Rapid iterative training platform for coal mine scene AI model

Abstract

The invention relates to the technical field of coal mine safety monitoring, in particular to a coal mine scene AI model rapid iterative training platform, which comprises a data center, a training center, an application center and an algorithm application mall, and is characterized in that the data center is used for providing data set management; the training center comprises algorithm management, standardized training and navigation type training, and in an algorithm management module, mirror images and algorithms are designed and uploaded; in standardized training, an algorithm template is formed after training and is released to a basic algorithm library of an application center; the application center selects an algorithm from the basic algorithm library, develops an algorithm application package, matches the algorithm application package with an actual application scene to form an application task template, and issues the application task template to an algorithm application shopping mall; in the navigation type training, an algorithm application package is selected from an algorithm application shopping mall through a guide type training task, automatic training is performed, and a required model file is obtained. High-efficiency production and centralized management of AI capability can be realized, and the landing threshold and maintenance cost of a coal mine AI scene are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mine safety monitoring, and in particular to a coal mine scenario AI model rapid iteration training platform. Background Art

[0002] With the development of artificial intelligence technology, the coal mining industry has ushered in a major change. Machine vision technology uses visual sensors to replace human eyes for various measurements and judgments, and is an important means of coal mine safety monitoring. Compared with industrial visual inspection, the underground environment of coal mines is more complex. The location, angle, and fill light conditions of camera installation are often difficult to achieve the best state. There are many external influencing factors, especially the influence of uneven lighting, dust and water mist, etc., which makes most algorithms require customized development and rapid update iterations. It is necessary to collect data and retrain the algorithm according to the on-site situation, and then deploy the optimal model to the corresponding edge equipment underground. The existing platform has high requirements on the professional quality of operators, which is not conducive to the operation of coal mine workers. There is an urgent need for a low-code platform to make AI training and deployment simpler, more intuitive, concise and faster, lower the technical threshold, and enable operation and maintenance personnel and business personnel to participate in development.

[0003] Most of the existing training platforms are specialized and detailed training platforms built by AI algorithm engineers. They require strong professional knowledge as support. The task flow requires various connections and lacks guidance. They are relatively complicated to use and the threshold is too high for coal miners. These platforms cannot be well applied in the coal mining field. Algorithm development, model training, format conversion, etc. in coal mine scenarios involve algorithm software, computing chips, image acquisition equipment, and mutual coordination. The technology stack involved is deep and the technical requirements are high, which makes it impossible for the most proficient on-site staff to quickly update and iterate the model based on the actual situation on site. Summary of the invention

[0004] The technical problem to be solved by the present invention is: in order to solve the problems existing in the prior art in the above-mentioned background technology, a rapid iteration training platform for AI models of coal mine scenarios is provided.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a coal mine scene AI model rapid iteration training platform, comprising: Data center, used to provide data set management; The training center includes algorithm management, standardized training and navigation training. In the algorithm management module, images and algorithms are designed and uploaded. In standardized training, algorithms and images in the algorithm management module are selected, and training parameters and data sets are configured. After training is completed and the model is evaluated, converted and verified, an algorithm template is formed and published to the basic algorithm library of the application center. The application center has a basic algorithm library. For a certain application scenario and specific device model, it selects algorithms from the basic algorithm library, develops algorithm application packages, matches them with actual application scenarios, forms application task templates, and publishes them to the algorithm application mall; In navigation training, you can select an algorithm application package from the algorithm application mall through guided training tasks, and then automatically train to obtain the required model file.

[0006] Furthermore, in the data center, users import image or video data through newly added data sets, perform data processing on the imported images, and publish them as a public data set for use by algorithm developers or engineers.

[0007] Furthermore, in the data center, users use local import to directly import images or videos or import compressed packages containing annotation information and images in a specific format.

[0008] Furthermore, the newly added data sets are first stored in the original data warehouse and then imported into the private data set for data processing. The locally imported data goes directly into the private data set. The data processing includes data enhancement, image annotation, and video frame extraction.

[0009] Furthermore, image annotation includes automatic annotation, manual annotation or multi-person annotation. If automatic annotation is used, the images in the data set are annotated according to a certain algorithm; if manual annotation is used, each image is annotated in the annotation interface; multi-person annotation is managed by a multi-person annotation team, which initiates multi-person annotation tasks and completes multi-person annotation according to the tasks.

[0010] Furthermore, private data sets are managed and published according to data versions, and the versions are published to public data sets. At the same time, the labeled data version is saved to the current data version, and a historical version is switched to the current version, replacing the data of a historical version for use by algorithm developers or engineers.

[0011] Furthermore, users develop post-processing codes in the application center, configure and bind the parameters of computing chips, device firmware, scenario names, and application names one by one, develop algorithm application packages including models, packaging scripts, information files, and post-processing files, and match algorithms, post-processing files, device models, and other actual application scenarios.

[0012] Furthermore, in the algorithm application mall, all published algorithm applications are displayed in the form of a list.

[0013] Furthermore, in the navigation training, first create a guided training task, select the algorithm application package, start the guided training, enter the model training, and then verify the model. After the verification is correct, the training model will be packaged for download.

[0014] Furthermore, while the model is training, you can view the training parameters, status, and logs, and you can also stop training.

[0015] Beneficial effects of the present invention: The platform constructed by the present invention includes two forms: development mode and engineering mode. The development mode is used by algorithm development engineers. According to business needs, they collect, clean, and annotate data, select or design appropriate algorithms, perform pre-processing and post-processing code development, build model format conversion images according to different frameworks and hardware, and complete the configuration of algorithm-related hyperparameters through standardized training to form training templates and store them in the algorithm application mall of the platform; the engineering mode is used by coal mine on-site engineering personnel. Based on the training templates in the algorithm application mall, it assists engineering personnel to complete model training and format conversion through guided operation processes, automatically evaluates the model accuracy, and deploys and updates to the corresponding smart terminals or edge devices with one click; The present invention can realize efficient production and centralized management of AI capabilities, reduce the threshold for implementation and maintenance costs of AI scenarios in coal mines, and improve the intelligence level of coal mine enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0017] Figure 1 It is a flow chart of the present invention.

[0018] Figure 2 It is a flow chart of data set management in a data center of the present invention.

[0019] Figure 3 It is a flow chart of model generation in the present invention. DETAILED DESCRIPTION

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0021] For a complete AI scenario application, the update and iteration of the algorithm model is divided into five steps: ① Data preparation, including data collection, cleaning, labeling, etc.; ② Algorithm design and training, parameters need to be adjusted continuously during the training process to improve the performance of the model; ③ Model compression, the trained model may be very large, and model compression is required to reduce the size of the model and increase the reasoning speed; ④ Format conversion, the smart devices used underground involve different hardware frameworks, and model format conversion is required; ⑤ Model deployment, the trained model needs to be packaged into a file and applied to the actual environment. For ordinary coal mine engineers, if there is a lack of training templates that match the application scenario, they will not be able to guide them to better carry out on-site engineering applications.

[0022] The coal mine scenario AI model rapid iteration training platform of this embodiment takes into account the usage needs of developers and on-site coal mine engineering personnel by introducing standardized training and navigation training methods, thereby reducing the application threshold and maintenance cost of coal mine AI scenarios.

[0023] like Figure 1 As shown in the figure, the coal mine scenario AI model rapid iteration training platform includes a data center, a training center, an application center and an algorithm application mall, wherein the data center is used to provide data set management; the training center includes algorithm management, standardized training and navigation training, and images and algorithms are designed and uploaded in the algorithm management module; in standardized training, the algorithm and image in the algorithm management module are selected, and the training parameters and data sets are configured. After the training is completed and the model is evaluated, converted and verified, an algorithm template is formed and published to the basic algorithm library of the application center; the application center has a basic algorithm library, and for a certain application scenario and a specific equipment model, an algorithm is selected from the basic algorithm library, an algorithm application package is developed, matched with the actual application scenario, an application task template is formed, and published to the algorithm application mall; in navigation training, an algorithm application package is selected from the algorithm application mall through a guided training task, and training is performed automatically to obtain the required model file.

[0024] like Figure 2 As shown in the figure, in the data center, users can import pictures or video data through the newly added data set, process the imported images, and publish them as a public data set for use by algorithm developers or engineers; users can also use local import to directly import pictures or videos or import compressed packages containing annotation information and pictures in a specific format. For the newly added data set, it is first stored in the original data warehouse, and then imported into the private data set for data processing; for the locally imported data, it is directly entered into the private data set, and the data processing includes data enhancement, image annotation and video frame extraction. Among them, image annotation includes automatic annotation, manual annotation or multi-person annotation. If automatic annotation is used, the annotation of the pictures in the data set is completed according to a certain algorithm; if manual annotation is used, each picture is annotated in the annotation interface; multi-person annotation is managed by the multi-person annotation team, and multi-person annotation tasks are initiated and completed according to the tasks. Private data sets are managed and published according to data versions, and versions are published to public data sets. At the same time, the data version after annotation is saved to the current data version, and a certain historical version is switched to the current version, replacing the data of a certain historical version for use by algorithm developers or engineers.

[0025] like Figure 3As shown in the figure, in standardized training, standardized training tasks are created based on the public data sets that have been annotated and published and the uploaded training images. Then, training can be performed. Training subtasks can be created to train the model, and the model can be verified by testing the verification set or by uploading pictures locally. After verification, the model is converted to obtain the training model and training parameters, etc., which are synchronized to the basic algorithm library, and then the algorithm application package is made to the application center and then published to the algorithm application mall; this process is operated by the algorithm development engineer, so it is a development mode. In navigation training, based on the training template of the algorithm application mall, the guided operation process assists the engineering personnel to complete the model training and format conversion, and automatically evaluates the model accuracy, and deploys and updates to the corresponding smart terminal or edge side device with one click; this process is operated by the on-site engineering personnel of the coal mine, so it is an engineering mode. The equipment mentioned here is specifically the smart camera or edge computing unit in the coal mine shaft, which can store the iterated model and ensure the accuracy of the model.

[0026] In actual applications, coal mine field staff or engineering personnel can quickly update and iterate the model through navigation training without writing any code, which reduces the threshold for use. When it is found that the model accuracy is not enough, the engineering personnel only need to collect pictures or video data through the camera equipment at the on-site work location, complete the labeling of the data set in the data center, and then select the algorithm application, device model, firmware version number, etc. from the algorithm application mall through the navigation training task, and then automatically load the training weights, training hyperparameters, computing power resource configuration, etc., and then automatically train, and finally get the required model file, so as to achieve the purpose of "fool-style operation". At the same time, a series of logs of the training process, quality inspection analysis process, and model conversion process can be viewed to facilitate the analysis of the training process.

[0027] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A coal mine scenario AI model rapid iteration training platform, characterized by: include: Data center, which provides data set management; The training center includes algorithm management, standardized training and navigation training. In the algorithm management module, images and algorithms are designed and uploaded. In standardized training, algorithms and images in the algorithm management module are selected, and training parameters and data sets are configured. After training is completed and the model is evaluated, converted and verified, an algorithm template is formed and published to the basic algorithm library of the application center. The application center has a basic algorithm library. For a certain application scenario and specific device model, it selects algorithms from the basic algorithm library, develops algorithm application packages, matches them with actual application scenarios, forms application task templates, and publishes them to the algorithm application mall; In navigation training, you can select an algorithm application package from the algorithm application mall through guided training tasks, and then automatically train to obtain the required model file.

2. The coal mine scenario AI model rapid iteration training platform according to claim 1 is characterized by: In the data center, users import image or video data through newly added data sets, process the imported images, and publish them as a public data set for use by algorithm developers or engineers.

3. The coal mine scenario AI model rapid iteration training platform according to claim 2 is characterized by: In the data center, users use local import to directly import images or videos or import compressed packages in a specific format that contain annotation information and images.

4. The coal mine scenario AI model rapid iteration training platform according to claim 2 or 3, characterized in that: The newly added data set is first stored in the original data warehouse, and then imported into the private data set for data processing. The locally imported data directly enters the private data set. Data processing includes data enhancement, image annotation, and video frame extraction.

5. The coal mine scenario AI model rapid iteration training platform according to claim 4 is characterized by: Image annotation includes automatic annotation, manual annotation or multi-person annotation. If automatic annotation is used, the images in the data set are annotated according to a certain algorithm; if manual annotation is used, each image is annotated in the annotation interface; multi-person annotation is managed by a multi-person annotation team, which initiates multi-person annotation tasks and completes multi-person annotation according to the tasks.

6. The coal mine scenario AI model rapid iteration training platform according to claim 4 is characterized by: Private data sets are managed and published according to data versions, and the versions are published to public data sets. At the same time, the labeled data version is saved to the current data version, and a historical version is switched to the current version, replacing the data of a historical version for use by algorithm developers or engineers.

7. The coal mine scenario AI model rapid iteration training platform according to claim 1 is characterized by: Users develop post-processing codes in the application center, configure and bind the parameters of computing chips, device firmware, scenario names, and application names one by one, develop algorithm application packages including models, packaging scripts, information files, and post-processing files, and match algorithms, post-processing files, device models, and other actual application scenarios.

8. The coal mine scenario AI model rapid iteration training platform according to claim 1 is characterized by: In the algorithm application store, all published algorithm applications are displayed in a list.

9. The coal mine scenario AI model rapid iteration training platform according to claim 1 is characterized by: In the guided training, you first create a guided training task, select the algorithm application package, start the guided training, enter the model training, and then verify the model. After the verification is correct, the training model will be packaged for download.

10. The coal mine scenario AI model rapid iteration training platform according to claim 9 is characterized by: While the model is training, you can view the training parameters, status, and logs, and you can also stop training.

Images