Online Continuous Learning System for Safety Warning Based on Loss Regularization

Abstract

The present invention discloses a secure risk indication online continuous learning system based on loss regularization, belonging to the technical field of continuous learning systems, including a data platform, an application platform, a data capture module, a data recognition module, a training module, a data reading module, a learning model module, a loss function module, and a differential data module. Among them, the output end of the data platform is connected to the data capture module, the output end of the data capture module is connected to the data recognition module, and the output end of the data recognition module is connected to the data reading module. The secure risk indication online continuous learning system based on loss regularization proposed by the present invention, based on loss regularization, splits new and old knowledge, minimizes the difference between the two, and prevents the problem of catastrophic forgetting. Based on sample reproduction, it retains some representative samples or knowledge bases and trains them together with new data. Based on network structure adjustment, part of the network is used to learn old tasks, and new knowledge is stored in other parts of the network.

Description

Technical Field

[0001] The present invention relates to the technical field of continuous learning systems, and particularly to a secure risk-indicating online continuous learning system based on loss regularization. Background Art

[0002] In traditional practices, after machine learning algorithms are deployed to the application side, each company still continuously generates new data, but the current model has been solidified and cannot acquire new knowledge from the new data. Traditional secondary training based on Fine-Tuning requires collecting new training data, which is quite unrealistic in the scenario of data isolation in federated learning. Catastrophic forgetting: Continuous training with continuously generated new data will gradually cause the old knowledge learned before to be forgotten. If knowledge highly relevant to the task is forgotten, it will lead to a catastrophic decline in performance. Interference between participants: Due to differences in the number and types of devices of each participant, there is an inherent data heterogeneity, so there is a certain interference between the new data generated by different participants. Summary of the Invention

[0003] The purpose of the present invention is to provide a secure risk-indicating online continuous learning system based on loss regularization. Based on loss regularization, the old and new knowledge is split, the difference between the two is minimized, and the problem of catastrophic forgetting is prevented. Based on sample reproduction, some representative samples or knowledge bases are retained and trained together with the new data. Based on network structure adjustment, part of the network is used to learn the old tasks, and the new knowledge is stored in other parts of the network, so as to solve the problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides the following technical solutions: A secure risk-indicating online continuous learning system based on loss regularization, including a data platform, an application platform, a data capture module, a data identification module, a training module, a data reading module, a learning model, a loss function module, and a differential data module. Among them, the output end of the data platform is connected to the data capture module, the output end of the data capture module is connected to the data identification module, the output end of the data identification module is connected to the training module, the output end of the training module is connected to the data reading module, the output end of the data reading module is connected to the learning model, the output end of the learning model is connected to the loss function module, the output end of the loss function module is connected to the differential data module, the output end of the differential data module is connected to the application platform, and the output end of the application platform is connected to the data platform.

[0005] Further, the data platform includes a server, an old database, a data management module, a new database, and a data distribution module. The input ends of the old database and the new database are connected to the data distribution module, the input end of the data distribution module is connected to the data management module, and the data management module is connected to the application platform through the server.

[0006] Further, the server is used to monitor the data reading, copying, and transferring of the application platform in real time, and update the real-time data management module;

[0007] The old database is used to store old data that has not been trained;

[0008] The data management module is used to manage the old data and new data of the application platform, and control the data distribution module to distribute the obtained old data and new data to the old database and the new database;

[0009] The new database is used to store new data continuously generated after the machine learning algorithm, and is new knowledge after training;

[0010] The data distribution module is used to determine the data sent by the application platform, determine its training data, and separately synchronize and send the data to the old database and the new database from the old data.

[0011] Further, the data capture module is used to capture new data continuously generated after the machine learning algorithm is deployed to the application side, and use the logs of the new database and the old database itself to judge the changed data. When performing insert, update, or delete operations on the source table, the changed data can be extracted, and the changed data is saved in the change table of the database and provided to the data recognition module.

[0012] Further, the data recognition module is used to retrieve the changed data obtained by the data capture module, collect, clean, transform the changed data and send it to the communication and training module, and synchronously record the change amplitude of the changed data and real-time statistical information.

[0013] Further, the training module is used to input the changed data into the training machine learning model. While training, establish datasets after training and before training. The data in the datasets is used to train or teach the labeled data of the machine model or algorithm. Based on sample reproduction, retain some representative samples OR knowledge bases, train together with new data, and test the training data to provide fair evaluation data, and can obtain training data with fast high precision and high reuse rate;

[0014] The data reading module is used to extract data from the dataset and upload the data, perform data unpacking, post-cleaning, and data association, and automatically generate data shared with the learning model according to the standard.

[0015] Further, the learning model provides data interaction control and supervision to establish its model. While making full use of the information increment contained in the new time series, control the time series data volume, so as to control the computing cost and computer storage capacity, and the data of the training module runs in the learning model.

[0016] Furthermore, the loss function module in the learning model runs machine learning algorithms that more or less rely on the process of maximizing or minimizing the objective function, finding the minimized function, splitting the two pieces of data according to the minimized function, and sending them to the differential data module respectively, minimizing the difference between the two, and preventing the problem of catastrophic forgetting in the optimization direction.

[0017] Furthermore, the differential data module processing module receives and processes two-way data. By comparing with the input initial new data, it finds the losses in the two-way data in the training module and the learning model, updates the current benefit score for the said feature, and generates the usefulness in the accurate prediction result for the training feature vector. It corrects the training module and the learning model according to the loss.

[0018] Furthermore, the application platform outputs the learned data to the data platform, and transfers the data stored in the new database to the old database every set time, and continues to supplement new data into the new database.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. The safety warning online continuous learning system based on loss regularization proposed by the present invention adjusts the old database based on the network structure to store part of the old tasks of network learning. The new knowledge is stored in the new database in the network. The old database and the new database are established based on the network architecture, and each data feature point is encoded one by one. The ports of its devices are directly connected to the database, and the databases established by the new data and the old data can be used directly.

[0021] 2. The safety warning online continuous learning system based on loss regularization proposed by the present invention uses the training module to input variable data into the training machine learning model. While training, it establishes the data sets after and before training. The data in the data set is used to train or teach the labeled data of the machine model or algorithm. Based on sample reproduction, it retains some representative samples OR knowledge bases, trains them together with the new data, and the data for testing the training provides fair evaluation data, and can obtain training data with high speed, high accuracy and high reuse rate.

[0022] 3. The online continuous learning system for safety warning based on loss regularization proposed by the present invention. The loss function module in the machine learning algorithms running the learning model all more or less rely on the process of maximizing or minimizing the objective function. Find the minimized function, split the two data according to the minimized function, and send them to the differential data module respectively, minimize the difference between the two, prevent the problem of catastrophic forgetting in the optimization direction, based on loss regularization, split the old and new knowledge, minimize the difference between the two, prevent the problem of catastrophic forgetting, based on sample reproduction, retain some representative samples or knowledge bases, and train them together with new data. Based on network structure adjustment, use part of the network to learn the old tasks, and store the new knowledge in other parts of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is the overall module diagram of the present invention;

[0024] Figure 2 is the schematic diagram of the data platform principle of the present invention;

[0025] Figure 3 is the topology diagram of the data recognition module of the present invention;

[0026] Figure 4 is the topology diagram of the loss function module of the present invention;

[0027] Figure 5 is the correction schematic diagram of the present invention;

[0028] Figure 6 is the knowledge network architecture diagram of the present invention;

[0029] Figure 7 is another embodiment diagram of the present invention.

[0030] In the figure: 1. Data platform; 11. Server; 12. Old database; 13. Data management module; 14. New database; 15. Data distribution module; 2. Application platform; 3. Data capture module; 4. Data recognition module; 5. Training module; 6. Data reading module; 7. Learning model; 8. Loss function module; 9. Differential data module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] Embodiment 1:

[0033] Please refer to Figure 1 , a security risk warning online continuous learning system based on loss regularization, including a data platform 1, an application platform 2, a data capture module 3, a data recognition module 4, a training module 5, a data reading module 6, a learning model 7, a loss function module 8, and a differential data module 9. Among them, the output end of the data platform 1 is connected to the data capture module 3, the output end of the data capture module 3 is connected to the data recognition module 4, the output end of the data recognition module 4 is connected to the training module 5, the output end of the training module 5 is connected to the data reading module 6, the output end of the data reading module 6 is connected to the learning model 7, the output end of the learning model 7 is connected to the loss function module 8, the output end of the loss function module 8 is connected to the differential data module 9, the output end of the differential data module 9 is connected to the application platform 2, and the output end of the application platform 2 is connected to the data platform 1.

[0034] Please refer to Figure 2 , the data platform 1 includes a server 11, an old database 12, a data management module 13, a new database 14, and a data distribution module 15. The input ends of the old database 12 and the new database 14 are connected to the data distribution module 15, the input end of the data distribution module 15 is connected to the data management module 13, and the data management module 13 is connected to the application platform 2 through the server 11.

[0035] The server 11 is used to monitor the data reading, copying, and transferring of the application platform 2 in real time and update the real-time data management module 13.

[0036] The old database 12 is used to store old data that has not been trained.

[0037] The data management module 13 is used to manage the old data and new data of the application platform 2 and control the data distribution module 15 to distribute the obtained old data and new data to the old database 12 and the new database 14.

[0038] The new database 14 is used to store new data continuously generated after machine learning algorithms and new knowledge that has been trained.

[0039] The data distribution module 15 is used to determine the data sent by the application platform 2, determine its training data, and synchronously transport the data to the old database 12 and the new database 14 separately from the old data.

[0040] Based on the network structure, the old database 12 stores some old tasks of network learning, and new knowledge is stored in the new database 14 in the network. The old database 12 and the new database 14 are established based on the network architecture and encoded one by one for their data feature points. The ports of the devices are directly connected to the database, and the databases established for new data and old data can be used directly.

[0041] The data capture module 3 is used to capture new data continuously generated after the machine learning algorithm is deployed to the application side, and to judge the changed data based on the logs of the new database 14 and the old database 12. When performing insert (insertion), update (update), or delete (deletion) operations on the source table, the changed data can be extracted, and the changed data is saved in the change table of the database and provided to the data recognition module 4.

[0042] Please refer to Figure 3 , the data recognition module 4 is used to retrieve the changed data obtained by the data capture module 3, collect, clean, transform the changed data and send it to the communication and training module 5, and simultaneously record the change range of the changed data and statistically analyze the information in real time.

[0043] The training module 5 is used to input the changed data into the machine learning model for training. While training, datasets after training and before training are established. The data in the datasets is used to train or teach the labeled data of the machine model or algorithm. Based on sample reproduction, some representative samples or knowledge bases are retained and trained together with the new data. And the trained data is tested to provide fair evaluation data, and training data with high speed, high accuracy, and high reuse rate can be obtained.

[0044] The data reading module 6 is used to extract the data in the dataset and upload the data, perform data unpacking, post-cleaning, and data association, automatically generate the data shared with the learning model 7 according to the standard, and clean the variable data read by the data reading module 6 against the existing data.

[0045] The learning model 7 provides data manipulation and supervision to establish its model. While making full use of the information increment contained in the new time series, it controls the time series data volume, thereby controlling the computing cost and computer storage capacity, and the data of the training module 5 runs in the learning model 7.

[0046] Please refer to Figure 4 , the loss function module 8 in the machine learning algorithm running in the learning model 7 more or less depends on the process of maximizing or minimizing the objective function, finds the minimized function, splits the two data according to the minimized function, and sends them to the differential data module 9 respectively, minimizing the difference between the two, and preventing the problem of catastrophic forgetting in the optimization direction.

[0047] The differential data module 9 processing module receives and processes the two-way data. By comparing with the input initial new data, it finds the loss in the two-way data in the training module 5 and the learning model 7, updates the current benefit score for the feature, and generates the usefulness in the accurate prediction result for the training feature vector, and corrects the training module 5 and the learning model 7 according to the loss.

[0048] Please refer to Figure 5 , the application platform 2 outputs the learned data to the data platform 1, and transfers the data stored in the new database 14 to the old database 12 every set time, continues to supplement new data into the new database 14. The application platform 2 is controlled by an external instruction, decomposes the network parameters, modifies the loss function module of the neural network, performs regularization constraints, and based on the regularization method, avoids storing the original input, gives priority to privacy and reduces memory requirements. An additional regularization term is introduced in the loss function module 8 to consolidate the previous knowledge when learning new data, alleviates catastrophic forgetting by continuous training and updating the imposed constraints, and combines the difference generated by the loss function module 8, and continuously corrects the training module 5 and the learning model 7 with different degrees of plasticity cascade states, which can protect the consolidated knowledge from being forgotten.

[0049] Embodiment 2:

[0050] Please refer to Figure 6 , establish a task schedule, and establish serial numbers t - n... t - 3, t - 2, t - 1, t, t + 1, t + 2... t + n according to the time of the task. The learning mode established during the t time period is Fe, and the learned knowledge is L. A cycle is established between the learning mode Fe and the learned knowledge L. The learned knowledge L is input into the learning mode Fe through previous learning tasks. The learning mode Fe stores the newly established knowledge for future use and extracts its essence from the existing knowledge to keep learning continuously.

[0051] Embodiment 3:

[0052] Please refer to Figure 7 , the data platform 1, the application platform 2, the data capture module 3, the data recognition module 4, the training module 5, the data reading module 6, the learning model 7, the loss function module 8, and the differential data module 9. Among them, the output end of the data platform 1 is connected to the data capture module 3, the output end of the data capture module 3 is connected to the data recognition module 4, the output end of the data recognition module 4 is connected to the training module (5), the output end of the training module (5) is connected to the data reading module 6, the output end of the data reading module 6 is connected to the loss function module 8, the output end of the loss function module 8 is connected to the differential data module 9, the output end of the differential data module 9 is connected to the application platform 2, and the output end of the application platform 2 is connected to the data platform 1.

[0053] The data platform 1 includes a server 11, an old database 12, a data management module 13, a new database 14, and a data distribution module 15. The input ends of the old database 12 and the new database 14 are connected to the data distribution module 15. The input end of the data distribution module 15 is connected to the data management module 13. The data management module 13 is connected to the application platform 2 through the server 11.

[0054] The server 11 is used to monitor the data reading, copying, and transferring of the application platform 2 in real time, and update the real-time data management module 13.

[0055] The old database 12 is used to store the old data that has not been trained.

[0056] The data management module 13 is used to manage the old data and new data of the application platform 2, and control the data distribution module 15 to distribute the obtained old data and new data to the old database 12 and the new database 14.

[0057] The new database 14 is used to store the new data continuously generated after the machine learning algorithm, and the new knowledge after training.

[0058] The data distribution module 15 is used to determine the data sent by the application platform 2, determine its training data, and synchronously transport the data to the old database 12 and the new database 14 separately from the old data.

[0059] Based on the network structure, the old database 12 stores some old tasks of network learning, and the new knowledge is stored in the new database 14 in the network. The old database 12 and the new database 14 are established based on the network architecture, and each data feature point is encoded. The ports of the devices are directly connected to the database, and the databases established by the new data and the old data can be used directly.

[0060] The data capture module 3 is used to capture the new data continuously generated after the machine learning algorithm is deployed to the application side, and judge the changed data based on the logs of the new database 14 and the old database 12 itself. When performing insert (insert), update (update), or delete (delete) operations on the source table, the changed data can be extracted, and the changed data is saved in the change table of the database and provided to the data recognition module 4.

[0061] The data recognition module 4 is used to retrieve the changed data obtained by the data capture module 3, collect, clean, and transform the changed data and send it to the communication and training module 5, and synchronously record the change range of the changed data and statistically analyze the information in real time.

[0062] The training module 5 is used to input varying data into the machine learning model being trained. While training, it establishes datasets before and after training. The data in the datasets are labeled data used to train or teach the machine model or algorithm. Based on sample reproduction, it retains some representative samples OR knowledge bases, trains them together with new data, and tests the training data to provide fair evaluation data, enabling the acquisition of training data with high speed, high precision, and high reuse rate.

[0063] The data reading module 6 is used to extract data from the dataset and upload the data, perform data unpacking, post - cleaning, and data association, and automatically generate data shared with the learning model 7.

[0064] The loss function module 8: In the machine learning algorithms run by the learning model 7, more or less, they rely on the process of maximizing or minimizing the objective function. It finds the minimized function, splits the two data according to the minimized function, and sends them to the differential data module 9 respectively, minimizing the difference between them and preventing the problem of catastrophic forgetting in the optimization direction.

[0065] The differential data module 9: The processing module receives and processes two - way data. By comparing with the initially input new data, it finds the losses in the two - way data in the training module 5, updates the current gain score for the said feature, and generates the usefulness in the accurate prediction result for the training feature vector. It corrects the training module 5 and the learning model 7 according to the losses.

[0066] The application platform 2 outputs the learned data into the data platform 1, and transfers the data stored in the new database 14 to the old database 12 for storage at set intervals, and continues to supplement new data into the new database 14. The application platform 2 is controlled by external instructions, decomposes the network parameters, modifies the loss function module of the neural network, performs regularization constraints, and based on the regularization method, avoids storing the original input, gives priority to privacy and reduces memory requirements. An additional regularization term is introduced in the loss function module 8 to consolidate previous knowledge when learning new data, alleviates catastrophic forgetting by continuously training and updating to impose constraints, combines the differences generated by the loss function module 8, and continuously corrects the training module 5 with different degrees of plasticity cascade states, which can protect the consolidated knowledge from being forgotten.

[0067] Based on loss regularization, split the old and new knowledge, minimize the difference between the two, prevent the problem of catastrophic forgetting. Based on sample reproduction, retain some representative samples OR knowledge bases, train them together with new data. Based on network structure adjustment, use part of the network to learn the old tasks, and store new knowledge in other parts of the network.

[0068] As described above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An online continuous learning system for security risk warning based on loss regularization, characterized in that, It includes a data platform (1), an application platform (2), a data capture module (3), a data recognition module (4), a training module (5), a data reading module (6), a learning model (7), a loss function module (8), and a differential data module (9). Among them, the output end of the data platform (1) is connected to the data capture module (3), the output end of the data capture module (3) is connected to the data recognition module (4), the output end of the data recognition module (4) is connected to the training module (5), the output end of the training module (5) is connected to the data reading module (6), the output end of the data reading module (6) is connected to the learning model (7), the output end of the learning model (7) is connected to the loss function module (8), the output end of the loss function module (8) is connected to the differential data module (9), the output end of the differential data module (9) is connected to the application platform (2), and the output end of the application platform (2) is connected to the data platform (1); the data capture module (3) is used to capture new data continuously generated after the machine learning algorithm is deployed to the application side; Among them, the data platform (1) includes a server (11), an old database (12), a data management module (13), a new database (14), and a data distribution module (15). The input ends of the old database (12) and the new database (14) are connected to the data distribution module (15), the input end of the data distribution module (15) is connected to the data management module (13), and the data management module (13) is connected to the application platform (2) through the server (11); the old database (12) is used to store old data that has not been trained, and the new database (14) is used to store new data continuously generated after the machine learning algorithm and is new knowledge after being trained.

2. The online continuous learning system for security warning based on loss regularization according to claim 1, wherein The server (11) is used to monitor the data reading, copying, and transferring of the application platform (2) in real time and update the real-time data management module (13); The data management module (13) is used to manage the old data and new data of the application platform (2) and control the data distribution module (15) to distribute the obtained old data and new data to the old database (12) and the new database (14); The data distribution module (15) is used to determine the data sent by the application platform (2), determine its training data, and synchronously transport the data to the old database (12) and the new database (14) separately from the old data.

3. The online continuous learning system for safety risk indication based on loss regularization as claimed in claim 1, wherein The new database (14) and the old database (12) judge the changed data through their own logs, extract the changed data while performing insert, update, or delete operations on the source table, and the changed data is saved in the change table of the database and provided to the data recognition module (4).

4. The online continuous learning system for security risk warning based on loss regularization according to claim 1, wherein, The data recognition module (4) is used to retrieve the changed data obtained by the data capture module (3), collect, clean, and transform the changed data and send it to the communication to the training module (5), and synchronously record the change range of the changed data and real-time statistical information.

5. The online continuous learning system for safety risk warning based on loss regularization according to claim 1, wherein The training module (5) is used to input the change data into the machine learning model for training. While training, it establishes the datasets after and before training. The data in the datasets are labeled data used to train or teach the machine model or algorithm. Based on sample reproduction, it retains part of the samples or the knowledge base, trains them together with the new data, and tests the training data to provide evaluation data, thereby obtaining the training data; The data reading module (6) is used to extract the data in the dataset and upload the data, perform data unpacking, post - cleaning, and data association, and automatically generate the data shared with the standard and learning model (7).

6. The online continuous learning system for security risk warning based on loss regularization according to claim 1, characterized in that The learning model (7) provides data manipulation and supervision to establish its model. While making full use of the information increment contained in the newly added time series, it controls the time series data volume, thereby controlling the computing cost and the computer storage capacity. And the data of the training module (5) runs in the learning model (7).

7. The online continuous learning system for security risk indication based on loss regularization according to claim 1, characterized in that, The loss function module (8) in the machine learning algorithms running in the learning model (7) all rely on the process of minimizing the objective function. It finds the minimized function, splits the two data according to the minimized function, and sends them to the differential data module (9) respectively, minimizing the difference between the two, and preventing the problem of catastrophic forgetting in the optimization direction.

8. The online continuous learning system for security risk warning based on loss regularization according to claim 1, characterized in that, The differential data module (9) processing module receives and processes the two - way data. By comparing with the input initial new data, it finds the losses in the two - way data in the training module (5) and the learning model (7), updates the current gain score for the said feature, and generates the usefulness in the accurate prediction result for the training feature vector. It corrects the training module (5) and the learning model (7) according to the loss.

9. The online continuous learning system for security risk warning based on loss regularization according to claim 1, characterized in that, The application platform (2) outputs the learned data into the data platform (1), and transfers the data stored in the new database (14) to the old database (12) for storage at regular intervals, and continues to supplement new data into the new database (14).

Images