Intelligent equipment visual data management and AI model development platform and method

Abstract

The invention provides an intelligent equipment visual data management and AI model development platform and method, and belongs to the technical field of image processing, and the platform specifically comprises a data annotation module, a model training module and a version management module, the model training module is responsible for carrying out training processing on the visual model according to needs by utilizing the annotation data, and the version management module is responsible for carrying out management on different historical versions of the visual model according to the training data and the updating training condition of the training data in the visual model, so that the reliability of management processing of the historical versions is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to an intelligent equipment vision data governance and AI model development platform and method. Background Art

[0002] In the existing intelligent algorithm training and integrated verification systems for scientific research systems and intelligent security, the data fusion method is too single. Currently, generally, algorithm processing is performed on a single sensor, and partial algorithms perform feature-level fusion. Finally, the results of the algorithm processing are fused at the decision level. This processing method is relatively simple, lacks effective utilization of multi-dimensional data, does not fully explore data association information, and thus cannot guarantee an improvement in algorithm accuracy.

[0003] Therefore, in order to solve the above technical problems, in the existing technical solutions, incremental training processing is performed through the acquisition and annotation of incremental data. However, the following problems exist in the existing technical solutions: In the process of visual data governance, in the existing technical solutions, training data is often generated through automatic annotation, which inevitably may lead to a deterioration in the accuracy of identifying certain targets after incremental training. Therefore, how to manage historical versions targeted so as to facilitate and quickly locate the positioning and processing of problems generated in visual training has become an urgent technical problem to be solved.

[0004] To solve the above technical problems, the present application provides an intelligent equipment vision data governance and AI model development platform and method. Summary of the Invention

[0005] To achieve the object of the present invention, the present invention adopts the following technical solutions: Specifically, the present application provides an intelligent equipment vision data governance and AI model development platform, which specifically includes: A data annotation module, a model training module, and a version management module; Wherein the data annotation module is responsible for performing annotation processing on the data to obtain annotated data; The model training module is responsible for using the annotated data to perform training processing on the vision model as required; The version management module is responsible for managing different historical versions of the vision model based on the training data and the updated training situation of the training data in the vision model.

[0006] A further technical solution lies in that the annotation processing of the data includes manual annotation and automatic annotation.

[0007] A further technical solution lies in that it further includes a data augmentation module, which is responsible for performing amplification processing on the annotated data.

[0008] A further technical solution lies in that the augmentation processing includes augmentation processing methods such as rotation, cutting, flipping, generative adversarial network, moving target real visual background synthesis, and real captured image + simulated visual interference synthesis.

[0009] A further technical solution lies in that when performing the training process of the visual model, the change situation of the model parameters of the visual model supports visual display.

[0010] A further technical solution lies in managing different historical versions of the visual model, specifically including: When the training data of the historical version has not been incrementally trained in the visual model within the most recent preset time period, then based on the verification result of the incrementally trained visual model and the optimized target version in the target database, it is determined whether forgetting management is required. Among them, when the deviation between the accuracy of the verification result of the optimized target version and the visual model is greater than the preset deviation value, then the optimized target version forgetting process is performed.

[0011] In a second aspect, the present application provides a method for managing historical versions, which is applied to the above-mentioned intelligent equipment visual data governance and AI model development platform, and specifically includes: S1 Based on the training data of the visual model, determine the composition of the automatically labeled data under different types of training data, and determine the risky training data in the training data based on the composition. S2 Determine the composition of the training risk data during incremental training for different historical versions, and combine the training risk data with the similarity of other later historical versions to determine the optimized target version in the historical versions. S3 Based on the training risk data, divide the historical versions into different version intervals, and based on the composition data of the historical versions within different version intervals and the change situation of the model parameters of the optimized target version and later historical versions, determine the type of optimization requirements within different version intervals. S4 When the types of optimization requirements of the optimized target version in different version intervals do not belong to the target requirement types, based on the updated training situation of the training risk data corresponding to the optimized target version in the visual model and the types of optimization requirements in different version intervals, determine the forgetting management method of the optimized target version.

[0012] The beneficial effects of the present invention are as follows: Based on the composition data of historical versions within different version intervals and the changes in model parameters between the optimized target version and subsequent historical versions, determine the types of optimization requirements within different version intervals. That is, considering the differences in problem localization requirements due to the differences in the number of historical versions within different version intervals, which lead to problems with the training risk data corresponding to the version intervals, and also considering the differences in the training risk data corresponding to the version intervals caused by the changes in the model parameters of subsequent historical versions, the determination of the types of optimization requirements is achieved from the perspective of localization requirements.

[0013] Based on the updated training situation of the training risk data corresponding to the optimized target version in the visual model and the types of optimization requirements within different version intervals, determine the forgetting management method for the optimized version. This realizes the determination of the forgetting management method for the training update situation of the training risk data and the types of optimization requirements within different version intervals. That is, it ensures the efficiency of the forgetting process for the optimized target version with a large forgetting processing requirement, and at the same time reduces the storage space requirement.

[0014] A further technical solution is that the composition of the automatic annotation under the training data includes the data volume and the data volume ratio of the automatically annotated training data during different incremental trainings.

[0015] Specifically, the type of the training data is determined according to the recognition target corresponding to the training data.

[0016] A further technical solution is that the method for determining the risk training data in the training data is as follows: Based on the composition of the training data of this type during the incremental training of the visual model in different historical versions, determine the data volume of the automatically annotated training data of this type under different incremental training times; According to the ratio of the data volume of the automatically annotated training data of this type under different incremental training times, determine the data volume ratio under different incremental training times; Based on the data volume ratio under different incremental training times, determine whether the training data is risk training data.

[0017] A further technical solution is that the method for determining the forgetting management method for the optimized target version is as follows: Use the training risk data corresponding to the optimized target version as the matching risk data. According to the updated training situation of the matching risk data in the visual model, determine the historical versions corresponding to different types of matching risk data during incremental training in the most recent preset period, and use them as the updated versions; Determine the version range belonging to the second - type requirement type according to the optimization requirement types within different version ranges, and use it as the second - type requirement range; Determine the forgetting management method of the optimization target version according to the updated versions of different types of matching risk data and the composition data of the second - type requirement range.

[0018] Other features and advantages will be described in the subsequent specification. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the specification and the drawings.

[0019] To make the above - mentioned objectives, features, and advantages of the present invention more obvious and understandable, the following specifically gives preferred embodiments and, in conjunction with the accompanying drawings, makes a detailed description as follows. Brief Description of the Drawings

[0020] By referring to the drawings and describing its exemplary embodiments in detail, the above - mentioned and other features and advantages of the present invention will become more obvious; Figure 1 is a framework diagram of an intelligent equipment vision data governance and AI model development platform; Figure 2 is a flowchart of a management method for historical versions; Figure 3 is a flowchart of a method for determining risk training data in training data; Figure 4 is a flowchart of a method for determining an optimization target version in a historical version; Figure 5 is a flowchart of a method for determining optimization requirement types within a version range. Detailed Description of the Embodiments

[0021] In order to enable those skilled in the art of this technology to better understand the technical solutions in this specification, the following will clearly and completely describe the technical solutions in the embodiments of this specification in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this specification.

[0022] Embodiment 1 As Figure 1 shown, this application provides an intelligent equipment vision data governance and AI model development platform, which specifically includes: A data annotation module, a model training module, and a version management module; Among them, the data annotation module is responsible for performing annotation processing on data to obtain annotated data; The model training module is responsible for using the labeled data to perform training processing on the visual model as needed; The version management module is responsible for managing different historical versions of the visual model based on the training data and the updated training status of the training data in the visual model.

[0023] Furthermore, the annotation processing of the data includes manual annotation and automatic annotation.

[0024] Specifically, it also includes a data augmentation module responsible for performing augmentation processing on the labeled data.

[0025] It should be noted that the augmentation processing includes augmentation processing methods such as rotation, cropping, flipping, generative adversarial network, real moving target visual background synthesis, and real captured image + simulated visual interference synthesis.

[0026] It can be understood that when performing training processing on the visual model, the change situation of the model parameters of the visual model supports visual display.

[0027] Specifically, the management of different historical versions of the visual model specifically includes: When the training data of the historical version has not been incrementally trained in the visual model within the most recent preset time period, then through the verification results of the incrementally trained visual model and the optimization target version in the target database, it is determined whether forgetting management is required. Among them, when the deviation between the accuracy of the verification results of the optimization target version and the visual model is greater than the preset deviation value, then optimization target version forgetting processing is performed.

[0028] In addition, it should be noted that when the training data of the historical version has been incrementally trained in the visual model within the most recent preset time period, when there is a visual model of the optimization target version with similar model parameters after incremental training, then the verification results of the visual model of the optimization target version with similar model parameters and the optimization target version in the target database are used to determine whether forgetting management is required. Among them, whether the model parameters are similar is determined according to whether the deviation of the model parameters is within the preset range. Among them, when the deviation between the accuracy of the verification results of the optimization target version and the visual model is greater than the second preset deviation value, then optimization target version forgetting processing is performed, where the preset deviation value is greater than the second preset deviation value.

[0029] Embodiment 2 According to the proportion of the data volume of the training data of the type of automatic annotation under different incremental training times, determine the data volume proportion under different incremental training times. When the average value of the data volume proportion under different incremental training times is greater than 0.1 or more, then determine that the training data is risk training data.

[0030] The optimized target version is that the data volume ratio of the training risk data during incremental training is less than 0.2, and there is a historical version in other later historical versions whose deviation from the data volume ratio of its training risk data meets the requirements. Specifically, the later version whose deviation from the data volume of the concerned risk data in different historical versions is within the preset deviation range is used as a similar version. When the number of similar versions is greater than the threshold, the historical version is determined as the optimized target version.

[0031] If the number of historical versions within the version interval is greater than the preset version number threshold and there is only one optimized target version, it is determined that the optimization requirement type of the optimized target version within the version interval is a type of requirement type. If the number of historical versions is not greater than the preset version number threshold, it is determined that the optimization requirement type of the optimized target version within the version interval is the target requirement type.

[0032] If the number of historical versions is greater than the preset version number threshold and there is more than one optimized target version, it is determined that it belongs to the type-two requirement type.

[0033] Use the historical versions corresponding to different types of matching risk data during incremental training within the most recent preset time period as updated versions, and use the version interval belonging to the type-two requirement type as the type-two requirement interval. When the number of type-two requirement intervals is greater than the preset requirement interval number or there is a matching risk data with the number of updated versions greater than the preset updated version number threshold, it is determined that the forgetting management method of the optimized target version is that no forgetting management is required.

[0034] Second aspect, as Figure 2 shown, the present application provides a method for managing historical versions, which is applied to the above-mentioned intelligent equipment vision data governance and AI model development platform, and specifically includes: S1 Based on the training data of the vision model, determine the composition of the automatically annotated data under different types of training data, and determine the risk training data in the training data based on the composition. Furthermore, the composition of the automatically annotated data under the training data includes the data volume and data volume ratio of the automatically annotated training data during different incremental trainings.

[0035] Specifically, the type of the training data is determined according to the recognition target corresponding to the training data.

[0036] Specifically, as Figure 3 shown, the method for determining the risk training data in the training data is: Determine the data volume of the automatically labeled training data of the said type at different incremental training times based on the composition of the training data of the said type during the incremental training of different historical versions of the visual model; Determine the data volume ratio at different incremental training times based on the ratio of the data volume of the automatically labeled training data of the said type at different incremental training times; Based on the data volume ratio at different incremental training times, determine whether the training data is risk training data.

[0037] It can be understood that when there are more than a preset number of incremental training times with data volume ratios not meeting the requirements, then determine that the training data is risk training data. Specifically, when the data volume ratio is greater than the preset threshold, then determine that it meets the requirements.

[0038] In addition, it should be noted that when there is no risk training data, then according to a preset cycle, use the verification results in the target database between historical versions with similar model parameters to determine whether forgetting management is required, where whether the model parameters are similar is determined according to whether the deviation amounts of the model parameters are all within the preset range, and the historical version with a lower accuracy rate of the verification results can be forgotten.

[0039] In another possible embodiment, the method for determining the risk training data in the training data is as follows: Determine the data volume of the automatically labeled training data of the said type at different incremental training times based on the composition of the training data of the said type during the incremental training of different historical versions of the visual model; Based on the data volume of the automatically labeled training data of the said type at different incremental training times, determine whether the training data is risk training data.

[0040] It can be understood that when there are more than a preset number of incremental training times with data volumes not meeting the requirements, then determine that the training data is risk training data.

[0041] S2 Determine the composition of the training risk data during the incremental training of different historical versions, and combine the training risk data with the similarity of other historical versions in the later stage to determine the optimized target version in the historical versions; Furthermore, the composition of the training risk data includes the type of the training risk data in the training data during the incremental training of the historical version and the data volume of different types of training risk data.

[0042] It can be understood that as Figure 4 shown, the method for determining the optimized target version in the historical versions is as follows: Based on the composition of the training risk data during incremental training of the historical version, determine the data volume of different types of training risk data in the historical version, and determine the concerned risk data in the training risk data based on the data volume. Take the historical version after the historical version as the later version, and determine the similarity of the data volume of different concerned risk data between the later version and the historical version according to the coincidence of the concerned risk data between the later version and the historical version. Determine whether the historical version is an optimized target version according to the similarity of the data volume of different concerned risk data between the later version and the historical version.

[0043] Specifically, the concerned risk data is the training risk data whose data volume does not meet the requirements. Specifically, the training risk data with a data volume greater than the preset data volume threshold is used as the concerned risk data.

[0044] In addition, it should be noted that the similarity of the data volume of different concerned risk data between the later version and the historical version includes the deviation amount of the data volume of different concerned risk data.

[0045] Furthermore, determining whether the historical version is an optimized target version according to the similarity of the data volume of different concerned risk data between the later version and the historical version specifically includes: Take the later version whose deviation amounts of the data volume of different concerned risk data from the historical version are all within the preset deviation range as the similar version. When the number of similar versions is large, that is, greater than the threshold, it is determined that the historical version is an optimized target version.

[0046] Optionally, the method for determining the optimized target version in the historical version is: Based on the composition of the training risk data during incremental training of the historical version, determine the data volume of different types of training risk data in the historical version, and determine the concerned risk data in the training risk data based on the data volume. In addition, it should be noted that in one possible embodiment, if the total data volume of different types of training risk data or the proportion of the data volume of training risk data in the training data does not meet the requirements, that is, when the data volume of training risk data is large, the probability of problems in subsequent problem location is relatively high. Therefore, it can be determined that it does not belong to the optimized target version.

[0047] In addition, it should be further noted that even when the total amount of data of different types of training risk data in the above steps meets the requirements, it is also necessary to determine whether the number of types of risk data to be concerned meets the requirements. Specifically, when the number of types of risk data to be concerned is greater than a certain threshold, it is determined that it does not belong to the optimized target version.

[0048] Even when the number of types of risk data to be concerned meets the requirements, if the amount of data of the risk data to be concerned does not meet the requirements, it can be directly determined that it does not belong to the optimized target version, and whether it meets the requirements can be determined by means of a threshold.

[0049] Take the historical version after the historical version as the later version, and determine the similarity of the data volume of the later version and the historical version under different risk data to be concerned according to the coincidence of the risk data to be concerned between the later version and the historical version, and determine the deviation amount of the data volume of the later version under different risk data to be concerned based on the similarity. It should be further noted that when there is no later version in the later version whose deviation amounts of the data volume under different risk data to be concerned with the historical version all meet the requirements, at this time, in the later training positioning process, due to the dissimilarity of the training data, the risk of anomalies caused by the intersection of multiple risk data to be concerned is relatively large, and the probability that it needs to be used as the positioning target is relatively large. In this case, it can be determined that the historical version does not belong to the optimized target version.

[0050] In addition, it can be understood that when there is a later version in which the deviation amounts of the data volume under different risk data to be concerned with the historical version all meet the requirements, it is taken as the similar version. When the number of the similar versions is relatively large, that is, greater than the threshold, it is determined that the historical version is the optimized target version.

[0051] If the number of similar versions is not large, it is necessary to proceed to the next step for further judgment.

[0052] Determine whether the historical version is the optimized target version according to the deviation amount of the data volume of different later versions under different risk data to be concerned and the data volume of the risk data to be concerned.

[0053] It can be understood that in a possible embodiment, based on the total amount of data of the risk data to be concerned, determine the number threshold of similar versions under the total amount of data. Only when the number of similar versions is greater than the number threshold of similar versions under the total amount of data, the historical version is taken as the optimized target version.

[0054] S3 divides historical versions into different version intervals based on training risk data, and determines the types of optimization requirements in different version intervals based on the composition data of historical versions in different version intervals and the changes in the model parameters between the optimized target version and subsequent historical versions. Specifically, as Figure 5 shown, the method for determining the type of optimization requirement in the version interval is as follows: Based on the composition data of historical versions in the version interval, determine the number of historical versions and the number of optimized target versions in the version interval. According to the changes in the model parameters between the optimized target version and subsequent historical versions, determine the similarity of the model parameters between the optimized target version and subsequent historical versions, and determine parameter-similar versions based on the similarity. Determine the type of optimization requirement of the optimized target version in the version interval based on the number of historical versions, the number of optimized target versions, and the number of parameter-similar versions of different optimized versions.

[0055] It can be understood that parameter-similar versions are subsequent historical versions whose deviation amounts of different model parameters from the optimized target version all meet the requirements, and specifically, it is determined whether the requirements are met by whether they are within a preset interval.

[0056] In a possible embodiment, if the number of historical versions is greater than the preset version number threshold and the number of optimized target versions is only one, then determine that the type of optimization requirement of the optimized target version in the version interval is a type of requirement.

[0057] In addition, it should be noted that if the number of historical versions is not greater than the preset version number threshold, then determine that the type of optimization requirement of the optimized target version in the version interval is the target requirement type.

[0058] Furthermore, if the number of historical versions is greater than the preset version number threshold and the number of optimized target versions is more than one, then determine the type of optimization requirement of the optimized target version according to the number of parameter-similar versions of the optimized target version in the version interval. Specifically, if the number of parameter-similar versions is above the preset reference version number, then determine that it belongs to a type of requirement, and if it is not above the preset reference version number, then it is a type of requirement II. In a possible embodiment, the value range of the preset reference version number is from 3 to 5.

[0059] It should be noted that when there is a version interval for the optimized target version that belongs to the target requirement type, since there are versions with a relatively high degree of association for anomaly investigation, it cannot be forgotten.

[0060] If there is no version range of the target requirement type and the version ranges all belong to the same requirement type, it can be determined to perform forgetting processing on it.

[0061] In another embodiment, the method for determining the optimization requirement type within the version range is as follows: Based on the composition data of the historical versions within the version range, determine the number of historical versions and the number of optimized target versions within the version range; In the above step, if the number of historical versions is not greater than the preset version number threshold, it is determined that the optimization requirement type of the optimized target version within the version range is the target requirement type.

[0062] Furthermore, if the number of historical versions is greater than the preset version number threshold, the number of historical versions is relatively large at this time. Therefore, it is necessary to determine whether there is a variation deviation parameter, and it is necessary to transfer to the next step to determine the variation deviation parameter.

[0063] According to the variation of the model parameters of the optimized target version and the subsequent historical versions, determine the variation trend of the model parameters of the optimized target version and the subsequent historical versions, and determine the variation deviation parameter based on the variation trend; It can be understood that if the number of variation deviation parameters is relatively large, that is, greater than a certain threshold, it means that abnormal traceability processing is required for the variation deviation parameters in different dimensions. Therefore, it can be directly determined that the optimization requirement type of the optimized target version within the version range is the target requirement type.

[0064] In addition, if among the variation deviation parameters, there is a variation deviation parameter whose deviation amount of the variation parameter between the latest version and the optimized target version is not within the preset parameter deviation amount range, that is, there is a variation deviation parameter with a large variation, there is a variation deviation parameter with a high demand degree for abnormal troubleshooting at this time. Therefore, it can be directly determined that the optimization requirement type of the optimized target version within the version range is the target requirement type.

[0065] If the deviation amount of the variation parameter between the latest version and the optimized target version is within the preset parameter deviation amount range for all variation deviation parameters. Specifically, if the number of variation deviation parameters does not meet the requirements or the number of variation deviation parameters whose variation situation is within the preset variation range does not meet the requirements, it can be determined whether the requirements are met by means of a threshold, and then it is determined that the optimization requirement type of the optimized target version within the version range is the target requirement type.

[0066] Only when there is no variation deviation parameter and the number of reference versions is above the preset number of reference versions, it is determined that it belongs to a type of requirement type. In other cases, it is necessary to comprehensively consider various factors to determine the optimized requirement type.

[0067] Specifically, the variation deviation parameter is a model parameter with a consistent variation trend between different adjacent historical versions in the later stage. Specifically, if the model parameters of the later historical version and the previous historical version both increase or decrease, it indicates that the model parameter is a variation deviation parameter.

[0068] Determine the optimized requirement type of the optimized target version within the version range according to the variation situation of the variation deviation parameter, the number of historical versions, and the number of optimized target versions.

[0069] It can be understood that the variation situation of the variation deviation parameter is determined according to the deviation amount of the latest version and the optimized target version in the variation parameter.

[0070] It should be noted that if there is only one optimized target version, it is determined that the optimized requirement type of the optimized target version within the version range is a type of requirement type. When there are multiple optimized target versions, it is necessary to determine the variation situation of the variation deviation parameters of different optimized target versions and the number of parameter-similar versions. Specifically, the optimized target with the deviation amount of the latest version and the optimized target version in different variation deviation parameters less than the preset variation threshold and the number of parameter-similar versions above the preset number of reference versions is regarded as a type of requirement type, and in other cases, it is a type II requirement type.

[0071] S4 When the optimized requirement types of the optimized target version in different version ranges do not belong to the target requirement type, determine the forgetting management method of the optimized target version based on the update training situation of the training risk data corresponding to the optimized target version in the visual model and the optimized requirement types in different version ranges.

[0072] Specifically, the method for determining the forgetting management method of the optimized target version is as follows: Take the training risk data corresponding to the optimized target version as the matching risk data, and determine the historical versions corresponding to different types of matching risk data during incremental training in the most recent preset period according to the update training situation of the matching risk data in the visual model, and use them as the updated versions; Determine the version range belonging to the type II requirement type based on the optimized requirement types in different version ranges, and use it as the type II requirement range; Determine the forgetting management method of the optimized target version according to the updated versions of different types of matching risk data and the composition data of the type II requirement range.

[0073] Specifically, when the number of second-class demand intervals is greater than the preset number of demand intervals or there is matching risk data where the number of updated versions is greater than the preset updated version quantity threshold, it is determined that the forgetting management method for the optimized target version does not require forgetting processing.

[0074] In addition, it should be noted that when the number of second-class demand intervals is not greater than the preset number of demand intervals and there is matching risk data where the number of updated versions is not greater than the preset updated version quantity threshold, based on the sum of the numbers of updated versions under different types of matching risk data, when the sum of the numbers of updated versions under different types of matching risk data is greater than the preset updated version quantity threshold, it is determined that the forgetting management method for the optimized target version does not require forgetting processing.

[0075] Furthermore, it should be further noted that if the sum of the numbers of updated versions under different types of matching risk data is not greater than the preset updated version quantity threshold, the update demand quantity is determined based on the sum of the numbers of updated versions under different types of matching risk data and the number of second-class demand intervals. When the update demand quantity is less than the preset demand quantity threshold, after only performing incremental training on the visual model, that is, based on the verification result between the incrementally trained visual model and the optimized target version in the target database, it is determined whether forgetting management is required. Among them, when the deviation between the accuracy of the verification result of the optimized target version and the visual model is greater than the preset deviation value, forgetting processing of the optimized target version is performed.

[0076] And if the update demand quantity is less than the preset demand quantity threshold, when there is a visual model of the optimized target version similar to the model parameters after incremental training, the verification result between the visual model of the optimized target version similar to the model parameters and the optimized target version in the target database is used to determine whether forgetting management is required. Among them, whether the model parameters are similar is determined according to whether the deviation of the model parameters is within the preset range. Among them, when the deviation between the accuracy of the verification result of the optimized target version and the visual model is greater than the second preset deviation value, forgetting processing of the optimized target version is performed, where the preset deviation value is greater than the second preset deviation value.

[0077] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments. In particular, for the embodiments of the device, equipment, and non-volatile computer storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0078] The specific embodiments of this specification have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0079] The above is only one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, various changes and modifications can be made to one or more embodiments of this specification. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of one or more embodiments of this specification shall be included within the scope of the claims of this specification.

Claims

1. An intelligent equipment vision data governance and AI model development platform, characterized in that Specifically include: Data annotation module, model training module, version management module; Among them, the data annotation module is responsible for performing annotation processing on the data to obtain annotated data; The model training module is responsible for using the annotated data to perform training processing on the visual model as needed; The version management module is responsible for managing different historical versions of the visual model based on the training data and the updated training situation of the training data in the visual model.

2. The intelligent equipment vision data governance and AI model development platform according to claim 1, wherein The annotation processing of the data includes manual annotation and automatic annotation.

3. The intelligent equipment vision data governance and AI model development platform according to claim 1, wherein It also includes a data augmentation module, which is responsible for performing amplification processing on the annotated data.

4. The intelligent equipment vision data governance and AI model development platform according to claim 3, characterized in that The amplification processing includes amplification processing methods such as rotation, cropping, flipping, generative adversarial network, synthetic real visual background of moving targets, and synthetic real captured images + simulated visual interference.

5. The intelligent equipment vision data governance and AI model development platform according to claim 1, characterized in that, Managing different historical versions of the visual model specifically includes: When the training data of the historical version has not been incrementally trained in the visual model within the most recent preset time period, determine whether forgetting management is required through the verification results of the incrementally trained visual model and the optimization target version in the target database. Among them, when the deviation between the accuracy of the verification results of the optimization target version and the visual model is greater than the preset deviation value, perform forgetting processing on the optimization target version.

6. A management method for historical versions, which is applied to an intelligent equipment vision data governance and AI model development platform according to any one of claims 1-5, and is characterized in that, Specifically include: Based on the training data of the visual model, determine the composition of the automatically annotated data under different types of training data, and determine the risky training data in the training data based on the composition; Determine the composition of the training risk data during incremental training in different historical versions, and combine the training risk data with the similarity of other later historical versions to determine the optimization target version in the historical versions; Based on the training risk data, divide the historical versions into different version intervals, and determine the types of optimization requirements in different version intervals based on the composition data of the historical versions in different version intervals and the changes in the model parameters of the optimization target version and later historical versions; When the types of optimization requirements of the optimization target version in different version intervals do not belong to the target requirement type, determine the forgetting management method of the optimization target version based on the updated training situation of the training risk data corresponding to the optimization target version in the visual model and the types of optimization requirements in different version intervals.

7. The management method of the historical version according to claim 6, characterized in that, The method for determining the risky training data in the training data is: Based on the composition of the training data of the type during incremental training of the visual model in different historical versions, determine the data volume of the automatically annotated training data of the type under different incremental training times; According to the proportion of the data volume of the automatically annotated training data of the type under different incremental training times, determine the data volume proportion under different incremental training times; Based on the data volume proportion under different incremental training times, determine whether the training data is risky training data.

8. The management method of the historical version according to claim 7, wherein, When there are more than a preset number of incremental training times with data volume proportions not meeting the requirements, determine that the training data is risky training data.

9. The management method of the historical version according to claim 6, wherein, The method for determining the forgetting management method of the optimization target version is: Take the training risk data corresponding to the optimized target version as the matching risk data, determine the historical versions corresponding to different types of matching risk data during incremental training in the most recent preset period according to the updated training situation of the matching risk data in the visual model, and use them as the updated versions; Determine the version interval belonging to the second - type demand type based on the types of optimization requirements in different version intervals, and use it as the second - type demand interval; Determine the forgetting management method of the optimized target version according to the updated versions of different types of matching risk data and the composition data of the second - type demand interval.

10. The management method of the historical version according to claim 9, characterized in that, When the number of second - type demand intervals is greater than the preset demand interval number or there is matching risk data with the number of updated versions greater than the preset updated version number threshold, then determine that the forgetting management method of the optimized target version is that no abnormal management is required.

Images