Method for combining intelligent computing center HMI time sequence optimization with large model data processing

Abstract

The invention discloses a method for combining intelligent computing center HMI time sequence optimization with large model data processing, which comprises the following steps: deploying a digital twin engine architecture in an intelligent computing module of an intelligent computing center, and performing log server stratified analysis and association relationship mining on data of a main engine and a sub-engine; an HMI time sequence optimization technology is adopted, a rule chain is formed according to implementation logic of task modes in the task pool, and task scheduling and reconfiguration based on the rule chain are achieved; adopting real-time federal learning to recombine the rule chains for multiple times, and taking the rule chain with the most appearing order after recombination as an optimal solution; and executing the rule chain, and adopting a ChatGPT size model as a main sub-engine for training. According to the method, dynamic parameters can be provided for training of a ChatGPT size model, the defect that tasks are executed in sequence in a digital twin engine architecture is overcome, and the problems that an intelligent calculation center is poor in availability and poor in expandability are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of large model application, and specifically relates to a method for combining intelligent computing center HMI timing optimization with large model data processing. Background Art

[0002] The intelligent computing platform can conveniently support the portrait analysis function based on the portrait data. The object of analysis can be a group of people or a single user. The data source of the analysis can be various label data that have been aggregated, or behavior detail data. The optimization of the intelligent computing platform involves all aspects, including data generation, function construction, and architecture design. In order to prevent the occurrence of various operational disasters, it is necessary to build an intelligent computing module including a digital twin engine. The most important thing is the deployment of the computing interaction layer, namely the intelligent computing module and the interactive driving module. The intelligent computing module can be divided into two parts: one that needs to rely on the big data framework and one that does not. The part that requires the big data software framework needs to deploy a dedicated server and a digital twin engine.

[0003] The existing intelligent computing platform has long execution times due to resource constraints, and the poor availability and scalability of the intelligent computing platform are particularly prominent. Poor user experience leads to inaccurate resource cost accounting. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for combining HMI timing optimization of an intelligent computing center with large model data processing in view of the deficiencies of the above-mentioned prior art.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0006] A method for combining HMI timing optimization of an intelligent computing center with large model data processing, comprising:

[0007] Deploy the digital twin engine architecture in the intelligent computing module of the intelligent computing center, and perform log server hierarchical analysis and correlation mining on the data of the main and sub-engines, thereby providing dynamic parameters for the training of ChatGPT large and small models;

[0008] Adopt HMI timing optimization technology to form a rule chain according to the implementation logic of the task mode in the task pool, so as to solve the drawbacks of sequential execution of tasks in the digital twin engine architecture through task scheduling and reconfiguration based on rule chains;

[0009] Real-time federated learning is used to reorganize the rule chain multiple times, and the rule chain with the most order after reorganization is taken as the optimal solution, overcoming the problems of poor availability and scalability of the intelligent computing center;

[0010] Execute rule chaining and use ChatGPT size model as the main engine for training.

[0011] To optimize the above technical solutions, the specific measures taken also include:

[0012] The above-mentioned digital twin engine architecture is deployed on different servers according to the application characteristics, and the digital twin engine architecture includes the digital twin engine data layer and the model layer; the deployment of the data layer includes the deployment of related data storage systems, and the data storage systems include Mysql, HBase, and HDFS. Mysql is used to implement relational data storage, HBase is used to implement non-relational data storage, and HDFS is used to implement file data storage; the deployment of the model layer includes a collection of specific algorithm models built according to specific needs.

[0013] The above-mentioned log server hierarchical analysis and correlation mining of the main engine data provide dynamic parameters for the training of ChatGPT size model, including:

[0014] First, the log server is used to perform size model operations, and performs the following two-layer analysis on the received data:

[0015] The first layer uses the log service to mine the application service topology based on the correlation between the database, middleware, server CPU, memory, disk, and process indicators in the database corresponding to the data link. The application service topology records the relationship between the database, middleware, server CPU, memory, disk, and process indicators between different servers.

[0016] At the second level, a data set based on the application service topology relationship corresponding to the data chain in the first level is mined from the log of each service. As the topological data set, it records the monitoring data and operating health status between the server and the database, middleware, server CPU, memory, disk and process in different time dimensions.

[0017] Then, the application service topology nodes associated with the database, middleware, server CPU, memory, disk and process are organized into sub-matrices, and multiple sub-matrices form a node matrix set, whose corresponding topology data set provides dynamic parameters for the training of ChatGPT size models.

[0018] The types of tasks in the above task pool include resource readiness detection tasks, crowd creation tasks, portrait calculation tasks, portrait analysis report output tasks, human rights data storage tasks and other crowd post-operation tasks.

[0019] In the above rule-chain-based task scheduling and reconfiguration process, computing resources and priorities are considered, and concurrency control is performed while monitoring the task status.

[0020] The above-mentioned real-time federated learning utilizes the training resources on the user terminal, and completes the reporting of the training results of the (N-1)th generation and the distribution of the global model and the configuration of the training parameters of the (N+1)th iteration during the training process of the Nth iteration of the user terminal; during the training process of the Nth iteration, the user terminal performs local training based on the global model downloaded from the FL server using the locally collected images / videos; then the terminal reports the intermediate training result of the Nth iteration - the gradient of the DNN to the FL server through the 5G line channel, and at the same time the FL server sends the global model and training parameter configuration of the (N+1)th iteration to the user terminal; when the FL server performs federated aggregation on the gradients collected in the Nth iteration, the terminal performs training for the (N+1)th iteration; the output of the federated aggregation is used for the updated global model of the next iteration training, and the model is distributed to the terminal together with the updated training configuration.

[0021] The above execution rule chain is trained using the ChatGPT size model as the main engine, including:

[0022] The keyword parameters of the rule chain trained by the ChatGPT large model are used as interval strings. Combined with the accuracy evaluation algorithm, the rule chain business data is trained for the first time through ChatGPT. The output results are marked with intervals in each interval string and the interval weight value is recorded. The interval marks with a weight value below 50% are removed to prevent synonyms or keywords with little meaning from interfering with the second training of the ChatGPT small model. The interval strings with a weight value above 50% are put into the ChatGPT small model as parameters for calculation, thereby completing the secondary training.

[0023] The above evaluation accuracy algorithm uses the LDAM loss function to weigh the sample interval of keywords using the output results of the first training of ChatGPT to achieve the optimal keyword retention, and then puts the filtered data matrix and the retained keyword parameters into the small model for calculation.

[0024] The present invention has the following beneficial effects:

[0025] The present invention deploys a digital twin engine architecture in the intelligent computing module of the intelligent computing center, and performs log server hierarchical analysis and association mining on the data of the master and sub-engines, providing dynamic parameters for the training of ChatGPT large and small models; adopts HMI timing optimization technology to form a rule chain according to the implementation logic of the task mode in the task pool, so as to solve the problem of sequential execution of tasks in the digital twin engine architecture through task scheduling and reconfiguration based on the rule chain; adopts real-time federated learning to reorganize the rule chain multiple times, and takes the rule chain with the most order after reorganization as the optimal solution, thereby overcoming the problems of poor availability and scalability of the intelligent computing center. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the logic principle diagram of the method in the present invention;

[0027] Figure 2 Form a rule chain diagram for the implementation logic of the task mode;

[0028] Figure 3 For real-time federated learning process. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. 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.

[0030] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0031] like Figure 1 As shown, a method of combining timing optimization of HMI in an intelligent computing center with large model data processing of the present invention includes:

[0032] 1. Deploy the digital twin engine architecture in the intelligent computing module of the intelligent computing center of the intelligent computing platform, and perform log server hierarchical analysis and correlation mining on the data of the master and sub-engines, so as to provide dynamic parameters for the training of ChatGPT large and small models; and adopt HMI timing optimization technology to form a rule chain according to the implementation logic of the task mode in the task pool, so as to solve the drawbacks of sequential execution of tasks in the digital twin engine architecture through task scheduling and reconfiguration based on the rule chain; adopt real-time federated learning to reorganize the rule chain multiple times, and take the rule chain with the most order after reorganization as the optimal solution, so as to overcome the problems of poor availability and scalability of the intelligent computing center;

[0033] In the embodiment, first, in order to prevent the occurrence of various operational disasters, it is necessary to deploy a digital twin engine architecture in the intelligent computing module, and match the log layered association relationships of the main and sub-engines of the intelligent computing engine module, so as to provide dynamic parameters for the training of large and small models.

[0034] Secondly, to solve the drawbacks of sequential execution, HMI (Human Machine Interface) technology is used to filter and verify the data that needs to be trained based on the application relationship of the human-machine interface and the custom rule events combined with the code library to establish a rule chain. Task scheduling and reconfiguration of the rule chain are performed.

[0035] Then, the real-time federated learning process is used to reorganize the rule chain multiple times, and the rule chain with the most order after reorganization is taken as the optimal solution. This ensures the flexible configuration and application of the functions of each link and overcomes the problems of poor availability and scalability of the intelligent computing platform.

[0036] The details are as follows:

[0037] (1) Deploy the digital twin engine architecture in the intelligent computing module of the intelligent computing center, and perform log server hierarchical analysis and correlation mining on the data of the main and sub-engines to provide dynamic parameters for the training of ChatGPT large and small models;

[0038] (1.1) Deploy the digital twin engine architecture and perform log-layered correlation matching on the main and sub-engines of the intelligent computing engine module, thereby providing dynamic parameters for the training of large and small models.

[0039] like Figure 1 The digital twin engine needs to be deployed on different servers according to the application characteristics.

[0040] The deployment of the digital twin engine data layer is mainly the deployment of related data storage systems. This patent selects three mature and representative data storage systems as the data layer of the digital twin engine. The relational data storage is implemented using MySQL, the non-relational data storage is implemented using HBase, and the file system is HDFS. The model layer is a collection of a series of specific algorithm models, which needs to be constructed according to specific needs.

[0041] (1.2) Dynamic parameters are formed by accessing log data based on the size model and performing correlation matching on the hierarchical data. The specific process is as follows:

[0042] First, the log server performs large and small model operations in some large systems that provide global services. The amount of daily log data may reach tens of terabytes (TBs). The received data is divided into two levels for analysis.

[0043] The first layer uses the log service to mine the application service topology based on the association between the database (database, middleware, and server basic monitoring indicators) in the corresponding data chain. The application service topology records the relationship between the three types of indicators (database, middleware, server CPU, memory, disk, and process) between different servers.

[0044] The second layer is to mine a data set based on the application service topology relationship corresponding to the data chain in the first layer from the log of each service, which is called the topology data set. These data record the monitoring data and operation health status between the server and the database, middleware, server CPU, memory, disk, and process in different time dimensions.

[0045] Then, the node matrix set is composed of sub-matrices that associate the current node with the nodes in the first layer (database, middleware, server CPU, memory, disk, process). Multiple sub-matrices form a node matrix set and update the rule chain. The matrix set is more conducive to model operation, and the data chain mark can quickly locate the call index mark of each link and related data.

[0046] (2) Using HMI timing optimization technology, a rule chain is formed according to the implementation logic of the task mode in the task pool to solve the drawbacks of sequential execution of tasks in the digital twin engine architecture through task scheduling and reconfiguration based on rule chains;

[0047] In the embodiment, the drawbacks of sequential execution are solved in the digital twin architecture, as follows:

[0048] HMI (Human Machine Interface) technology is used to filter and verify the data that needs to be trained from the application relationship of the human-machine interface and the custom rule events combined with the code library to establish a rule chain. Task scheduling and reconfiguration of the rule chain are performed. This ensures the flexible configuration and application of the functions of each link to overcome the problems of poor availability and scalability of the intelligent computing platform. The implementation logic of the task mode forms a rule chain such as Figure 2 shown.

[0049] (3) Real-time federated learning is used to reorganize the rule chain multiple times, and the rule chain with the most order after reorganization is taken as the optimal solution, overcoming the problems of poor availability and scalability of the intelligent computing center.

[0050] In the embodiment, the real-time federated learning process is used to reorganize the rule chain multiple times, and the rule chain with the most order after reorganization is taken as the optimal solution and passed, as follows:

[0051] In the Nth training iteration, the user terminal performs local training based on the global model downloaded from the FL server using the locally collected images / videos. The terminal then reports the intermediate training results of the Nth iteration (such as the gradient of the DNN) to the FL server through the 5G line channel. At the same time, the FL server sends the (N+1)th global model and training configuration to the user terminal. While the FL server aggregates the gradients collected in the Nth iteration, the terminal can perform training for the (N1)th iteration. Federated Aggregation outputs an updated global model for the next iteration of training, and distributes the model to the terminal together with the updated training configuration.

[0052] In order to make full use of the training resources on the user terminal and minimize the training delay, Figure 3 The training sequence shown requires that the training result report of the (N-1)th iteration and the global model distribution and training parameter configuration of the (N+1)th iteration be completed during the training process of the UE's Nth iteration. Of course, in the actual training process, it is also possible to consider using a relaxed sequence at the cost of sacrificing the convergence speed of the training, that is, performing non-real-time federated learning.

[0053] 2. Execute the rule chain and use the ChatGPT size model as the main engine for training.

[0054] In the embodiment, after the rule chain starts to execute, the rule chain keyword parameters trained by the large model are used as interval strings and combined with the accuracy evaluation algorithm to put in the rule chain business data through the large and small models as the main engine (ChatGPT) of the intelligent computing engine module for the first training. The output result is to mark each interval string with an interval weight value between 0-100, and remove the interval marks with a weight value below 50% to prevent synonyms or keywords with little meaning from interfering with the second training of the small model. Then the interval strings with a weight value of more than 50% are put into the small model as parameters for calculation, thereby completing the secondary training, which improves the query accuracy and avoids the problems of resource waste and poor timeliness.

[0055] Among them, the evaluation accuracy algorithm is:

[0056] Label distribution-aware margin loss: The LDAM function is an interval loss function that achieves the optimal margin trade-off between the head and tail categories by adjusting the interval between the category and the decision boundary.

[0057] The output result of the first training of ChatGPT is used to weigh the sample interval of keywords through the LDAM loss function to achieve the optimal keyword retention, and then the filtered data matrix and the retained keyword parameters are put into the small model for calculation.

[0058] First, for a training sample x, its corresponding label is y, then the interval from the sample to y is

[0059]

[0060] Among them, f(x)) is the probability prediction of sample x on each category sample, and j is the category closest to sample x except the correct label y. Then the interval corresponding to category y is the smallest sample interval in this category:

[0061]

[0062] Among them, S y is the sample set of category y. By optimizing the generalization error boundary of each category, the interval of category y in the multi-small model category scenario is obtained as

[0063]

[0064] Where C is a hyperparameter. According to the interval calculation formula, when n j When it is very small, the interval γ y The value of will be large, otherwise it will be small. This just meets the requirement that the interval is biased towards the tail category, and at the same time, the interval satisfies the optimal interval trade-off between categories.

[0065] This patent is represented by Hadoop. The part that relies on local algorithm calculations only needs to be deployed on the computing platform where the virtual model is located, and can be implemented using C# and Unity. However, both real scenes and virtual scenes have the problem of inaccurate regional positioning, mainly due to the large differences in computing power, hardware and data. In addition, there are few methods to use large models to measure the operation of portrait wide-table business data from a data-thinking perspective. Large models promote data storage and calls.

[0066] The names mentioned above are explained as follows:

[0067] The technology involved in the intelligent computing platform is not complicated. It mainly relies on existing big data technology to realize related functions. However, in order for the intelligent computing platform to realize its functions, it needs continuous polishing by the production and research team. Especially in large models, the accuracy requirements for the intelligent computing platform are more stringent.

[0068] ChatGPT (Chat Generative Pre-trained Transformer) is a chatbot program developed by OpenAI. ChatGPT is a natural language processing tool driven by artificial intelligence technology. It can generate answers based on the patterns and statistical laws seen in the pre-training stage, and can also interact according to the context of the chat, truly chatting and communicating like a human, and can even complete tasks such as writing papers, emails, scripts, copywriting, translation, and code.

[0069] Small models usually refer to machine learning or deep learning models with fewer parameters and shallower layers. They are lightweight, efficient, easy to deploy, and can run efficiently in resource-limited environments such as mobile devices or embedded systems.

[0070] Federated machine learning (Federated Machine Learning / Federated Learning), also known as federated learning, joint learning, and alliance learning, is a machine learning framework that can effectively help multiple institutions use data and conduct machine learning modeling while meeting the requirements of user privacy protection, data security, and government regulations.

[0071] HMI (Human Machine Interface) refers to the interface for interaction between people and machines. HMI configuration software is usually suitable for small, single-machine, simple industrial automation systems, such as machine tools, cold storage, etc. Common HMI configuration software includes EASYBuilder, TouchWin, Pro-face, etc. These software have the characteristics of simple use, flexible configuration, and beautiful interface, but their functions are not as comprehensive as SCADA.

[0072] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0073] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for combining HMI timing optimization with large model data processing in an intelligent computing center, characterized in that: include: Deploy the digital twin engine architecture in the intelligent computing module of the intelligent computing center, and perform log server hierarchical analysis and correlation mining on the data of the main and sub-engines, thereby providing dynamic parameters for the training of ChatGPT large and small models; Adopt HMI timing optimization technology to form a rule chain according to the implementation logic of the task mode in the task pool, so as to solve the drawbacks of sequential execution of tasks in the digital twin engine architecture through task scheduling and reconfiguration based on rule chains; Real-time federated learning is used to reorganize the rule chain multiple times, and the rule chain with the most order after reorganization is taken as the optimal solution, overcoming the problems of poor availability and scalability of the intelligent computing center; Execute rule chaining and use ChatGPT size model as the main engine for training.

2. The method of combining timing optimization of HMI in intelligent computing center with large model data processing according to claim 1 is characterized in that: The digital twin engine architecture is deployed on different servers according to application characteristics, and the digital twin engine architecture includes a digital twin engine data layer and a model layer; the deployment of the data layer includes the deployment of related data storage systems, and the data storage systems include Mysql, HBase, and HDFS. Mysql is used to implement relational data storage, HBase is used to implement non-relational data storage, and HDFS is used to implement file data storage; the deployment of the model layer includes a collection of specific algorithm models built according to specific needs.

3. The method of combining timing optimization of HMI in intelligent computing center with large model data processing according to claim 1 is characterized in that: The log server hierarchical analysis and association mining of the data of the main engine are performed to provide dynamic parameters for the training of the ChatGPT size model, including: First, the log server is used to perform size model operations, and performs the following two-layer analysis on the received data: The first layer uses the log service to mine the application service topology based on the correlation between the database, middleware, server CPU, memory, disk, and process indicators in the database corresponding to the data link. The application service topology records the relationship between the database, middleware, server CPU, memory, disk, and process indicators between different servers. At the second level, a data set based on the application service topology relationship corresponding to the data chain in the first level is mined from the log of each service. As the topological data set, it records the monitoring data and operating health status between the server and the database, middleware, server CPU, memory, disk and process in different time dimensions. Then, the application service topology nodes associated with the database, middleware, server CPU, memory, disk and process are organized into sub-matrices, and multiple sub-matrices form a node matrix set, whose corresponding topology data set provides dynamic parameters for the training of ChatGPT size models.

4. The method of combining timing optimization of HMI in intelligent computing center with large model data processing according to claim 1 is characterized in that: The types of tasks in the task pool include resource readiness detection tasks, crowd creation tasks, portrait calculation tasks, portrait analysis report output tasks, human rights data storage tasks, and other crowd post-operation tasks.

5. The method of combining timing optimization of HMI in intelligent computing center with large model data processing according to claim 1 is characterized in that: During the rule-chain-based task scheduling and reconfiguration process, computing resources and priorities are considered, and task status is monitored while concurrent control is performed.

6. The method of combining timing optimization of HMI in intelligent computing center with large model data processing according to claim 1 is characterized in that: The real-time federated learning utilizes the training resources on the user terminal, and completes the reporting of the training results of the (N-1)th iteration and the distribution of the global model and the configuration of the training parameters of the (N+1)th iteration during the training process of the Nth iteration of the user terminal; during the training process of the Nth iteration, the user terminal performs local training based on the global model downloaded from the FL server using the locally collected images / videos; then the terminal reports the intermediate training result of the Nth iteration - the gradient of the DNN to the FL server through the 5G line channel, and at the same time the FL server sends the global model and training parameter configuration of the (N+1)th iteration to the user terminal; When the FL server performs federated aggregation on the gradients collected in the Nth iteration, the terminal performs training for the (N+1)th iteration; the output of the federated aggregation is used for an updated global model for the next iteration training, and the model is distributed to the terminal together with the updated training configuration.

7. The method of combining timing optimization of HMI in intelligent computing center with large model data processing according to claim 1 is characterized in that: The execution rule chain is trained using the ChatGPT size model as the main engine, including: The keyword parameters of the rule chain trained by the ChatGPT large model are used as interval strings. Combined with the accuracy evaluation algorithm, the rule chain business data is trained for the first time through ChatGPT. The output results are marked with intervals in each interval string and the interval weight value is recorded. The interval marks with a weight value below 50% are removed to prevent synonyms or keywords with little meaning from interfering with the second training of the ChatGPT small model. The interval strings with a weight value above 50% are put into the ChatGPT small model as parameters for calculation, thereby completing the secondary training.

8. The method of combining timing optimization of HMI in intelligent computing center with large model data processing according to claim 7 is characterized in that: The evaluation accuracy algorithm uses the LDAM loss function to weigh the sample interval of keywords using the output results of the first training of ChatGPT to achieve the optimal keyword retention, and then puts the filtered data matrix and the retained keyword parameters into the small model for calculation.

Images