An enterprise intelligence generation platform based on multi-model fusion technology

Abstract

The present invention provides an enterprise intelligence generation platform based on multi-model fusion technology, including a model layer for deploying and running enterprise master data models, proprietary models, open source models and business models; P Different models are quantitatively evaluated. The distribution layer selects the optimal model and assigns tasks based on task characteristics and model performance. A model selection formula is introduced to calculate the scores of each model, and the model with the highest score is selected for task processing. The platform layer performs business processing and data storage based on task requests. During data processing, task requests from the distribution layer are received through an application server cluster and business data is stored in a database server. The portal layer includes a web server cluster, CDN acceleration nodes, and security equipment. By integrating multiple large language model resources, this invention achieves dynamic model scheduling, capability combination, and resource optimization, improving the overall performance and adaptability of the system.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence, large language models, and enterprise information management technology, and in particular to an enterprise intelligence generation platform based on multi-model fusion technology. Background Art

[0002] In today's digital age, artificial intelligence (AI) technology is developing at an unprecedented pace, profoundly transforming the operations of various industries. As key players in economic activity, businesses are facing an increasingly urgent need for intelligent knowledge management, data analytics, and automated workflows. Intelligent knowledge management can help companies better integrate, store, and utilize internal knowledge resources, improving employee productivity and decision-making quality. Data analytics can unlock the potential value of enterprise data, providing strong support for strategic planning and business development. Automated workflows can reduce manual intervention, improve the efficiency and accuracy of business processes, and reduce operating costs.

[0003] However, the enterprise intelligence solutions currently available on the market face many challenges in practical applications, which seriously hinder the process of enterprise digital transformation and reduce the application value of artificial intelligence technology in enterprises.

[0004] Most existing enterprise intelligent systems rely heavily on a single, large language model. Different business scenarios have unique characteristics and requirements. For example, in customer service scenarios, the model must possess excellent language understanding and generation capabilities to accurately answer customer questions; in financial risk assessment scenarios, the model must possess strong data analysis and logical reasoning capabilities to accurately assess risk. Due to the limitations of its structure and training data, a single, large language model cannot simultaneously meet these diverse requirements. When faced with complex or domain-specific business scenarios, the model's capabilities are often limited, making it unable to provide accurate and effective solutions. This forces enterprises to seek alternative solutions for different business scenarios, increasing operational costs and management difficulties.

[0005] Existing enterprise intelligence systems typically deploy independent model resources for each application scenario. While this approach can meet the needs of specific scenarios to a certain extent, it leads to significant waste of computing resources and complex management. Different application scenarios may have similar or overlapping functional requirements. For example, multiple business departments may need to perform text classification tasks. However, due to the use of independent model resources, each department must deploy and maintain the corresponding model separately, resulting in duplicate computing resources. Furthermore, as enterprise application scenarios continue to expand, the number of model resources will also increase dramatically, posing significant challenges to model management, updates, and maintenance, increasing operational costs and technical difficulties for enterprises.

[0006] In actual business operations, a wide variety of data types are involved, including multimodal data such as text, images, and tables. However, existing enterprise intelligent systems have limited capabilities in processing this multimodal data. For example, in the medical field, doctors need to simultaneously review a patient's medical records, medical images (such as X-rays and CT images), and examination report forms to diagnose a disease. However, existing intelligent diagnostic systems may only be able to process one or two of these data types and cannot provide an integrated solution. This forces doctors to rely on multiple different systems to obtain and analyze information, increasing the difficulty and time cost of diagnosis, and also affecting the accuracy and efficiency of diagnosis.

[0007] As enterprise IT environments diversify, so do their needs for system deployment methods. Some companies may prefer to deploy their systems in the cloud to benefit from the elastic computing resources and convenient management services provided by cloud service providers. Others, driven by data security and compliance concerns, may prefer to deploy their systems on local servers. Still others prefer a hybrid deployment approach, deploying some services in the cloud and others on-premises. However, most current enterprise intelligent systems struggle to simultaneously meet the requirements of cloud, on-premises, and hybrid deployments, failing to adapt to diverse enterprise IT environments, limiting their application and adoption.

[0008] In summary, the current enterprise intelligence solutions on the market have many problems and cannot meet the growing demand for intelligence in enterprises. Therefore, developing an enterprise intelligence solution that can overcome these problems and possess multiple advantages has important practical significance and application value. Summary of the Invention

[0009] In response to the shortcomings of the existing technology, the present invention provides an enterprise intelligent generation platform based on multi-model fusion technology. By integrating multiple large language model resources, it realizes dynamic model scheduling, capability combination and resource optimization, and improves the overall performance and adaptability of the system.

[0010] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0011] An enterprise intelligence generation platform based on multi-model fusion technology, including:

[0012] The model layer is used to deploy and run enterprise master data models, proprietary models, open source models, and commercial models. During the model training process, computing server clusters are used for large-scale data training, and the training data is stored in storage devices. The model performance evaluation indicators are used to evaluate the performance of the model. P = α × A + β × R- c × CQuantitative evaluation of different models; P is the comprehensive performance score of the model, A is the model accuracy, R is the model recall rate, C Calculate the complexity of the model, α 、 β 、 c is the weight coefficient;

[0013] The distribution layer selects the optimal model and assigns tasks based on task characteristics and model performance; in the model selection process, the model selection formula is introduced S = oh 1× F ( t )+ oh 2× U ( t )1+ oh 3× P Calculate the scores of each model and select the model with the highest score to process the task; S Select a score for the model, F ( t ) is the task feature matching function, U ( t ) is the resource utilization function at the current moment, P is the comprehensive performance score of the model calculated at the model layer, oh 1. oh 2. oh 3 is the weight coefficient;

[0014] The platform layer performs business processing and data storage based on task requests. During data processing, it receives task requests from the distribution layer through the application server cluster and stores business data in the database server.

[0015] The portal layer includes a Web server cluster, CDN acceleration nodes, and security devices. The Web server cluster establishes a connection with the CDN acceleration node, caches static resources on the CDN acceleration node, and authenticates and authorizes user access through security devices.

[0016] According to the present invention, an enterprise intelligence generation platform based on multi-model fusion technology is provided. The distribution layer includes a load balancing device, an API gateway server, a task scheduler, a resource management module, and a network monitoring device. After receiving a request, the API gateway server extracts traffic feature information and shares it with the load balancing device in real time. The load balancing device dynamically adjusts the load balancing strategy based on this feature information. The load balancing device uses historical load data and real-time traffic trends, employing time series analysis or machine learning algorithms to predict future load changes. The load prediction results are shared with the task scheduler, which uses the load prediction information to reasonably allocate tasks in advance.

[0017] Before the task scheduler distributes tasks to the model, it accurately predicts the resources required for the task based on the task characteristics and the model's historical processing performance data. The resource demand prediction results are fed back to the resource management module, which then reserves or adjusts resources in advance based on the prediction results.

[0018] During task execution, the task scheduler monitors the execution status and resource usage of tasks in real time. If a task's actual resource usage deviates significantly from its predicted value, or if a task's execution efficiency is low due to insufficient resources, the task scheduler triggers the resource management module to dynamically adjust resources. Based on the task scheduler's trigger signals and in conjunction with Kubernetes HPA mechanisms, the resource management module quickly adjusts resource allocation for the model instance.

[0019] According to the present invention, an enterprise intelligence generation platform based on multi-model fusion technology is provided. The model layer includes a computing server cluster, a storage device and a network switch. The computing server cluster is internally connected through a high-speed network to achieve high-speed data transmission and parallel computing collaboration between servers; the computing server cluster is connected to the storage device through a high-speed fiber channel or network file system protocol. During the model training process, the training data is read from the storage device, and the intermediate results and final model parameters of the training process are stored back to the storage device; wherein, the storage device adopts a distributed storage architecture, which is composed of multiple storage nodes, and the nodes are connected through a network for data synchronization and backup; the network switch divides different devices into different network areas by configuring switch ports and virtual local area networks.

[0020] According to the enterprise intelligence generation platform based on multi-model fusion technology provided by the present invention, the deployment and operation of the enterprise master data model, proprietary model, open source model and business model specifically include:

[0021] Enterprise master data models, proprietary models, open source models, and business models are classified and stored according to their application scenarios, data dependencies, and computational complexity, including:

[0022] Abstract the application scenario into different business dimensions and assign a weight to each business dimension w si ,in, i Indicates the business dimension number, ∑ i w si =1, and score the applicability of each model in each business dimension. s mj,si , j Indicates the model number, 0≤ s mj,si ≤1, then the model m j Comprehensive score of application scenarios S mj The calculation formula is: S mj =∑ i w si × s mj,si ;

[0023] Define the dependency strength coefficient between the model and different data sources d mj,Dk , k Indicates the data source number, 0≤ d mj,Dk ≤1, and the priority weight of the data source w Dk ,∑ k w Dk =1, model m j Data dependency composite score D mj The calculation formula is: D mj =∑ k w Dk × d mj,Dk ;

[0024] According to the number of calculation steps of the model n mj , data size v mj And the time complexity of the algorithm O mj , quantifying computational complexity into a comprehensive indicator C mj , expressed as:

[0025] Cmj = α × n mj + β × v mj + c × f ( O mj )

[0026] in α 、 β 、 c is the weight coefficient ( α + β + c =1), f ( O mj ) is a numerical value converted according to the time complexity of the algorithm.

[0027] According to the enterprise intelligence generation platform based on multi-model fusion technology provided by the present invention, the classification thresholds of application scenarios, data dependencies and computational complexity are set. TS 、 TD 、 TC , the models are classified according to the following rules:

[0028] Business Owner Data Model: If S mj ≥ TS and D mj If the sum of the dependency coefficients on the main data source within the enterprise is greater than a certain ratio, the model mj Classified as Enterprise Master Data Model;

[0029] Own model: If S mj High, that is, close to but not reaching the threshold of the enterprise master data model, and D mj Relying on the company's own internal data, C mj If it is within a certain range, it is classified as a proprietary model;

[0030] Open source model: If S mj At an average level, that is, lower than the typical application scenario scores of enterprise master data models and proprietary models. D mj The sum of the dependency intensity coefficients with open source data sources is greater than a certain ratio, and C mj If it is a low value, it is classified as an open source model;

[0031] Business model: IfS mj High and closely related to specific business scenarios, Dmj Relying on data from commercial data providers, while C mj If the value is high, it is classified as a business model.

[0032] According to the present invention, an enterprise intelligent generation platform based on multi-model fusion technology is provided. The enterprise master data model is built around the enterprise's core business data and stored in a specific data-related storage area. This area establishes a high-speed data channel with the enterprise's core business database in a large-capacity storage device; the proprietary model is developed based on the enterprise's own business needs and stored in a customized storage partition. A special model version management directory is set up in the partition to record change information of different versions of the model; the open source model is stored in the open source model resource library, which has an automatic update mechanism, regularly obtains the latest model version from the open source community and performs compatibility testing; the business model is stored in a securely isolated business model storage area, which uses encrypted storage technology to ensure the confidentiality of the business model.

[0033] According to the present invention, an enterprise intelligence generation platform based on multi-model fusion technology uses containerization technology to encapsulate and deploy various models, creating an independent container image for each model, specifically including:

[0034] Set the compute server node set to N ={ n 1, n 2,..., nl}, l Indicates the number of nodes, each node nl Available resources include the number of CPU cores CPU nl available , memory size MEM nl available , number of GPUs GPU nl available , storage space STORAGE nl available and network bandwidth BANDWIDTH nl available ;

[0035] For the model m j , calculate its relationship with each node n l Resource matching MATCH mj,nl , the formula is:

[0036] MATCH mj,nl = oh 1×min(1, CPU nl available / CPU mj )+ oh 2×min(1, MEM nl available / MEM mj )+ oh 3×min(1, GPU nl available / GPU mj )+ oh 4×min(1, STORAGE nl available / STORAGE mj )+ oh 5×min(1, BANDWIDTH nl available / BANDWIDTH mj );

[0037] in oh 1. oh 2. oh 3. oh 4. oh 5 is the weight coefficient, oh 1+ oh 2+ oh 3+ oh 4+ oh 5=1, used to adjust the importance of different resources in the matching calculation;

[0038] Then, select the resource matching MATCH mj,nl The largest computing server node n best Deploy the container image, namely: n best =argmax nl∈N MATCH mj,nl , argmax is the optimization function, that is, it returns the input variable that makes the objective function achieve the maximum value, nl∈N It means traversing all nodes nl Collection N .

[0039] According to the present invention, an enterprise intelligence generation platform based on multi-model fusion technology uses containerization technology to encapsulate and deploy various models, creating an independent container image for each model. The image contains at least all elements required for the model code, dependent libraries, and configuration files to run. During the deployment process, the container image is deployed to the corresponding computing server node based on the model classification and resource pre-allocation results. In particular, for scenarios where multiple models need to work together, clustered deployment of models is achieved through container orchestration tools to ensure efficient communication and collaborative computing between models.

[0040] Among them, when it is detected that the resource utilization of the server node where a model is located is too high or a fault occurs, the dynamic migration mechanism is triggered, and the container instance of the model is migrated from the current node to a node with sufficient resources through the container orchestration tool. Incremental synchronization technology is used during the migration process to only transmit the model's operating status and data changes, reducing migration time and impact on the business.

[0041] According to the enterprise intelligence generation platform based on multi-model fusion technology provided by the present invention, a task feature analysis and intelligent distribution module is deployed at the front end of the computing server cluster. This module first extracts multi-dimensional features of the input tasks. Based on the extracted features, it uses a machine learning algorithm to build a task-server mapping model. This model comprehensively considers the current load of each computing server, its historical processing performance, and its compatibility with the task, and intelligently distributes the tasks to the computing servers.

[0042] A parallel processing framework based on a computing server cluster is built. Complex, decomposable tasks are split into multiple subtasks and assigned to different computing servers for simultaneous processing. A dynamic load balancing mechanism is introduced to monitor the task processing progress and resource usage of each server in real time. When a server is found to be overloaded while other servers are idle, some tasks are automatically migrated from the high-load server to the low-load server.

[0043] The enterprise intelligence generation platform based on multi-model fusion technology provided by the present invention designs a multi-model data fusion storage strategy for different types of data generated by the enterprise master data model, proprietary models, open source models, and business models. Structured data is stored in a high-performance relational database and unstructured data is stored in a distributed file system. At the same time, a data association mapping table is established to record the association relationships between data from different models, thereby realizing the fusion storage and unified management of multi-model data.

[0044] An intelligent traffic scheduling algorithm is implemented on the network switch. The distribution of network traffic is dynamically adjusted according to the data traffic characteristics and task requirements between different computing servers in the model layer. By real-time monitoring of the bandwidth utilization, latency, and packet loss rate indicators of the network links, network traffic is preferentially allocated to links with good performance, taking into account the priority and real-time requirements of the tasks. For the traffic generated by model inference tasks with high real-time requirements, links with low latency are selected for transmission. For large-volume model training data transmission tasks, load balancing is used to distribute them to multiple links.

[0045] According to the enterprise intelligence generation platform based on multi-model fusion technology provided by the present invention, a multi-level task feature extraction module is built in the API gateway server to perform a comprehensive analysis of input tasks, including:

[0046] For text tasks, natural language processing technology is used to extract features including at least semantic depth, sentiment tendency, and domain relevance;

[0047] For image tasks, convolutional neural networks are used to extract at least the texture features, shape features, and scene features of the image;

[0048] For speech tasks, we use Mel-frequency cepstral coefficients (MFCCs) to extract the spectral features of speech, and combine them with speech recognition technology to obtain the speech content and further analyze the speech features.

[0049] For cross-modal tasks, a multimodal fusion model is used to extract correlation features between different modalities.

[0050] According to the enterprise intelligent generation platform based on multi-model fusion technology provided by the present invention, the web server cluster is internally connected through a network to achieve load balancing and failover between servers, and user requests are evenly distributed to multiple web servers through a reverse proxy server. When a user accesses a portal website, static resources are preferentially obtained from the CDN node closest to the user. The web server cluster communicates with the application server cluster at the platform layer through the HTTP / HTTPS protocol to obtain dynamic data and render the page back to the user.

[0051] CDN acceleration nodes are distributed at the edge of the network in different geographical locations and connected to the portal layer's web server cluster via a high-speed network. When a user initiates a request, the CDN node selects the node closest to the user based on the user's geographical location and network conditions to provide services and return the cached static resources to the user.

[0052] Among them, the security device authenticates and authorizes user access, adopts the role-based access control RBAC model, and the role authority calculation formula is: Ruser =U r∈Roles(user) Rr ,in R user A set of user permissions. Roles ( user ) is the set of roles to which the user belongs, Rr For the role r The set of permissions owned by U r It is a union operation on the role permission set, used to aggregate the permissions of all roles to which the user belongs.

[0053] This shows that compared to existing technologies, the model layer, distribution layer, platform layer, and portal layer of the present invention achieve high-speed data connections through network switches, enabling data transmission and interaction between the layers. During platform operation, each layer works closely together. The model layer provides the distribution layer with a variety of model options. The distribution layer selects the optimal model and assigns tasks based on task characteristics and model performance. The platform layer performs business processing and data storage based on task requests. The portal layer provides users with a friendly interactive interface and permission control, thus achieving efficient operation and intelligent application of the enterprise intelligence generation platform based on multi-model fusion technology.

[0054] Furthermore, the platform deploys and runs enterprise master data models, proprietary models, open source models and business models, fully leveraging the advantages of different types of models. The enterprise master data model can deeply fit the enterprise's own business logic and data characteristics, the proprietary model can be customized and developed according to the specific needs of the enterprise, the open source model can quickly acquire advanced technologies with the help of the open source community, and the business model usually has mature and stable performance. The integration of multiple models enables the platform to cope with various complex and changing enterprise business scenarios and provide more comprehensive and accurate solutions.

[0055] Furthermore, during the model training process, the model performance evaluation index is used P Quantitatively evaluate different models by comprehensively considering the model accuracy A , recall rate R and computational complexity C , and assign corresponding weight coefficients α 、 β 、 c It can comprehensively and objectively measure the comprehensive performance of the model, which not only helps to timely discover the advantages and disadvantages of the model during the model training process, but also provides a clear direction for model optimization and adjustment, improves the efficiency and quality of model training, and ensures that the model finally put into use has a high performance level.

[0056] Furthermore, in the model selection process, the model selection formula is introduced to calculate the scores of each model, and the model with the highest score is selected for task processing, taking into account the task feature matching function.F ( t ), current moment resource utilization function U ( t ) and the comprehensive performance score of the model calculated by the model layer P , and assign corresponding weight coefficients. In this way, the distribution layer can accurately select the most suitable model to handle tasks based on the specific characteristics of the tasks and the current system resource status, thereby effectively improving the quality and efficiency of task processing and ensuring that business needs are accurately met.

[0057] Furthermore, the resource utilization function in the model selection formula U ( t ) enables the distribution layer to perceive the system's resource usage in real time and dynamically adjust task allocation strategies based on resource availability. When system resources are limited, models with lower resource usage and better performance are prioritized to handle tasks, avoiding system performance degradation caused by excessive resource concentration. When system resources are sufficient, more resources can be appropriately allocated to complex tasks to speed up task processing, thereby fully utilizing system resources, improving overall system performance and stability, and reducing the company's operating costs.

[0058] Furthermore, the platform layer receives task requests from the distribution layer through the application server cluster, which has high availability and fault tolerance. When some application servers fail, other servers can automatically take over task processing to ensure the continuity of business processing, avoid business interruptions caused by single point failures, and provide enterprises with a stable and reliable business processing environment.

[0059] Furthermore, the web server cluster in the portal layer establishes connections with CDN acceleration nodes, caching static resources there. When users access the platform, the CDN acceleration nodes quickly deliver static resources based on their geographic location and network conditions, significantly reducing data transmission latency and improving user access speed and experience. Regardless of their location, users can quickly load platform pages and access the information they need, enhancing their satisfaction and loyalty to the platform.

[0060] Furthermore, security devices authenticate and authorize user access, effectively preventing unauthorized users from accessing the platform and protecting sensitive enterprise data and business information from being leaked. Security devices employ a variety of security technologies, such as firewalls, intrusion detection systems, and encryption, to provide comprehensive user protection, ensuring the secure and stable operation of the platform and reducing security risks faced by enterprises.

[0061] In summary, through the collaborative work and innovative design of various layers, the present invention has achieved remarkable beneficial effects in model selection and training, task allocation, business processing and data storage, user experience and system security. It can provide enterprises with efficient, intelligent and secure intelligent generation services, helping enterprises to achieve better development in the digital age.

[0062] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a schematic diagram of an embodiment of an enterprise intelligent generation platform based on multi-model fusion technology of the present invention.

[0064] Figure 2 This is a flowchart of deploying and running enterprise master data models, proprietary models, open source models and business models in an embodiment of an enterprise intelligent generation platform based on multi-model fusion technology of the present invention.

[0065] Figure 3 This is a flowchart of an embodiment of an enterprise intelligent generation platform based on multi-model fusion technology in the present invention, which uses containerization technology to encapsulate and deploy various models.

[0066] Figure 4 This is a schematic diagram of a multi-level task feature extraction module constructed in an API gateway server in an embodiment of an enterprise intelligent generation platform based on multi-model fusion technology of the present invention. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0068] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0069] See also Figure 1 This embodiment provides an enterprise intelligence generation platform based on multi-model fusion technology, including:

[0070] The model layer is used to deploy and run enterprise master data models, proprietary models, open source models, and commercial models. During the model training process, computing server clusters are used for large-scale data training, and the training data is stored in storage devices. The model performance evaluation indicators are used to evaluate the performance of the model. P = α × A + β × R- c × C Quantitative evaluation of different models; P is the comprehensive performance score of the model, A is the model accuracy, R is the model recall rate, C Calculate the complexity of the model, α 、 β 、 c is the weight coefficient.

[0071] The distribution layer selects the optimal model and assigns tasks based on task characteristics and model performance; in the model selection process, the model selection formula is introduced S = oh 1× F ( t )+ oh 2× U ( t )1+ oh 3× P Calculate the scores of each model and select the model with the highest score to process the task; S Select a score for the model, F ( t ) is the task feature matching function, U ( t ) is the resource utilization function at the current moment, P is the comprehensive performance score of the model calculated at the model layer, oh 1. oh 2. oh 3 is the weight coefficient.

[0072] The platform layer performs business processing and data storage based on task requests. During the data processing process, it receives task requests from the distribution layer through the application server cluster and stores the business data in the database server.

[0073] The portal layer includes a Web server cluster, CDN acceleration nodes, and security devices. The Web server cluster establishes a connection with the CDN acceleration node, caches static resources on the CDN acceleration node, and authenticates and authorizes user access through security devices.

[0074] As can be seen, in this embodiment, the model, distribution, platform, and portal layers are connected via high-speed network switches, enabling data transmission and interaction between the layers. During platform operation, each layer works closely together. The model layer provides the distribution layer with a variety of models to choose from. The distribution layer selects the optimal model and assigns tasks based on task characteristics and model performance. The platform layer performs business processing and data storage based on task requests. The portal layer provides users with a user-friendly interactive interface and permission control, thus achieving efficient operation and intelligent application of the enterprise intelligence generation platform based on multi-model fusion technology.

[0075] Furthermore, the platform layer also includes building capabilities, operating capabilities and underlying support to realize knowledge base construction, database connection, plug-in generation, intelligent question and answer, workflow orchestration, data dashboard and account and enterprise management functions.

[0076] Among them, the platform layer's building capabilities include: RAG knowledge base module, which is used for unstructured data material management, material vectorization and index establishment, hybrid retrieval strategy implementation and intelligent sorting of retrieval results; DataBase database module, which is used for built-in database support, remote database direct connection capabilities, natural language to SQL conversion and data visualization display; Tools plug-in generation unit: built-in image generation tools, image recognition, AI search, content condensation and other functions, and provides a custom plug-in framework to meet the personalized needs of enterprises.

[0077] Among them, the operational capabilities of the platform layer include: access management module, which is used for client configuration, dialogue management and customer management; data dashboard module, which is used for RAG search monitoring, usage statistics, transfer to manual monitoring and high-frequency question statistics.

[0078] Among them, the underlying support of the platform layer includes: LLMs Platform module, which is used for large language model aggregation and multimodal support; RAG Engine module, which is used for question processing and hybrid retrieval engine; WorkFlow Engine module, which is used for visual process engine.

[0079] Furthermore, the portal layer, as a user interaction interface, is also used to publish platform functions by department and functional classification, supports the creation of dialogue applications, API services, dialogue workflows, and form workflows, and provides official creation of organizations and individual creation permission control of members to meet the diverse usage needs of enterprises.

[0080] In this embodiment, the distribution layer includes a load balancing device (such as an F5 or Nginx load balancer), an API gateway server, a task scheduler, a resource management module, and a network monitoring device. Upon receiving a request, the API gateway server not only performs basic routing and forwarding but also conducts in-depth analysis of the request traffic, extracting traffic characteristics, including but not limited to the regional distribution of request sources, request time distribution patterns, and the proportion of different request types. This extracted traffic characteristics are shared in real time with the load balancing device, which dynamically adjusts the load balancing strategy based on this information. For example, if a region experiences a sharp increase in request volume within a specific time period, the load balancing device can proactively direct requests from that region to a backend server cluster with more sufficient resources, thereby avoiding local overload.

[0081] The API gateway server has a built-in security detection module that performs real-time security checks on requests, identifying potential security threats such as SQL injection, cross-site scripting (XSS), and DDoS attacks. When a security threat is detected, the API gateway server immediately synchronizes the threat information (including threat type, request source, and attack characteristics) to the load balancer. Based on this information, the load balancer restricts or blocks requests from suspicious sources and adjusts the load balancing policy to avoid routing subsequent requests to potentially vulnerable backend servers, thereby ensuring overall system security.

[0082] The load balancing device uses historical load data and real-time traffic trends, employing time series analysis or machine learning algorithms to predict future load changes. The load prediction results are shared with the task scheduler, which then uses this information to allocate tasks appropriately in advance. For example, if a model instance is overloaded and another model instance with similar functionality but a lower load exists, the load balancing device recommends to the task scheduler that the new task be assigned to the lower-loaded model instance. This ensures a more balanced load distribution and improves the system's overall processing capacity.

[0083] Before dispatching tasks to a model, the task scheduler accurately predicts the resources required for the task based on the task's characteristics and the model's historical processing performance data. This prediction is fed back to the resource management module, which then reserves or adjusts resources in advance based on the prediction. For example, if a computationally intensive task is predicted to require a significant amount of GPU resources, the resource management module can allocate sufficient GPU resources in advance to ensure smooth execution.

[0084] During task execution, the task scheduler monitors the task's execution status and resource usage in real time. If a task's actual resource usage deviates significantly from its predicted value, or if a task's execution efficiency is low due to insufficient resources, the task scheduler triggers the resource management module to dynamically adjust resources. Based on the task scheduler's trigger signals and in conjunction with Kubernetes HPA mechanisms, the resource management module rapidly adjusts the resource allocation of the model instance, such as increasing or decreasing the quotas for CPU, memory, GPU, and other resources, to ensure efficient task execution.

[0085] In this embodiment, the model layer includes a computing server cluster, a storage device, and a network switch. The computing server cluster is internally connected through a high-speed network to achieve high-speed data transmission and parallel computing collaboration between servers; the computing server cluster is connected to the storage device through a high-speed fiber channel or network file system protocol. During the model training process, the training data is read from the storage device, and the intermediate results and final model parameters of the training process are stored back to the storage device; wherein, the storage device adopts a distributed storage architecture, which is composed of multiple storage nodes, and the nodes are connected through a network for data synchronization and backup; the network switch divides different devices into different network areas by configuring switch ports and virtual local area networks.

[0086] In this embodiment, the platform layer includes an application server cluster, a cache server, and a database server; the application server cluster is used to deploy and run various functional modules of the platform layer, such as knowledge base construction, database connection, plug-in generation, intelligent question and answer, workflow orchestration, data dashboard, and account and enterprise management; the cache server improves data access speed and system responsiveness; the database server stores various data and log information generated by the platform layer; the platform layer is connected to the distribution layer and the portal layer through a network switch, receives processed data from the distribution layer, and provides users with various intelligent services.

[0087] The application server cluster is connected through a network to achieve collaborative work between servers. For example, when processing complex business logic, multiple application servers can divide the work and cooperate to complete the task together. Task allocation and coordination can be performed through a distributed task scheduling framework (such as ZooKeeper + Dubbo). During the data processing process, the application server cluster receives task requests from the distribution layer, improves data access speed through the cache server, and reduces the pressure on the database server. For the intelligent question and answer function, a question and answer matching algorithm based on semantic similarity calculation is adopted. The semantic similarity calculation formula is

[0088] in Sim ( Q , A ) is the problemQ With answers A The semantic similarity of with is the vocabulary weight, count( with ∈ Q ) and count( with ∈ A ) are vocabulary w i The number of times the question and answer appear in the question and answer, and the accurate answer is provided to the user based on the semantic similarity calculation results.

[0089] The model layer, distribution layer, platform layer, and portal layer are connected via high-speed network switches, enabling data transmission and interaction between layers. During platform operation, each layer works closely together. The model layer provides the distribution layer with a variety of models to choose from. The distribution layer selects the optimal model and assigns tasks based on task characteristics and model performance. The platform layer performs business processing and data storage based on task requests. The portal layer provides users with a user-friendly interactive interface and permission control.

[0090] In this embodiment, the deployment and operation of the enterprise master data model, proprietary model, open source model, and business model specifically include:

[0091] Enterprise master data models, proprietary models, open source models and business models are classified and stored according to their application scenarios, data dependencies and computational complexity, such as Figure 2 Shown, including:

[0092] Abstract the application scenario into different business dimensions and assign a weight to each business dimension w si ,in, i Indicates the business dimension number, ∑ i w si =1, and score the applicability of each model in each business dimension. s mj,si , j Indicates the model number, 0≤ s mj,si ≤1, then the model m j Comprehensive score of application scenarios S mj The calculation formula is: S mj =∑ i w si × s mj,si ;

[0093] Define the dependency strength coefficient between the model and different data sources dmj,Dk , k Indicates the data source number, 0≤ d mj,Dk ≤1, and the priority weight of the data source w Dk ,∑ k w Dk =1, model m j Data dependency composite score D mj The calculation formula is: D mj =∑ k w Dk × d mj,Dk ;

[0094] According to the number of calculation steps of the model n mj , data size v mj And the time complexity of the algorithm O mj , quantifying computational complexity into a comprehensive indicator C mj , expressed as:

[0095] C mj = α × n mj + β × v mj + c × f ( O mj )

[0096] in α 、 β 、 c is the weight coefficient ( α + β + c =1), f ( O mj ) is a numerical value converted according to the time complexity of the algorithm.

[0097] Set classification thresholds based on application scenarios, data dependencies, and computational complexity TS 、 TD 、 TC , the models are classified according to the following rules:

[0098] Business Owner Data Model: IfS mj ≥ TS and D mj If the sum of the dependency coefficients on the main data source within the enterprise is greater than a certain ratio, the model mj Classified as Enterprise Master Data Model;

[0099] Own model: If S mj High, that is, close to but not reaching the threshold of the enterprise master data model, and D mj Rely on the company's own internal data (data that is not the main data source but is collected independently by the company), and C mj If it is within a certain range (set according to the characteristics of the company's own business), it is classified as a proprietary model;

[0100] Open source model: If S mj At an average level, that is, lower than the typical application scenario scores of enterprise master data models and proprietary models. D mj The sum of the dependency intensity coefficients with open source data sources is greater than a certain ratio, such as 60%, and C mj If the value is low (open source models are usually not computationally complex), it is classified as an open source model;

[0101] Business model: If S mj High and closely related to specific business scenarios, Dmj Relying on data from commercial data providers, while C mj If the value is high (commercial models are usually powerful and computationally complex), it is classified as a commercial model.

[0102] The enterprise master data model is specifically designed for enterprise master data (such as data on core business entities like customers, products, and suppliers). It ensures that a company's most core business data is properly processed and managed, providing consistent, accurate, and reliable master data to support various business operations and decision-making. A proprietary model is defined as a model developed by an enterprise for specific business functions or scenarios that does not directly process master data. Examples include marketing analysis models (used to analyze market trends and customer behavior to support marketing strategy development) and customer churn prediction models (used to predict the likelihood of customer churn so that the company can take measures to retain customers). Therefore, the enterprise master data model is applicable to all business scenarios that require accurate master data, while proprietary models are applicable to specific business functions or scenarios, such as marketing analysis and customer churn prediction. The enterprise master data model provides the underlying data for proprietary models, which in turn perform further processing and analysis based on the enterprise master data model to meet specific business needs.

[0103] In this embodiment, the enterprise master data model is built around the enterprise's core business data and stored in a specific data-related storage area, which establishes a high-speed data channel with the enterprise's core business database in a large-capacity storage device; the proprietary model is developed based on the enterprise's own business needs and stored in a customized storage partition, and a special model version management directory is set up in the partition to record change information of different versions of the model; the open source model is stored in the open source model resource library, which has an automatic update mechanism, regularly obtains the latest model version from the open source community and performs compatibility testing; the business model is stored in a securely isolated business model storage area, which uses encrypted storage technology to ensure the confidentiality of the business model.

[0104] In this embodiment, containerization technology is used to package and deploy various models, and an independent container image is created for each model, such as Figure 3 As shown, specifically including:

[0105] For each model mj , according to its computational complexity C mj and historical operating data to estimate the number of CPU cores required CPU mj , memory size MEM mj and the number of GPUs (if GPUs are required) GPU mj .

[0106] According to the data dependencies of the model Dmj and data size v mj , estimate the storage space required for the model STORAGE mjConsidering the real-time requirements of the model and the frequency of interaction with external data sources, estimate the network bandwidth required by the model BANDWIDTH mj .

[0107] Set the compute server node set to N ={ n 1, n 2,..., nl}, l Indicates the number of nodes, each node nl Available resources include the number of CPU cores CPU nl available , memory size MEM nl available , number of GPUs GPU nl available , storage space STORAGE nl available and network bandwidth BANDWIDTH nl available ;

[0108] For the model m j , calculate its relationship with each node nl Resource matching MATCH mj,nl , the formula is:

[0109] MATCH mj,nl = oh 1×min(1, CPU nl available / CPU mj )+ oh 2×min(1, MEM nl available / MEM mj )+ oh 3×min(1, GPU nl available / GPU mj )+ oh 4×min(1, STORAGE nl available / STORAGE mj )+ oh 5×min(1, BANDWIDTH nlavailable / BANDWIDTH mj );

[0110] in oh 1. oh 2. oh 3. oh 4. oh 5 is the weight coefficient, oh 1+ oh 2+ oh 3+ oh 4+ oh 5=1, used to adjust the importance of different resources in the matching calculation;

[0111] Then, select the resource matching MATCH mj,nl The largest computing server node n best Deploy the container image, namely: n best =argmax nl∈N MATCH mj,nl , argmax is the optimization function, that is, returns the objective function ( MATCH mj,nl ) to obtain the maximum value of the input variable (node nl ), nl∈N It means traversing all nodes nl Collection N ,Right now n 1, n 2,..., nl .

[0112] The above formula enables classified storage of models and rational deployment of container images to computing server nodes, thereby improving the resource utilization and operational efficiency of the enterprise intelligence generation platform.

[0113] This embodiment uses containerization technology to encapsulate and deploy various models, creating an independent container image for each model. The image contains at least all elements required for the model code, dependent libraries, and configuration files to run. During the deployment process, the container image is deployed to the corresponding computing server node based on the model classification and resource pre-allocation results. For scenarios where multiple models need to work together, clustered deployment of the model is achieved through container orchestration tools to ensure efficient communication and collaborative computing between models.

[0114] When it is detected that the resource utilization of the server node where a model is located is too high or a failure occurs, the dynamic migration mechanism is triggered. The container orchestration tool is used to migrate the container instance of the model from the current node to a node with sufficient resources. Incremental synchronization technology is used during the migration process to only transmit the model's operating status and data changes, reducing migration time and impact on the business.

[0115] In this embodiment, the multi-model fusion and dynamic collaboration strategy is implemented, including the following:

[0116] A task feature analysis and intelligent distribution module is deployed at the front end of the computing server cluster. This module first extracts multi-dimensional features from the input tasks, including task type (such as image recognition, natural language processing, data mining), data size, processing complexity, and real-time requirements. Based on these extracted features, a task-server mapping model is constructed using a machine learning algorithm. This model intelligently distributes tasks to computing servers by comprehensively considering each computing server's current load (CPU usage, memory utilization, GPU computing resource utilization, etc.), historical processing performance, and compatibility with the task. For example, large-scale image recognition tasks are preferentially distributed to servers equipped with high-performance GPUs and currently low loads; natural language processing tasks with high real-time requirements are distributed to servers with low processing latency.

[0117] A parallel processing framework based on a computing server cluster is constructed. Complex, decomposable tasks are split into multiple subtasks and assigned to different computing servers for simultaneous processing. A dynamic load balancing mechanism is introduced to monitor the task processing progress and resource usage of each server in real time. When a server's load is too high while other servers are idle, some tasks are automatically migrated from the high-loaded server to a low-loaded one, ensuring full utilization of the entire computing server cluster's resources and improving task processing efficiency. For example, a heartbeat detection mechanism is used to obtain real-time status information on each server, and a task migration algorithm is used to dynamically adjust the load.

[0118] To enable collaborative computing between servers in a computing server cluster, efficient communication protocols and communication middleware are designed, employing low-latency, high-bandwidth network communication technologies to reduce data transmission delays between servers. Furthermore, data transmission formats and compression algorithms are optimized to reduce network bandwidth usage. During collaborative computing, different parallel computing modes—such as pipeline parallelism, data parallelism, or model parallelism—are employed to improve computational efficiency, depending on the characteristics of the task and the data dependencies between servers. For example, for deep learning model training tasks, a data parallel mode is employed to distribute training data to different servers for parallel computing. The model parameters on each server are then synchronized through the communication middleware.

[0119] A multi-model data fusion storage strategy is designed for the diverse data types generated by enterprise master data models, proprietary models, open source models, and commercial models. Structured data (such as core business data and model parameters) is stored in a high-performance relational database, leveraging the database's transaction processing and indexing mechanisms to ensure data consistency and fast query speed. Unstructured data (such as images, audio, and text) is stored in a distributed file system, employing data sharding and redundant storage technologies to improve data reliability and scalability. Furthermore, a data association mapping table is established to record the associations between data from different models, enabling the fusion and unified management of multi-model data. For example, customer information stored in the enterprise master data model is associated with customer behavior data stored in proprietary models to support comprehensive model analysis.

[0120] To improve data access speed, a multi-level data caching mechanism is established between storage devices and computing servers. Frequently used data is cached in local memory or high-speed cache devices on the computing server based on data access frequency and popularity. Furthermore, a data prefetching algorithm is used to pre-fetch potentially needed model data from storage devices into the cache based on the task's historical access patterns and current processing progress, reducing data access latency. For example, common datasets frequently accessed by multiple models are cached long-term. Furthermore, data for the next batch likely to be used during model training is pre-fetched in advance.

[0121] A strict data security and privacy protection architecture is established for the storage and access of multi-model data fusion. Encryption technology is used to encrypt the storage and transmission of sensitive data, ensuring data security during storage and transmission. Furthermore, an access control mechanism is established to restrict access to data from different models based on user roles and permissions. For core data from commercial models and proprietary models, additional encryption algorithms and access auditing mechanisms are implemented to prevent data leaks and unauthorized access. For example, data is encrypted using a combination of symmetric and asymmetric encryption, and user permissions are managed through a role-based access control (RBAC) model.

[0122] An intelligent traffic scheduling algorithm is implemented on the network switch. The distribution of network traffic is dynamically adjusted according to the data traffic characteristics and task requirements between different computing servers in the model layer. By real-time monitoring of the bandwidth utilization, latency, and packet loss rate indicators of the network links, network traffic is preferentially allocated to links with good performance, taking into account the priority and real-time requirements of the tasks. For the traffic generated by model inference tasks with high real-time requirements, links with low latency are selected for transmission. For large-volume model training data transmission tasks, load balancing is used to distribute them to multiple links.

[0123] To adapt to the ever-changing computing tasks and resource requirements of the model layer, a dynamic network topology adjustment mechanism is designed. Based on the deployment of the computing server cluster and task processing requirements, the network switch connections are automatically adjusted to build an optimal network topology. Furthermore, a network fault tolerance mechanism is established. When a network failure (such as a link interruption or switch failure) occurs, the fault is quickly detected and automatically switched to a backup link or network device, ensuring that communication between components in the model layer is not affected. For example, the Spanning Tree Protocol (STP) or Rapid Spanning Tree Protocol (RSTP) is used to prevent network loops, and redundant links and devices are used to achieve network fault tolerance.

[0124] In this embodiment, a multi-level task feature extraction module is built in the API gateway server to perform a comprehensive analysis of the input task, such as Figure 4 Shown, including:

[0125] For text tasks, natural language processing technology is used to extract at least features including semantic depth (a comprehensive evaluation based on lexical, syntactic, and semantic analysis, such as using dependency syntactic analysis to calculate sentence structure complexity and combining word vector similarity to measure semantic richness), sentiment tendency (using a combination of sentiment dictionaries and deep learning models to determine the polarity and intensity of the emotions expressed in the text), and domain relevance (by constructing domain knowledge graphs and calculating the similarity between the text and the graphs of each domain).

[0126] For image tasks, a convolutional neural network (CNN) is used to extract at least the image's texture features (such as gray-level co-occurrence matrix to calculate texture roughness, contrast, etc.), shape features (obtaining object shape information through edge detection and contour analysis), and scene features (using a pre-trained scene classification model to identify the scene type to which the image belongs).

[0127] For speech tasks, Mel-frequency cepstral coefficients (MFCCs) are used to extract the spectral features of speech, and speech recognition technology is used to obtain the speech content. The speech is then further analyzed for emotion (through a comprehensive judgment of acoustic features and semantic information), speaking speed, intonation, and other characteristics.

[0128] For cross-modal tasks (such as image-text association tasks), a multimodal fusion model is used to extract association features between different modalities, such as the degree of semantic alignment between images and texts, and the matching degree between visual elements and text descriptions.

[0129] In the distribution layer, a combination of principal component analysis (PCA) and linear discriminant analysis (LDA) is used to reduce the dimensionality of the extracted multimodal task features, remove redundant information, and retain key features.

[0130] An autoencoder is used to encode the reduced-dimensional features, mapping them to a unified low-dimensional space to generate a standardized task feature vector. At the same time, a normalization method is used to scale the values ​​of each dimension of the feature vector to the range [0, 1] to eliminate the influence of different feature dimensions.

[0131] Based on the standardized task feature vectors, a dynamic feature library is constructed. A clustering algorithm (such as the DBSCAN clustering algorithm) is used to cluster the task feature vectors to form different task feature clusters. Each cluster represents a class of tasks with similar characteristics.

[0132] Monitor the features of newly received tasks in real time and assign them to corresponding task feature clusters. At the same time, based on business development and changes in model capabilities, regularly update and optimize the feature library, such as adjusting clustering parameters and merging or splitting task feature clusters.

[0133] Use network monitoring equipment to collect network performance metrics for the server where the model resides, including network latency (using the ping command or specialized network monitoring tools to measure the time it takes for a data packet to be sent and received), bandwidth utilization (using traffic monitoring tools to calculate the ratio of the actual traffic flow to the maximum bandwidth of the network interface), etc. Use load balancing equipment to record the model's load, such as the number of tasks currently being processed and the length of the task processing queue.

[0134] At the model layer, a performance monitoring agent is deployed for each model to collect business performance indicators of the model, such as accuracy (calculated by comparing with labeled data), recall rate, F1 value (the harmonic mean of accuracy and recall rate), etc., as well as resource performance indicators such as CPU usage, memory usage, and GPU computing resource utilization (if a GPU is available).

[0135] A time series prediction algorithm (such as a long short-term memory (LSTM) network) is used to model the model's historical performance data and predict future performance trends. For example, based on the model's accuracy trends over the past period, the accuracy can be predicted at several future time points. A dynamic performance evaluation model is constructed, combining factors such as the current system load, task feature distribution, and business priorities. This model uses a weighted scoring approach to assign weights to different performance indicators and calculate the model's overall performance score. For example, for tasks with high real-time requirements, the response time indicator is given a higher weight; for tasks with high accuracy requirements, the accuracy indicator is given a higher weight.

[0136] Based on the evaluation results of the dynamic performance evaluation model, models are classified into multiple performance levels, such as excellent, good, good, fair, and poor. Corresponding performance thresholds are set for each level. A performance warning mechanism is established to trigger an alert when a model's performance indicator falls below the corresponding threshold. Alerts are sent to system administrators via email, text message, or system notification. Alert information, including the time of the alert, model name, performance indicator, and current value, is recorded for subsequent analysis and processing.

[0137] A multi-objective optimization matching algorithm based on a genetic algorithm is employed, comprehensively considering multiple objectives, including task feature vectors, model performance grading, and business priorities. The task-model matching problem is transformed into an optimization problem, with objective functions such as minimizing task processing time, maximizing task processing accuracy, and balancing system resource utilization. In the genetic algorithm, a chromosome encoding scheme is defined, representing the matching relationship between tasks and models as chromosomes. Through operations such as selection, crossover, and mutation, the chromosomes are continuously optimized to find the optimal task-model matching solution.

[0138] Based on the results of the multi-objective optimization matching algorithm, a dynamic task allocation strategy is formulated. When allocating tasks, high-priority tasks are preferentially assigned to models with excellent performance and low load. For tasks with extremely high real-time requirements, even if a model's current load is slightly higher, it will be given priority if its response time meets the requirements and resource availability is good. A load balancing strategy is adopted to prevent some models from being overly busy while others are idle. When it is detected that the load of a model is about to reach the upper limit, subsequent tasks are assigned to other models with comparable performance and lower load. At the same time, the geographical location of the model is taken into consideration (if the model is deployed in a distributed environment), and models that are closer to the task data source are given priority to reduce network transmission delays.

[0139] The distribution layer, through its interactive interface with the model layer, obtains real-time model capability information, including supported task types, input and output formats, and maximum data processing volume. It also monitors model updates and promptly updates model capability information in the distribution layer whenever a model upgrade or parameter adjustment occurs. It records historical task assignments and executions, analyzing the relationship between different task characteristics and model selection. Machine learning algorithms (such as decision trees and random forests) are used to construct a model selection preference model to predict the optimal model selection based on new task characteristics. As historical data accumulates, the model selection preference model is continuously optimized to improve model selection accuracy.

[0140] When the system load changes significantly or unexpected tasks occur, the distribution layer and the model layer collaborate to elastically allocate resources. For example, when the task volume suddenly increases, the distribution layer sends a resource request to the model layer. The model layer dynamically adjusts the model's resource allocation based on the request, such as increasing the model's computing resources or starting a backup model instance, to meet the task processing needs.

[0141] Through the above architecture and strategy, the distribution layer can accurately and efficiently select the optimal model and assign tasks based on task characteristics and model performance, achieve deep collaboration with the model layer, and significantly improve the performance, stability and resource utilization of the entire system.

[0142] In this embodiment, the web server cluster is internally connected via a network, achieving load balancing and failover between servers. User requests are evenly distributed across multiple web servers via a reverse proxy server. When a user accesses the portal website, static resources (such as images, CSS, and JavaScript files) are preferentially retrieved from the CDN node closest to the user. The web server cluster communicates with the platform-level application server cluster via HTTP / HTTPS protocols, obtaining dynamic data and rendering pages for the user.

[0143] Among them, CDN acceleration nodes are distributed at the edge of the network in different geographical locations and are connected to the Web server cluster at the portal layer through a high-speed network. When a user initiates a request, the CDN node selects the node closest to the user to provide services to the user based on the user's geographical location and network conditions, and returns the cached static resources to the user.

[0144] Among them, security devices authenticate and authorize user access and are deployed at the front end of the Web server cluster to provide security protection for user access. Firewall rules filter illegal network traffic, and intrusion detection and prevention systems (IDS / IPS) are used to monitor and prevent malicious attacks in real time, such as SQL injection and cross-site scripting (XSS). Among them, the role-based access control RBAC model is adopted, and the role permission calculation formula is R user =U r∈Roles(user) Rr ,in R user A set of user permissions. Roles ( user ) is the set of roles to which the user belongs, Rr For the role r The set of permissions owned by U r It is a union operation on the role permission set, used to aggregate the permissions of all roles to which the user belongs.

[0145] Furthermore, the platform deploys and runs enterprise master data models, proprietary models, open source models and business models, fully leveraging the advantages of different types of models. The enterprise master data model can deeply fit the enterprise's own business logic and data characteristics, the proprietary model can be customized and developed according to the specific needs of the enterprise, the open source model can quickly acquire advanced technologies with the help of the open source community, and the business model usually has mature and stable performance. The integration of multiple models enables the platform to cope with various complex and changing enterprise business scenarios and provide more comprehensive and accurate solutions.

[0146] Furthermore, during the model training process, the model performance evaluation index is used P Quantitatively evaluate different models by comprehensively considering the model accuracy A , recall rate R and computational complexity C , and assign corresponding weight coefficients α 、 β 、 c It can comprehensively and objectively measure the comprehensive performance of the model, which not only helps to timely discover the advantages and disadvantages of the model during the model training process, but also provides a clear direction for model optimization and adjustment, improves the efficiency and quality of model training, and ensures that the model finally put into use has a high performance level.

[0147] Furthermore, in the model selection process, the model selection formula is introduced to calculate the scores of each model, and the model with the highest score is selected for task processing, taking into account the task feature matching function. F ( t ), current moment resource utilization function U ( t ) and the comprehensive performance score of the model calculated by the model layer P , and assign corresponding weight coefficients. In this way, the distribution layer can accurately select the most suitable model to handle tasks based on the specific characteristics of the tasks and the current system resource status, thereby effectively improving the quality and efficiency of task processing and ensuring that business needs are accurately met.

[0148] Furthermore, the resource utilization function in the model selection formula U ( t ) enables the distribution layer to perceive the system's resource usage in real time and dynamically adjust task allocation strategies based on resource availability. When system resources are limited, models with lower resource usage and better performance are prioritized to handle tasks, avoiding system performance degradation caused by excessive resource concentration. When system resources are sufficient, more resources can be appropriately allocated to complex tasks to speed up task processing, thereby fully utilizing system resources, improving overall system performance and stability, and reducing the company's operating costs.

[0149] Furthermore, the platform layer receives task requests from the distribution layer through the application server cluster, which has high availability and fault tolerance. When some application servers fail, other servers can automatically take over task processing to ensure the continuity of business processing, avoid business interruptions caused by single point failures, and provide enterprises with a stable and reliable business processing environment.

[0150] Furthermore, the web server cluster in the portal layer establishes connections with CDN acceleration nodes, caching static resources there. When users access the platform, the CDN acceleration nodes quickly deliver static resources based on their geographic location and network conditions, significantly reducing data transmission latency and improving user access speed and experience. Regardless of their location, users can quickly load platform pages and access the information they need, enhancing their satisfaction and loyalty to the platform.

[0151] Furthermore, security devices authenticate and authorize user access, effectively preventing unauthorized users from accessing the platform and protecting sensitive enterprise data and business information from being leaked. Security devices employ a variety of security technologies, such as firewalls, intrusion detection systems, and encryption, to provide comprehensive user protection, ensuring the secure and stable operation of the platform and reducing security risks faced by enterprises.

[0152] In summary, through the collaborative work and innovative design of various layers, the present invention has achieved remarkable beneficial effects in model selection and training, task allocation, business processing and data storage, user experience and system security. It can provide enterprises with efficient, intelligent and secure intelligent generation services, helping enterprises to achieve better development in the digital age.

[0153] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] The above embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and replacements made by technicians in this field on the basis of the present invention fall within the scope of protection required by the present invention.

Claims

1. An enterprise intelligence generation platform based on multi-model fusion technology, characterized by: include: The model layer is used to deploy and run enterprise master data models, proprietary models, open source models, and business models; During the model training process, computing server clusters are used for large-scale data training, and the training data is stored in storage devices. The comprehensive performance score of the model is used to P = α × A + β × R-γ × C Quantitative evaluation of different models; A is the model accuracy, R is the model recall rate, C Calculate the complexity of the model, α 、 β 、 γ is the weight coefficient; The distribution layer selects the optimal model and assigns tasks based on task characteristics and model performance; in the model selection process, the model selection formula is introduced S = ω 1× F ( t )+ ω 2× U ( t )+ ω 3× P Calculate the scores of each model and select the model with the highest score to process the task; in S Select a score for the model, F ( t ) is the task feature matching function, U ( t ) is the resource utilization function at the current moment, P is the comprehensive performance score of the model calculated at the model layer, ω 1. ω 2. ω 3 is the weight coefficient; The platform layer performs business processing and data storage based on task requests. During data processing, it receives task requests from the distribution layer through the application server cluster and stores business data in the database server. The portal layer includes a web server cluster, CDN acceleration nodes, and security devices. The web server cluster establishes a connection with the CDN acceleration nodes, caches static resources on the CDN acceleration nodes, and authenticates and authorizes user access through security devices. The deployment and operation of enterprise master data models, proprietary models, open source models, and commercial models specifically include: Enterprise master data models, proprietary models, open source models, and business models are classified and stored according to their application scenarios, data dependencies, and computational complexity, including: Abstract the application scenario into different business dimensions and assign a weight to each business dimension w si ,in, i Indicates the business dimension number, ∑ i w si =1, and score the applicability of each model in each business dimension. s mj,si , j Indicates the model number, 0≤ s mj,si ≤1, then the model m j Comprehensive score of application scenarios S mj The calculation formula is: S mj =∑ i w si × s mj,si ; Define the dependency strength coefficient between the model and different data sources d mj,Dk , k Indicates the data source number, 0≤ d mj,Dk ≤1, and the priority weight of the data source w Dk ,∑ k w Dk =1, model m j Data dependency composite score D mj The calculation formula is: D mj =∑ k w Dk × d mj,Dk ; According to the number of calculation steps of the model n mj , data size v mj And the time complexity of the algorithm O mj , quantifying computational complexity into a comprehensive indicator C mj , expressed as: C mj = α 1× n mj + β 1× v mj + γ 1× f ( O mj ) in α 1. β 1. γ 1 is the weight coefficient, α 1+ β 1+ γ 1=1, f ( O mj ) is the value converted according to the time complexity of the algorithm; Set classification thresholds based on application scenarios, data dependencies, and computational complexity TS 、 TD 、 TC , the models are classified according to the following rules: Business Owner Data Model: If S mj ≥ TS and D mj If the sum of the dependency coefficients on the main data source within the enterprise is greater than a certain ratio, the model mj Classified as Enterprise Master Data Model; Own model: If S mj High, that is, close to but not reaching the threshold of the enterprise master data model, and D mj Relying on the company's own internal data, C mj If it is within a certain range, it is classified as a proprietary model; Open source model: If S mj At an average level, that is, lower than the typical application scenario scores of enterprise master data models and proprietary models. D mj The sum of the dependency intensity coefficients with open source data sources is greater than a certain ratio, and C mj If it is a low value, it is classified as an open source model; Business model: If S mj High and closely related to specific business scenarios, Dmj Relying on data from commercial data providers, while C mj If it is a high value, it is classified as a business model; Containerization technology is used to encapsulate and deploy various models, creating independent container images for each model, including: Set the compute server node set to N ={ n 1, n 2,..., nl }, l Indicates the number of nodes, each node nl Available resources include the number of CPU cores CPU nl available , memory size MEM nl available , number of GPUs GPU nl available , storage space STORAGE nl available and network bandwidth BANDWIDTH nl available ; For the model m j , calculate its relationship with each node nl Resource matching MATCH mj,nl , the formula is: MATCH mj,nl = ω 1×min(1, CPU nl available / CPU mj )+ ω 2×min(1, MEM nl available / MEM mj )+ ω 3×min(1, GPU nl available / GPU mj )+ ω 4×min(1, STORAGE nl available / STORAGE mj )+ ω 5×min(1, BANDWIDTH nl available / BANDWIDTH mj ); in ω 1. ω 2. ω 3. ω 4. ω 5 is the weight coefficient, ω 1+ ω 2+ ω 3+ ω 4+ ω 5=1, used to adjust the importance of different resources in the matching calculation; Then, select the resource matching MATCH mj,nl The largest computing server node n best Deploy the container image, namely: n best =argmax nl∈N MATCH mj,nl , argmax is the optimization function, that is, it returns the input variable that makes the objective function achieve the maximum value, nl∈N It means traversing all nodes nl Collection N .

2. The platform according to claim 1, characterized in that: The distribution layer includes a load balancing device, an API gateway server, a task scheduler, a resource management module, and a network monitoring device. After receiving a request, the API gateway server extracts traffic feature information and shares it with the load balancing device in real time. The load balancing device dynamically adjusts the load balancing strategy based on this feature information. The load balancing device uses historical load data and real-time traffic trends, employing time series analysis or machine learning algorithms to predict future load changes. The load prediction results are shared with the task scheduler, which uses the load prediction information to reasonably allocate tasks in advance. Before the task scheduler distributes tasks to the model, it accurately predicts the resources required for the task based on the task characteristics and the historical processing performance data of the model, and feeds the resource demand prediction results back to the resource management module. The resource management module reserves or adjusts resources in advance based on the prediction results.

3. The platform according to claim 1, characterized in that: The model layer includes a computing server cluster, storage devices, and network switches. The computing server cluster is internally connected via a high-speed network to achieve high-speed data transmission and parallel computing collaboration between servers. The computing server cluster is connected to the storage device via a high-speed fiber channel or network file system protocol. During the model training process, the training data is read from the storage device, and the intermediate results and final model parameters of the training process are stored back to the storage device. The storage device adopts a distributed storage architecture and consists of multiple storage nodes. Data synchronization and backup are performed between the nodes via a network connection. The network switch divides different devices into different network areas by configuring switch ports and virtual local area networks.

4. The platform according to claim 3, characterized in that: A task feature analysis and intelligent distribution module is deployed at the front end of the computing server cluster. This module first extracts multi-dimensional features from the input tasks. Based on the extracted features, it uses a machine learning algorithm to build a task-server mapping model. This model comprehensively considers the current load of each computing server, its historical processing performance, and its compatibility with the task, and intelligently distributes tasks to the computing servers. A parallel processing framework based on a computing server cluster is constructed. For complex tasks that can be decomposed, they are split into multiple subtasks and assigned to different computing servers for simultaneous processing. At the same time, a dynamic load balancing mechanism is introduced to monitor the task processing progress and resource usage of each server in real time. When it is found that the load on a server is too high and the resources of other servers are idle, some tasks are automatically migrated from the high-load server to the low-load server.

5. The platform according to claim 4, characterized in that: Design a multi-model data fusion storage strategy for different types of data generated by the enterprise's master data model, proprietary models, open source models, and commercial models. Store structured data in a high-performance relational database and unstructured data in a distributed file system. At the same time, establish a data association mapping table to record the associations between data from different models to achieve fusion storage and unified management of multi-model data. An intelligent traffic scheduling algorithm is implemented on the network switch. The distribution of network traffic is dynamically adjusted according to the data traffic characteristics and task requirements between different computing servers in the model layer. By real-time monitoring of the bandwidth utilization, latency, and packet loss rate indicators of the network links, network traffic is preferentially allocated to links with good performance, taking into account the priority and real-time requirements of the tasks. For the traffic generated by model inference tasks with high real-time requirements, links with low latency are selected for transmission. For large-volume model training data transmission tasks, load balancing is used to distribute them to multiple links.

6. The platform according to any one of claims 2 to 5, characterized in that: A multi-level task feature extraction module is built in the API gateway server to perform a comprehensive analysis of input tasks, including: For text tasks, natural language processing technology is used to extract features including at least semantic depth, sentiment tendency, and domain relevance; For image tasks, convolutional neural networks are used to extract at least the texture features, shape features, and scene features of the image; For speech tasks, we use Mel-frequency cepstral coefficients (MFCCs) to extract the spectral features of speech, and combine them with speech recognition technology to obtain the speech content and further analyze the speech features. For cross-modal tasks, a multimodal fusion model is used to extract correlation features between different modalities.

7. The platform according to any one of claims 1 to 5, characterized in that: The web server cluster is connected internally via a network to achieve load balancing and failover between servers, and evenly distributes user requests to multiple web servers via a reverse proxy server. When a user accesses the portal website, static resources are preferentially obtained from the CDN node closest to the user. The web server cluster communicates with the application server cluster at the platform layer via HTTP / HTTPS protocols to obtain dynamic data and render pages for users. CDN acceleration nodes are distributed at the edge of the network in different geographical locations and connected to the portal layer's web server cluster via a high-speed network. When a user initiates a request, the CDN node selects the node closest to the user based on the user's geographical location and network conditions to provide services and return the cached static resources to the user. Among them, the security device authenticates and authorizes user access, adopts the role-based access control RBAC model, and the role authority calculation formula is: R user =U r∈Roles(user) Rr ,in R user A set of user permissions. Roles ( user ) is the set of roles to which the user belongs, Rr For the role r The set of permissions owned by U r It is a union operation on the role permission set, used to aggregate the permissions of all roles to which the user belongs.

Images