A semi-automatic artificial intelligence model optimization system and method

Abstract

The present invention provides a semi-automatic artificial intelligence model optimization system and method, the system includes: a control unit, a storage unit and a training model, the training model has: a main neural network unit, a monitoring module, a manual interaction interface, an identification module, multiple sub-neural network units and a revision module. After the training model is built, the present application randomly extracts the training results obtained by the training model, and forms monitoring samples after extraction. The monitoring samples are sent to the manual interaction end, and the monitoring samples are verified by manual experts to obtain verification results. The verification results are compared with the corresponding monitoring samples, and the verification results are adaptively labeled according to the comparison results. The sub-neural network unit loads the adaptive annotation of the verification results to iteratively train the verification results to form revised results; the revised results are compared with the corresponding original monitoring data to form a difference, and the corresponding experience parameters are updated based on the difference.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence application technology, and in particular to a semi-automatic artificial intelligence model optimization system and method. Background Art

[0002] Traditional production process monitoring basically uses human monitoring, and major safety accidents are generally caused by human negligence. Therefore, current production process monitoring basically uses artificial intelligence processing. In the early stages of the construction of the artificial intelligence system, its training model cannot be optimized and still has large errors. It can only be deployed in the control device after continuous optimization. Summary of the invention

[0003] In view of this, the main purpose of the present invention is to provide a semi-automatic artificial intelligence model optimization system and method.

[0004] The technical solution adopted by the present invention is as follows:

[0005] A semi-automated artificial intelligence model optimization system, comprising:

[0006] Control Department;

[0007] A storage unit, used to send the monitoring data collected during the production process to the storage unit, wherein the storage unit has a database for storing monitoring data and an instruction unit for storing executable instructions;

[0008] A training model with:

[0009] A main neural network unit is used to import historical monitoring data in a database and load manually reviewed empirical parameters and to perform training based on training instructions from a control unit to obtain training results;

[0010] A monitoring module, comprising a plurality of monitoring units, wherein the plurality of monitoring units are connected to the main neural network unit, and is used for the control unit to load monitoring instructions from the instruction unit based on the setting of monitoring parameters and to randomly extract training results as monitoring samples according to the progress of the training task of the main neural network unit;

[0011] A manual interaction interface is connected to the manual interaction terminal, and the manual interaction interface sends the monitoring sample to the manual interaction terminal, and the monitoring sample is manually verified to obtain the verification result, and the verification result is compared with the corresponding monitoring sample, and the verification result is adaptively marked according to the comparison result;

[0012] An identification module, used to identify the adaptability annotation of the verification results, classify the verification results based on the adaptability annotation status of the verification results, and store the classification results in the database accordingly;

[0013] A plurality of sub-neural network units, wherein the plurality of sub-neural network units are selectively activated based on the classification result of the recognition module, and the sub-neural network units are used to load the adaptive annotation of the verification result to iteratively train the verification result to form a revised result;

[0014] The revision module is used to compare the revision result with the original monitoring data corresponding to the verification result to form a difference, and update the empirical parameters accordingly based on the difference.

[0015] Furthermore, the monitoring unit has a monitoring field that responds to the training results, and the monitoring field is used to extract and call the training results of the main neural network model under the sampling instruction to form a monitoring sample, and based on the monitoring sample, establish a traceability path between the monitoring sample and the training result, and between the training result and the original monitoring data, and store the traceability path in the traceability list of the database.

[0016] Furthermore, the identification module has:

[0017] An identification unit, used to identify the adaptive annotation and annotation code of the verification result;

[0018] Classification unit, used to classify the adaptive annotations of the verification results based on the annotation codes;

[0019] The tracing unit is used to establish two-segment tracing paths between the classification results and the verification results, and between the verification results and the monitoring samples, and store the two-segment tracing paths correspondingly under the tracing list of one tracing path.

[0020] Furthermore, each of the sub-neural network units has an activation mark, the control unit is connected to the recognition module and generates an activation instruction corresponding to the classification result of the recognition module, the activation instruction corresponds to the selected sub-neural network unit and the activation mark of the sub-neural network unit is activated, and the adaptive annotation of the verification result is loaded to the sub-neural network unit through the loading unit to iteratively train the verification result to form a revised result;

[0021] Each of the sub-neural network units also has a tracing executor, which responds to the activation instruction to activate the activation flag to track the learning path of the verification result through the set tracing execution event during the process of the sub-neural network unit loading the adaptive annotation of the verification result to perform iterative training on the verification result, and matches and stores the learning path with the two-segment tracing path of the verification result;

[0022] A behavior monitor is provided, wherein the behavior monitor is capable of monitoring an activation field in an activation tag, wherein the activation field links the activation field into a loading unit and a tracing executor in response to an activation instruction and independently controls execution operations of the loading unit and the tracing executor.

[0023] Furthermore, the revision module is based on the learning path of the revision result, loads the learning path through two traceability paths and matches one traceability path to compare the revision result with the original monitoring data corresponding to the verification result and form a difference, and updates the experience parameters accordingly based on the difference.

[0024] The present invention also provides a semi-automatic artificial intelligence model optimization method, comprising the following steps:

[0025] A1): Send the monitoring data collected during the production process to the database of the storage department;

[0026] A2): The main neural network unit imports historical monitoring data from the database and loads manually reviewed empirical parameters and performs training based on the training instructions of the control unit to obtain training results;

[0027] A3): The monitoring unit loads the monitoring instructions based on the setting of the monitoring parameters and randomly extracts the training results as monitoring samples according to the progress of the training task of the main neural network unit;

[0028] A4): The manual interaction interface sends the monitoring samples to the manual interaction terminal, and the monitoring samples are manually verified to obtain verification results, and the verification results are compared with the corresponding monitoring samples, and the verification results are adaptively labeled according to the comparison results;

[0029] A5): Identify the adaptability annotations and annotation codes of the verification results, and classify the adaptability annotations of the verification results based on the annotation codes; establish two-segment traceability paths between the classification results and the verification results, and between the verification results and the monitoring samples, and store the two-segment traceability paths in the traceability list of the first traceability path;

[0030] A6): The control unit connects to the identification module and generates an activation instruction corresponding to the classification result of the identification module, selects a sub-neural network unit corresponding to the activation instruction and activates the activation flag of the sub-neural network unit, and loads the adaptive annotation of the verification result to the sub-neural network unit through the loading unit to iteratively train the verification result to form a revised result; the tracking executor responds to the activation instruction to activate the activation flag to track the learning path of the verification result in the process of iteratively training the verification result by loading the adaptive annotation of the verification result in the sub-neural network unit through the set tracking execution event, and matches and stores the learning path with the two-stage tracing path of the verification result;

[0031] A7): The revision module is based on the learning path of the revision result. The learning path is loaded and matched through two traceability paths and one traceability path to compare the revision result with the original monitoring data corresponding to the verification result and form a difference. Based on the difference, the corresponding experience parameters are updated.

[0032] In step A2), the monitoring unit has a monitoring field that responds to the training results, and the monitoring field is used to extract and call the training results of the main neural network model under the sampling instruction to form a monitoring sample, and based on the monitoring sample, establish a traceability path between the monitoring sample and the training result, and between the training result and the original monitoring data, and store the traceability path in the traceability list of the database.

[0033] After the training model is built, the present application randomly extracts the training results obtained by the training model to form monitoring samples, which are sent to the manual interaction terminal. Manual experts verify the monitoring samples to obtain verification results, compare the verification results with the corresponding monitoring samples, and adaptively annotate the verification results based on the comparison results. The sub-neural network unit then loads the adaptive annotations of the verification results to iteratively train the verification results to form revised results. The revised results are compared with the corresponding original monitoring data to form a difference, and the empirical parameters are updated accordingly based on the difference.

[0034] In the present invention, a tracing system is also established for tracing the entire training process, including establishing a tracing path between the monitoring sample and the training result, and between the training result and the original monitoring data based on the monitoring sample, establishing a two-segment tracing path between the classification result and the verification result, and between the verification result and the monitoring sample, and a learning path constructed based on the tracing executor; the one-segment tracing path, the two-segment tracing path and the learning path can be consistently matched, and a monitoring data can be completely traced during the processing process, and tracing is conducive to matching the learning results with the original monitoring data. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The following drawings are only used to illustrate and explain the present invention, and are not used to limit the scope of the present invention, wherein:

[0036] Figure 1 It is a system configuration diagram of the present invention;

[0037] Figure 2 It is a schematic diagram of the detailed framework principle of the system of the present invention;

[0038] Figure 3 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution, design method and advantages of the present invention more clear, the present invention is further described in detail by specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] Reference Figures 1 to 3 The present invention provides a semi-automatic artificial intelligence model optimization system. After the training model is built, the training results obtained by the training model are randomly extracted to form monitoring samples. The monitoring samples are sent to the manual interaction terminal, and the monitoring samples are verified by manual experts to obtain verification results. The verification results are compared with the corresponding monitoring samples, and the verification results are adaptively labeled according to the comparison results. The adaptive annotations of the verification results are then loaded by the sub-neural network unit to iteratively train the verification results to form revised results; the revised results are compared with the corresponding original monitoring data to form a difference, and the corresponding experience parameters are updated based on the difference.

[0041] Specifically, a semi-automatic artificial intelligence model optimization system includes:

[0042] Control Department;

[0043] A storage unit, used to send the monitoring data collected during the production process to the storage unit, wherein the storage unit has a database for storing monitoring data and an instruction unit for storing executable instructions;

[0044] A training model with:

[0045] A main neural network unit is used to import historical monitoring data in a database and load manually reviewed empirical parameters and to perform training based on training instructions from a control unit to obtain training results;

[0046] A monitoring module, comprising a plurality of monitoring units, wherein the plurality of monitoring units are connected to the main neural network unit, and is used for the control unit to load monitoring instructions from the instruction unit based on the setting of monitoring parameters and to randomly extract training results as monitoring samples according to the progress of the training task of the main neural network unit;

[0047] A manual interaction interface is connected to the manual interaction terminal, and the manual interaction interface sends the monitoring sample to the manual interaction terminal, and the monitoring sample is manually verified to obtain the verification result, and the verification result is compared with the corresponding monitoring sample, and the verification result is adaptively marked according to the comparison result;

[0048] An identification module, used to identify the adaptability annotation of the verification results, classify the verification results based on the adaptability annotation status of the verification results, and store the classification results in the database accordingly;

[0049] A plurality of sub-neural network units, wherein the plurality of sub-neural network units are selectively activated based on the classification result of the recognition module, and the sub-neural network units are used to load the adaptive annotation of the verification result to iteratively train the verification result to form a revised result;

[0050] The revision module is used to compare the revision result with the original monitoring data corresponding to the verification result to form a difference, and update the empirical parameters accordingly based on the difference.

[0051] In the above, the monitoring unit has a monitoring field that responds to the training results, and the monitoring field is used to extract and call the training results of the main neural network model under the sampling instruction to form a monitoring sample, and based on the monitoring sample, establish a traceability path between the monitoring sample and the training result, and between the training result and the original monitoring data, and store the traceability path in the traceability list of the database.

[0052] In the above, the identification module has:

[0053] An identification unit, used to identify the adaptive annotation and annotation code of the verification result;

[0054] Classification unit, used to classify the adaptive annotations of the verification results based on the annotation codes;

[0055] The tracing unit is used to establish two-segment tracing paths between the classification results and the verification results, and between the verification results and the monitoring samples, and store the two-segment tracing paths correspondingly under the tracing list of one tracing path.

[0056] In the above, each of the sub-neural network units has an activation mark, the control unit is connected to the recognition module and generates an activation instruction corresponding to the classification result of the recognition module, the activation instruction corresponds to the selected sub-neural network unit and activates the activation mark of the sub-neural network unit, and the adaptive annotation of the verification result is loaded to the sub-neural network unit through the loading unit to iteratively train the verification result to form a revised result;

[0057] Each of the sub-neural network units also has a tracing executor, which responds to the activation instruction to activate the activation flag to track the learning path of the verification result through the set tracing execution event during the process of the sub-neural network unit loading the adaptive annotation of the verification result to perform iterative training on the verification result, and matches and stores the learning path with the two-segment tracing path of the verification result;

[0058] A behavior monitor is provided, wherein the behavior monitor is capable of monitoring an activation field in an activation tag, wherein the activation field links the activation field into a loading unit and a tracing executor in response to an activation instruction and independently controls execution operations of the loading unit and the tracing executor.

[0059] Furthermore, the revision module is based on the learning path of the revision result, loads the learning path through two traceability paths and matches one traceability path to compare the revision result with the original monitoring data corresponding to the verification result and form a difference, and updates the experience parameters accordingly based on the difference.

[0060] The present invention also provides a semi-automatic artificial intelligence model optimization method, comprising the following steps:

[0061] A1): Send the monitoring data collected during the production process to the database of the storage department;

[0062] A2): The main neural network unit imports historical monitoring data from the database and loads manually reviewed empirical parameters and performs training based on the training instructions of the control unit to obtain training results;

[0063] A3): The monitoring unit loads the monitoring instructions based on the setting of the monitoring parameters and randomly extracts the training results as monitoring samples according to the progress of the training task of the main neural network unit;

[0064] A4): The manual interaction interface sends the monitoring samples to the manual interaction terminal, and the monitoring samples are manually verified to obtain verification results, and the verification results are compared with the corresponding monitoring samples, and the verification results are adaptively labeled according to the comparison results;

[0065] A5): Identify the adaptability annotations and annotation codes of the verification results, and classify the adaptability annotations of the verification results based on the annotation codes; establish two-segment traceability paths between the classification results and the verification results, and between the verification results and the monitoring samples, and store the two-segment traceability paths in the traceability list of the first traceability path;

[0066] A6): The control unit connects to the identification module and generates an activation instruction corresponding to the classification result of the identification module, selects a sub-neural network unit corresponding to the activation instruction and activates the activation flag of the sub-neural network unit, and loads the adaptive annotation of the verification result to the sub-neural network unit through the loading unit to iteratively train the verification result to form a revised result; the tracking executor responds to the activation instruction to activate the activation flag to track the learning path of the verification result in the process of iteratively training the verification result by loading the adaptive annotation of the verification result in the sub-neural network unit through the set tracking execution event, and matches and stores the learning path with the two-stage tracing path of the verification result;

[0067] A7): The revision module is based on the learning path of the revision result. The learning path is loaded and matched through two traceability paths and one traceability path to compare the revision result with the original monitoring data corresponding to the verification result and form a difference. Based on the difference, the corresponding experience parameters are updated.

[0068] In step A2), the monitoring unit has a monitoring field that responds to the training results, and the monitoring field is used to extract and call the training results of the main neural network model under the sampling instruction to form a monitoring sample, and based on the monitoring sample, establish a traceability path between the monitoring sample and the training result, and between the training result and the original monitoring data, and store the traceability path in the traceability list of the database.

[0069] With reference to the following embodiments:

[0070] A semi-automated artificial intelligence model optimization method comprises the following steps:

[0071] A1): Send the monitoring images collected during the production process to the database of the storage department;

[0072] A2): The main neural network unit imports the historical monitoring images in the database and loads the manually reviewed empirical parameters and performs training based on the training instructions of the control unit to obtain training results; in this step, the manually reviewed empirical parameters are the identification features that affect safe production that are manually marked in the historical monitoring images, and empirical parameters can be formed by marking these identification features.

[0073] A3): The monitoring unit loads the monitoring instructions based on the set monitoring parameters and randomly extracts the training results as monitoring samples according to the progress of the training task of the main neural network unit; wherein, the monitoring parameters can collect different training results as samples according to the set period, and the monitoring parameters can set the collection frequency and collection quantity.

[0074] A4): The manual interaction interface sends the monitoring samples to the manual interaction terminal, which manually verifies the monitoring samples to obtain verification results, compares the verification results with the corresponding monitoring samples, and adaptively annotates the verification results based on the comparison results; this step is to review the monitoring samples collected by random inspections again by manual experts. Since the monitoring samples have been screened to a certain extent, marking them on the screening results is conducive to re-screening the preliminary training results. In this process, in order to facilitate efficiency, the identification features identified in the training results can be extracted, and the training results with identification features can be manually annotated again.

[0075] A5): Identify the adaptive annotations and annotation codes of the verification results, and classify the adaptive annotations of the verification results based on the annotation codes; establish two-segment traceability paths between the classification results and the verification results, and between the verification results and the monitoring samples, and store the two-segment traceability paths correspondingly under the traceability list of the first traceability path; in this step, the adaptive annotation of the verification results is to re-annotate the characteristic positions and forms of safety hazards in the training results.

[0076] A6): The control unit connects to the identification module and generates an activation instruction corresponding to the classification result of the identification module, selects a sub-neural network unit corresponding to the activation instruction and activates the activation flag of the sub-neural network unit, and loads the adaptive annotation of the verification result to the sub-neural network unit through the loading unit to iteratively train the verification result to form a revised result; the tracking executor responds to the activation instruction to activate the activation flag to track the learning path of the verification result in the process of iteratively training the verification result by loading the adaptive annotation of the verification result in the sub-neural network unit through the set tracking execution event, and matches and stores the learning path with the two-stage tracing path of the verification result;

[0077] A7): The revision module is based on the learning path of the revision result. The learning path is loaded and matched through two traceability paths and one traceability path to compare the revision result with the original monitoring data corresponding to the verification result and form a difference. Based on the difference, the corresponding experience parameters are updated.

[0078] The present invention establishes a tracing system for tracing the entire training process, including establishing a first tracing path between monitoring samples and training results, and between training results and original monitoring data based on monitoring samples, establishing two tracing paths between classification results and verification results, and between verification results and monitoring samples, and a learning path constructed based on a tracing executor; the first tracing path, the second tracing path and the learning path can be consistently matched, and a monitoring data can be completely traced during the processing process, and tracing is conducive to matching the learning results with the original monitoring data.

[0079] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A semi-automatic artificial intelligence model optimization system, characterized in that: include: Control Department; A storage unit, used to send the monitoring data collected during the production process to a database of the storage unit, wherein the storage unit has a database for storing monitoring data and an instruction unit for storing executable instructions; A training model with: A main neural network unit is used to import historical monitoring data in a database and load manually reviewed empirical parameters and to perform training based on training instructions from a control unit to obtain training results; A monitoring module, comprising a plurality of monitoring units, wherein the plurality of monitoring units are connected to the main neural network unit, and is used for the control unit to load monitoring instructions from the instruction unit based on the setting of monitoring parameters and to randomly extract training results as monitoring samples according to the progress of the training task of the main neural network unit; A manual interaction interface is connected to the manual interaction terminal, and the manual interaction interface sends the monitoring sample to the manual interaction terminal, and the monitoring sample is manually verified to obtain the verification result, and the verification result is compared with the corresponding monitoring sample, and the verification result is adaptively marked according to the comparison result; an identification module, used to identify the adaptability annotation and annotation code of the verification result, and classify the adaptability annotation status of the verification result based on the annotation code, establish a second traceability path between the classification result and the verification result, and between the verification result and the monitoring sample, and store the second traceability path in the traceability list of the first traceability path; A plurality of sub-neural network units, wherein the plurality of sub-neural network units are used to generate activation instructions corresponding to the classification results of the recognition module, select a sub-neural network unit corresponding to the activation instruction and activate the activation flag of the sub-neural network unit, and load the adaptive annotation of the verification result to the sub-neural network unit through a loading unit to iteratively train the verification result to form a revised result; The tracing executor responds to the activation instruction to activate the activation flag to trace the learning path of the verification result in the process of iterative training of the verification result by loading the adaptive annotation of the verification result in the sub-neural network unit through the set tracing execution event, and matches and stores the learning path with the second tracing path of the verification result; A revision module is used for comparing the revision result with the original monitoring data corresponding to the verification result based on the learning path of the revision result, loading the learning path and matching the first tracing path through the second tracing path, and forming a difference, and updating the experience parameter accordingly based on the difference; The monitoring unit has a monitoring field that responds to the training results. The monitoring field is used to extract and call the training results of the main neural network model under the sampling instruction to form a monitoring sample, and establish a first traceability path between the monitoring sample and the training result, and between the training result and the original monitoring data based on the monitoring sample, and store the first traceability path in the traceability list of the database.

2. The semi-automatic artificial intelligence model optimization system according to claim 1, characterized in that: The sub-neural network unit also includes: A behavior monitor is provided, wherein the behavior monitor is capable of monitoring an activation field in an activation tag, wherein the activation field links the activation field into a loading unit and a tracing executor in response to an activation instruction and independently controls execution operations of the loading unit and the tracing executor.

3. A semi-automatic artificial intelligence model optimization method, characterized in that: The steps include: A1): Send the monitoring data collected during the production process to the database of the storage department; A2): The main neural network unit imports historical monitoring data from the database and loads manually reviewed empirical parameters and performs training based on the training instructions of the control unit to obtain training results; A3): The monitoring unit loads the monitoring instructions based on the setting of the monitoring parameters and randomly extracts the training results as monitoring samples according to the progress of the training task of the main neural network unit; A4): The manual interaction interface sends the monitoring samples to the manual interaction terminal, and the monitoring samples are manually verified to obtain verification results, and the verification results are compared with the corresponding monitoring samples. The verification results are adaptively labeled according to the comparison results; A5): Identify the adaptability annotations and annotation codes of the verification results, and classify the adaptability annotations of the verification results based on the annotation codes; establish a second traceability path between the classification results and the verification results, and between the verification results and the monitoring samples, and store the second traceability path in the traceability list of the first traceability path; A6): The control unit connects to the recognition module and generates an activation instruction corresponding to the classification result of the recognition module, selects a sub-neural network unit corresponding to the activation instruction and activates the activation flag of the sub-neural network unit, and loads the adaptive annotation of the verification result to the sub-neural network unit through the loading unit to iteratively train the verification result to form a revised result; The tracing executor responds to the activation instruction to activate the activation flag to trace the learning path of the verification result in the process of iterative training of the verification result by loading the adaptive annotation of the verification result in the sub-neural network unit through the set tracing execution event, and matches and stores the learning path with the second tracing path of the verification result; A7): The revision module is based on the learning path of the revision result, loads the learning path, matches the first traceability path through the second traceability path, compares the revision result with the original monitoring data corresponding to the verification result, and forms a difference, and updates the experience parameter accordingly based on the difference; In step A2), the monitoring unit has a monitoring field that responds to the training results, and the monitoring field is used to extract and call the training results of the main neural network model under the sampling instruction to form a monitoring sample, and establish a first traceability path between the monitoring sample and the training result, and between the training result and the original monitoring data based on the monitoring sample, and store the first traceability path in the traceability list of the database.

Images