model sent by the model training unit, calibrate the trained model according to the model storage instruction, and send the calibrated model to the control unit, the model index submodule and the model update submodule respectively; wherein the calibration content includes an identifier, a version number, model structure information, model training time and model parameter values;

[0037] The model indexing submodule is used to receive the model indexing instruction of the control unit, and is also used to receive the calibrated model sent by the model storage submodule, perform language analysis on the calibrated model using language indexing technology, determine the keyword information of each calibrated model, send the keyword information to the control unit, and provide the control unit with an indexing function according to the model indexing instruction; wherein the keyword information includes the function of the model, the application scenario of the model and the training data characteristics of the model;

[0038] The model update submodule is used to receive the model update instruction from the control unit, and is also used to receive the trained model sent by the model training unit, retrain the trained model according to the model update instruction to obtain a new trained model, send the new trained model to the model storage submodule and the model index submodule, and send the progress information and result information of the model update to the control unit.

[0039] Optionally, the data storage and indexing module includes: a data storage submodule and a data indexing submodule;

[0040] The data storage submodule is used to receive the data storage instruction of the control unit, and is also used to receive the process data of the data processing instruction sent by the data processing unit, and the process data of the training instruction sent by the model training unit, obtain the process type of each process data, classify and store each process data according to the data storage instruction, obtain the classified storage result, and send the classified storage result to the data index submodule and the control unit respectively;

[0041] The data indexing submodule is used to receive the data indexing instruction of the control unit, and is also used to receive the classified storage results sent by the data storage submodule, determine the attribute characteristic information of each data in the classified storage results, determine the hash value corresponding to the attribute characteristic information of each data according to the distributed hash table technology, establish a hash table through multiple hash values, input the data indexing instruction of the control unit into the hash table, and complete the data indexing.

[0042] The data management system based on model training provided by the embodiment of the present application comprises: a control unit, a data processing unit, a model training unit and a storage unit; the control unit is used to send instructions to each unit and receive feedback information from each unit; the data processing unit is used to receive the data processing instructions of the control unit, complete data processing according to the data processing instructions, send the progress information and completion result of the data processing to the control unit, send the processed data to the model training unit, and send the process data of the data processing instructions to the storage unit; the model training unit is used to receive the training instructions of the control unit and the processed data sent by the data processing unit, complete the model training for the target model according to the training instructions and the processed data, send the progress information and performance index information of the model training to the control unit, send the trained model related information and the process data of the training instructions to the storage unit; the storage unit is used to receive the process data sent by the data processing unit, and is also used to receive the model related information trained by the model training unit and the process data of the training instructions, and is also used to store the received information. Through the interaction between multiple units, effective management of data in model training is achieved, and the efficiency of data management is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0044] Figure 1 Schematic diagram of the structure of the data management system based on model training provided in this application Figure 1 ;

[0045] Figure 2 Schematic diagram of the structure of the data management system based on model training provided in this application Figure 2 ;

[0046] Figure 3 Schematic diagram of the structure of the data management system based on model training provided in this application Figure 3 .

[0047] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0048] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of systems and systems consistent with some aspects of the present application as detailed in the appended claims.

[0049] In the prior art, the core functions of the data management system in model training usually rely on the collaborative work of the data acquisition system, the preprocessing system and the model training system to complete the training process of the entire model. With the continuous updating of data and models, the data management system in the prior art has certain drawbacks in data management. On the one hand, the rapid growth and diversification of data have put forward higher requirements for storage and retrieval capabilities; on the other hand, users' demand for flexibility, security and privacy protection of data access is also increasing. Traditional data management methods often find it difficult to meet these requirements at the same time, resulting in inefficient data management and poor user experience. Therefore, the data management system based on model training in the prior art has technical problems of low management efficiency, which not only hinders the improvement of system performance, but also limits its application potential in the fields of artificial intelligence and big data.

[0050] To solve the above problems, the data management system based on model training provided by the embodiment of the present application includes: a control unit, a data processing unit, a model training unit and a storage unit; the control unit is used to send instructions to each unit and receive feedback information from each unit; the data processing unit is used to receive the data processing instructions of the control unit, complete data processing according to the data processing instructions, send the progress information and completion result of the data processing to the control unit, send the processed data to the model training unit, and send the process data of the data processing instructions to the storage unit; the model training unit is used to receive the training instructions of the control unit and the processed data sent by the data processing unit, complete the model training for the target model according to the training instructions and the processed data, send the progress information and performance indicator information of the model training to the control unit, send the trained model related information and the process data of the training instructions to the storage unit; the storage unit is used to receive the process data sent by the data processing unit, and is also used to receive the model related information trained by the model training unit and the process data of the training instructions, and is also used to store the received information. Through the interaction between multiple units, effective management of data in model training is achieved, and the efficiency of data management is improved.

[0051] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0052] Figure 1 Schematic diagram of the structure of the data management system based on model training provided in this application Figure 1 ,like Figure 1 As shown, the data management system 100 based on model training in this embodiment includes: a control unit 101, a data processing unit 102, a model training unit 103 and a storage unit 104.

[0053] The control unit 101 is used to send instructions to each unit and receive feedback information from each unit;

[0054] In this embodiment, data processing instructions are sent to the data processing unit 102, training instructions are sent to the model training unit 103, and information fed back by each unit is received, including progress information, result information and storage status. The control unit provides a user interface for the user end to log in to the system, and adopts a multi-factor authentication method. The authentication methods include but are not limited to passwords, fingerprint recognition, facial recognition and hardware token authentication; ordinary users on the user end can set model training parameters, view model training progress and results, and manage data in the user interface. Administrator users can comprehensively configure and manage the entire system, including but not limited to adding or deleting data sources, adjusting data collection plans and modifying the system's storage strategy; at the same time, the control unit sets access rights for the user end, combining role-based access control with attribute-based access control to achieve more flexible permission management.

[0055] The data processing unit 102 is used to receive the data processing instruction of the control unit 101, complete the data processing according to the data processing instruction, send the progress information and the completion result of the data processing to the control unit 101, send the processed data to the model training unit 103, and send the process data of the data processing instruction to the storage unit 104;

[0056] In this embodiment, the data processing unit 102 processes the data by receiving instructions from the control unit 101, and feeds back the progress of the processing process and the completed results to the control unit 101. At the same time, the processed data is sent to the next unit, namely the model training unit 103, and the process data is sent to the storage unit 104. This process ensures important information and historical records in the data processing process. These process data are of great reference value for subsequent data analysis, model tuning, and troubleshooting.

[0057] The model training unit 103 is used to receive the training instructions of the control unit 101 and the processed data sent by the data processing unit 102, complete the model training for the target model according to the training instructions and the processed data, send the progress information and performance indicator information of the model training to the control unit 101, and send the trained model related information and the process data of the training instructions to the storage unit 104;

[0058] In this embodiment, the model training unit 103 completes the training of the model according to the training instructions sent by the control unit 101 and the processed data sent by the data processing unit 102, and sends the training information and the generated performance indicator information to the control unit 101, and at the same time sends the model and the data of the training process to the storage unit 104, so that the various units of the data processing system form a close collaborative relationship.

[0059] The storage unit 104 is used to receive process data sent by the data processing unit 102, and is also used to receive model-related information trained by the model training unit 103 and process data of training instructions, and is also used to store the received information.

[0060] In this embodiment, the storage unit 104 receives process data fed back by various units and stores the data, thereby ensuring the integrity of the process data in data processing and model training for subsequent data management.

[0061] The data management system based on model training provided by the embodiment of the present application includes: a control unit, a data processing unit, a model training unit and a storage unit; the control unit sends a data processing instruction to the data processing unit, sends a training instruction to the model training unit, and receives information fed back by each unit; the data processing unit processes the data by receiving the instruction of the control unit, and feeds back the progress of the processing process and the completed result to the control unit, and sends the processed data to the next unit, namely the model training unit, and sends the process data to the storage unit, the model training unit completes the training of the model according to the training instruction sent by the control unit and the processed data sent by the data processing unit, sends the information and the generated performance indicator information to the control unit 101, and sends the data of the model and the training process to the storage unit, so that the various units of the data processing system form a close collaborative relationship, and further, the storage unit is used to receive the process data fed back by each unit, and store these data, so as to ensure the integrity of the process data in the data processing and model training, so as to facilitate the subsequent data management. Through the interaction between multiple units, the effective management of data in model training is realized, and the efficiency of data management is improved.

[0062] Figure 2Schematic diagram of the structure of the data management system based on model training provided in this application Figure 2 ,like Figure 2 As shown, in this embodiment Figure 1 Based on the embodiment, a data management system based on model training is described in detail, wherein the data processing unit 102 includes: a data acquisition module 201, a data cleaning and annotation module 202, and a feature engineering module 203;

[0063] The data acquisition module 201 is used to receive the data acquisition instruction of the control unit 101, complete the data acquisition according to the data acquisition instruction, send the progress information and acquisition status of the acquired data to the control unit 101, send the acquired data to the data cleaning and annotation module 202, and send the acquired process data to the storage unit 104;

[0064] In this embodiment, the module includes multiple databases, and the types of databases include structured databases and unstructured databases. According to the data collection instruction, the corresponding database is selected for data collection, and the collected data is sent to the data cleaning and annotation module 202, and the progress information and collection status are sent to the control unit 101. The collection status includes collection not started, collection in progress, and collection completed; the collected process data is sent to the storage unit 104.

[0065] Optionally, the data acquisition module 201 includes: a data acquisition scheduler 301 and a data screening submodule 302; Figure 3 A schematic diagram of the structure of a data management system based on model training provided in an embodiment of the present application Figure 3 .like Figure 3 As shown, the data acquisition module 201 includes:

[0066] The data collection scheduler 301 is used to receive the data collection instruction of the control unit 101, determine the priority information of the data collection instruction, and also to obtain the resource status information of the model training and data management system, collect data from multiple databases according to the resource status information of the model training and data management system and the priority information of the data collection instruction, obtain the collected data, send the progress information and collection status of the collected data to the control unit 101, and send the collected data to the storage unit 104;

[0067] In this embodiment, the data acquisition scheduler 301 uses its multiple data interfaces to establish connections with multiple databases. For structured databases, connections are established through database drivers and corresponding connection strings; for unstructured file storage systems, file system read interfaces are used to traverse folders and files; the web crawler interface starts crawling data according to the set crawler rules and target URLs; the IoT sensor network receives real-time sensor data through the sensor communication protocol. During the data collection process, the data acquisition scheduler 301 determines the priority information of the data collection instructions, and also needs to obtain the resource status information of the model training and data management system. The resource status information includes CPU usage and memory remaining. According to the resource status information of the model training and data management system, as well as the priority information of the data collection instructions, the data in multiple databases are reasonably arranged for collection tasks to ensure the efficiency and stability of the collection process.

[0068] The data filtering submodule 302 is used to receive the filtering instructions sent by the control unit 101, and is also used to receive the collected data sent by the data acquisition scheduler 301, filter the collected data according to the filtering instructions, preset data type, preset data scale, and preset data quality standard, obtain the filtered data, and send the filtered data to the data cleaning and labeling module 202.

[0069] In this embodiment, the data screening module 302 can adaptively learn the data change rules in the database. If the structure of a database changes, the data screening module 302 can dynamically adjust the screening criteria to ensure that the collected data still meets the requirements.

[0070] The data cleaning and labeling module 202 is used to receive the data cleaning and labeling instructions of the control unit 101, and is also used to receive the collected data sent by the data acquisition module 201, perform data cleaning and data labeling on the collected data according to the data cleaning and data labeling instructions, obtain the data after data cleaning and data labeling, send the progress information and result information of the data cleaning and data labeling to the control unit 101, send the data after data cleaning and data labeling to the feature engineering module 203, and send the process data of data cleaning and data labeling to the storage unit 104;

[0071] In this embodiment, the collected data is cleaned and labeled according to the cleaning and labeling instructions to obtain the data after data cleaning and data labeling, and the progress information and result information are sent to the control unit 101. The result information includes the cleaning results and the labeling results. The data after data cleaning and data labeling is sent to the feature engineering module 203, and the process data of data cleaning and data labeling is sent to the storage unit 104.

[0072] Optionally, the data cleaning and labeling module 202 includes: a data cleaning submodule 303 and a data labeling submodule 304; Figure 3 A schematic diagram of the structure of a data management system based on model training provided in an embodiment of the present application Figure 3 ,like Figure 3 As shown, the data cleaning and labeling module 202 includes:

[0073] The data cleaning submodule 303 is used to receive the data cleaning instruction of the control unit 101, and is also used to receive the collected data sent by the data acquisition module 201, clean the collected data according to the data cleaning instruction and the preset cleaning rules, obtain the cleaned data, send the progress information and result information of the data cleaning to the control unit 101, send the cleaned data to the data annotation submodule 304, and send the process data of the data cleaning to the storage unit 104;

[0074] In this embodiment, the data cleaning submodule 303 uses a variety of statistical methods and predefined rules to remove duplications, errors, missing values ​​and outliers in the data. Taking the processing of a mixed data set containing temperature data collected by sensors and text descriptions of device status as an example, for temperature data, by statistically analyzing information such as the mean and standard deviation, outliers that are obviously deviated from the normal range are identified and removed; for the text description of device status, text processing rules are used to check grammatical errors and words that do not conform to the predefined vocabulary. For data points with missing values, interpolation methods can be used to process them according to the distribution and correlation of the data to ensure the purity of the data.

[0075] The data labeling submodule 304 is used to receive the data labeling instructions from the control unit 101, and is also used to receive the cleaned data sent by the data cleaning submodule 303. According to the data labeling instructions, an active learning algorithm is used to label the cleaned data to obtain labeled data, and the progress information and result information of the labeling are sent to the control unit 101, the labeled data is sent to the feature engineering module 203, and the labeled process data is sent to the storage unit 104.

[0076] In this embodiment, the data annotation submodule 304 uses an automatic annotation algorithm to add categories and attribute labels to the data, and the manual reviewer further confirms the accuracy of the annotation. During the annotation process, an active learning algorithm is used to give priority to annotating data that has the greatest gain for model training. For example, the algorithm will identify the category data that is most difficult to distinguish for the current model based on feedback during the model training process, and give priority to annotating these data. At the same time, multi-user collaborative annotation is supported. When a manual annotation method is used, different annotators can annotate the same batch of data, and the data annotation submodule 304 will fuse and resolve conflicts between the annotation results of different users. In one possible implementation, most annotators annotate a medical image as pathological tissue, while a few annotators annotate it as normal tissue. The data annotation submodule 304 will prompt the manual reviewer to focus on this difference and make the final judgment and adjustment.

[0077] The feature engineering module 203 is used to receive the data feature parameter setting instruction of the control unit 101, and is also used to receive the data after data cleaning and data labeling sent by the data cleaning and labeling module 202, set the feature parameters of the data after data cleaning and data labeling according to the data feature parameter setting instruction, obtain the data after the feature parameter setting, send the progress information and result information of the feature parameter setting of the data after data cleaning and data labeling to the control unit 101, send the data after the feature parameter setting to the model training unit 103, and send the process data of the feature parameter setting of the data after data cleaning and data labeling to the storage unit 104.

[0078] In this embodiment, according to the data feature parameter setting instruction, feature parameters are set for the received data after data cleaning and data annotation, the progress information and result information of the feature parameter setting are sent to the control unit 101, the data after feature parameter setting is sent to the model training unit 103, and the process data is sent to the storage unit 104.

[0079] Optionally, the feature engineering module 203 includes: a feature selection submodule 305 and a feature transformation submodule 306; Figure 3 A schematic diagram of the structure of a data management system based on model training provided in an embodiment of the present application Figure 3 .like Figure 3 As shown, the feature engineering module 203 includes:

[0080] The feature selection submodule 305 is used to receive the feature selection instruction of the control unit 101, and is also used to receive the data after data cleaning and data labeling sent by the data cleaning and labeling module 202, screen the data after data cleaning and data labeling according to the feature selection instruction, information gain and chi-square test algorithm, determine multiple data features, select multiple data features according to a method combining deep learning and traditional feature selection, determine multiple selected data features, send progress information and result information of feature selection to the control unit 101, send the selected multiple data features to the feature transformation submodule 306, and send the process data of feature selection to the storage unit 104;

[0081] In this embodiment, the feature selection submodule 305 selects features that have a significant impact on model training from the data after data cleaning and data annotation based on information gain and chi-square test algorithms. The feature selection submodule 305 combines the feature representation ability of deep learning and traditional feature selection methods to more accurately mine potential effective features. During the entire model training process, the feature selection submodule 305 can dynamically adjust the strategy according to model feedback.

[0082] The feature transformation submodule 306 is used to receive the feature transformation instruction of the control unit 101, and is also used to receive the selected multiple data features sent by the feature selection submodule 305, and perform data standardization operations on the selected multiple data features according to the feature transformation instruction to obtain standardized data, determine the data type of the standardized data, perform feature transformation on data of different data types to obtain transformed data, send the progress information and result information of the feature transformation to the control unit 101, send the transformed data to the model training unit 103, and send the process data of the feature transformation to the storage unit 104; wherein the data type includes numerical data and non-numerical data, the numerical data is normalized to obtain normalized data, and the non-numerical data is discretized to obtain discretized data.

[0083] In this embodiment, the feature transformation submodule 306 can perform operations such as data standardization, normalization, and discretization. Determine the data type of the standardized data, perform feature transformation on data of different data types, and for numerical data, normalize sensor data of different ranges so that they are within the range of 0-1 to facilitate model training. For categorical data, perform a one-hot encoding discretization operation to convert categorical variables into numerical forms that can be used for model training, and perform feature encoding on text and image non-numerical data to obtain discretized data.

[0084] The model training unit 103 includes: a model selection module 204, a model parameter initialization module 205, a model training module 206, a model verification module 207, a model evaluation module 208 and a visualization display module 209;

[0085] The model selection module 204 is used to receive the model selection instruction of the control unit 101 and also to receive the processed data sent by the data processing unit 102. The model selection module 204 includes multiple models, selects multiple models according to the model selection instruction and the data features of the processed data, determines the target model, sends the result information of the model selection to the control unit 101, and sends the target model and the processed data to the model parameter initialization module 205;

[0086] In this embodiment, the model selection module 204 includes multiple models, including but not limited to deep neural networks, convolutional neural networks, recurrent neural networks, support vector machines, and decision trees. The appropriate model architecture is automatically selected according to the model selection instruction, the data characteristics of the processed data, and the application scenario. For example, for image classification tasks, based on the feature representation of the image data after feature engineering, the convolutional neural network architecture is selected because it has advantages in processing the local correlation and hierarchical structure of the image data; for time series data, a recurrent neural network or a long short-term memory network is selected to capture the sequence information in the data.

[0087] The model parameter initialization module 205 is used to receive the parameter initialization instruction of the control unit 101, and is also used to receive the target model and processed data sent by the model selection module 204, initialize the parameters of the target model according to the parameter initialization instruction and the data characteristics of the processed data, obtain the model after parameter initialization, send the model after parameter initialization and the processed data to the model training module 206, and send the process data of parameter initialization to the storage unit 104;

[0088] In this embodiment, the model parameter initialization module 205 initializes the parameters of the target model according to the distribution of the data features of the processed data and the model type of the selected model. For the neural network model, the weight parameters are initialized using a suitable initialization method based on the mean and standard deviation of the input data to avoid the gradient disappearance or gradient explosion problems in the early stages of training; for the support vector machine model, the relevant parameters are initialized based on the feature range and category distribution of the data to ensure the stability of the model training.

[0089] The model training module 206 is used to receive the model training instructions of the control unit 101, and is also used to receive the parameter-initialized model and processed data sent by the model parameter initialization module 205. According to the model training instructions, the processed data is input into the parameter-initialized model, and the parameter-initialized model is trained according to multiple optimization algorithms. The speed of the training process is improved through distributed computing technology, and the trained model is obtained. The progress information and result information of the training are sent to the control unit 101, and the process data in training is sent to the storage unit 104.

[0090] In this embodiment, the model training module 206 supports multiple optimization algorithms such as gradient descent and stochastic gradient descent for model training. During the training process, the training execution unit uses distributed computing technology to increase the training speed.

[0091] The model verification module 207 is used to receive the model verification instruction of the control unit 101, and is also used to receive the trained model sent by the model training module 206, and verify the trained model through a cross-validation method according to the model verification instruction to obtain multiple verification results and the verified model, and send the verified model to the model evaluation module 208, send the process data of the model verification to the storage unit 104, and send the progress information of the model verification and multiple verification results to the control unit 101; wherein the verification results include accuracy, recall rate, mean square error information and robustness index;

[0092] In this embodiment, during the training process, the model verification module 207 divides the training set and the validation set according to a preset ratio, and verifies the model performance by cross-validation and other methods. The model verification module calculates the accuracy, recall, mean square error information and robustness index of the trained model through the validation set. When the accuracy of the model on the validation set does not meet expectations, the model training unit 102 will automatically adjust the hyperparameters and model architecture of the training algorithm according to the feedback information of the model verification module 207, so that the model training module 206 retrains the model.

[0093] The model evaluation module 208 is used to receive the model evaluation instruction of the control unit 101, and is also used to receive the verified model sent by the model verification module 207, determine the model type of the verified model according to the model evaluation instruction, determine the corresponding evaluation index according to the model type of the verified model, evaluate the verified model through multiple evaluation indexes, obtain multiple evaluation results, send the multiple evaluation results to the visualization display module 209, send the progress information and result information of the evaluation to the control unit 101, and send the process data of the evaluation to the storage unit 104; wherein, the model type of the verified model includes a classification model and a regression model, the evaluation indexes corresponding to the classification model include accuracy, recall rate and F1 value, and the evaluation indexes corresponding to the regression model include mean square error information and mean absolute error information;

[0094] In this embodiment, after the model training is completed, the model evaluation module 208 determines the model type of the verified model, and determines the corresponding evaluation indicators according to the model type of the verified model. For the classification model, the accuracy (the proportion of the number of samples predicted correctly to the total number of samples), the recall rate (the proportion of the samples predicted to be positive examples that are actually positive examples), and the F1 value (an indicator that combines the accuracy and recall rates) are calculated; for the regression model, the mean square error (the average value of the square of the difference between the predicted value and the true value) and the mean absolute error (the average value of the absolute value of the difference between the predicted value and the true value) are calculated.

[0095] The visualization display module 209 is used to receive the display instructions of the control unit 101, and is also used to receive multiple evaluation results sent by the model evaluation module 208, determine the display form of each evaluation indicator according to the display instructions, and send the display results corresponding to each evaluation result to the control unit 101 and the storage unit 104 respectively.

[0096] In this embodiment, the visualization display module 209 displays the evaluation results in an intuitive chart form, determines the display form of each evaluation indicator, and for the accuracy, recall rate and F1 value indicators, a bar chart can be used to compare the performance of different models or the same model at different training stages; for the change of the loss value of the model during the training process, a line chart is used to display it, so that the user can intuitively understand the convergence of the model training process. These visualization results are displayed to the user through the user interface of the user interaction and control unit, so that the user can quickly and clearly grasp the model performance.

[0097] The storage unit 104 includes: a model storage and management module 210, and a data storage and indexing module 211;

[0098] The model storage and management module 210 is used to receive the model storage management instruction of the control unit 101, and is also used to receive the trained model sent by the model training unit 103, store and manage the trained model according to the model storage management instruction, and send the storage status of the trained model to the control unit 101;

[0099] In this embodiment, the model storage and management module 210 mainly stores the trained model and the relevant data of the model. When the model training unit 102 trains a new model, the model storage and management module 210 stores the structural information of the model, the trained parameter values, the evaluation indicators on the validation set and the test set, and the relevant information of the training time.

[0100] Optionally, the model storage and management module 210 includes: a model storage submodule 307, a model index submodule 308 and a model update submodule 309; Figure 3 A schematic diagram of the structure of a data management system based on model training provided in an embodiment of the present application Figure 3 .like Figure 3 As shown, the model storage and management module 210 includes:

[0101] The model storage submodule 307 is used to receive the model storage instruction of the control unit 101, and is also used to receive the trained model sent by the model training unit 103, calibrate the trained model according to the model storage instruction, and send the calibrated model to the control unit 101, the model index submodule 308 and the model update submodule 309 respectively;

[0102] In this embodiment, the model storage submodule 307 stores the trained models in a versioned manner and calibrates each model, and the calibration content includes an identifier, a version number, and model structure information. The model index submodule 308 is used to receive the model index instruction of the control unit 101, and is also used to receive the calibrated model sent by the model storage submodule 307, use language indexing technology to perform language analysis on the calibrated model, determine the keyword information of each calibrated model, send the keyword information to the control unit 101, and provide the control unit 101 with an indexing function according to the model index instruction; wherein the keyword information includes the function of the model, the application scenario of the model, and the training data characteristics of the model;

[0103] In this embodiment, the model index submodule 308 establishes an index based on the model's functions, application scenarios, and training data features to facilitate rapid retrieval. Using semantic indexing technology, semantic analysis is performed on the functional description of the model to determine the keyword information of each calibrated model. If the model is a CNN model for tumor detection in medical images, the model index submodule 308 will extract keywords such as "medical image", "tumor detection", "CNN", and match the model with related query requests through semantic understanding and association. When the user end of the control unit 101 searches for a model for medical image analysis, the system can quickly and accurately retrieve the model. The model index submodule 308 can automatically update the model index information based on the feedback of the model in actual application.

[0104] The model update submodule 309 is used to receive the model update instruction from the control unit 101, and is also used to receive the trained model sent by the model training unit 103, retrain the trained model according to the model update instruction to obtain a new trained model, send the new trained model to the model storage submodule 307 and the model index submodule 308, and send the progress information and result information of the model update to the control unit 101.

[0105] In this embodiment, when new data or new requirements appear, the model update submodule 309 can perform incremental training or retraining based on the existing model and update the stored model. When new training data is obtained, the model update submodule 309 can determine the similarity of the new data to the model. When the new data is similar to the existing data distribution, the incremental training method can be used to fine-tune the model parameters based on the original model using the new data; when the characteristics of the new data are significantly different from the original data, the retraining process is triggered, and the model information and index information in the model storage database are updated after the model is retrained.

[0106] The data storage and indexing module 211 is used to receive the data storage indexing instructions of the control unit 101, and is also used to receive the process data sent by the data processing unit 102 and the process data sent by the model training unit 103, and manage multiple process data according to the data storage indexing instructions.

[0107] In this embodiment, the data storage and indexing module 211 is used to receive process information sent by various modules and sub-modules, and manage multiple process data.

[0108] Optionally, the data storage and indexing module 211 includes: a data storage submodule 310 and a data indexing submodule 311; Figure 3 A schematic diagram of the structure of a data management system based on model training provided in an embodiment of the present application Figure 3 .like Figure 3 As shown, the data storage and indexing module 211 includes:

[0109] The data storage submodule 310 is used to receive the data storage instruction of the control unit 101, and is also used to receive the process data of the data processing instruction sent by the data processing unit 102, and the process data of the training instruction sent by the model training unit 103, obtain the process type of each process data, classify and store each process data according to the data storage instruction, obtain the classified storage result, and send the classified storage result to the data index submodule 311 and the control unit 101 respectively;

[0110] In this embodiment, the data storage submodule 310 stores each type of process data by category, and stores the original data, cleaned data, feature data, etc. in different storage nodes. For example, the large amount of original collected data and low access frequency data are stored in a large-capacity, low-cost storage node; the cleaned data and feature data that are often used for model training and analysis are stored in a high-performance storage node to improve the data reading speed.

[0111] The data indexing submodule 311 is used to receive the data indexing instruction of the control unit 101, and is also used to receive the classified storage results sent by the data storage submodule 310, determine the attribute feature information of each data in the classified storage results, determine the hash value corresponding to the attribute feature information of each data according to the distributed hash table technology, establish a hash table through multiple hash values, input the data indexing instruction of the control unit 101 into the hash table, and complete the data indexing.

[0112] In this embodiment, the data index submodule 311 determines the attribute feature information of each data in the classification storage result, establishes an index for the data feature information, realizes fast data query and extracts feature information including but not limited to attributes such as labels, feature values ​​and timestamps, and uses distributed hash table technology to map the attribute feature information of the data to hash values, and quickly locates the data storage location through a hash table established by multiple hash values. When the user needs to query specific image data or image data within a specific time period, the data index submodule 311 quickly finds the storage location of the corresponding data in the distributed storage system through the index, thereby improving the query efficiency. The data index submodule 311 supports dynamic update and maintenance of the index to adapt to the dynamic changes of data. When new data is added to the storage unit 104 or the data is updated, the data index submodule 311 will update the index information accordingly.

[0113] The data management system based on model training provided by the embodiment of the present application includes a control unit, a data processing unit, a model training unit and a storage unit, wherein the data processing unit includes: a data acquisition module, a data cleaning and annotation module, and a feature engineering module; the model training unit includes: a model selection module, a model parameter initialization module, a model training module, a model verification module, a model evaluation module and a visualization display module; the storage unit includes: a model storage and management module, and a data storage and indexing module; the data acquisition module includes: a data acquisition scheduler and a data screening submodule; the data cleaning and annotation module includes: a data cleaning submodule and a data annotation submodule; the feature engineering module includes: a feature selection submodule and a feature transformation submodule; the model storage and management module includes: a model storage submodule, a model indexing submodule and a model updating submodule; the data storage and indexing unit module includes: a data storage submodule and a data indexing submodule; by establishing a complete data management system, the process data in the model training is stored, and the data processing flow of the model training is improved. At the same time, the control unit provides a visual interface for the user terminal, so that the user terminal can index the process data and the training model in real time, and dynamically display the whole process of the model training, thereby improving the efficiency and timeliness of data management during the model training process.

[0114] It should be noted that, although the various steps in the structure diagram are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear description in this article, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in the structure diagram may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0115] It should be understood that the above system embodiments are only illustrative, and the system of the present application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units, modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed.

[0116] In addition, unless otherwise specified, each functional unit / module in each embodiment of the present application may be integrated into one unit / module, each unit / module may exist physically separately, or two or more units / modules may be integrated together. The above-mentioned integrated unit / module may be implemented in the form of hardware or in the form of a software program module.

[0117] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0118] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A data management system based on model training, characterized in that: The data management system comprises: a control unit, a data processing unit, a model training unit and a storage unit; The control unit is used to send instructions to each unit and receive feedback information from each unit; The data processing unit is used to receive the data processing instruction of the control unit, complete the data processing according to the data processing instruction, send the progress information and completion result of the data processing to the control unit, send the processed data to the model training unit, and send the process data of the data processing instruction to the storage unit; The model training unit is used to receive the training instruction of the control unit and the processed data sent by the data processing unit, complete the model training for the target model according to the training instruction and the processed data, send the progress information and performance indicator information of the model training to the control unit, and send the trained model related information and the process data of the training instruction to the storage unit; The storage unit is used to receive the process data sent by the data processing unit, and is also used to receive the model-related information trained by the model training unit and the process data of the training instructions, and is also used to store the received information.

2. The system according to claim 1, characterized in that The data processing unit includes: a data acquisition module, a data cleaning and annotation module, and a feature engineering module; The data acquisition module is used to receive the data acquisition instruction of the control unit, complete data acquisition according to the data acquisition instruction, send the progress information and acquisition status of the acquired data to the control unit, send the acquired data to the data cleaning and annotation module, and send the acquired process data to the storage unit; The data cleaning and labeling module is used to receive the data cleaning and labeling instructions of the control unit, and is also used to receive the collected data sent by the data acquisition module, perform data cleaning and data labeling on the collected data according to the data cleaning and labeling instructions, obtain the data after data cleaning and data labeling, send the progress information and result information of the data cleaning and data labeling to the control unit, send the data after data cleaning and data labeling to the feature engineering module, and send the process data of data cleaning and data labeling to the storage unit; The feature engineering module is used to receive the data feature parameter setting instruction of the control unit, and is also used to receive the data after data cleaning and data annotation sent by the data cleaning and data annotation module, set feature parameters for the data after data cleaning and data annotation according to the data feature parameter setting instruction, obtain the data after feature parameter setting, send the progress information and result information of the feature parameter setting for the data after data cleaning and data annotation to the control unit, send the data after feature parameter setting to the model training unit, and send the process data of the feature parameter setting for the data after data cleaning and data annotation to the storage unit.

3. The system according to claim 2, characterized in that The data acquisition module includes: a data acquisition scheduler and a data screening submodule; The data acquisition scheduler is used to receive the data acquisition instruction of the control unit, determine the priority information of the data acquisition instruction, and obtain the resource status information of the model training and data management system, collect data from multiple databases according to the resource status information of the model training and data management system and the priority information of the data acquisition instruction, obtain the collected data, send the progress information of the collected data and the collection status to the control unit, and send the collected data to the storage unit; The data screening submodule is used to receive the screening instructions sent by the control unit, and is also used to receive the collected data sent by the data acquisition scheduler, screen the collected data according to the screening instructions, preset data type, preset data scale, and preset data quality standard, obtain the screened data, and send the screened data to the data cleaning and annotation module.

4. The system according to claim 2, characterized in that The data cleaning and labeling module includes: a data cleaning submodule and a data labeling submodule; The data cleaning submodule is used to receive the data cleaning instruction of the control unit, and is also used to receive the collected data sent by the data acquisition module, clean the collected data according to the data cleaning instruction and the preset cleaning rules to obtain cleaned data, send the progress information and result information of the data cleaning to the control unit, send the cleaned data to the data annotation submodule, and send the process data of the data cleaning to the storage unit; The data labeling submodule is used to receive the data labeling instructions of the control unit, and is also used to receive the cleaned data sent by the data cleaning submodule. According to the data labeling instructions, an active learning algorithm is used to label the cleaned data to obtain labeled data, and the progress information and result information of the labeling are sent to the control unit, the labeled data is sent to the feature engineering module, and the labeled process data is sent to the storage unit.

5. The system according to claim 2, characterized in that The feature engineering module includes: a feature selection submodule and a feature transformation submodule; The feature selection submodule is used to receive the feature selection instruction of the control unit, and is also used to receive the data after data cleaning and data annotation sent by the data cleaning and annotation module, screen the data after data cleaning and data annotation according to the feature selection instruction, information gain and chi-square test algorithm, determine multiple data features, select the multiple data features according to a method combining deep learning and traditional feature selection, determine the multiple data features after selection, send the progress information and result information of feature selection to the control unit, send the multiple data features after selection to the feature transformation submodule, and send the process data of feature selection to the storage unit; The feature transformation submodule is used to receive the feature transformation instruction of the control unit, and is also used to receive the selected multiple data features sent by the feature selection submodule, and according to the feature transformation instruction, perform data standardization operation on the selected multiple data features to obtain standardized data, determine the data type of the standardized data, perform feature transformation on data of different data types to obtain transformed data, send progress information and result information of the feature transformation to the control unit, send the transformed data to the model training unit, and send process data of the feature transformation to the storage unit; wherein the data types include numerical data and non-numerical data, perform normalization processing on numerical data to obtain normalized data, and perform discretization processing on non-numerical data to obtain discretized data.

6. The system according to claim 1, characterized in that The model training unit includes: a model selection module, a model parameter initialization module and a model training module; The model selection module is used to receive the model selection instruction of the control unit and also to receive the processed data sent by the data processing unit. The model selection module includes multiple models, selects the multiple models according to the model selection instruction and the data features of the processed data, determines the target model, sends the result information of the model selection to the control unit, and sends the target model and the processed data to the model parameter initialization module; The model parameter initialization module is used to receive the parameter initialization instruction of the control unit, and is also used to receive the target model and the processed data sent by the model selection module, initialize the parameters of the target model according to the parameter initialization instruction and the data characteristics of the processed data, obtain the model after parameter initialization, send the model after parameter initialization and the processed data to the model training module, and send the process data of parameter initialization to the storage unit; The model training module is used to receive the model training instructions of the control unit, and is also used to receive the parameter-initialized model and the processed data sent by the model parameter initialization module. According to the model training instructions, the processed data is input into the parameter-initialized model, and the parameter-initialized model is trained according to multiple optimization algorithms. The speed of the training process is improved through distributed computing technology, and the trained model is obtained. The progress information and result information of the training are sent to the control unit, and the process data in training is sent to the storage unit.

7. The system according to claim 6, characterized in that The model training unit also includes: a model verification module, a model evaluation module and a visualization display module; The model verification module is used to receive the model verification instruction of the control unit, and is also used to receive the trained model sent by the model training module, and according to the model verification instruction, the trained model is verified by a cross-validation method to obtain multiple verification results and a verified model, and the verified model is sent to the model evaluation module, the process data of the model verification is sent to the storage unit, and the progress information of the model verification and the multiple verification results are sent to the control unit; wherein the verification result includes accuracy, recall rate, mean square error information and robustness index; The model evaluation module is used to receive the model evaluation instruction of the control unit, and is also used to receive the verified model sent by the model verification module, determine the model type of the verified model according to the model evaluation instruction, determine the corresponding evaluation index according to the model type of the verified model, evaluate the verified model through multiple evaluation indexes, obtain multiple evaluation results, send the multiple evaluation results to the visualization display module, send the progress information and result information of the evaluation to the control unit, and send the process data of the evaluation to the storage unit; wherein the model types of the verified model include classification models and regression models, the evaluation indexes corresponding to the classification model include accuracy, recall rate and F1 value, and the evaluation indexes corresponding to the regression model include mean square error information and mean absolute error information; The visualization display module is used to receive the display instructions of the control unit, and is also used to receive the multiple evaluation results sent by the model evaluation module, determine the display form of each evaluation indicator according to the display instructions, and send the display results corresponding to each evaluation result to the control unit and the storage unit respectively.

8. The system according to claim 1, characterized in that The storage unit includes: a model storage and management module, and a data storage and indexing module; The model storage and management module is used to receive the model storage management instruction of the control unit, and is also used to receive the trained model sent by the model training unit, store and manage the trained model according to the model storage management instruction, and send the storage status of the trained model to the control unit; The data storage and indexing module is used to receive the data storage indexing instruction of the control unit, and is also used to receive the process data sent by the data processing unit and the process data sent by the model training unit, and manage multiple process data according to the data storage indexing instruction.

9. The system according to claim 8, characterized in that The model storage and management module includes: a model storage submodule, a model index submodule and a model update submodule; The model storage submodule is used to receive the model storage instruction of the control unit, and is also used to receive the trained model sent by the model training unit, calibrate the trained model according to the model storage instruction, and send the calibrated model to the control unit, the model index submodule and the model update submodule respectively; wherein the calibration content includes an identifier, a version number, model structure information, model training time and model parameter values; The model indexing submodule is used to receive the model indexing instruction of the control unit, and is also used to receive the calibrated model sent by the model storage submodule, perform language analysis on the calibrated model using language indexing technology, determine keyword information of each calibrated model, send the keyword information to the control unit, and provide an indexing function for the control unit according to the model indexing instruction; wherein the keyword information includes the function of the model, the application scenario of the model and the training data characteristics of the model; The model update submodule is used to receive the model update instruction of the control unit, and is also used to receive the trained model sent by the model training unit, retrain the trained model according to the model update instruction to obtain a new trained model, send the new trained model to the model storage submodule and the model index submodule, and send the progress information and result information of the model update to the control unit.

10. The system according to claim 8, characterized in that The data storage and indexing module includes: a data storage submodule and a data indexing submodule; The data storage submodule is used to receive the data storage instruction of the control unit, and is also used to receive the process data sent by the data processing unit for performing the data processing instruction, and the process data sent by the model training unit for performing the training instruction, obtain the process type of each process data, classify and store each process data according to the data storage instruction, obtain the classified storage result, and send the classified storage result to the data index submodule and the control unit respectively; The data indexing submodule is used to receive the data indexing instruction of the control unit, and is also used to receive the classified storage result sent by the data storage submodule, determine the attribute feature information of each data in the classified storage result, determine the hash value corresponding to the attribute feature information of each data according to the distributed hash table technology, establish a hash table through multiple hash values, input the data indexing instruction of the control unit into the hash table, and complete the data indexing.

Images